text
stringlengths 4
2.78M
| meta
dict |
|---|---|
\
[^1]\
[ September 6, 2017 ]{}
[**Abstract**]{}
Based upon the unique and simple starting point of the continuous flow of time a physical theory is derived through an analysis of the elementary arithmetic composition and symmetries of this one-dimensional progression. We describe how the explicit development of the theory leads to a prediction of the unique and largest exceptional Lie group $\ee$ as the full ‘symmetry of time’, and hence as the unification group for the physical theory. This proposal results from the identification of a series of esoteric properties of the Standard Model of particle physics from a series of intermediate augmentations in the ‘multi-dimensional form of time’. These physical properties derive from the breaking of the full symmetry of time through the necessary interposition of an external 4-dimensional spacetime arena, itself constructed from a 4-dimensional form of time, as the background to all observations. The basic conceptual picture is presented together with reviews of a number of references regarding $\ee$ structures which may provide a significant guide in pursuing the goal of converging upon a complete unified theory.
[^1]: email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
[ $\mbox{\bf Seiichiro Tani}^{\ast\dagger}$ ]{}\
[[email protected]]{}
- |
[ $\mbox{\bf Hirotada Kobayashi}^{\ddagger}$ ]{}\
[[email protected]]{}
- |
[ $\mbox{\bf Keiji Matsumoto}^{\ddagger \dagger}$ ]{}\
[[email protected]]{}
title: |
Exact Quantum Algorithms\
for the Leader Election Problem[^1]
---
${}^{\ast}$NTT Communication Science Laboratories, NTT Corporation\
${}^{\dagger}$Quantum Computation and Information Project, ERATO-SORST, JST\
${}^{\ddagger}$Principles of Informatics Research Division, National Institute of Informatics
[10]{}
Yehuda Afek and Yossi Matias. Elections in anonymous networks. , 113(2):312–330, 1994.
Masami Amano and Kazuo Iwama. Undecidability on quantum finite automata. In [*Proceedings of the Thirty-First Annual ACM Symposium on Theory of Computing*]{}, pages 368–375, 1999.
Andris Ambainis. A new protocol and lower bounds for quantum coin flipping. , 68(2):398–416, 2004.
Andris Ambainis. Quantum walk algorithm for element distinctness. In [*Proceedings of the Forty-Fifth Annual Symposium on Foundations of Computer Science*]{}, pages 22–31, 2004.
Andris Ambainis, Harry M. Buhrman, Yevgeniy Dodis, and Hein Röhrig. Multiparty quantum coin flipping. In [*Proceedings of the Nineteenth Annual [IEEE]{} Conference on Computational Complexity*]{}, pages 250–259, 2004.
Andris Ambainis and Rusins Freivalds. 1-way quantum finite automata: Strengths, weaknesses and generalizations. In [*Proceedings of the Thirty-Ninth Annual IEEE Symposium on Foundations of Computer Science*]{}, pages 332–341, 1998.
Andris Ambainis and John Watrous. Two-way finite automata with quantum and classical state. , 287(1):299–311, 2002.
Dana Angluin. Local and global properties in networks of processors (extended abstract). In [*Proceedings of the Twelfth Annual ACM Symposium on Theory of Computing*]{}, pages 82–93, 1980.
Ziv Bar-Yossef, Thathachar S. Jayram, and Iordanis Kerenidis. Exponential separation of quantum and classical one-way communication complexity. In [*Proceedings of the Thirty-Sixth Annual ACM Symposium on Theory of Computing*]{}, pages 128–137, 2004.
Howard Barnum, Claude Crépeau, Daniel Gottesman, Adam D. Smith, and Alain Tapp. Authentication of quantum messages. In [*Proceedings of the Forty-Third Annual IEEE Symposium on Foundations of Computer Science*]{}, pages 449–458, 2002.
Michael Ben-[O]{}r and Avinatan Hassidim. Fast quantum [B]{}yzantine agreement. In [*Proceedings of the Thirty-Seventh Annual ACM Symposium on Theory of Computing*]{}, pages 481–485, 2005.
Charles H. Bennett. Quantum cryptography using any two nonorthogonal states. , 68(21):3121–3124, 1992.
Charles H. Bennett and Gilles Brassard. Quantum cryptography: Public key distribution and coin tossing. In [*Proceedings of IEEE International Conference on Computers, Systems and Signal Processing*]{}, pages 175–179, 1984.
Gilles Brassard, Peter H[ø]{}yer, Michele Mosca, and Alain Tapp. Quantum amplitude amplification and estimation. In [*Quantum Computation and Quantum Information: A Millennium Volume*]{}, volume 305 of [*AMS Contemporary Mathematics Series*]{}, pages 53–74. 2002.
Randal E. Bryant. Graph-based algorithms for [Boolean]{} function manipulation. , 35(8):677–691, 1986.
Harry M. Buhrman, Richard E. Cleve, John H. Watrous, and Ronald de Wolf. Quantum fingerprinting. , 87(16):167902, 2001.
Harry M. Buhrman, Richard E. Cleve, and Avi Wigderson. Quantum vs. classical communication and computation. In [*Proceedings of the Thirtieth Annual ACM Symposium on the Theory of Computing*]{}, pages 63–68, 1998.
A. Childs, R. Cleve, E. Deotto, E. Farhi, S. Gutmann, and D. Spielman. Exponential algorithmic speedup by quantum walk. In [*Proceedings of the Thirty-Fifth Annual ACM Symposium on Theory of Computing*]{}, pages 59–68, 2003. quant-ph/0209131.
Claude Crépeau, Daniel Gottesman, and Adam D. Smith. Secure multi-party quantum computation. In [*Proceedings of the Thirty-Fourth Annual ACM Symposium on Theory of Computing*]{}, pages 643–652, 2002.
Claude Crépeau, Frédéric Légaré, and Louis Salvail. How to convert the flavor of a quantum bit commitment. In [*Proceedings of International Conference on the Theory and Application of Cryptographic Techniques (EUROCRYPT 2001)*]{}, volume 2045 of [*Lecture Notes in Computer Science*]{}, pages 60–77, 2001.
Ellie D’Hondt and Prakash Panangaden. The computational power of the w and ghz states. , 6(2), 2006.
Danny Dolev, Maria M. Klawe, and Michael Rodeh. An [$O(n\log n)$]{} unidirectional distributed algorithm for extrema finding in a circle. , 3(3):245–260, 1982.
Paul Dumais, Dominic Mayers, and Louis Salvail. Perfectly concealing quantum bit commitment from any quantum one-way permutation. In [*Proceedings of International Conference on the Theory and Application of Cryptographic Techniques (EUROCRYPT 2000)*]{}, volume 1807 of [*Lecture Notes in Computer Science*]{}, pages 300–315, 2000.
Greg N. Frederickson and Nancy A. Lynch. Electing a leader in a synchronous ring. , 34(1):98–115, 1987.
Robert G. Gallager, Pierre A. Humblet, and Philip M. Spira. A distributed algorithm for minimum-weight spanning trees. , 5(1):66–77, 1983.
Lov. K. Grover. A fast quantum mechanical algorithm for database search. In [*Proceedings of the Twenty-Eighth Annual [ACM]{} Symposium on Theory of Computing*]{}, pages 212–219, 1996.
Alon Itai and Michael Rodeh. Symmetry breaking in distributive networks. In [*Proceedings of the Twenty-Second Annual IEEE Symposium on Foundations of Computer Science*]{}, pages 150–158, 1981.
Alon Itai and Michael Rodeh. Symmetry breaking in distributed networks. , 88(1):60–87, 1990.
Alexei Yu. Kitaev, Alexander H. Shen, and Mikhail N. Vyalyi. , volume 47 of [*Graduate Studies in Mathematics*]{}. AMS, 2002.
Evangelos Kranakis, Danny Krizanc, and Jacob van den Berg. Computing [B]{}oolean functions on anonymous networks. , 114(2):214–236, 1994.
Nancy A. Lynch. . Morgan Kaufman Publishers, 1996.
Dominic Mayers. Unconditional security in quantum cryptography. , 48(3):351–406, 2001.
Michael A. Nielsen and Isaac L. Chuang. . Cambridge University Press, 2000.
Nancy Norris. Universal covers of graphs: Isomorphism to depth n-1 implies isomorphism to all depths. , 56(1):61–74, 1995.
M. Mosca P. H[ø]{}yer and R. de Wolf. Quantum search on bounded-error inputs. In [*Proceedings of the Thirtieth International Colloquium on Automata, Languages and Programming(ICALP’03)*]{}, volume 2719 of [*Lecture Notes in Computer Science*]{}, pages 291–299. Springer, 2003.
Gary L. Peterson. An [$O(n \log n)$]{} unidirectional algorithm for the circular extrema problem. , 4(4):758–762, 1982.
Ran Raz. Exponential separation of quantum and classical communication complexity. In [*Proceedings of the Thirty-First Annual ACM Symposium on Theory of Computing*]{}, pages 358–367, 1999.
Baruch Schieber and Marc Snir. Calling names on nameless networks. , 113(1):80–101, 1994.
Peter W. Shor. Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer. , 26(5):1484–1509, 1997.
Peter W. Shor and John Preskill. Simple proof of security of the [BB84]{} quantum key distribution protocol. , 85(2):441–444, 2000.
Kiyoshi Tamaki, Masato Koashi, and Nobuyuki Imoto. Security of the [Bennett]{} 1992 quantum-key distribution protocol against individual attack over a realistic channel. , 67(3):032310, 2003.
Kiyoshi Tamaki, Masato Koashi, and Nobuyuki Imoto. Unconditionally secure key distribution based on two nonorthogonal states. , 90(16):167904, 2003.
Kiyoshi Tamaki and Norbert Lütkenhaus. Unconditional security of the [Bennett]{} 1992 quantum key-distribution protocol over a lossy and noisy channel. , 69(3):032316, 2004.
Seiichiro Tani, Hirotada Kobayashi, and Keiji Matsumoto. Exact quantum algorithms for the leader election problem. In [*Proceedings of the Twenty-Second Symposium on Theoretical Aspects of Computer Science (STACS 2005)*]{}, volume 3404 of [*Lecture Notes in Computer Science*]{}, pages 581–592. Springer, 2005.
Jan van Leeuwen and Richard B. Tan. An improved upperbound for distributed election in bidirectional rings of processors. , 2(3):149–160, 1987.
Tomohiro Yamasaki, Hirotada Kobayashi, and Hiroshi Imai. Quantum versus deterministic counter automata. , 334(1-3):275–297, 2005.
Tomohiro Yamasaki, Hirotada Kobayashi, Yuuki Tokunaga, and Hiroshi Imai. One-way probabilistic reversible and quantum one-counter automata. , 289(2):963–976, 2002.
Masafumi Yamashita and Tsunehiko Kameda. Computing on anonymous networks: Part [I]{} – characterizing the solvable cases. , 7(1):69–89, 1996.
Masafumi Yamashita and Tsunehiko Kameda. Computing on anonymous networks: Part [II]{} – decision and membership problems. , 7(1):90–96, 1996.
Masafumi Yamashita and Tsunehiko Kameda. Leader election problem on networks in which processor identity numbers are not distinct. , 10(9):878–887, 1999.
[^1]: A preliminary version of this paper appeared in [@TanKobMatSTACS05].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we investigate the coexistence of two technologies that have been put forward for the fifth generation (5G) of cellular networks, namely, network-assisted device-to-device (D2D) communications and massive MIMO (multiple-input multiple-output). Potential benefits of both technologies are known individually, but the tradeoffs resulting from their coexistence have not been adequately addressed. To this end, we assume that D2D users reuse the downlink resources of cellular networks in an underlay fashion. In addition, multiple antennas at the BS are used in order to obtain precoding gains and simultaneously support multiple cellular users using multiuser or massive MIMO technique. Two metrics are considered, namely the average sum rate (ASR) and energy efficiency (EE). We derive tractable and directly computable expressions and study the tradeoffs between the ASR and EE as functions of the number of BS antennas, the number of cellular users and the density of D2D users within a given coverage area. Our results show that both the ASR and EE behave differently in scenarios with low and high density of D2D users, and that coexistence of underlay D2D communications and massive MIMO is mainly beneficial in low densities of D2D users.'
author:
- 'Serveh Shalmashi, , Emil Björnson, , Marios Kountouris, , Ki Won Sung, , and Mérouane Debbah, [^1] [^2] [^3] [^4]'
bibliography:
- 'IEEEabrv.bib'
- 'refs.bib'
- 'serveh.bib'
title: 'Energy Efficiency and Sum Rate Tradeoffs for Massive MIMO Systems with Underlaid Device-to-Device Communications'
---
D2D communications, massive MIMO, coexistence, energy efficiency, stochastic geometry.
Introduction
============
The research on future mobile broadband networks, referred to as the fifth generation (5G), has started in the past few years. In particular, stringent key performance indicators (KPIs) and tight requirements have been introduced in order to handle higher mobile data volumes, reduce latency, increase the number of connected devices and at the same time increase the energy efficiency (EE) [@Osseiran-2014-COMM; @Bjornson-2014-b-SPM]. The current network and infrastructure cannot cope with 5G requirements—fundamental changes are needed to handle future non-homogeneous deployments as well as new trends in user behavior such as high quality video streaming and future applications like augmented reality. 5G technology is supposed to evolve existing networks and at the same time integrate new dedicated solutions to meet the KPIs [@Bjornson-2014-b-SPM]. The new key concepts for 5G include massive MIMO (multiple-input multiple-output), ultra dense networks (UDN), device-to-device (D2D) communications, and huge number of connected devices, known as machine-type communications (MTC). The potential gains and properties of these different solutions have been studied individually, but the practical gains when they coexist and share network resources are not very clear so far. In this paper, we study the coexistence of two of these main concepts, namely massive MIMO and D2D communication.
Massive MIMO is a type of multiuser MIMO (MU-MIMO) technology where the base station (BS) uses an array with hundreds of active antennas to serve tens of users on the same time/frequency resources by coherent transmission processing [@Marzetta2010a; @Rusek-2013-SPM]. Massive MIMO techniques are particularly known to be very spectral efficient, in the sense of delivering high sum rates for a given amount of spectrum [@Bjornson2016a]. This comes at the price of deploying more transceiver hardware, but the solution is still likely to improve the energy efficiency of networks [@Ngo2013a; @Bjornson-2014-arxiv]. On the other hand, in a D2D communication, user devices can communicate directly with each other and the user plane data is not sent through the BS [@Doppler-2009-COMM]. D2D communications are considered for close proximity applications which have the potential to achieve high data rates with little amount of transmission energy, if interference is well-managed. In addition, D2D communications can be used to decrease the load of the core network. D2D users either have their own dedicated time/frequency resources (overlay approach) which in turn leads to elimination of the cross-tier interference between the two types of users (i.e., cellular and D2D users), or they transmit simultaneously with cellular users in the same resource (underlay approach).
We consider two network performance metrics in this work: The average sum rate (ASR) in $\mathrm{bit/s}$ and the EE which is defined as the number of bits transmitted per Joule of energy consumed by the transmitted signals and the transceiver hardware. It is well-known that these metrics depend on the network infrastructure, radio interface, and underlying system assumptions [@Tombaz-2011-WCM; @Bjornson-2014-arxiv; @Auer-2013-WCM]. The motivation behind our work is to study how the additional degrees of freedom resulting from high number of antennas in the BS can affect the ASR and EE of a multi-tier network where a D2D tier is bypassing the BS, and how a system with massive MIMO is affected by adding a D2D tier. We focus on the downlink since majority of the payload data and network energy consumption are coupled to the downlink [@Tombaz-2011-WCM]. We assume that each D2D pair is transmitting simultaneously with the BS in an underlay fashion. In addition, we assume that the communication mode of each user (i.e., D2D or cellular mode) has already been decided by higher layers.
Related Work
------------
The relation between the number of BS antennas, ASR and EE in cellular networks has been studied in [@Ngo2013a; @Yang-2013-OnlineGreenCom; @Bjornson-2014-arxiv; @Bjornson-2013-ICT] among others. The tradeoff between ASR and EE was described in [@Ngo2013a] for massive MIMO systems with negligible circuit power consumption. This work was continued in [@Yang-2013-OnlineGreenCom] where radiated power and circuit power were considered. In [@Bjornson-2014-arxiv], joint downlink and uplink design of a cellular network was studied in order to maximize EE for a given coverage area. The maximal EE was achieved by having a hundred BS antennas and serving tens of users in parallel, which matches well with the massive MIMO concept. Furthermore, the study [@Bjornson-2013-ICT] considered a downlink scenario in which a cellular network has been overlaid by small cells. It was shown that by increasing the number of BS antennas, the array gain allows for decreasing the radiated signal energy while maintaining the same ASR. However, the energy consumed by the transceiver chains increases. Maximizing the EE is thus a complicated problem where several counteracting factors need to be balanced. This stands in contrast to maximization of the ASR, which is relatively straightforward since the sum capacity is the fundamental upper bound.
There are only a few works in the D2D communication literature where the base stations have multiple antennas [@Min-2011-TWC-b; @Yu-2012-GLOBECOM; @Fodor-2011-GLOBECOM; @Shalmashi-2014a-WCNC; @Xingqin-2014-arxiv]. In [@Min-2011-TWC-b], uplink MU-MIMO with one D2D pair was considered. Cellular user equipments (CUEs) were scheduled if they are not in the interference-limited zone of the D2D user. The study [@Yu-2012-GLOBECOM] compared different multi-antenna transmission schemes. In [@Fodor-2011-GLOBECOM], two power control schemes were proposed for a multi-cell MIMO network. Two works that are more related to our work are [@Shalmashi-2014a-WCNC] and [@Xingqin-2014-arxiv]. The former investigates the mode selection problem in the uplink of a network with potentially many antennas at the BS. The impact of the number of antennas on the quality-of-service and transmit power was studied when users need to decide their mode of operation (i.e., D2D or cellular). The latter study, [@Xingqin-2014-arxiv], only employs extra antennas in the network to protect the CUEs from interference of D2D users in the uplink.
The ASR in D2D communications is mostly studied in the context of interference and radio resource management [@Shalmashi-2013-PIMRC; @Zulhasnine-2010-WiMob]. There are a few works that consider EE in D2D communications, but only for single antenna BSs, e.g., [@Yaacoub-2012-GCW; @Mumtaz-2014-ICC], and [@Wang-2013-ICC], where the first one proposed a coalition formation method, the second one designed a resource allocation scheme, and the third one aimed at prolonging the battery life of user devices.
The spatial degrees of freedom offered by having multiple antennas at BSs are very useful in the design of future mobile networks, because the spatial precoding enables dense multiplexing of users while keeping the inter-user interference under control. In particular, the performance for cell edge users, which have almost equal signal-to-noise ratios (SNRs) to several BSs, can be greatly improved since only the desired signals are amplified by the transmit precoding [@Baldemair-2013-VTM; @Bjornson2013d; @Gesbert-2007-SPM]. In order to model the random number of users and random user positions, we use mathematical tools from stochastic geometry [@Haenggi-Stochastic] which are powerful in analytically quantifying certain metrics in closed-form.
Contributions
-------------
Our main contributions in this paper can be summarized as follows:
- *A tractable model for underlaid D2D communication in massive MIMO systems*: We model a two-tier network with two different user types. The first tier users, i.e., CUEs, are served in the downlink by a BS using massive multiuser MIMO precoding to cancel interference. The second tier users, i.e., D2D users, exploit their close proximity and transmit simultaneously with the downlink cellular transmissions bypassing the BS. The number of D2D transmitters and their locations are modeled according to a homogeneous Poisson point process (PPP) while a fixed number of CUEs are randomly distributed in the network.
- *Tractable and directly computable expressions*: We derive tightly approximated expressions for the coverage probability of D2D users and CUEs. These expressions are directly used to compute our main performance metrics, namely, the ASR and EE. We verify the tightness of these approximations by Monte-Carlo simulations. Furthermore, we provide analytical insights on the behavior of these metrics for both CUEs and D2D users.
To the best of our knowledge, the energy efficiency analysis for underlay D2D communications in a network with large number of BS antennas has not been carried out before.
- *Performance analysis*: Based on extensive simulations, we characterize the typical relation between the ASR and EE metrics in terms of the number of BS antennas, the number of CUEs, and the D2D user density for a given coverage area and study the incurred tradeoffs in two different scenarios.
System Model {#sec:Sys_Mod}
============
![System model where a multi-antenna BS communicates in the downlink with multiple CUEs, while multiple user pairs communicate in D2D mode. The CUEs are distributed uniformly in the coverage area and the D2D users are distributed according to a PPP. The D2D users that are outside the coverage area are only considered as interferers.[]{data-label="figure:sysMod_massiveMIMO_D2D"}](./figures/sysMod_massiveMIMO_D2D){width="0.6\columnwidth"}
We consider a single-cell scenario where the BS is located in the center of the cell and its coverage area is a disc of radius $R$. The BS serves $U_c$ single-antenna CUEs which are uniformly distributed in the coverage area. These are simultaneously served in the downlink using an array of $T_c$ antennas located at the BS. It is assumed that $1 \leq U_c \leq T_c$ so that the precoding can be used to control the interference caused among the CUEs [@Bjornson-2014-a-SPM].
In addition to the CUEs, there are other single-antenna users that bypass the BS and communicate pairwise with each other using a D2D communication mode. The locations of the D2D transmitters (D2D Tx) are modeled by a homogeneous PPP $\Phi$ with density $\lambda_d$ in $\mathbb{R}^2$.[^5] This means that the average number of D2D Tx per unit area is $\lambda_d$ and these users are uniformly distributed in that area. The D2D receiver (D2D Rx) is randomly located in an isotropic direction with a fixed distance away from its corresponding D2D Tx—a model that is similar to the one considered in [@Lee-2014-JSAC]. The system setup is illustrated in Fig. \[figure:sysMod\_massiveMIMO\_D2D\].
Let $R_{k,j}$ denote the distance between the $j$-th D2D Tx to the $k$-th D2D Rx. The performance analysis for D2D users is carried out for a typical D2D user, which is denoted by the index $0$. The typical D2D user is an arbitrary D2D user located in the cell and its corresponding receiver is positioned in the origin. The results for a typical user show the statistical average performance of the network [@Haenggi-Stochastic]. Therefore, for any performance metric derivation, the D2D users inside the cell are considered and the ones outside the cell are only taken into account as sources of interference. Note that we neglect potential interference from other BSs and leave the multi-cell case for future work. This is because the interference from D2D transmissions is likely to be much stronger than the interference from other BSs. We assume equal power allocation for both CUEs and D2D users. Let $P_c$ denote the total transmit power of the BS, then the transmit power per CUE is $\frac{P_c}{U_c}$. The transmit power of the D2D Tx is denoted by $P_d$.
Let ${\mathbf{h}}_j \in \mathbb{C}^{T_c \times 1}$ be the normalized channel response between the BS and the $j$-th CUE, for $j \in \{0, \dots, U_c-1 \}$. These channels are modeled as Rayleigh fading such that ${\mathbf{h}}_j \sim\mathcal{CN}(\mathbf{0},\mathbf{I})$, where $\mathcal{CN}(\cdot,\cdot)$ denotes a circularly symmetric complex Gaussian distribution. Perfect instantaneous channel state information (CSI) is assumed in this work for analytic tractability, but imperfect CSI is a relevant extension. Linear downlink precoding is considered at the BS based on the zero-forcing (ZF) scheme that cancels the interference between the CUEs [@Bjornson-2014-a-SPM]. The precoding matrix is denoted by $\mathbf{V}=[{\mathbf{v}}_0, \dots, {\mathbf{v}}_{Uc-1}]\in \mathbb{C}^{T_c \times U_c}$ in which each column ${\mathbf{v}}_j$ is the normalized transmit precoding vector assigned to the CUE $j$. Let ${\mathbf{f}}_{0,\textrm{BS}} \in \mathbb{C}^{T_c \times 1}$ be the channel response from the BS to D2D Rx and let it be Rayleigh fading as ${\mathbf{f}}_{0,\textrm{BS}} \sim \mathcal{CN}(\mathbf{0},\mathbf{I})$. Moreover, let $r_j \in \mathbb{C}$ and ${\mathbf{s}} \in \mathbb{C}^{U_c \times 1}$ denote the transmitted data signals intended for a D2D Rx and the CUEs, respectively. Since each user requests different data, the transmitted signals can be modeled as zero-mean and uncorrelated with $\mathbb{E}\big[|r_j|^2\big]=P_d$ and $\mathbb{E}\big[||{\mathbf{s}}||^2\big]=P_c$. The fading channel response between the $j$-th D2D Tx and the $k$-th D2D Rx is denoted by $g_{k,j} \in \mathbb{C}$ where $g_{k,j} \sim\mathcal{CN}(0,1)$. Moreover, $R_{0,\textrm{BS}}$ denotes the random distance between the typical D2D Rx and the BS. The pathloss is modeled as $A_i d^{-\alpha_i}$ with $i \in \{c,d\}$, where index $c$ indicates the pathloss between a user and the BS and index $d$ gives the pathloss between any two users. $A_i$ and $\alpha_i$ are the pathloss coefficient and exponent, respectively, where we assume $\alpha_i>2$. The received signal at the typical D2D Rx is $$\begin{aligned}
y_{d,0} &= \sqrt{A_d } R_{0,0}^{-\alpha_d/2} g_{0,0} r_0 +\underbrace{ \sqrt{A_c } R_{0,\textrm{BS}}^{-\alpha_c/2} {\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V} {\mathbf{s}} }_{\textrm{Interference from the BS}} + \underbrace{\sqrt{A_d } \sum_{j\neq 0} R_{0,j}^{-\alpha_d/2} g_{0,j} r_j}_{\textrm{Interference from other D2D users}} + \eta_d,\end{aligned}$$ where $\eta_d$ is zero-mean additive white Gaussian noise with power $N_0 = \tilde{N}_0 B_w$, $\tilde{N}_0$ is the power spectral density of the white Gaussian noise, and $B_w$ is the channel bandwidth. For given channel realizations, the signal-to-interference-plus-noise ratio (SINR) at the typical D2D Rx is $$\begin{aligned}
\label{eq:SINR_d2d}
\mathrm{SINR}_d = \frac{P_d R_{0,0}^{-\alpha_d}|g_{0,0}|^2}{I_{\textrm{BS},0} + I_{d,0} + \frac{N_0}{A_d}},\end{aligned}$$ in which both the numerator and the denominator have been normalized by $A_d$. $I_{\textrm{BS},0}$ is the received interference power from the BS and $I_{d,0}$ is the received interference power from other D2D users that transmit simultaneously which are defined as $$\begin{aligned}
I_{\textrm{BS},0} &\triangleq \frac{\zeta R_{0,\textrm{BS}}^{-\alpha_c}}{A_d}\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2, \label{eq:I_BS0} \\
I_{d,0} &\triangleq \sum_{j\neq 0} P_d R_{0,j}^{-\alpha_d} |g_{0,j}|^2, \label{eq:I_d0}\end{aligned}$$ where $$\zeta \triangleq A_c \frac{P_c}{U_c}. \label{eq:zeta-def}$$
Let $D_{0,k}$ and $e_{0,k}\in \mathbb{C}$ with $e_{0,k}\sim\mathcal{CN}(0,1)$ be the distance and fading channel response between a typical CUE and the $k$-th D2D Tx, respectively, and let $D_{0,\textrm{BS}}$ denote the distance between a typical CUE and the BS. Then, the received signal at the typical CUE is $$\begin{aligned}
y_{c,0} &= \sqrt{A_c} D_{0,\textrm{BS}}^{-\alpha_c/2} {\mathbf{h}}_{0}^H \mathbf{V} {\mathbf{s}} + \underbrace{\sqrt{A_d } \sum_{j} D_{0,j}^{-\alpha_d/2} e_{0,j} r_j}_{\textrm{Interference from all D2D users}} + \eta_c,\end{aligned}$$ where $\eta_c$ is zero-mean additive white Gaussian noise with power $N_0$. Then, the corresponding SINR for the typical CUE is $$\begin{aligned}
\mathrm{SINR}_c = \frac{|{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2}{\frac{A_d }{\zeta} D_{0,\textrm{BS}}^{\alpha_c}(I_{d,c} + \frac{N_0}{A_d })}, \label{eq:SINR_CUE}\end{aligned}$$ where $$\label{eq:inf_dc}
I_{d,c} \triangleq \sum_{j} P_d D_{0,j}^{-\alpha_d} |e_{0,j}|^2$$ is the received interference power from all D2D users (normalized by $A_d$).
Performance Analysis {#sec:performance_analysis}
====================
In this section, we first introduce the performance metrics that are considered in this paper. Then we proceed to derive the coverage probability for both CUEs and D2D users which are needed to compute these metrics.
Performance Metrics {#sec:metrics}
-------------------
In this paper, two main performance metrics for the network are considered: the average sum rate (ASR) and energy efficiency (EE). The ASR is obtained from total rates of both D2D users and CUEs as $$\text{ASR}= U_c \bar{R}_c + \pi R^2 \lambda_d \bar{R}_d,
\label{eq:ASR}$$ where $\pi R^2 \lambda_d$ is the average number of D2D users in the cell and $\bar{R}_t$ with $t\in\{c,d\}$ denotes the average rates of the CUEs and D2D users, respectively. $\bar{R}_t$ for both cellular and D2D users is computed as the successful transmission rate by $$\label{eq:Avg_Rate1}
\bar{R}_t = \underset{\beta_t \geq 0}{\mathrm{sup}}~B_w \log_{2}(1 + \beta_t) \mathrm{P}^t_{\mathrm{cov}} (\beta_t)$$ where $$\label{eq:PsuccDef}
\mathrm{P}^t_{\mathrm{cov}} (\beta_t) = \mathrm{Pr}\big\{\mathrm{SINR}_{t} \geq \beta_t \big\}$$ is the coverage probability when the received SINR is higher than a specified threshold $\beta_t$ needed for successful reception. Note that $\mathrm{SINR}_{t}$ contains random channel fading and random user locations. Finding the supremum guarantees the best constant rate for the D2D users and the CUEs. If we know the coverage probability ($\mathrm{P}^t_{\mathrm{cov}} (\beta_t)$), can easily be computed by using line search for each user type independently. Moreover, is easily achievable in practice since the modulation and coding is performed without requiring that every transmitter knows the interference characteristics at its receiver.
Energy efficiency is defined as the benefit-cost ratio between the ASR and the total consumed power: $$\text{EE} = \frac{\text{ASR}}{\text{Total power}}. \label{eq:EE}$$ For the total power consumption, we consider a detailed model described in [@Bjornson-2014-arxiv]: $$\begin{aligned}
\text{Total power}&=\frac{1}{\eta} \big(P_c +\lambda_d \pi R^2 P_d\big) + C_0 + T_c C_1 + \big(U_c + 2 \lambda_d \pi R^2\big)C_2,\label{eq:P_tot}\end{aligned}$$ where $P_c +\lambda_d \pi R^2 P_d$ is the total transmission power averaged over the number of D2D users, $\eta$ is the amplifier efficiency ($0 < \eta \leq 1$), $C_0$ is the load independent power consumption at the BS, $C_1$ is the power consumption per BS antenna, $C_2$ is the power consumption per user device, and $U_c + 2 \lambda_d \pi R^2$ is the average number of active users.
In order to calculate the ASR and EE, we need to derive the coverage probability for both cellular and D2D users. The analytic derivation of these expressions is one of the main contributions of this paper.
Coverage Probability of D2D Users
---------------------------------
We first derive the expression for the coverage probability of D2D users.
\[proposition:P\_succ\_d2d\] The approximate coverage probability for a typical D2D user is given by $$\begin{aligned}
\mathrm{P}^d_{\mathrm{cov}}(\beta_d) &= \frac{ (\kappa\beta_d)^{2/\alpha_c}}{R^2}\left(y^{U_c + \frac{2}{\alpha_c}-1} (1-y)^{- \frac{2}{\alpha_c}} - \Big(U_c + \frac{2}{\alpha_c}-1\Big) \mathcal{B}\Big(y; U_c + \frac{2}{\alpha_c}-1, 1-\frac{2}{\alpha_c}\Big) \right) \notag \\
&\;\quad \cdot \exp\bigg(- \frac{\pi \lambda_d R_{0,0}^2}{\mathrm{sinc}(\frac{2}{\alpha_d})} \beta_d^{2/\alpha_d}\bigg) \exp\bigg(-\frac{\beta_d} {\bar{\gamma}_d}\bigg),
\label{eq:P_succ_d2d}\end{aligned}$$ where $\kappa \triangleq \frac{\zeta}{P_d A_d R_{0,0}^{-\alpha_d}}$ with $\zeta$ defined in , $y \triangleq \frac{1}{\kappa \beta_d R^{-\alpha_c} + 1}$, $\mathrm{sinc}(x) = \frac{\sin(\pi x)}{\pi x}$, $\bar{\gamma}_d = \frac{A_d R_{0,0}^{-\alpha_d}P_d}{N_0}$ is the average D2D SNR, and $\mathcal{B}(x;a,b)$ is the incomplete Beta function.
The proof is given in Appendix \[sec:proof\_Pcov\_d2d\].
The coverage probability expression in Proposition \[proposition:P\_succ\_d2d\] allows us to compute the average data rate of a typical D2D user in . We note that is actually a tight approximation and its tightness is evaluated in Section \[sec:results\]. From the expression in , we make several observations as listed below.
\[remark:P\_succ\_d2d\_highSNR\] In the high-SNR regime for the D2D users where $\bar{\gamma}_d \gg \beta_d$, the last term in converges to one, i.e., $\exp\left(- \frac{\beta_d} {\bar{\gamma}_d}\right) \to 1$, and we have $$\begin{aligned}
\mathrm{P}^d_{\mathrm{cov}}(\beta_d) &= \frac{ (\kappa\beta_d)^{2/\alpha_c}}{R^2}\left(y^{U_c + \frac{2}{\alpha_c}-1} (1-y)^{- \frac{2}{\alpha_c}} - \Big(U_c + \frac{2}{\alpha_c}-1\Big) \mathcal{B}\Big(y; U_c + \frac{2}{\alpha_c}-1, 1-\frac{2}{\alpha_c}\Big) \right) \notag \\
&\;\quad \cdot \exp\bigg(- \frac{\pi \lambda_d R_{0,0}^2}{\mathrm{sinc}(\frac{2}{\alpha_d})} \beta_d^{2/\alpha_d}\bigg).
\label{eq:P_succ_d2d_highSNR}\end{aligned}$$ This can also be referred to as the interference-limited regime.
\[remark:P\_succ\_d2d\_lambda\_d\] The coverage probability of a typical D2D user is a decreasing function of the D2D density $\lambda_d$. Because higher $\lambda_d$ results in more interference among D2D users. In particular, it can be seen that $\mathrm{P}^d_{\mathrm{cov}}$ in is a function of $\lambda_d$ through $\exp(-C\lambda_d)$ with $C\triangleq\frac{\pi R_{0,0}^2 \beta_d^{2/\alpha_d}}{\mathrm{sinc}(\frac{2}{\alpha_d})} > 0$. Thus, if $\lambda_d \to \infty$, $\mathrm{P}^d_{\mathrm{cov}} \to 0$.
Recall that in our model, the D2D Rx is associated to the D2D Tx which is located at a fixed distance away. However, if we had assumed that the D2D Rx’s association to a D2D Tx is based on, for example, the shortest distance or the maximum SINR, then the $\mathrm{P}^d_{\mathrm{cov}}$ would have been unaffected by the D2D density (in the high-interference regime).
Now, considering the number of BS antennas or the number of CUEs as variables, we have the following behavior of the D2D coverage probability.
\[remark:P\_succ\_d2d\_Tc\_Uc\] $\mathrm{P}^d_{\mathrm{cov}}$ is not affected by the number of BS antennas $T_c$. The BS antennas are used to cancel out the interference among CUEs and they do not have any impact on D2D users’ performance as long as the number of CUEs $U_c$ is constant and does not vary with the number of BS antennas $T_c$. The coverage probability of a typical D2D user $\mathrm{P}^d_{\mathrm{cov}}$ is a decreasing function of $U_c$. However, increasing the number of CUEs have a small effect on D2D users’ performance. This is due to the fact that the resulting interference from the BS to D2D users does not change significantly by increasing the number of CUEs as the transmit power of the BS is the same irrespective of the number of users and the precoding is independent of the D2D channels. Thus, a change of $U_c$ will only change the distribution of the interference but not its average.
Next we comment on how changes in the transmit powers of the BS and D2D Tx as well as the distance between D2D user pairs affect the coverage probability of D2D users.
\[remark:P\_succ\_d2d\_PcPd\] $\mathrm{P}^d_{\mathrm{cov}}$ is a decreasing function of the ratio between the transmit power of the BS and of the D2D users, i.e., $\frac{P_c}{P_d}$, which is part of the first term in and corresponds to the interference from the BS. For instance, if we fix $P_c$ and decrease $P_d$, the coverage probability for D2D users decreases as the interference from the BS would be the dominating factor. At the same time, if we decrease $P_c$, it would improve the coverage of D2D users.
\[remark:P\_succ\_d2d\_Uc\] $\mathrm{P}^d_{\mathrm{cov}}$ is a decreasing function of the distance between D2D Tx-Rx pairs $R_{0,0}$ and the cell radius $R$. Increasing the cell radius with the same D2D user density reduces the effect of the interference from the BS. Also by decreasing the distance between D2D Tx-Rx pairs, it is evident that a better performance for D2D users can be obtained.
Using Proposition \[proposition:P\_succ\_d2d\], the following corollary provides the optimal D2D user density that maximizes the D2D ASR, i.e., $\pi R^2 \lambda_d \bar{R}_d$, where $\bar{R}_d$ is given in .
\[corollary:ASR\_d2d\_lambda\_d\_max\] For a given SINR threshold $\beta_d$, the optimal density of D2D users $\lambda_d^*$ that maximizes the D2D ASR is $$\begin{aligned}
\lambda_d^*(\beta_d) = \frac{\mathrm{sinc}(\frac{2}{\alpha_d})}{\pi R_{0,0}^2}\beta_d^{-2/\alpha_d}. \label{eq:lambda_max}\end{aligned}$$
Given the SINR threshold $\beta_d$ and using –, the D2D ASR is $$\pi R^2 \lambda_d B_w\log_2 (1+\beta_d) \mathrm{P}^d_{\mathrm{cov}}(\beta_d), \label{eq:D2D_ASR_optLambda_d}$$ where $\mathrm{P}^d_{\mathrm{cov}}(\beta_d)$ is given in and depends on $\lambda_d$ through an exponential function. Taking the derivative of with respect to $\lambda_d$ and setting it to zero yields the optimal D2D user density $\lambda_d^*(\beta_d)$ given in that maximizes the D2D ASR.
Coverage Probability of Cellular Users
--------------------------------------
Next, we compute the coverage probability for CUEs.
\[proposition:P\_succ\_cue\] The coverage probability for a typical cellular user is given by $$\begin{aligned}
\mathrm{P}^c_{\mathrm{cov}}(\beta_c) &= \mathbb{E}_{D_{0,\textrm{BS}}}\left[e^{-\frac{N_0}{A_d}s} \sum_{k=0}^{T_c - U_c} \frac{s^k}{k!} \sum_{i=0}^{k} \binom{k}{i} \left(\frac{N_0}{A_d }\right)^{k-i}(-1)^i
\;\Upsilon(\lambda_d,s,i)\right],
\label{eq:P_succ_cue}\end{aligned}$$ with $$\begin{aligned}
\Upsilon(\lambda_d,s,i) &= \exp\left(-C_d \lambda_d s^{2/\alpha_d} \right) \sum_{(j_1,\ldots,j_i)\in\mathcal{J}} i! \prod_{\ell=1}^i\frac{1}{j_\ell!(\ell!)^{j_\ell}}\left(-C_d \lambda_d s^{\frac{2}{\alpha_d}-\ell}\prod_{q=0}^{\ell-1}\Big(\frac{2}{\alpha_d}-q\Big)\right)^{j_\ell}, \label{eq:Upsilon}\end{aligned}$$ where $s\triangleq \frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c}\beta_c$ with $\zeta$ defined in , $C_d \triangleq \frac{\pi P_d^{2/\alpha_d}}{\mathrm{sinc}(\frac{2}{\alpha_d})}$, and $$\mathcal{J}\triangleq \bigg\{(j_1,\ldots,j_i): j_\ell \in \mathbb{Z}_{\geq 0}, \; \sum_{\ell=1}^i \ell j_\ell = i \bigg\}.$$
The proof is given in Appendix \[sec:proof\_Pcov\_cue\].
This proposition gives an expression for the coverage probability of CUEs in which there is only one random variable left. The expectation in with respect to $D_{0,\textrm{BS}}$ is intractable to derive analytically but can be computed numerically. The analytical results of Proposition \[proposition:P\_succ\_d2d\] and Proposition \[proposition:P\_succ\_cue\] have been verified by Monte-Carlo simulations in Section \[sec:results\]. A main benefit of the analytic expressions (as compared to pure Monte-Carlo simulations with respect to all sources of randomness) is that they can be computed much more efficiently, which basically is a prerequisite for the multi-variable system analysis carried out in Section \[sec:results\].
Next, we present some observations from the result in Proposition \[proposition:P\_succ\_cue\] as follows.
\[remark:P\_succ\_cue\_inflim\] In the interference-limited regime where where $I_{d,c}\gg N_0$, the coverage probability in for a typical cellular user is simplified to $$\begin{aligned}
\mathrm{P}^c_{\mathrm{cov}}(\beta_c) &= \mathbb{E}_{D_{0,\textrm{BS}}}\left[\sum_{k=0}^{T_c - U_c} \frac{(-s)^k}{k!}\Upsilon(\lambda_d,s,k)\right].
\label{eq:P_succ_cue_inflim}\end{aligned}$$
The result obtained in Remark \[remark:P\_succ\_cue\_inflim\] has a lower computational complexity compared to the expression in Proposition \[proposition:P\_succ\_cue\] and at the same time it is a tight approximation for Proposition \[proposition:P\_succ\_cue\]. This can be observed from the denominator of the where the term $\frac{N_0}{A_d }\approx 0$.
\[remark:P\_succ\_cue\_lambda\_dinf\] The coverage probability of a typical CUE $\mathrm{P}^c_{\mathrm{cov}}(\beta_c)$ is a decreasing function of the D2D user density $\lambda_d$. From Proposition \[proposition:P\_succ\_cue\], only $\Upsilon(\lambda_d,s,i)$ is a function of $\lambda_d$ which is composed of an exponential term in $\lambda_d$ multiplied by a polynomial term in $\lambda_d$. Thus, if $\lambda_d \to \infty$, the exponential term which has a negative growth dominates the polynomial term and $\mathrm{P}^c_{\mathrm{cov}}(\beta_c) \to 0$.
We proceed to analyze the behavior of Proposition \[proposition:P\_succ\_cue\] by considering a number of special cases.
\[corollary:col\_P\_succ\_cue\_TcequalUc\] If $T_c = U_c$, the coverage probability for a typical cellular user is given by $$\begin{aligned}
\mathrm{P}^c_{\mathrm{cov}}(\beta_c) &= \mathbb{E}_{D_{0,\textrm{BS}}}\left[\exp\bigg(-\frac{N_0}{A_d}s - C_d \lambda_d s^{2/\alpha_d}\bigg)\right], \label{eq:P_succ_cue_TcequalUc}\end{aligned}$$ where $s = \frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c} \beta_c$ and $C_d=\frac{\pi P_d^{2/\alpha_d}}{\mathrm{sinc}(\frac{2}{\alpha_d})}$.
follows directly from by setting $T_c-U_c = 0$.
\[corollary:P\_succ\_cue\_TcUcinf\] If $(T_c - U_c) \to \infty$, the coverage probability for a typical cellular user tends to one, that is, $$\begin{aligned}
\lim_{(T_c - U_c) \to \infty}~\mathrm{P}^c_{\mathrm{cov}}(\beta_c) = 1.
\label{eq:P_succ_cue_TcUcinf}\end{aligned}$$
Let $m=T_c - U_c$. Substituting $\mathrm{SINR}_c$ from into , we have $$\begin{aligned}
\lim_{m \to \infty}~\mathrm{P}^c_{\mathrm{cov}}(\beta_c) &= \lim_{m \to \infty}~\mathrm{Pr}\left\{ |{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2 \geq \frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c}\Big(I_{d,c} + \frac{N_0}{A_d}\Big)\beta_c \right\} \\
&{\overset{(a)}{=}} \lim_{m \to \infty}~\mathbb{E}_{D_{0,\textrm{BS}}, I_{d,c}}\left[e^{-\frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c}(I_{d,c} + \frac{N_0}{A_d})\beta_c } \sum_{k=0}^{m} \frac{1}{k!}\bigg(\frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c}\Big(I_{d,c} + \frac{N_0}{A_d}\Big)\beta_c\bigg)^k\right] \\
&{\overset{(b)}{=}} \lim_{m \to \infty}~\mathbb{E}_{D_{0,\textrm{BS}}, I_{d,c}}\left[e^{-z} \sum_{k=0}^{m} \frac{z^k}{k!}\right] \\
&{\overset{(c)}{=}} \mathbb{E}_{D_{0,\textrm{BS}}, I_{d,c}}\left[\lim_{m \to \infty} e^{-z} \sum_{k=0}^{m} \frac{z^k}{k!}\right] \\
&{\overset{(d)}{=}} \mathbb{E}_{D_{0,\textrm{BS}}, I_{d,c}}\left[e^{-z} e^{z} \right] = 1,\end{aligned}$$ where $(a)$ follows from the CCDF of $|{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2$ with $2|{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2 \sim \chi^2_{2}$ given $D_{0,\textrm{BS}}$ and $I_{d,c}$. Step $(b)$ follows from setting $z = \frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c}(I_{d,c} + \frac{N_0}{A_d})\beta_c$. Step $(c)$ is obtained from the dominated convergence theorem which allows for an interchange of limit and expectation and step $(d)$ is due to the fact that $\sum_{k=0}^{\infty} \frac{z^k}{k!} = e^z$.
In the results so far, we have discussed the case where there exist some D2D users as underlay to the cellular network, that is, $\lambda_d \neq 0$, However, it is interesting to see what can be achieved without D2D users.
\[corollary:col\_P\_succ\_cue\_lambda\_dzero\] If $\lambda_d= 0$, the coverage probability for a typical cellular user is given by $$\begin{aligned}
\mathrm{P}^c_{\mathrm{cov}}(\beta_c) &= \frac{2}{\alpha_c R^2}\Gamma\left(\frac{2}{\alpha_c}\right)\left(\frac{N_0}{\zeta} \beta_c \right)^{-2/\alpha_c}
\sum_{k=0}^{T_c-U_c}\binom{\frac{2}{\alpha_c}+k-1}{k}, \label{eq:P_succ_cue_lambda_dzero}\end{aligned}$$ where $\Gamma(\cdot)$ is the Gamma function and $\zeta$ is defined in .
Substituting $\mathrm{SINR}_c$ from into and setting $\lambda_d= 0$, we have $$\begin{aligned}
\mathrm{P}^c_{\mathrm{cov}}(\beta_c) &= \mathrm{Pr}\left\{ |{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2 \geq \frac{D_{0,\textrm{BS}}^{\alpha_c}}{\zeta} N_0 \beta_c \right\} \nonumber\\
&{\overset{(a)}{=}} \mathbb{E}_{z}\Bigg[\sum_{k=0}^{T_c-U_c}\frac{l^k}{k!}z^k e^{-lz} \Bigg] \nonumber\\
&{\overset{(b)}{=}}\frac{2}{\alpha_c R^2}\Gamma\left(\frac{2}{\alpha_c}\right)
\sum_{k=0}^{T_c-U_c}\frac{(-l)^k}{k!} ~\frac{\mathrm{d}^k}{{\mathrm{d}l}^k} ~l^{-2/\alpha_c},\end{aligned}$$ where $(a)$ follows from the CCDF of $|{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2$ with $2|{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2 \sim \chi^2_{2}$ given $D_{0,\textrm{BS}}$ and setting $l= \frac{N_0}{\zeta} \beta_c$ and $z = D_{0,\textrm{BS}}^{\alpha_c}$ with PDF $f(z)=\frac{2}{\alpha_cR^2}z^{\frac{2}{\alpha_c}-1}$. Step $(b)$ follows from taking the expectation with respect to $z$ which is similar to the expression in with the Laplace transform $ \mathcal{L}_{z}(l) = \frac{2}{\alpha_c R^2}\Gamma\big(\frac{2}{\alpha_c}\big)l^{-2/\alpha_c}$. Simplifying the $k$-th derivative to $\frac{\mathrm{d}^k}{\mathrm{d}l^k} ~l^{-2/\alpha_c} = (-1)^k l^{-\frac{2}{\alpha_c}-k} \prod_{i=0}^{k-1}\big(\frac{2}{\alpha_c}+i\big)$ and using the identity $\frac{1}{k!}\prod_{i=0}^{k-1}\big(\frac{2}{\alpha_c}+i\big) = \binom{\frac{2}{\alpha_c}+k-1}{k}$, follows.
The closed-form results in Corollary \[corollary:col\_P\_succ\_cue\_lambda\_dzero\] for $\lambda_d= 0$ depends only on noise rather than interference and perhaps can result in higher ASR for CUEs. The ASR for $\lambda_d>0$ also depends on noise but its impact is much smaller. However, we note that this result is obtained for a single cell scenario. Thus, comparing Proposition \[proposition:P\_succ\_cue\] and Corollary \[corollary:col\_P\_succ\_cue\_lambda\_dzero\] and evaluating the potential performance gain/loss due to introducing D2D communications would make more sense in a multi-cell scenario.
Using the results from Proposition \[proposition:P\_succ\_d2d\] and Proposition \[proposition:P\_succ\_cue\], we proceed to evaluate the network performance in terms of the ASR and EE from and , respectively.
Numerical Results {#sec:results}
=================
Description Parameter Value
-------------------------------------- ------------ ------------
D2D TX power $P_d$ $6$ dBm
BS TX power $P_c$ $30$ dBm
Cell radius $R$ $500$ m
Bandwidth $B_w$ $20$ MHz
Thermal noise power $N_0$ $-131$ dBm
Noise figure in UE $F$ $5$ dB
Carrier frequency $f_c$ $2$ GHz
D2D pair distance $R_{0,0}$ $35$ m
Pathloss exponent betw. devices $\alpha_d$ $3$
Pathloss exponent betw. BS–device $\alpha_c$ $3.67$
Pathloss coefficient betw. devices $A_d$ $38.84$ dB
Pathloss coefficient betw. BS–device $A_c$ $30.55$ dB
Amplifier efficiency $\eta$ $0.3$
Load-independent power in BS $C_0$ $5$ W
Power per BS antenna $C_1$ $0.5$ W
Power per UE handset $C_2$ $0.1$ W
Monte-Carlo runs MC $5000$
: System and simulation parameters.[]{data-label="table:Sim_param"}
In this section, we assess the performance of the setup in Fig. \[figure:sysMod\_massiveMIMO\_D2D\] in terms of ASR and EE using numerical evaluations. As we pointed out in , many parameters affect these performance metrics. Initially, we consider the EE and the ASR as functions of three key parameters, namely, the number of BS antennas $T_c$, the density of D2D users $\lambda_d$, and the number of cellular users $U_c$. We show the individual effect of these system parameters on the two performance metrics while other parameters such as BS transmit power $P_c$, D2D transmit power $P_d$, and distance between D2D Tx-Rx pair $R_{0,0}$ are fixed. Later on, we also comment on the choice of these fixed parameters. The system and simulation parameters are given in Table \[table:Sim\_param\].
Before we proceed to the performance evaluation, we verify the analytical results of Proposition \[proposition:P\_succ\_d2d\] and Proposition \[proposition:P\_succ\_cue\] by Monte-Carlo simulations. As depicted in , simulation results closely follow the analytical derivations. The small gap in Fig. \[fig:P\_cov\_d2d\] is due to the spatial interference correlation resulting from the fact that multiple interfering streams are coming from the same location, hence, the Chi-squared distribution in is an approximation. This is a quite standard approximation in analyzing MIMO systems [@Dhillon-2013-TWC]. Moreover, in the simulations, the locations of the D2D Tx are generated in an area with radius $10R$ according to the PPP as opposed to our analytical assumption that they are located in the whole $\mathbb{R}^2$ region. This assumption reduces the interference as compared to our analytical results and thus improves the coverage probability as can be seen in Fig. \[fig:P\_cov\_d2d\].
We consider two scenarios corresponding to the number of CUEs $U_c$ in our evaluations. First, we assume that $U_c$ is chosen as a function of the number of BS antennas $T_c$. Then, we move on to the case where we fix the number of CUEs and study the tradeoffs among other parameters. Both scenarios are relevant in the design of massive MIMO systems. In order to speed up the numerical computations, we neglected the terms that are very small.
Number of CUEs as a Function of the Number of BS Antennas
---------------------------------------------------------
In this scenario, we assume that there is a fixed ratio between the number of CUEs $U_c$ and the number of BS antennas $T_c$. We assume this ratio to be $\frac{T_c}{U_c}=5$. Simply put, to serve one additional user, we add five more antennas at the BS since the main gains from massive MIMO come from multiplexing of many users rather than only having many antennas.
shows the ASR as a function of the density of D2D users $\lambda_d$ and the number of CUEs $U_c$, which is scaled by $T_c$. It is observed that increasing $U_c$, or equivalently $T_c$, always increases the ASR. In contrast, there is an optimal value of $\lambda_d$ as derived in Corollary \[corollary:ASR\_d2d\_lambda\_d\_max\] which results in the maximum ASR for all values of $U_c$ and appears approximately at $\lambda_d = 10^{-4}$. However, there is a difference in the shape of the ASR between the lower and higher values of $U_c$. In order to clarify this effect, we plot the ASR versus $\lambda_d$ in a 2-D plot with $U_c \in \{1,14\}$ equivalent to $T_c \in \{5,70\}$ in .
![ASR $\mathrm{[Mbit/s]}$ as a function of the number of CUEs $U_c$ and the D2D user density $\lambda_d$ for a fixed ratio $\frac{T_c}{U_c}= 5$.[]{data-label="fig:ASR_ld_uc_phi5"}](ASR_ld_uc_phi5){width="0.65\columnwidth"}
As seen in , for $U_c = 1$ user and $T_c =5$ antennas, the rate contributed from the CUEs to the sum rate is low as there is only one CUE. This rate is in a comparable level as the contribution of D2D users sum rate to the total ASR. Adding D2D users to the network (i.e., increasing $\lambda_d$), which may cause interference, will nevertheless leads to an increase in the ASR. This increase in the ASR continues until reaching a certain density that gives the maximum ASR. By further increasing $\lambda_d$, the interference between D2D users reduces their coverage probability as previously observed in Remark \[remark:P\_succ\_d2d\_lambda\_d\]. This limits the per link data rate and even a high number of D2D users cannot compensate for the D2D rate loss. At the same time, increasing $\lambda_d$ tremendously affects the CUEs sum rate (cf. Remark \[remark:P\_succ\_cue\_lambda\_dinf\]). Consequently, as $\lambda_d$ increases, the ASR decreases.
By increasing the number of CUEs and BS antennas to $U_c = 14$ users and $T_c = 70$ antennas, respectively, in , the average rates of the CUEs become higher than the case with $U_c = 1$ user and $T_c =5$ antennas as expected from Corollary \[corollary:P\_succ\_cue\_TcUcinf\] and the multiplexing gain from having many CUEs. However, by introducing a small number of D2D users, there is a substantial probability that the interference from the D2D users reduces the CUEs’ rates per link as observed in Remark \[remark:P\_succ\_cue\_lambda\_dinf\]. The reduction in these rates are not compensated in the ASR by the contribution of the D2D users’ rates. Note that, as we stated in Remark \[remark:P\_succ\_d2d\_Tc\_Uc\], when $U_c$ is scaled with $T_c$, it impacts the D2D coverage probability, but the decrease in the performance of D2D users is not significant. Furthermore, if we keep increasing $\lambda_d$, even though the rate per link decreases for both CUEs and D2D users, there is a local minima after which the aggregate D2D rate over all D2D users becomes higher and the ASR increases again. The second turning point follows from the same reasoning as for the case of $U_c=1$ user and $T_c = 5$ antennas, i.e., in higher D2D densities, the interference from D2D users are the limiting factor for the ASR. This effect can also be observed in where the ASR performance is depicted versus different number of CUEs (and BS antennas) for two D2D densities. At the lower density, the ASR is linearly increasing with $U_c$ (and $T_c)$, however, in the interference-limited regime (higher $\lambda_d$), increasing the number of CUEs and BS antennas do not impact the network ASR performance.
The reasoning in and can be well understood from which explains the tradeoff between the ASR of CUEs and D2D users in the network. In the scenario in which we have $T_c = 70$ antennas and $U_c =14$ users, the cellular network contributes more to the total ASR for the low D2D density regime (e.g., $\lambda_d=10^{-6}$) due to high number of CUEs and BS antennas. In this region, the ASR gains from massive MIMO is large. By increasing $\lambda_d$, the gain from massive MIMO vanishes as the interference added by the D2D users dominates and degrades the performance that was achieved by interference cancellation between CUEs. Therefore, with medium D2D user density, if there is a fixed rate constraint for CUEs, the network can still benefit (from the ASR perspective) from underlay D2D communications. However, in the high D2D density regime (e.g., $\lambda_d = 10^{-4}$), the cellular ASR is too small and it is better that the cellular and D2D tiers use the overlay approach for communication instead of the underlay approach.
In , we show the network performance in terms of the EE as a function of the parameters $\lambda_d$ and $U_c$ with $\frac{T_c}{U_c}= 5$. It is observed that the EE is a decreasing function of $U_c$ and $T_c$. In contrast, there is a maximum point in the EE based on different values of $\lambda_d$. To study this result further, similar to the ASR, we first plot the EE versus $\lambda_d$ for $U_c \in \{1,14\}$ and $T_c \in \{4,70\}$ in . We can see that the pattern for both low and high number of BS antennas are similar to . The higher EE is achieved with $U_c =1$ user and $T_c=5$ antennas as opposed to $U_c =14$ users and $T_c=70$ antennas. This is because the extra circuit power of the cellular tier with $U_c =14$ users and $T_c=70$ antennas does not bring any substantial ASR improvement over the case with $U_c =1$ user and $T_c=5$ antennas.
![Cellular ASR vs. D2D ASR $\mathrm{[Mbit/s]}$ for a fixed ratio $\frac{T_c}{U_c}= 5$. The curves are obtained by varying the value of $\lambda_d$ from $10^{-6}$ to $10^{-2}$.[]{data-label="fig:asrC_asrD_ld"}](asrC_asrD_ld_phi5){width="0.65\columnwidth"}
Furthermore, if we plot the EE versus $U_c$, we see a different behavior for low and high D2D densities. illustrates that in the low D2D density regime ($\lambda_d=10^{-6}$), even though the ASR increases linearly, the EE almost stays the same as the number of CUEs, and correspondingly the number of BS antennas, increases. From , we can observe that for a fixed $\lambda_d$, only the circuit power is changed by increasing $U_c$ and $T_c$. At the same time, the circuit power dominates the the total power consumption and increases almost linearly leading to an (almost) constant EE. The network performance in terms of the EE is poor with high density of D2D users ($\lambda_d=10^{-4}$). This is due to the fact that the sum rate contributed by the CUEs is already degraded by the interference from high number of D2D users, and additionally, increasing $U_c$ (and accordingly $T_c$) increases the circuit power without any gain in the total ASR. Consequently, the EE decreases. Thus, massive MIMO can only improve the EE if the D2D user density is small, otherwise dedicated resources or underlaying with fewer BS antennas is beneficial.
![EE $\mathrm{[Mbit/Joule]}$ as a function of the number of CUEs $U_c$ and the D2D user density $\lambda_d$ for a fixed ratio $\frac{T_c}{U_c}= 5$.[]{data-label="fig:EE_ld_uc_phi5"}](EE_ld_uc_phi5){width="0.65\columnwidth"}
Fixed Number of CUEs
--------------------
In this section, we evaluate the system performance when the number of CUEs is fixed with $U_c = 4$ users. The general trend of the network performance is the same as the case with $\frac{T_c}{U_c} =5$ in the previous section. However, there are some differences which are highlighted in and for the ASR and EE, respectively. As it is shown in , in the low D2D user density regime (i.e., $\lambda_d=10^{-6}$) the ASR is increasing in $T_c$, however, with a lower slope as compared to the case of $\frac{T_c}{U_c} =5$. By increasing the number of BS antennas for the fixed number of CUEs, better performance per user can be achieved, however in this case, as the number of CUEs is not high, the ASR increases with a small slope. For high D2D user density (i.e., $\lambda_d = 10^{-4}$), the ASR is almost flat.
illustrates that when the D2D user density is low, the EE benefits from adding extra BS antennas until the sum of the circuit power consumption of all antennas dominates the performance and leads to a gradual decrease in the EE. As the figure implies, there exists an optimal number of BS antennas which is relatively small since the main massive MIMO gains come from multiplexing rather than just having many antennas. However, in high density D2D scenario, which is the interference-limited scenario, the EE decreases monotonically with $T_c$. Increasing the number of BS antennas in this region cannot improve the ASR significantly, as shown in ; at the same time the circuit power consumption increases as a result of the higher number of BS antennas, which in turn leads to decreasing network EE.
The conclusion is that the D2D user density has a very high impact on a network that employs the massive MIMO technology. In the downlink, these two technologies can only coexist in low density of D2D users with careful interference coordination. The number of CUEs should be a function of the number of BS antennas in order to benefit from high number of BS antennas in terms of the ASR and EE. Otherwise, in high density of D2D users, the D2D communication should use the overlay approach rather than the underlay, that is, dedicated time/frequency resources should be allocated to the D2D tier.
The Effect of Other System Parameters
-------------------------------------
So far, we have discussed the results based on constant transmit power $P_c$, D2D transmit power $P_d$, and distance between D2D Tx-Rx pairs $R_{0,0}$ given in Table \[table:Sim\_param\]. Now we comment on the choice of these parameters and study their effects on the system performance. From Proposition \[proposition:P\_succ\_d2d\], Proposition \[proposition:P\_succ\_cue\], and Remark \[remark:P\_succ\_d2d\_PcPd\], it is evident that the coverage probability for both D2D and cellular tiers, and consequently the network ASR and EE, depend on the ratio of $P_d$ and $P_c$. Therefore, we fix $P_c$ and vary $P_d$.
shows the ASR as function of $\lambda_d$ under two different power levels, i.e., $P_d = 6~\mathrm{dBm}$ and $P_d = 13~\mathrm{dBm}$ in a scenario where the number of CUEs $U_c$ is scaled by $T_c$. We see that higher $P_d$ degrades the ASR at higher number of CUEs (and BS antennas) when the D2D user density is low, but has negligible impact at lower number of CUEs. The reason is that increasing $P_d$, on the one hand, boosts the D2D user rates, and on the other hand, causes more interference to CUEs which deteriorates their rates. Consequently, at low D2D user densities and high number of CUEs and BS antennas where the cellular sum rate is the main contributer to the total ASR, the interference caused by higher D2D transmit power is the dominant factor leading to lower total ASR. However, as $\lambda_d$ increases, the contribution of the D2D sum rate to the total ASR increases, and thus with higher $P_d$, the increase in the D2D sum rates compensates the decrease in CUEs sum rate and the difference in terms of the total ASR between the different power levels vanishes. When the number of CUEs is small, i.e., $U_c=1$ user and $T_c=5$ antennas, the CUE and D2D users have almost the same contributions to the ASR and increasing $P_d$ has negligible impact on the performance.
depicts the EE as a function of $\lambda_d$ under the same two levels of D2D transmit power. It is observed that lower $P_d$ is more beneficial in terms of the EE in both cases of $U_c = 1$ user and $U_c=14$ users. This is particularly visible in higher density of D2D users (e.g., $\lambda_d=3\times 10^{-5}$) with $U_c = 1$ user and $T_c=5$ antennas when the interference is the limiting factor. With $U_c = 14$ users and $T_c=70$ antennas, the CUEs have higher impact on the ASR, and as a consequence, the system benefits from lower transmit power of D2D users in terms of the EE. Therefore, we have chosen $P_d = 6~\mathrm{dBm}$ in the previous performance evaluation, as it has a better impact on the ASR as well as EE, especially in higher number of BS antennas.
Another important parameter that impacts the ASR is the distance between D2D Tx-Rx pairs, i.e., $R_{0,0}$. The effect of this parameter is only on the coverage probability of D2D users as seen in Proposition \[proposition:P\_succ\_d2d\] and Proposition \[proposition:P\_succ\_cue\]. illustrates the cellular ASR versus the D2D ASR for different values of $\lambda_d$ and $R_{0,0}$. The figure verifies that by decreasing $R_{0,0}$ only the ASR of D2D tier increases and as Remark \[remark:P\_succ\_d2d\_Uc\] implies increasing $R_{0,0}$ decreases the coverage probability of D2D users leading to lower ASR and EE. Since D2D communications are mostly meant for close proximity applications, we have chosen $R_{0,0} =35~\mathrm{m}$ in our performance study. Moreover, by decreasing the distance between D2D users, more D2D users can coexist simultaneously. This is observed in that with $R_{0,0} =35~\mathrm{m}$ the maximum ASR (of the D2D tier as well as the network) is achieved at the D2D density $\lambda_d=10^{-4}$ while with $R_{0,0} =50~\mathrm{m}$, it is achieved at the D2D density $\lambda_d=3.98 \times 10^{-5}$.
![Cellular ASR vs. D2D ASR $\mathrm{[Mbit/s]}$ for different distances between D2D Tx and D2D Rx with $U_c=4$ users and $T_c=70$ antennas. The curves are obtained by varying the value of $\lambda_d$ from $10^{-6}$ to $10^{-2}$.[]{data-label="fig:asrC_asrD_diff-d2d-distance"}](asrC-asrD-Uc4-Tc70-diff-d2d-distance){width="0.65\columnwidth"}
Conclusions {#sec:conclusion}
===========
We studied the coexistence of two key 5G concepts: device-to-device (D2D) communication and massive MIMO. We considered two performance metrics, namely, the average sum rate in $\mathrm{bit/s}$ and the energy efficiency in $\mathrm{bit/Joule}$. We considered a setup with a number of uniformly distributed cellular users in the cell, while the D2D transmitters are distributed according to a Poisson point process. We derived tractable expressions for the coverage probabilities of both cellular and D2D users which led to computation of the average sum rate and energy efficiency. We then studied the tradeoff between the number of base station antennas, the number of cellular users, and the density of D2D users for a given coverage area in the downlink. Our results showed that both the average sum rate and energy efficiency behave differently in scenarios with low and high density of D2D users.
Underlay D2D communications and massive MIMO can only coexist in low densities of D2D users with careful interference coordination, because the massive MIMO gains vanish when the interference from the D2D tier becomes too large. The number of cellular users should be a function of the number of base station antennas in order to benefit from high number of base station antennas in terms of the average sum rate and energy efficiency. If there is a high density of D2D users, the D2D communication should use the overlay approach rather than the underlay or the network should only allow a subset of the D2D transmissions to be active at a time.
Proof of Proposition \[proposition:P\_succ\_d2d\] {#sec:proof_Pcov_d2d}
=================================================
The proof follows by substituting the definition of $\mathrm{SINR}_d$ from into where we obtain $$\begin{aligned}
\mathrm{P}^d_{\mathrm{cov}}(\beta_d) &= \mathrm{Pr}\big\{\mathrm{SINR}_d \geq \beta_d \big\} \nonumber \\
&= \mathrm{Pr} \left\{ P_{d} R_{0,0}^{-\alpha_d} |g_{0,0}|^2 \geq \beta_d \Big(I_{\textrm{BS},0} + I_{d,0} + \frac{N_0}{A_d }\Big) \right\} \nonumber \\
&= \mathrm{Pr} \left\{ |g_{0,0}|^2 \geq \frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} \Big(I_{\textrm{BS},0} + I_{d,0} + \frac{N_0}{A_d }\Big) \right\} \nonumber \\
&{\overset{(a)}{=}} \mathbb{E}_{I_{\textrm{BS},0}, I_{d,0}} \left[\exp\bigg(- \frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} \Big(I_{\textrm{BS},0} + I_{d,0} + \frac{N_0}{A_d }\Big)\bigg)\right]\nonumber \\
&{\overset{(b)}{=}} \mathbb{E}_{I_{\textrm{BS},0}}\left[\exp\bigg(- \frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} I_{\textrm{BS},0}\bigg)\right] \mathbb{E}_{I_{d,0}}\left[\exp\bigg(- \frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} I_{d,0}\bigg)\right] \exp\bigg(-\frac{\beta_d }{\bar{\gamma}_d}\bigg) \nonumber \\
&{\overset{(c)}{=}} \mathcal{L}_{I_{\textrm{BS},0}}\bigg(\frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}}\bigg)\mathcal{L}_{I_{d,0}}\bigg(\frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}}\bigg) \exp\bigg(-\frac{\beta_d}{\bar{\gamma}_d}\bigg). \label{eq:Pd2dProof}\end{aligned}$$ Step $(a)$ comes from the fact that $|g_{0,0}|^2 \sim \exp(1)$ and $(b)$ follows since the noise and interference terms are mutually independent. In step $(c)$, the Laplace transform defined as $\mathcal{L}_{x}(s) = \mathbb{E}_{x}\big[e^{- sx}\big]$ is identified.
The first Laplace transform in is with respect to $I_{\textrm{BS},0}$ in which is a function of two random variables, namely $\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2$ and $R_{0,\textrm{BS}}$. This Laplace transform is calculated as $$\begin{aligned}
\mathcal{L}_{I_{\textrm{BS},0}}&\bigg(\frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}}\bigg) = \mathbb{E}_{I_{\textrm{BS},0}} \left[\exp\bigg( - \frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} I_{\textrm{BS},0}\bigg)\right]\nonumber \\
&= \mathbb{E}_{R_{0,\textrm{BS}}} \left[\mathbb{E}_{\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2}\bigg[\exp\bigg(- \frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} \frac{\zeta R_{0,\textrm{BS}}^{-\alpha_c}}{A_d}\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2 \bigg) \Big| R_{0,\textrm{BS}}\bigg] \right] \nonumber \\
&=\mathbb{E}_{R_{0,\textrm{BS}}} \left[ \mathcal{L}_{\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2}\bigg(\frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} \frac{\zeta R_{0,\textrm{BS}}^{-\alpha_c}}{A_d}\bigg)\right]\nonumber \\
&{\overset{(a)}{=}} \mathbb{E}_{R_{0,\textrm{BS}}} \left[\frac{1}{(\kappa \beta_d R_{0,\textrm{BS}}^{-\alpha_c} + 1)^{U_c}}\right]\nonumber \\
&{\overset{(b)}{=}} \int_0^R \frac{2r}{R^2(\kappa \beta_d r^{-\alpha_c}+1)^{U_c}} \mathrm{d}r\nonumber \\
&{\overset{(c)}{=}} \frac{2(\kappa\beta_d)^{2/\alpha_c}}{\alpha_c R^2}\int_0^{y} \frac{t^{U_c+\frac{2}{\alpha_c}-1}}{(1-t)^{\frac{2}{\alpha_c}+1}}\mathrm{d}t \nonumber \\
&{\overset{(d)}{=}} \frac{ (\kappa\beta_d)^{2/\alpha_c}}{R^2}\left(y^{U_c + \frac{2}{\alpha_c}-1} (1-y)^{- \frac{2}{\alpha_c}}- \Big(U_c + \frac{2}{\alpha_c}-1\Big) \mathcal{B}\Big(y; U_c + \frac{2}{\alpha_c}-1, 1-\frac{2}{\alpha_c}\Big) \right) \label{eq:LT_I_BS0}\end{aligned}$$ for $\alpha_c>2$, where $(a)$ follows by introducing the notation $$\kappa = \frac{\zeta}{P_d A_d R_{0,0}^{-\alpha_d}}$$ and from the Laplace transform of the probability density function (PDF) of $\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2$ which, by neglecting the spatial correlation, is tightly approximated by a Chi-squared distribution as $2\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2 \sim \chi^2_{2U_c}$ [@Dhillon-2013-TWC]. Note that $$\begin{aligned}
\label{eq:normfHv}
\big\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\big\|^2 &= \big\|{\mathbf{f}}_{0,\textrm{BS}}^H [{\mathbf{v}}_{0},\ldots,{\mathbf{v}}_{{U_c}-1}]\big\|^2 \nonumber\\
&=\sum_{i=0}^{U_c-1} |{\mathbf{f}}_{0,\textrm{BS}}^H{\mathbf{v}}_{i}|^2,\end{aligned}$$ where ${\mathbf{f}}_{0,\textrm{BS}}^H{\mathbf{v}}_{i}$, $i = \{0, \ldots, U_c-1\}$, are zero-mean circular symmetric complex Gaussian random variables with unit variance. Therefore, $\sum_{i=0}^{U_c-1} |{\mathbf{f}}_{0,\textrm{BS}}^H{\mathbf{v}}_{i}|^2$ is the summation of $U_c$ i.i.d. exponential random variables which has an $\mathrm{Erlang}(U_c,1)$ distribution. Equivalently, the sum scaled down by $\frac{\sigma^2}{2}$ (i.e., multiplied by $\frac{2}{\sigma^2}$) has a (standard) Chi-squared distribution with $2U_c$ degrees of freedom. Hence, the PDF of $\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2$ is $$\label{eq:pdf_IBS0}
f_{\|{\mathbf{f}}_{0,\textrm{BS}}^H \mathbf{V}\|^2} (x) = \frac{x^{Uc-1}e^{-x}}{(U_c-1)!}.$$ From Laplace transform theory we know that $\mathcal{L}\big[t^n e^{-\alpha t}\big] = \frac{n!}{(s+\alpha)^{n+1}}$ and with some simplifications, we obtain the result in step $(a)$. Step $(b)$ in follows from the PDF of $R_{0,\textrm{BS}}$ which is $$f_{R_{0,\textrm{BS}}}(r) = \left\{\begin{array}{ll}
\frac{2r}{R^2}, & \textrm{if } 0\leq r \leq R,\\
0, & \textrm{otherwise}, \end{array} \right.$$ as the typical D2D Rx is uniformly distributed over the cell area and the BS is located in the cell center. Step $(c)$ in is obtained by the change of variable $\frac{1}{\kappa \beta_d r^{-\alpha_c} + 1}\rightarrow t$ which leads to the integral boundary $y \triangleq \frac{1}{\kappa \beta_d R^{-\alpha_c} + 1} $. Finally, $(d)$ follows by integration by part where $\mathcal{B}(x;a,b)$ is the incomplete Beta function defined as $$\label{eq:beta_func}
\mathcal{B}(x;a,b) = \int_0^x t^{a-1} (1-t)^{b-1} \mathrm{d}t,$$ for $a,b>0$.
Next, we proceed to calculate the second Laplace transform in . This transform is with respect to $I_{d,0}$ in which is a function of two random variables, that is $|g_{0,j}|^2$ and $R_{0,j}$. Therefore, we have $$\begin{aligned}
\mathcal{L}_{I_{d,0}} \bigg(\frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}}\bigg) &= \mathbb{E}_{I_{d,0}}\left[\exp\Big(- \frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} I_{d,0}\Big)\right]\nonumber \\
&= \mathbb{E}_{R_{0,j},|g_{0,j}|^2} \left[\exp\Big(- \frac{\beta_d}{P_dR_{0,0}^{-\alpha_d}} \sum_{j\neq 0} P_d R_{0,j}^{-\alpha_d} |g_{0,j}|^2\Big)\right] \nonumber \\
&= \mathbb{E}_{R_{0,j}} \left[\prod_j \mathbb{E}_{|g_{0,j}|^2}\bigg[\exp\Big(-\frac{\beta_d}{R_{0,0}^{-\alpha_d}} R_{0,j}^{-\alpha_d} |g_{0,j}|^2\Big)\bigg]\right]\nonumber \\
&{\overset{(a)}{=}} \exp\left( -2 \pi \lambda_d \int_0^{\infty} \bigg(1-\mathbb{E}_{G}\bigg[\exp\Big(-\frac{\beta_d}{R_{0,0}^{-\alpha_d}} r^{-\alpha_d} G\Big)\bigg]\bigg)r\,\mathrm{d}r \right) \nonumber \\
&{\overset{(b)}{=}}\exp\left(-2 \pi \lambda_d \int_0^{\infty} \frac{r}{\frac{R_{0,0}^{\alpha_d}}{\beta_d} r^{\alpha_d}+1}\mathrm{d}r \right)\nonumber \\
&{\overset{(c)}{=}} \exp\left(-\frac{\pi \lambda_d}{\mathrm{sinc}(\frac{2}{\alpha_d})}\Big(\frac{\beta_d}{R_{0,0}^{-\alpha_d}}\Big)^{2/\alpha_d}\right),
\label{eq:LT_I_d0}\end{aligned}$$ where $(a)$ is based on the probability generating functional (PGFL) [@NET-032], and $(b)$ follows from the fact that $G \sim \exp(1)$ and $\mathcal{L}\big[e^{-t}\big] = \frac{1}{s+1}$. Step $(c)$ follows by solving the integral in step $(b)$ and using $\mathrm{sinc}(x) = \frac{\sin(\pi x)}{\pi x}$.
Substituting and in concludes the proof of Proposition \[proposition:P\_succ\_d2d\].
Proof of Proposition \[proposition:P\_succ\_cue\] {#sec:proof_Pcov_cue}
=================================================
Substituting $\mathrm{SINR}_c$ from into , we get $$\begin{aligned}
\mathrm{P}^c_{\mathrm{cov}}(\beta_c) &= \mathrm{Pr}\left\{ |{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2 \geq \frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c}\Big(I_{d,c} + \frac{N_0}{A_d }\Big)\beta_c \right\} \nonumber\\
&{\overset{(a)}{=}} \mathbb{E}_{D_{0,\textrm{BS}}, I_{d,c}}\left[e^{-\frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c}(I_{d,c} + \frac{N_0}{A_d})\beta_c } \sum_{k=0}^{T_c - U_c} \frac{1}{k!}\bigg(\frac{A_d}{\zeta} D_{0,\textrm{BS}}^{\alpha_c}\Big(I_{d,c} + \frac{N_0}{A_d}\Big)\beta_c\bigg)^k\right] \nonumber \\
&{\overset{(b)}{=}} \mathbb{E}_{D_{0,\textrm{BS}}, I_{d,c}}\left[e^{-\frac{N_0}{A_d}s} \sum_{k=0}^{T_c - U_c} \frac{s^k}{k!} \sum_{i=0}^{k} \binom{k}{i} \left(\frac{N_0}{A_d }\right)^{k-i} I_{d,c}^i e^{-s I_{d,c}} \right] \nonumber \\
&{\overset{(c)}{=}} \mathbb{E}_{D_{0,\textrm{BS}}}\left[e^{-\frac{N_0}{A_d}s} \sum_{k=0}^{T_c - U_c} \frac{s^k}{k!} \sum_{i=0}^{k} \binom{k}{i} \left(\frac{N_0}{A_d }\right)^{k-i} \mathbb{E}_{I_{d,c}}\Big[I_{d,c}^i e^{-s I_{d,c}} \Big] \right]\nonumber\\
&{\overset{(d)}{=}} \mathbb{E}_{D_{0,\textrm{BS}}}\Bigg[e^{-\frac{N_0}{A_d}s} \sum_{k=0}^{T_c - U_c} \frac{s^k}{k!} \sum_{i=0}^{k} \binom{k}{i} \left(\frac{N_0}{A_d }\right)^{k-i} (-1)^i\frac{\mathrm{d}^i}{{\mathrm{d}s}^i}\mathcal{L}_{I_{d,c}}(s)\Bigg], \label{eq:proof_Prob_cov_CUE}\end{aligned}$$ where $(a)$ follows from the CCDF of $|{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2$ with $2|{\mathbf{h}}_{0}^H {\mathbf{v}}_{0}|^2 \sim \chi^2_{2(T_c-U_c+1)}$ given $D_{0,\textrm{BS}}$ and $I_{d,c}$. In $(b)$, we use Binomial expansion as $$\begin{aligned}
\Big(I_{d,c} + \frac{N_0}{A_d}\Big)^k &= \sum_{i=0}^k \binom{k}{i} \left(\frac{N_0}{A_d}\right)^{k-i} I_{d,c}^i,\end{aligned}$$ and $(c)$ follows by taking the expectation with respect to the interference $I_{d,c}$. Step $(d)$ follows from $$\begin{aligned}
\mathbb{E}_{I_{d,c}}\Big[I_{d,c}^i e^{-s I_{d,c}} \Big] = (-1)^i \frac{\mathrm{d}^i}{{\mathrm{d}s}^i} \mathcal{L}_{I_{d,c}}(s),\label{eq:E_Idc}\end{aligned}$$ where $\mathcal{L}_{I_{d,c}}(s)$ is obtained using similar steps as in the derivation of $\mathcal{L}_{I_{d,0}}$ in : $$\begin{aligned}
\mathcal{L}_{I_{d,c}}(s) =
\exp\left(-\frac{\pi \lambda_d P_d^{2/\alpha_d}}{\mathrm{sinc}(\frac{2}{\alpha_d})} s^{2/\alpha_d}\right).\label{eq:L_dc}\end{aligned}$$ Substituting in and using the Faà di Bruno’s formula for the $i$-th derivative of a composite function $f(g(s))$ with $f(s) = e^s$ and $g(s) = - \frac{\pi \lambda_d P_d^{2/\alpha_d}}{\mathrm{sinc}(\frac{2}{\alpha_d})} s^{2/\alpha_d}$, Proposition \[proposition:P\_succ\_cue\] follows.
[^1]: Serveh Shalmashi and Ki Won Sung are with the Dept. of Communication Systems, KTH Royal Institute of Technology, Stockholm, Sweden (emails: {serveh,sungkw}@kth.se).
[^2]: Emil Björnson is with the Dept. of Electrical Engineering (ISY), Linköping University, Linköping, Sweden, (email: [email protected]).
[^3]: Marios Kountouris and Mérouane Debbah are with the Mathematical and Algorithmic Sciences Lab, France Research Center, Huawei Technologies Co. Ltd. (emails: {marios.kountouris,merouane.debbah}@huawei.com).
[^4]: Part of the material in this paper will be presented at IEEE International Conference on Communications (ICC) Workshop on Device-to-Device Communication for Cellular and Wireless Networks, London, UK, June 2015 [@Shalmashi-2015-ICC].
[^5]: The assumption that the D2D Tx are distributed in the whole $\mathbb{R}^2$ plane removes any concern about the boundary effects and makes the model more mathematically tractable. The boundary effects are local effects in which users at the network boundary experience less interference than the ones closer to the center, because they have fewer neighbors.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In recent years, convolutional neural networks have demonstrated promising performance in a variety of medical image segmentation tasks. However, when a trained segmentation model is deployed into the real clinical world, the model may not perform optimally. A major challenge is the potential poor-quality segmentations generated due to degraded image quality or domain shift issues. There is a timely need to develop an automated quality control method that can detect poor segmentations and feedback to clinicians. Here we propose a novel deep generative model-based framework for quality control of cardiac MRI segmentation. It first learns a manifold of good-quality image-segmentation pairs using a generative model. The quality of a given test segmentation is then assessed by evaluating the difference from its projection onto the good-quality manifold. In particular, the projection is refined through iterative search in the latent space. The proposed method achieves high prediction accuracy on two publicly available cardiac MRI datasets. Moreover, it shows better generalisation ability than traditional regression-based methods. Our approach provides a real-time and model-agnostic quality control for cardiac MRI segmentation, which has the potential to be integrated into clinical image analysis workflows.'
author:
- |
Shuo Wang(), Giacomo Tarroni , Chen Qin ,\
Yuanhan Mo , Chengliang Dai , Chen Chen ,\
Ben Glocker , Yike Guo, Daniel Rueckert, and Wenjia Bai
bibliography:
- 'ref.bib'
title: 'Deep Generative Model-based Quality Control for Cardiac MRI Segmentation'
---
Introduction
============
Cardiovascular diseases (CVDs) are the leading cause of death globally, taking more than 18 million lives every year [@WHOScaleStroke.]. Cardiac magnetic resonance imaging (MRI) has been widely used in clinical practice for evaluating cardiac structure and function. To derive quantitative measures from cardiac MRI, accurate segmentation is of great importance. Over the past few years, various architectures of convolutional neural networks (CNNs) have been developed to deliver state-of-the-art performance in the task of automated cardiac MRI segmentation [@bai2018automated; @tao2019deep; @bernard2018deep; @zheng20183]. Although satisfactory performance has been achieved on specific datasets, care must be taken when deploying these models into clinical practice. In fact, it is inevitable for automated segmentation algorithms (not limited to CNN-based) to generate a number of poor-quality segmentations in real-world scenarios, due to differences in scanner models and acquisition protocols as well as potential poor image quality and motion artifacts. Therefore, reliable quality control (QC) of cardiac MRI segmentation on a per-case basis is highly desired and of great importance for successful translation into clinical practice.
### Related work:
Numerous efforts have been devoted into quality control of medical images [@tarroni2018learning; @carapella2016towards; @zhang2016automated] and segmentations [@robinson2019automated; @alba2018automatic]. In this work, we focus on the latter, i.e. segmentation quality control. Existing literature can be broadly classified into two categories:
**Learning-based quality control:** These methods consider quality control as a regression or classification task where a quality metric is predicted from extracted features. [@kohlberger2012evaluating] proposed 42 hand-crafted features based on intensity and appearance and achieved an accuracy of 85% in detecting segmentation failure. [@robinson2018real] developed a CNN-based method for real-time regression of the Dice similarity metric from image-segmentation pairs. [@hann2019quality] integrated quality control into the segmentation network by regressing the Dice metric. Most of these methods require poor-quality segmentations as negative samples to train the regression or classification model. This makes quality control specific to the segmentation model and the type of poor-quality segmentations used for training. [@liu2019alarm] used a variational auto-encoder (VAE) for learning the shape features of segmentation in an unsupervised manner and proposed to use the evidence lower bound (ELBO) as a predictor. This model-agnostic structure provides valuable insights and an elegant theoretical framework for quality control.
**Registration-based quality control:** These methods perform image registration between the test image with a set of pre-selected template images with known segmentations. Then the quality metric can be evaluated by referring to the warped segmentations of these template images. Following this direction, [@valindria2017reverse] proposed the concept of reverse classification accuracy (RCA) to predict segmentation quality and [@robinson2019automated] achieved good performance on a large-scale cardiac MRI dataset. These methods can be computationally expensive due to the cost of multiple image registrations, which could potentially be reduced by using GPU acceleration and learning-based registration tools [@Haskins2020DeepSurvey].
### Contributions:
There are three major contributions of this work. Firstly, we propose a generic deep generative model-based framework which learns the manifold of good-quality segmentations for quality control on a per-case basis. Secondly, we implement the framework with a VAE and propose an iterative search strategy in the latent space. Finally, we compare the performance of our method with regression-based methods on two different datasets, demonstrating both the accuracy and generalisation ability of the method.
Methodology
===========
Problem Formulation
-------------------
Let $F$ denote an arbitrary type of segmentation model to be deployed. Given a test image $I$, the segmentation model provides a predicted segmentation $\hat{S}=F(I)$. The ground-truth quality of $\hat{S}$ is defined as $q(S_{gt}, \hat{S})$ where $S_{gt}$ is the ground-truth segmentation and $q$ is a chosen quality metric (e.g. Dice metric). The aim of quality control is to develop a model $Q$ so that $Q( \hat{S};I)\approx q(S_{gt}, \hat{S})$.
Deep Generative Model-based Quality Control
-------------------------------------------
Quality control would be trivial if the ground-truth segmentation $S_{gt}$ was available. Intuitively, the proposed framework aims to find a good-quality segmentation $S_{sur}$ as a surrogate for ground truth so that $q(S_{sur}, \hat{S}) \approx q(S_{gt}, \hat{S})$. This is realised through iterative search on the manifold of good-quality segmentations (Fig.\[figure1\]).
![Overview of the deep generative model-based quality control framework. The generative model $G$ is trained to learn a mapping $G(z)$ from the low-dimensional latent space $D_{z}$ to the good-quality manifold $\Sigma$. The input image-segmentation pair $(I, \hat{S})$ is projected to $(S_{sur}, I_{sur})$ on the manifold through iterative search, which is in turn used as surrogate ground truth for quality prediction. $z_0$ is the initial guess in the latent space and it converges to $z_{sur}$. []{data-label="figure1"}](Figures/Fig1.png){width="\textwidth"}
### Good-quality manifold:
The core component of this framework is a deep generative model $G$ which learns how to generate good-quality image-segmentation pairs. Formally, let $X=(I, S)\in D_I \times D_S$ represent an image-segmentation pair, where $D_I$ and $D_S$ are the domains of images and possible segmentations. The key assumption of this framework is that good-quality pairs $(I, S_{gt})$ are distributed on a manifold $\Sigma \subset D_I \times D_S$, named as *good-quality manifold*. The generator $G$ learns to construct a low-dimensional latent space $D_z \subseteq \mathbb{R}^m$ and a mapping to the good-quality manifold: $$G(z): D_z \ni z \mapsto X=G(z) \in D_I \times D_S$$ where $z$ denotes the latent variable with dimension $m$. The mapping $G(z)$ is usually intractable but can be approximated using generative models such as generative adversarial networks (GANs) or VAEs.
### Iterative search in the latent space:
To incorporate the generator into the quality control framework, we develop an iterative search scheme in the latent space to find a surrogate segmentation for a given image-segmentation pair as input. This surrogate segmentation is used for quality prediction. Finding the closest surrogate segmentation (i.e. projection) on the good-quality manifold is formulated as an optimisation problem, $$z_{sur} = \operatorname*{argmin}_{z \in D_z} \mathcal{L}(G(z), (I,\hat{S}))$$ which minimises the distance metric $\mathcal{L}$ between the reconstructed $G(z)$ and the input image-segmentation pair $(I,\hat{S})$. This problem can be solved using the gradient descent method as explained in Algorithm \[iter\_scheme\].
A trained generator $G: D_z \ni z \mapsto G(z) = (I,S) \in \Sigma$ Image-segmentation pair $(I,\hat{S})$ Quality prediction $Q(\hat{S};I)$ **Initialization** $z=z_0 \in D_z$ $L = \mathcal{L}(G(z),(I,\hat{S}))$ $grad = \nabla_{z}L$ $z = z - \alpha \cdot grad$ $S_{sur}=G(z_{sur})$ $Q(\hat{S};I)=q(S_{sur},\hat{S})$
Generative Model using VAE
--------------------------
The proposed framework can be implemented with different generative models as long as a good-quality segmentation generator with smooth latent space is available. In this paper, we employ the VAE (Fig \[figure2\]) which includes an encoder $E_\varphi$ and a decoder $D_\phi$, where $\varphi$ and $\phi$ denote the model parameters [@kingma2013auto]. The image-segmentation pair $(I, S)$ is encoded by $E_{\varphi}$ to follow a Gaussian distribution $\mathcal{N}(\mu_z, \sigma_z^2)$ in the latent space, where $\mu_z$ and $\sigma_z^2$ denote the mean and variance respectively. A probabilistic reconstruction of the image-segmentation pair $(I', S')$ is generated from the decoder $D_\phi$.
![Framework implementation using the variational autoencoder (VAE). In the training stage, the ground-truth image-segmentation pairs are used. In the application stage, the VAE decoder is used as the generator for iterative search of the surrogate segmentation on the good-quality manifold. Initial guess $z_0$ is from the encoder. []{data-label="figure2"}](Figures/Fig2.png){width="\textwidth"}
At the training stage, the ground-truth image-segmentation pairs are used to train the VAE. The loss function includes a reconstruction loss and a KL divergence term for regularisation [@higgins2017beta]: $$\mathcal{L}_{VAE}=\mathcal{L}_{recon}+\beta \cdot D_{KL}(\mathcal{N}(\mu_z, \sigma_z^2)||\mathcal{N}(0, \textbf{I}))$$ $$\mathcal{L}_{recon}=BCE(S_{GT}, S_{GT}') + MSE(I, I')$$ where $\beta$ is the hyperparameter that balances between reconstruction loss and regularisation. The reconstruction loss is evaluated using the binary cross-entropy (BCE) for segmentation and the mean square error (MSE) for image, respectively. The effects of the weight $\beta$ and the latent space dimension $m$ will be evaluated in the ablation study.
At the application stage, the VAE decoder $D_\phi$ is used as the generator, reconstructing image-segmentation pairs from the latent space. The initial guess $z_0$ in the latent space is obtained from the encoder $E_{\varphi}$. Following Algorithm \[iter\_scheme\], the surrogate segmentation $S_{sur}$ can be found via iterative search (Fig. \[figure2\]b). Finally, the quality metric is evaluated by $q(S_{sur},\hat{S})$, e.g. Dice metric.
Experiments
===========
Datasets
--------
**UK Biobank dataset:** Short-axis cardiac images at the end-diastolic (ED) frame of 1,500 subjects were obtained from UK Biobank and split into three subsets for training (800 cases), validation (200 cases) and test (500 cases). The in-plane resolution is 1.8x1.8 mm with slice thickness of 8 mm and slice gap of 2 mm. A short-axis image stack typically consists of 10 to 12 image slices. Ground-truth good segmentations were generated from a publicly available fully-convolutional network (FCN) that has demonstrated a high performance [@bai2018automated], with manual quality control by an experienced cardiologist.
**ACDC dataset:** 100 subjects including a normal group and four pathology groups were obtained from ACDC dataset [@bernard2018deep] and resampled to the same spatial resolution as UK Biobank data. The ground-truth segmentations at the ED frame were provided by the ACDC challenge organisers.
Experimental Design
-------------------
In this study, we evaluate the performance of our proposed method and compare with two regression-based methods for quality control of cardiac MRI segmentation. Specifically, we focus on the myocardium, which is a challenging cardiac structure to segment and of high clinical relevance.
### VAE implementation and training: {#implementation}
The VAE encoder is composed of four convolutional layers (each was followed by $ReLU$ activation), and one fully connected layer. The decoder has a similar structure with reversed order and the last layer is followed by $Sigmoid$ function. The latent space dimension $m$ was set to 8 and the hyperparameter $\beta$ was set to 0.01 from ablation study results. The architecture is shown in Appendix Fig. A1. The model was implemented in PyTorch and trained using the Adam optimiser with learning rate 0.0001 and batch size 16. It was trained for 100 epochs and an early stopping criterion was used based on the validation set performance. To improve the computational efficiency, the VAE was trained on a region of interest (ROI) centered around the myocardium with the window size of 96x96 pixel, which was heuristically determined to include the whole cardiac structure. The cropped image intensity was normalised to the $[0, 1]$ range and stacked with the binary segmentation.
### Baseline methods:
Two regression-based methods were used as baselines: a) a support vector regression (SVR) model with 42 hand-crafted features about shape and appearance [@kohlberger2012evaluating] and b) a CNN regression network (ResNet-18 backbone) with the image-segmentation pair as input [@robinson2018real]. Both baseline methods use the Dice metric as the regression target.
### Experiment 1: UK Biobank
Besides the ground-truth segmentations, we generated poor-quality segmentations by attacking the segmentation model. White noise with different variance level was added to the original images, resulting in a dataset of poor-quality segmentations with uniform Dice distribution. The quality prediction was performed on the test set of the attacked segmentations.
**Experiment 2: ACDC** We deployed a UK Biobank trained segmentation model on ACDC dataset without fine-tuning. This reflects a real-world clinical setting, where segmentation failures would occur due to domain shift issues. **Ablation study:** We adjusted the dimensionality of the latent space $m$ and the hyperparameter $\beta$ and performed a sensitivity analysis on the UK Biobank validation dataset. The result is reported in the Appendix Table A1.
Results and Discussion
======================
Quality control performance is assessed in terms of Dice metric prediction accuracy. Pearson correlation coefficient $r$ and mean absolute error (MAE) between predicted Dice and real Dice are calculated. Table 1 compares the quantitative performance of the methods and Fig.\[figure3\] visualises the predictions. On UK Biobank dataset, the hand-crafted feature method performed the worst. The proposed method achieved a similar performance ($r$=0.96, MAE=0.07) as the CNN regression method ($r$=0.97, MAE=0.06). However, on ACDC dataset, the proposed method ($r$=0.97, MAE=0.03) outperformed the CNN regression method ($r$=0.97, MAE=0.17) with a smaller MAE. As shown in Fig. \[figure3\], on ACDC dataset, the prediction of the proposed method aligns well with the identity line, whereas the CNN regression method clearly deviates from the line, even though the $r$ coefficient is still high.
![Comparison of the performance of different quality control methods. The x-axis is the real Dice of each subject and the y-axis is the predicted Dice by each method. The dashed line is the $y=x$, plotted for reference. Top row: UK Biobank data ($n=500$). Bottom row: ACDC data ($n=100$), with five subgroups plotted in different colors. []{data-label="figure3"}](Figures/Fig3.png){width="300pt"}
A possible explanation for this is that the proposed generative method works by learning the good-quality manifold and proposing the surrogate ground truth for quality assessment. Thus it is agnostic to the types of poor-quality segmentations. The training of hand-crafted features and CNN-based regression methods require poor-quality segmentations and may be overfitted to the UK Biobank data. When they are deployed onto ACDC dataset, there is a shift not only for image appearance (e.g. difference between 1.5T and 3.0T MRI scanner) but also for types of segmentation failures. In addition, the ACDC dataset consists of more pathological cases, whereas the UK Biobank comes from a general healthy population. Due to the domain shift, the performance of regression-based methods degraded. In contrast, the proposed method maintained a high prediction accuracy against domain shift. This indicates the advantage of a generative model-based framework for generalisation. It also can be potentially used as a system to monitor the performance of deployed segmentation models over time.
\[table1\]
To gain insights into our proposed method, we visualised several searching paths within a two-dimensional latent space and corresponding image-segmentation pairs reconstructed by our generative model (Fig. \[figure4\]). The poor-quality segmentation could be projected onto the good-quality manifold and refined iteratively to obtain the surrogate segmentation. The surrogate segmentation on the good-quality manifold is more plausible and can potentially be used as $prior$ to correct poor-quality segmentations. It is also expected that the performance could be improved using advanced generative models and better manifold learning [@bojanowski2017optimizing].
![Visualisation of searching path in a two-dimensional latent space. Left: searching paths for five exemplar samples (green point: initial guess from the VAE encoder; black point: intermediate state during iterative search; red point: convergence point for surrogate segmentation). Right: the input image and segmentation, reconstructed segmentations along the searching path and the ground-truth segmentation.[]{data-label="figure4"}](Figures/Fig4.png){width="280pt"}
Conclusion
==========
Here we propose a generative-model based framework for cardiac image segmentation quality control. It is model-agnostic, in the sense that it does not depend on specific segmentation models or types of segmentation failures. It can be potentially extended for quality control for different anatomical structures.
Appendix
========
![Network architecture of the VAE.[]{data-label="figureS1"}](Figures/FigS1.png){width="\textwidth"}
---- ------- ------- ----------- -------
0 1E-3 1E-2 1E-1
2 0.095 0.094 0.105 0.502
4 0.086 0.092 0.114 0.416
8 0.088 0.072 **0.068** 0.229
16 0.147 0.141 0.095 0.091
---- ------- ------- ----------- -------
: Ablation study of the latent space dimension $m$ and the hyperparameter $beta$. The mean absolute error (MAE) between predicted and true Dice metrics on UKBB validation set are reported. $m$=8 and $\beta=0.01$ were selected as the optimal parameters according to the results.
\[tableS1\]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'E. Bozzo'
- 'L. Oskinova'
- 'A. Lobel'
- 'W.-R. Hamann'
bibliography:
- 'superorbital.bib'
date: 'Submitted: -; Accepted -'
title: 'On the super-orbital modulation of supergiant high mass X-ray binaries'
---
Introduction {#sec:intro}
============
Supergiant high mass X-ray binaries (SgXBs) are a sub-class of high mass X-ray binaries (HMXBs) hosting a compact object and an OB supergiant star [see @walter15 for a recent review]. SgXBs are typically divided in two groups, i.e., the classical systems, which show a virtually persistent high X-ray luminosity, and the supergiant fast X-ray transients (SFXTs), which feature a still highly debated and peculiarly prominent variability in X-rays [@nunez17]. The bulk of the X-ray radiation from the SgXBs can be reasonably well explained as being due to the accretion of the stellar wind from the OB supergiant onto a highly magnetized neutron star (NS), with no evidence supporting the presence of long-lived accretion disks [@bozzo08; @shakura12; @romano15; @hu17]. No clear indication has yet been reported of systematic differences between the properties of the supergiant stellar winds in classical systems and SFXTs (Hainich et al. 2017, in preparation), which also share similar orbital periods [@lutovinov13; @bozzo15] and super-orbital modulations [see Table \[tab:tab1\] and @corbet13].
[@llll@]{} Source Name & class & $P_{\rm NS}$ & $P_{\rm SO}$\
& & (days) & (days)\
IGRJ16479-4514 & SFXT & 3.3199$\pm$0.0005 & 11.880$\pm$0.002\
IGRJ16418-4532 & SgXB & 3.7389$\pm$0.0001 & 14.730$\pm$0.006\
4U1909+07 & SFXT & 4.4003$\pm$0.0004 & 15.180$\pm$0.003\
IGRJ16493-4348 & SgXB & 6.782$\pm$0.001 & 20.07$\pm$0.01\
2S0114+650 & SgXB & 11.591$\pm$0.003 & 30.76$\pm$0.03\
The super-orbital variability in disk accreting X-ray binaries is usually ascribed to the precession of the disk or to the precession of the compact object in its center [see, e.g., the cases of HerX-1, SMCX-1, and LMCX4; @petterson75; @ogilvie01; @pot13], but a solid interpretation of the same phenomenon in wind-fed binaries is still lacking. The lightcurves of the latter folded on their super-orbital period display a large variety of morphologies and modulations that are stable over years, although the sources are generally thought to be accreting from a much less regularly structured environment compared to that provided by an accretion disk. In the case of the classical SgXB 2S0114+650, @farrell08 reported on the detection of spectral slope changes as a function of the super-orbital phase but could not detect corresponding variations in the absorption column density [also due to the limited coverage at energies $\lesssim$3 keV provided by the instruments on-board RXTE; @bradt93]. These authors suggested that the most likely cause of the super-orbital variability was a modulation of the mass loss rate from the supergiant star. For all other systems showing a super-orbital modulation, the only X-ray data providing coverage on different super-orbital phases are those collected with /BAT [@gehrels05; @barthelmy05]. The relatively low signal-to-noise ratio of these data and the energy band-pass limited to $\gtrsim$15 keV hampered so far any investigation of spectral variability as a function of the super-orbital phase. @koeni06 proposed that the mass accretion rate modulation could be produced as a consequence of tidal interaction-driven oscillations of the supergiant star, but the authors showed that such mechanism only works for strictly circular orbits. @corbet13 also discussed the possibility that the super-orbital modulations are due to a third body orbiting the inner massive binary. However, the same authors highlighted that a stable three body solution requires a hierarchical system with the third body in a very distant orbit, while all super-orbital modulations discovered so far in SgXBs are not longer than roughly 3 times the orbital periods of these sources.
We propose here that the super-orbital periodicities in classical SgXBs and SFXTs are produced as a consequence of the interaction between the compact object with the so-called “corotating interaction regions” threading the winds of OB supergiants.
Corotating interaction regions around OB supergiants {#sec:cir}
====================================================
The stellar winds of OB supergiants are well known to be characterized by complex velocity and density structures . The smaller structures, i.e. “clumps”, are typically endowed with an increased density of a factor of $\gtrsim$10 compared to a smooth wind and can be as large as $\sim$0.1 $R_*$, where $R_*$ is the OB supergiant radius. Clumps are usually invoked to interpret the stochastic X-ray variability displayed by SgXBs on time scales of 10-1000 s [see, e.g., @nunez17 and references therein].
The existence of larger structures in the OB supergiant winds was suggested in the early 80s [@mullan84], and confirmed by the detection of discrete absorption components [DACs; see, e.g., @new]. These features are observed to propagate blue-ward on time-scales comparable with the stellar rotation through the profiles of UV resonance lines in OB supergiants [@massa95; @prinja98]. @cranmer96 used hydrodynamic models to show that irregularities on the stellar surface related either to dark/bright spots, magnetic loops, or non-radial pulsations can lead to the formation of corotating interaction regions (CIRs) causing spiral-shaped density and velocity perturbations in the stellar wind up to several tens of stellar radii. The CIRs are invoked to explain modulations of the X-ray emission observed in single OB stars [@Oskinova2001; @Naze2013; @Massa2014].
An advanced model to reproduce the observational properties of DACs in OB supergiants with CIRs was developed by @lobel08, who also investigated the dependence of the extension, velocity, and density profile variations of these structures as a function of the different properties of the stellar surface spots from which they originate. The intensity of the spot, its size, and its rotational velocity (that could in principle be different from that of the star) are the main parameters regulating the density/velocity contrast of the CIR compared to the surrounding unperturbed stellar wind, typically limited to a factor of a few. A larger rotational velocity of the spot also increases the winding of the spiral arms. The available observations of the UV variability of massive star resonance lines can be used to constrain all the spot free parameters of the model.
We show in Fig. \[fig:fig1\] a hydrodynamic simulation of @lobel08 obtained from the application of the Zeus-3D code for a radiatively-driven rotating wind in the B0.5 Ib supergiant JPup (HD64760). This star is characterized by a mass of 20 $M_{\odot}$, a radius of 22 $R_{\odot}$, a mass loss rate of 9$\times$10$^{-7}$ $M_{\odot}$ yr$^{-1}$, and a terminal wind velocity of 1500 km s$^{-1}$ [see Table 1 of @lobel08].
![An example of hydrodynamic calculations of the CIR in the wind of the B0.5 Ib supergiant JPup. There is a bright spot on the stellar surface producing a single CIR. The spot is assumed to be 20% brighter than the stellar surface, and is characterized by an angular diameter of 20$^{\circ}$. A full revolution of the CIR occur in $\sim$10.3 days. The colors code the wind density of the CIR model, relative to the density of the unperturbed, smooth wind. The maximum over-density in the CIR is by a factor 1.22 (red colors). In wind regions shown in yellow this ratio hardly deviates from unity, while in the green regions the wind is rarefied by a factor below 0.97. The black solid drawn lines show 20 overplotted contours of equal radial velocity in the hydrodynamic rotating wind model. The outermost curve marks the isovelocity line at an outflow velocity of 1460 km s$^{-1}$, while the curves at smaller distances from the stellar surface are shown for decreasing steps of 73 km s$^{-1}$. The case with two CIRs mentioned later in this paper is produced by assuming an additional spot that is 8% brighter than the surface of the supergiant star and has an angular diameter of 30$^{\circ}$.[]{data-label="fig:fig1"}](cir.eps)
The CIR induced super-orbital modulation in SgXBs {#sec:so}
=================================================
In order to show how the presence of CIRs around the supergiant star can introduce a super-orbital modulation of the X-ray luminosity from a SgXB, we assume here a simplified NS accretion scenario, following the treatment in @oskinova12. The cross section of the NS for the capture of the stellar wind material is provided by the so-called accretion radius, $r_{\rm accr}=\frac{2Gm_{\rm X}}{\varv_{\rm rel}^2}$, where $m_{\rm X}$ is the NS mass and $\varv_{\rm rel}$ is the relative velocity between the NS and the massive star wind. For the CIR, both the radial and tangential components of the velocity at the NS location are taken into account in the computation. The mass accretion rate onto the NS is given by $S_{\rm accr}=4\pi \zeta \frac{(Gm_{\rm X})^2}{\varv^3_{\rm rel}}\rho$, and the correspondingly released X-ray luminosity is $L^{\rm d}_{\rm X}=\eta S_{\rm accr} c^2$. In the equations above, we indicated with $c$ the speed of light, with $\rho$ the local wind density (derived from the outputs of the hydrodynamic model), and considered for all cases of interest $\zeta\sim 1$ and $\eta\sim 0.1$ [$\zeta$ is a parameter included to take into account corrections related to the contribution of the radiation pressure and the finite cooling time of the gas, while $\eta$ parametrizes the efficiency of accretion onto a NS; see, e.g., @ostriker]. It is clear from this simplified treatment that both the density and velocity contrasts of the CIR compared to the smooth wind can affect the resulting X-ray luminosity, as they can significantly alter the mass accretion rate and the size of the NS cross section for the capture of the wind material.
Let’s consider first a case in which the NS orbital period is given by $P_{\rm NS}$ and a single CIR rotating with a period $P_{\rm CIR}$ is present in the wind of the supergiant companion. A difference between $P_{\rm NS}$ and $P_{\rm CIR}$ can be expected in a not-synchronously rotating binary [see, e.g., @koeni06 and references therein], or in case the stellar spot does not rotate with the same velocity of the supergiant star [@lobel08]. The mass accretion rate onto the NS is altered every time the compact object encounters the CIR along its orbit and the amplitude of the variation is regulated by the CIR velocity/density contrast compared to the rest of the stellar wind. The period of the super-orbital modulation, $P_{\rm SO}$, is thus given by the difference between the NS and the CIR angular velocities: $$P_{\rm SO}=\frac{1}{\left|\frac{1}{P_{\rm NS}}-\frac{1}{P_{\rm CIR}}\right|}=
\left|\frac{P_{\rm NS} P_{\rm CIR}}{P_{\rm CIR}-P_{\rm NS}}\right|.
\label{eq:po}$$
As an example, we consider in detail the case of the classical SgXB IGRJ16493-4348. The donor star in this system has the same spectral type as JPup and the estimated distance is 6-26 kpc [see @nunez17 and references therein]. Given the typical parameters of a B0.5 Ib star (Sect. \[sec:cir\]) and the measured orbital period of the NS in IGRJ16493-4348 (Table \[tab:tab1\]), the separation between the compact object and the supergiant is $\sim$1.8 $R_*$. According to Eq. \[eq:po\], a single CIR rotating with a period of $\sim$10.3 days is thus expected to give rise to a super-orbital period of $\sim$20 days, as indeed observed in this source. The X-ray lightcurve simulated using the output of the hydrodynamic model for the CIR in JPup (Fig. \[fig:fig2\]) is characterized by an average luminosity which is in good agreement with the observations of IGRJ16493-4348[^1]. The amplitude of the modulation is smaller than that observed from IGRJ16493-4348, but the exact value depends on, e.g., the brightness of the spot on the stellar surface responsible for the CIR generation (in the present case the hydrodynamical simulations were tuned to reproduce the results UV spectroscopic monitoring for J Pup).
![Simulated long term lightcurve of IGRJ16493-4348, assuming a single CIR in this system rotating with a period of $\sim$10.3 days.[]{data-label="fig:fig2"}](dac1cir1.8.ps)
In a situation in which multiple CIRs are present and intercept the plane of the NS orbit, it would be naturally expected that their density/velocity contrasts are significantly different, reflecting the distinct properties of the stellar surface spots from which they originated. Assuming that $n$ CIRs cross the NS orbit, the overall period of the super-orbital modulation will be in this case $n$$\times$$P_{\rm SO}$. Several peaks of variable intensities can thus be originated in the modulation profile depending on the different CIR density/velocity contrast at the NS location. We show in the bottom plot of Fig. \[fig:fig3\] an example in which a second CIR is included in our calculations, giving rise to a double-peaked super-orbital modulation (see also the caption of Fig. \[fig:fig1\]). A multiple-peaked super-orbital modulation seems particularly interesting to reproduce structured profiles of the folded X-ray lightcurves displayed by, e.g., the classical SgXB 4U1909+07 and the SFXT IGRJ16479-4514 [see Fig. 8 and 10 of @corbet13]. Figure \[fig:fig3\] shows that the super-orbital variability is critically depending on the difference between $P_{\rm NS}$ and $P_{\rm CIR}$. Although CIRs seem to be an ubiquitous property of all supergiant stars [see, e.g., @massa15 and references therein], the detectability of a super-orbital modulation in the currently proposed model could be hampered in all those unfavorable cases where $P_{\rm NS}\simeq P_{\rm CIR}$ and $P_{\rm SO}$ exceeds the available observational time-span for a SgXB.
Although all the free parameters on the number and properties of the CIRs can be fine tuned to obtain a reasonable match with the properties of the super-orbital modulations of all sources in Table \[tab:tab1\], there is not an obvious way to explain the empirical relation connecting the system orbital and super-orbital period as shown by @corbet13. If the relation will be confirmed by future observations, a better understanding of the CIR formation in the SgXBs is required to investigate the possibility of explaining this observational finding in the current model.
![[*Top*]{}: Same as Fig. \[fig:fig2\] but assuming an orbital separation of 2.5 $R_*$, corresponding to an orbital period of $\sim$10.6 days. As this is only slightly larger than the CIR period ($\sim$10.3 days), the super-orbital modulation occurs on a much longer time-scale ($\sim$330 days). [*Bottom*]{}: Same as above but using 2 CIRs. The super-orbital modulation is now characterized by a double asymmetric peak repeating every $\sim$660 days, corresponding to the fact that the NS is crossing two CIRs with different density/velocity contrasts along its orbit.[]{data-label="fig:fig3"}](dac1cir2.5.ps "fig:") ![[*Top*]{}: Same as Fig. \[fig:fig2\] but assuming an orbital separation of 2.5 $R_*$, corresponding to an orbital period of $\sim$10.6 days. As this is only slightly larger than the CIR period ($\sim$10.3 days), the super-orbital modulation occurs on a much longer time-scale ($\sim$330 days). [*Bottom*]{}: Same as above but using 2 CIRs. The super-orbital modulation is now characterized by a double asymmetric peak repeating every $\sim$660 days, corresponding to the fact that the NS is crossing two CIRs with different density/velocity contrasts along its orbit.[]{data-label="fig:fig3"}](dac2cir2.5.ps "fig:")
We neglected in the present simplified approach the role of the NS spin period and magnetic field, as well as the effect of eccentric orbits and the interaction between the X-rays from the NS and the stellar wind. A strong NS magnetic field and slow spin period can induce large modulations of the X-ray luminosity, due to the onset of magnetic and/or centrifugal gates [@grebenev07; @bozzo08]. While this is relevant to explain the short time-scale X-ray variability ($\sim$10-1000 s) displayed by classical SgXBs and SFXTs [@bozzo16], we expect this variability to be largely smoothed out when considering the much longer integration times corresponding to the super-orbital modulations (tens of days).
As described in @bozzo16, the lack of a proper treatment of the X-ray illumination of the stellar wind by the accreting compact object limits the validity of the outcomes of the calculations to low luminosity SgXBs ($L_{\rm X}$$\lesssim$10$^{35}$ erg s$^{-1}$) and can only provide indications in case of brighter systems. This effect is thus unlikely to be critical in the case of the SFXTs, as their average X-ray luminosity is generally far below the critical level required to produce a systematic disruption of the stellar wind on scales that are as large as those expected for the CIRs [see, e.g., @ducci10 and references therein]. However, this might not apply to bright persistent classical wind-fed SgXBs with average X-ray luminosities $\gg$10$^{36}$-10$^{37}$ erg s$^{-1}$ [e.g., VelaX-1; @watanabe06; @sander17], which we predict to not display super-orbital modulations. Note that all SgXBs discovered so far to display a super-orbital modulation are characterized by relatively low long-term luminosities, the brightest being IGRJ16493-4348 with an estimated average X-ray luminosity of $\sim$1.5$\times$10$^{36}$ erg s$^{-1}$ when the largest allowed distance of 26 kpc is considered.
Finally, the usage of circular orbits in our calculation was adopted to limit the number of required hydrodynamic runs and to provide more intuitive examples to promote the proposed scenario. However, we do not expect that our conclusions will change significantly with the introduction of a relatively small eccentricity [$\lesssim$0.2; @corbet13 and references therein], as the range of orbital separations spanned by the NSs would be limited and the density/velocity contrasts within each CIRs are expected to undergo major variations only on several stellar radii (for reasonable assumptions of the model parameters).
Discussion and conclusions {#sec:conclusion}
==========================
We proposed in this letter that the still poorly understood super-orbital variability displayed by several classical SgXBs and SFXTs is related to the presence of CIRs in the winds of their OB supergiants. The mechanisms leading to the formation of CIRs are not yet fully understood and the model considered in Sect. \[sec:cir\] exploits the most widely accepted idea that CIRs are driven by spots on the supergiant surface with a number of free parameters that can be adjusted to reproduce the observational properties of DACs in many isolated OB supergiants. At present, no observations of DACs are available for the donor stars in classical SgXBs and SFXTs, and it is thus difficult to present an exhaustive exploration of the entire model parameter space (this is beyond the scope of this work, given also the computationally expensive runs of the hydrodynamic simulations). However, the examples provided in Sect. \[sec:so\] show that the number and physical characteristics of the CIRs can be adjusted within reasonable boundaries to reproduce the main observed properties of the super-orbital modulations displayed by all sources in Table \[tab:tab1\]. This paves the way to future theoretical and simulation efforts exploring the proposed model in more details and overcoming the simplifications adopted in the current approach, including the presence of the X-ray irradiation of the stellar wind and eccentric orbits.
We remark that an important open question in the context of the present interpretation of the super-orbital modulations is if CIRs can be stable for years. Indeed, most of the DAC observations leading to the concept of CIRs were carried out as part of the MEGA campaign [@mega] performed with the IUE satellite [@iue] and no longer repeated afterwards. Long dedicated monitoring campaigns of the UV spectroscopic variability of the previously observed supergiant stars and the other supergiants hosted in the SgXBs are thus critically required.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the anonymous referee for constructive comments which helped to improve the paper. This publication was motivated by a team sponsored by the ISSI in Bern, Switzerland. EB and LO thank ISSI for the financial support during their stay in Bern. EB is grateful for the hospitality of the Institut für Physik und Astronomie (Universität Potsdam) during part of this work. EB acknowledges financial traveling contribution from the Swiss Society for Astronomy and Astrophysics. LO acknowledges support by the DLR grant 50OR1302 and partial support by the Russian Government Program of Competitive Growth of Kazan Federal University. AL acknowledges partial financial support by the Belgian Federal Science policy Office under contract No. BR/143/A2/BRASS and by the ESA-Gaia Prodex Programme 2015-2017.
[^1]: The absolute value of the average X-ray luminosity from IGRJ16493-4348 ranges from $\sim$10$^{35}$ erg s$^{-1}$ to $\sim$10$^{36}$ erg s$^{-1}$, depending on the poorly known distance to the source (6-26 kpc).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we classify all blocks with defect group $C_{2^n}\times C_2\times C_2$ up to Morita equivalence. Together with a recent paper of Wu, Zhang and Zhou, this completes the classification of Morita equivalence classes of $2$-blocks with abelian defect groups of rank at most $3$. The classification holds for blocks over a suitable discrete valuation ring as well as for those over an algebraically closed field. The case considered in this paper is significant because it involves comparison of Morita equivalence classes between a group and a normal subgroup of index $2$, so requires novel reduction techniques which we hope will be of wider interest. We note that this also completes the classification of blocks with abelian defect groups of order dividing $16$ up to Morita equivalence. A consequence is that Broue’s abelian defect group conjecture holds for all blocks mentioned above.'
author:
- 'Charles Eaton[^1] and Michael Livesey[^2]'
date: 13th October 2017
title: 'Classifying blocks with abelian defect groups of rank $3$ for the prime $2$ [^3]'
---
Introduction
============
Let $p$ be a prime and $(K,\mathcal{O},k)$ a $p$-modular with $k$ algebraically closed. Let $P$ be a finite $p$-group. Donovan’s conjecture states that there are only finitely many Morita equivalence classes amongst blocks of $kG$ for finite groups $G$ with defect groups isomorphic to $P$, and it is natural to strengthen this conjecture to blocks with respect to ${\mathcal{O}}$. Advances in our understanding of blocks of finite groups of Lie type in non-defining characteristic open the possibility of using the classification of finite simple groups to tackle this conjecture and further to classify Morita equivalence classes of blocks. For $p=2$ this process has been started in [@ekks14], where Donovan’s conjecture (with respect to $k$) has been proved for all elementary abelian $2$-groups. For elementary abelian $2$-groups of order at most $16$ the Morita equivalence classes have further been classified, with respect to ${\mathcal{O}}$ (see [@ea17]). Abelian $p$-groups with a cyclic factor of order strictly great than $p$ present a significant problem to the extension of these results. This is because the case of a group generated by a normal subgroup and a defect group is especially difficult to to study with respect to Morita equivalence, and required the application of [@kk96] which applies only to split extensions by a factor of the defect group, and further only to blocks defined over $k$. In [@eatliv17] a partial generalization of [@kk96] was given (generalized further in Proposition \[prop:grunit\] below) which was sufficient to work with the Loewy length of blocks with arbitrary abelian defect groups. In this paper we combine this result with the existence of a certain perfect isometry from [@wa05] to prove Donovan’s conjecture for blocks (defined over ${\mathcal{O}}$) with defect groups $C_{2^n} \times C_2 \times C_2$ for $n>1$ and further show that for each $n$ there are precisely three Morita equivalence classes of such blocks. This completes the classification of Morita equivalence classes of blocks with abelian defect groups of order dividing $16$ (see [@li94], [@ea16] and [@ea17] for the elementary abelian $2$-groups and [@ekks14] for $C_4 \times C_4$, noting that in all other cases ${\rm Aut}(D)$ is a $2$-group and so all blocks with that defect group are nilpotent). In [@wzz17] it is shown that for $m,n \in {\mathbb{N}}$ with $n \geq 2$, if a block has defect group $D \cong C_{2^n} \times C_{2^n} \times C_{2^m}$, then it is Mortia equivalent to it Brauer correspondent in $N_G(P)$ and so to one of ${\mathcal{O}}D$, ${\mathcal{O}}(D \rtimes C_3)$, ${\mathcal{O}}(D \rtimes C_7)$ or ${\mathcal{O}}(D \rtimes (C_7 \rtimes C_3))$. Hence the classification of Morita equivalence classes of $2$-blocks with abelian defect groups of rank at most $3$ is complete.
We refer the reader to [@ea17] for a survey of progress on the problem of classifying Morita equivalence classes of blocks with a given defect group.
Recall the definition of the inertial quotient of a block $B$ of ${\mathcal{O}}G$ with defect group $D$, where $G$ is a finite group. Let $b_D$ be a block of $C_G(D)$ with Brauer correspondent $(b_D)^G=B$. The stabilizer of $b_D$ in $N_G(D)$ under conjugation is written $N_G(D,b_D)$. Then $N_G(D,b_D)/DC_G(D)$ is a $p'$-group, and is called the inertial quotient of $B$ (unique up to isomorphism). $|E|$ is called the inertial index. $(D,b_D)$ is called a maximal $B$-subpair.
\[inertial\_quotient\_lem\] Let $G$ be a finite group and $B$ be a block of ${\mathcal{O}}G$ with defect group $D \cong C_{2^n}\times C_2\times C_2$ for $n>1$. There are two possible fusion systems for $B$, given by $D$ and $C_{2^n} \times A_4$. In particular the possible inertial quotients for $B$ are $1$ and $C_3$.
We refer the reader to [@ako] for background on fusion systems. Since $D$ is abelian, the fusion systems on $D$ are given by groups $D \rtimes E$, where $E$ is an odd-order subgroup of ${\rm Aut}(D)$. We have ${\rm Aut}(D) \cong S_3$, so the possibilities are $E=1$ or $C_3$.
The main result is as follows (see Theorem \[thm:main\]):
Let $G$ be a finite group and $B$ a block of $\mathcal{O}G$ with defect group $D\cong C_{2^n}\times C_2\times C_2$ for $n>1$. Then $B$ is Morita equivalent to the principal block of $\mathcal{O}(C_{2^n}\times C_2\times C_2)$, $\mathcal{O}(C_{2^n}\times A_4)$ or $\mathcal{O}(C_{2^n}\times A_5)$.
Combining with the results of [@ea16], [@ekks14] and [@wzz17] we conclude:
Let $G$ be a finite group and $B$ a block of $\mathcal{O}G$ with defect group $D$ of $2$-rank at most $3$. Then $B$ is Morita equivalent to the principal block of one of:
\(i) ${\mathcal{O}}D$;
\(ii) ${\mathcal{O}}(D \rtimes C_3)$;
\(iii) ${\mathcal{O}}(C_{2^n}\times A_5)$ for $n \geq 0$;
\(iv) ${\mathcal{O}}(D \rtimes C_7)$;
\(v) ${\mathcal{O}}SL_2(8)$;
\(vi) ${\mathcal{O}}(D \rtimes (C_7 \rtimes C_3))$;
\(vii) ${\mathcal{O}}J_1$;
\(viii) ${\mathcal{O}}{\rm Aut}(SL_2(8))$.
Observe that this means that every block with this defect group is Morita equivalent to a principal block, and so in particular the Morita Frobenius number as defined in [@ke05] is one. Note that if $B$ above is not nilpotent, then the number $l(B)$ of irreducible Brauer characters of $B$ is $3$ and the number of irreducible characters $k(B)=2^{n+2}$ (by Proposition \[per\_isom\], which does not use the classification of finite simple groups).
Another consequence of Theorem \[thm:main\] is the following (see Corollary \[cor:derived\]):
Let $G$ be a finite group and $B$ a block of $\mathcal{O}G$ with defect group $D\cong C_{2^n}\times C_2\times C_2$ for $n>1$. Let $b$ be the unique block of $N_G(D)$ with $b^G=B$. Then $B$ and $b$ are derived equivalent.
Gathering together previous results, this completes the proof of Broué’s conjecture for $2$-blocks of defect at most $4$ and also for those of rank at most $3$ (see Corollaries \[cor:derived16\] and \[cor:derivedrank3\]).
The paper is structured as follows. In Section 2 we address the problem of lifting a Morita equivalence from a normal subgroup of index $2$. We obtain a perfect isometry from [@wa05] and show that this may be modified by a perfect self-isometry to produce a Morita equivalence using the central unit described in [@eatliv17]. In Section 3 we apply the classification in [@ekks14] to prove the main theorem, using the results of Section 2 to help reduce to quasisimple groups.
Normal subgroups of index 2
===========================
We first introduce some notation.
Let $G$ be a finite group and $N \lhd G$. For a block $B$ of ${\mathcal{O}}G$ we write ${\rm Irr}(B)$ for the set of irreducible characters of in $B$ (with respect to $K$) and ${\rm Irr}(B,\psi)$ for the set of irreducible characters in $B$ covering $\psi$ (that is, whose restriction has $\psi$ as a summand). Write ${\rm IBr}(B)$ for the set of irreducible Brauer characters. Write ${\rm prj}(B)$ for the set characters of projective indecomposable $B$-modules.
Suppose $B$ has defect group $D$. Let $(D,b_D)$ be a maximal $B$-subpair (see [@alp86] for background on subpairs). Note that all maximal $B$-subpairs are $G$-conjugate. If $u \in D$ and $b_u$ is a block of ${\mathcal{O}}C_G(u)$ with $(b_D)^{C_G(u)}=b_u$, then we call $(u,b_u)$ a subsection in $(D,b_D)$, and write $(u,b_u) \in (D,b_D)$.
Let $G$ be a finite group and $B$ a block of $\mathcal{O}G$. We denote by ${\rm CF}(G,B,K)$ the $K$-subspace of class functions on $G$ spanned by ${\rm Irr}(B)$, by ${\rm CF}(G,B,\mathcal{O})$ the $\mathcal{O}$-submodule $$\begin{aligned}
\{\phi\in {\rm CF}(G,B,K):\phi(g)\in\mathcal{O}\text{ for all }g\in G\}\end{aligned}$$ of ${\rm CF}(G,B,K)$ and by ${\rm CF}_{p'}(G,B,\mathcal{O})$ the $\mathcal{O}$-submodule $$\begin{aligned}
\{\phi\in {\rm CF}(G,B,\mathcal{O}):\phi(g)=0\text{ for all }g\in G\backslash G_{p'}\}\end{aligned}$$ of ${\rm CF}(G,B,\mathcal{O})$. Now in addition let $H$ be a finite group and $C$ a block of $\mathcal{O}H$. A **perfect isometry** between $B$ and $C$ is an isometry $$\begin{aligned}
I:\mathbb{Z}{\rm Irr}(B)\to\mathbb{Z}{\rm Irr}(C),\end{aligned}$$ such that $$\begin{aligned}
I_K:=K\otimes_{\mathbb{Z}}I:K{\rm Irr}(B)\to K{\rm Irr}(C),\end{aligned}$$ induces an ${\mathcal{O}}$-module isomorphism between ${\rm CF}(G,B,\mathcal{O})$ and ${\rm CF}(H,C,\mathcal{O})$ and also between ${\rm CF}_{p'}(G,B,\mathcal{O})$ and ${\rm CF}_{p'}(H,C,\mathcal{O})$. (Note that by an isometry we mean an isometry with respect to the usual inner products on $\mathbb{Z}{\rm Irr}(B)$ and $\mathbb{Z}{\rm Irr}(C)$, so for all $\chi\in{\rm Irr}(B)$, $I(\chi)=\pm\psi$ for some $\psi\in{\rm Irr}(C)$).
An alternative way of phrasing the condition that $I_K$ induces an isomorphism between ${\rm CF}_{p'}(G,B,\mathcal{O})$ and ${\rm CF}_{p'}(H,C,\mathcal{O})$ is that $I$ induces an isomorphism $\mathbb{Z}{\rm prj}(B) \cong \mathbb{Z}{\rm prj}(C)$.
We will be using the following well-known result frequently, so we include a proof.
\[index\_p\_background\_lem\] Let $G$ be a finite group and $N$ a normal subgroup of $G$ of index a power of $p$. Let $B$ be a block of $\mathcal{O}G$ with defect group $D$ covering a block $b$ of ${\mathcal{O}}N$. Then $B$ is the unique block of ${\mathcal{O}}G$ covering $b$, $D \cap N$ is a defect group for $b$ and the stabilizer of $b$ in $G$ is $DN$.
That $B$ is the unique block of ${\mathcal{O}}G$ covering $b$ is [@feit V.3.5]. The rest follows from [@alp86 15.1], noting that there is a one to one correspondence between blocks defined over $k$ and blocks defined over ${\mathcal{O}}$.
\[prop:grunit\] Let $G$ be a finite group and $N$ a normal subgroup of $G$ of index $p$. Now let $B$ be a block of $\mathcal{O}G$ with abelian defect group $D$ such that $G=ND$. Then there exists a block $b$ of $\mathcal{O}N$ with the same block idempotent as $B$ and defect group $D\cap N$. Moreover there exists a $G/N$-graded unit $a\in Z(B)$, in particular $B=\bigoplus_{j=0}^{p-1}a^jb$.
By [@eatliv17 Theorem 2.1] there exists a $G/N$-graded unit $\overline{a}\in Z(kB)$. Now as every element of $Z(kG)$ can be lifted to an element of $Z(\mathcal{O}G)$ we can lift $\overline{a}$ to $a\in Z(B)$. As the block idempotent of $B$ lies in $\mathcal{O}N$ and $Z(\mathcal{O}G)$ is $G/N$-graded, then we can assume $a$ is $G/N$-graded. Finally, as $Z(B)$ is a local ring and $a$ certainly does not lie in its maximal ideal ($\overline{a}$ is a unit), we have that $a$ is a unit.
\[rem:inner\] In the setting of the previous Proposition \[prop:grunit\] we note that conjugation by $G$ induces only inner automorphisms of $b$, and so in particular $Z(b) \subseteq Z(B)$.
\[prop:index\_p\] Let $G$, $N$, $B$ and $b$ be as in the Proposition \[prop:grunit\]. Then
\(i) Every irreducible character of $b$ is $G$-stable and extends to $p$ distinct irreducible characters of $B$.
\(ii) Induction ${\rm Ind}_N^G$ gives a bijection between the projective indecomposable $b$-modules and the projective indecomposable $B$-modules.
\(iii) Now let $G'$, $N'$, $B'$ and $b'$ be another quadruple satisfying Proposition \[prop:grunit\]. For each $\chi \in {\rm Irr}(b)$ write ${\rm Irr}(B,\chi)= \{\chi_1,\ldots, \chi_p \}$.
Suppose $I:\mathbb{Z}{\rm Irr}(B)\to\mathbb{Z}{\rm Irr}(B')$ is a perfect isometry such that for each $\chi \in {\rm Irr}(b)$ there is $\psi \in {\rm Irr}(b')$ and $\epsilon_\chi \in\{\pm1\}$ such that $I(\chi_i)=\epsilon_\chi \psi_i$ for $i=1, \ldots, p$ where ${\rm Irr}(B',\psi)= \{\psi_1, \ldots , \psi_p \}$. Then the isometry $I_{N,N'}:\mathbb{Z}{\rm Irr}(b)\to\mathbb{Z}{\rm Irr}(b')$ defined by $I_{N,N'}(\chi):=\epsilon_\chi \psi$ is perfect.
$ $ (i) By Remark \[rem:inner\] every character $\chi$ of $b$ is $G$-stable. Therefore, as $G/N$ is cyclic, $\chi$ extends to $G$. Taking the product with the $p$ distinct linear characters of $G/N$ inflated to $G$ gives the $p$ extensions of $\chi$ to $G$. Now every constituent of an irreducible character of $B$ restricted to $N$ must lie in $b$ and so every irreducible character of $B$ is the extension of some irreducible character of $b$.
\(ii) Certainly every projective indecomposable $B$-module is a summand of some projective indecomposable $b$-module induced to $G$ and Green’s indecomposability ensures ${\rm Ind}_N^G$. To prove we have a bijection we need ${\rm Ind}^G_N(P)$ and ${\rm Ind}^G_N(Q)$ to be non-isomorphic for non-isomorphic projective indecomposable $b$-modules $P$ and $Q$. However, this is true since by \[prop:grunit\], ${\rm Res}^G_N{\rm Ind}^G_N(P)\cong P^{\oplus p}$.
\(iii) For $\chi \in {\rm Irr}(b)$ and $\psi \in {\rm Irr}(b')$, write $\tilde{\chi}=\sum_{i=1}^p \chi_i$ and $\tilde{\psi}=\sum_{i=1}^p \psi_i$.
Let $$\begin{aligned}
\phi=\sum_{\chi\in{\rm Irr}(b)}\alpha_\chi\chi\in{\rm CF}(N,b,\mathcal{O}),\end{aligned}$$ for some $\alpha_\chi\in K$ and consider $$\begin{aligned}
\tilde{\phi}:=\sum_{\chi\in {\rm Irr}(b)} \alpha_\chi\tilde{\chi}.\end{aligned}$$ Note that $\tilde{\phi}(g)=p\phi(g)$ for all $g\in N$ and $\tilde{\phi}(g)=0$ for all $g\in G\backslash N$, so $\tilde{\phi}\in p{\rm CF}(G,B,\mathcal{O})$. Now note that $$\begin{aligned}
(I_{N,N'})_K(\phi)=\sum_{\psi\in{\rm Irr}(b')}\beta_\psi\psi,\end{aligned}$$ where $$\begin{aligned}
I_K(\tilde{\phi})=\sum_{\psi\in{\rm Irr}(b')} \beta_\psi\tilde{\psi}.\end{aligned}$$ As $\tilde{\phi}\in p{\rm CF}(G,B,\mathcal{O})$, we have $I_K(\tilde{\phi})\in p{\rm CF}(G',B',\mathcal{O})$. Again $I_K(\tilde{\phi})(g)=p(I_{N,N'})_K(\phi)(g)$ for all $g\in N'$ and so $(I_{N,N'})_K(\phi)\in{\rm CF}(N',b',\mathcal{O})$. So $(I_{N,N'})_K({\rm CF}(N,b,\mathcal{O}))\subseteq{\rm CF}(N',b',\mathcal{O})$ and by an identical argument $(I_{N,N'})_K^{-1}({\rm CF}(N',b',\mathcal{O}))\subseteq{\rm CF}(N,b,\mathcal{O})$. Now suppose in addition that $\phi\in{\rm CF}_{p'}(N,b,\mathcal{O})$. Then $\tilde{\phi}(g)=p\phi(g)=0$ for all $g\in N\backslash N_{p'}$ and $\tilde{\phi}(g)=0$ for all $g\in G\backslash N$ and so $\tilde{\phi}\in{\rm CF}_{p'}(G,B,\mathcal{O})$. Therefore $I_K(\tilde{\phi})\in{\rm CF}_{p'}(G',B',\mathcal{O})$ and by the previous paragraph $(I_{N,N'})_K(\phi)\in{\rm CF}_{p'}(N',b',\mathcal{O})$ and so $(I_{N,N'})_K$ induces an isomorphism between ${\rm CF}_{p'}(N,b,\mathcal{O})$ and ${\rm CF}_{p'}(N',b',\mathcal{O})$ and hence it satisfies both the properties of a perfect isometry.
\[lem:isomcent\] Let $G$ and $G'$ be finite groups, $B$ and $B'$ blocks of ${\mathcal{O}}G$ and ${\mathcal{O}}G'$ respectively and $I:\mathbb{Z}{\rm Irr}(B)\to\mathbb{Z}{\rm Irr}(B')$ a perfect isometry.
\(i) The $K$-algebra isomorphism between $Z(KB)$ and $Z(KB')$ given by the bijection of character idempotents induced by $I$ induces an ${\mathcal{O}}$-algebra isomorphism $\phi_I:Z(B)\to Z(B')$.
\(ii) Suppose further that $I$ satisfies the conditions of Proposition \[prop:index\_p\]. Then $\phi_{I_{N,N'}}=\phi_I|_{Z(b)}$.
$ $ (i) This is proved in [@br90].
\(ii) Let $\chi\in{\rm Irr}(b)$ and $\pm I(\chi)=\psi\in{\rm Irr}(b')$. Then, adopting the notation of Proposition \[prop:index\_p\], we have that $$\begin{aligned}
e_\chi=e_{\chi_1}+\dots+e_{\chi_p}\text{ and }e_\psi=e_{\psi_1}+\dots+e_{\psi_p},\end{aligned}$$ where $e_\varphi$ is the block idempotent of $\varphi$ in the appropriate group algebra over $K$. The statement now follows from the definitions of $I_{N,N'}$ and $\phi_I$.
Let $n$ be a positive integer. We now work towards constructing all the perfect self-isometries of $\mathcal{O}(C_{2^n}\times A_4)$. These will ultimately be used in Theorem \[index2theorem\]. From now on we assume $p=2$. Let $\omega\in K$ be a primitive $3^{\operatorname{rd}}$ root of unity. We recall the character table of $A_4$, where we also set up some labelling of characters. $$\begin{aligned}
\begin{tabular}{|c||c|c|c|c|} \hline
& $()$ & $(12)(34)$ & $(123)$ & $(132)$ \\ \hline
$\chi_1$ & $1$ & $1$ & $1$ & $1$ \\
$\chi_2$ & $1$ & $1$ & $\omega$ & $\omega^2$ \\
$\chi_3$ & $1$ & $1$ & $\omega^2$ & $\omega$ \\
$\chi_4$ & $3$ & $-1$ & $0$ & $0$ \\ \hline
\end{tabular}\end{aligned}$$
\[prop:self\_A4\] The perfect self-isometries of $\mathcal{O}A_4$ are precisely the isometries of the form: $$\begin{aligned}
I_{\sigma,\epsilon}:\mathbb{Z}{\rm Irr}(A_4)&\to\mathbb{Z}{\rm Irr}(A_4)\\
\chi_j&\mapsto\epsilon\delta_j\delta_{\sigma(j)}\chi_{\sigma(j)}\end{aligned}$$ for $1\leq j\leq 4$, where $\sigma\in S_4$, $\epsilon\in\{\pm1\}$ and $\delta_1=\delta_2=\delta_3=-\delta_4=1$. Hence the group of perfect self-isometries is isomorphic to $C_2 \times S_4$.
We first note that $$\begin{aligned}
{\rm prj}(\mathcal{O}A_4))=\{\chi_{P_1},\chi_{P_2},\chi_{P_3}\},\text{ where }\chi_{P_j}=\chi_j+\chi_4\text{ for }j=1,2,3.\end{aligned}$$ Therefore $$\begin{aligned}
\mathbb{Z}{\rm prj}(\mathcal{O}A_4))=\left\{\sum_{j=1}^4a_j\chi_j:a_j\in\mathbb{Z},\sum_{j=1}^3a_j=a_4\right\}.\end{aligned}$$ So the isometries $\mathbb{Z}{\rm Irr}(\mathcal{O}A_4)\to\mathbb{Z}{\rm Irr}(\mathcal{O}A_4)$ that leave $\mathbb{Z}{\rm prj}(\mathcal{O}A_4))$ invariant are precisely the permutations of $\{\chi_1,\chi_2,\chi_3,-\chi_4\}$ together with their negatives. These are precisely the $I_{\sigma,\epsilon}$’s. We now describe ${\rm CF}(A_4,\mathcal{O}A_4,\mathcal{O})$. Let $\sum_{j=1}^4a_j\chi_j\in{\rm CF}(A_4,\mathcal{O}A_4,K)$. Now by evaluating at various elements of $A_4$ we get that $\sum_{j=1}^4a_j\chi_j\in{\rm CF}(A_4,\mathcal{O}A_4,\mathcal{O})$ if and only if $$\begin{aligned}
a_1+a_2+a_3+3a_4&\in\mathcal{O},\\
a_1+a_2+a_3-a_4&\in\mathcal{O},\\
a_1+\omega a_2+\omega^2a_3&\in\mathcal{O}\\
\text{and }a_1+\omega^2a_2+\omega a_3&\in\mathcal{O}.\end{aligned}$$ These conditions are equivalent to $$\begin{aligned}
4a_1\in\mathcal{O},4a_2\in\mathcal{O},4a_3\in\mathcal{O},4a_4\in\mathcal{O}\\
\text{and }\delta_ja_j-\delta_la_l\in\mathcal{O}\text{ for all }1\leq j,l\leq 4.\end{aligned}$$ One can now check that all the $I_{\sigma,\epsilon}$’s leave ${\rm CF}(A_4,\mathcal{O}A_4,\mathcal{O})$ invariant and the proposition is proved.
Let $\zeta\in K$ be a primitive $(2^{n})^{\operatorname{th}}$ root of unity.
\[lem:roots\_in\_o\] Let $m$ be a positive integer and suppose $\sum_{i=0}^{2^m-1}\zeta^{l_i}\in2^m\mathcal{O}$, where $l_i\in\mathbb{Z}$ for $0\leq i<2^m$. Then either $\zeta^{l_0}=\dots=\zeta^{l_{2^m-1}}$ or $\sum_{i=0}^{2^m-1}\zeta^{l_i}=0$.
We consider $\mathbb{Q}$ as embedded in $K$. First note that $$\begin{aligned}
\prod_{\sigma\in\operatorname{Gal}(\mathbb{Q}(\zeta)/\mathbb{Q})}\left(\sigma\left(\sum_{i=0}^{2^m-1}\zeta^{l_i}\right)\right)\in2^{m|\operatorname{Gal}(\mathbb{Q}(\zeta)/\mathbb{Q})|}\mathcal{O}\cap\mathbb{Z}=2^{m|\operatorname{Gal}(\mathbb{Q}(\zeta)/\mathbb{Q})|}\mathbb{Z}.\end{aligned}$$ However, for each $\sigma\in\operatorname{Gal}(\mathbb{Q}(\zeta)/\mathbb{Q})$ we have $$\begin{aligned}
\left|\sigma\left(\sum_{i=0}^{2^m-1}\zeta^{l_i}\right)\right|\leq2^m\end{aligned}$$ with equality if and only if $\zeta^{l_0}=\dots=\zeta^{l_{2^m-1}}$, where $|\quad|$ denotes the usual norm in $K$. The claim now follows.
Let $x$ be a generator of $C_{2^n}$. For $0\leq i<2^n$ we define $\theta_i\in{\rm Irr}(C_{2^n})$ by $\theta_i(x)=\zeta^i$.
\[prop:self\_cyclic\] The perfect self-isometries of $\mathcal{O}C_{2^n}$ are precisely the isometries of the form: $$\begin{aligned}
I_{j,l,\epsilon}:\mathbb{Z}{\rm Irr}(C_{2^n})&\to\mathbb{Z}{\rm Irr}(C_{2^n})\\
\theta_i&\mapsto\epsilon\theta_{i+l}^j\end{aligned}$$ for $0\leq i<2^n$, where $0\leq j,l<2^n$ with $j$ odd, $\epsilon\in\{\pm1\}$ and $i+l$ is considered modulo $2^n$. Moreover each $I_{j,l,1}$ is induced by the $\mathcal{O}$-algebra automorphism $x\mapsto\zeta^{-l}x^{\frac{1}{j}}$. Hence the group of perfect self-isometries is isomorphic to $C_2\times[(\mathbb{Z}/2^n)^\times\ltimes(\mathbb{Z}/2^n)]$, where the action of $(\mathbb{Z}/2^n)^\times$ on $\mathbb{Z}/2^n$ is given by multiplication.
We know ${\rm prj}(\mathcal{O}C^{2^n})=\left\{\sum_{i=0}^{2^n-1}\theta_i\right\}$ so any perfect isometry $I:\mathbb{Z}{\rm Irr}(\mathcal{O}C_{2^n})\to\mathbb{Z}{\rm Irr}(\mathcal{O}C_{2^n})$ must have all signs the same. Therefore we need only check what permutations of the $\theta_i$’s leave ${\rm CF}(C_{2^n},\mathcal{O}C_{2^n},\mathcal{O})$ invariant. We first note that $$\begin{aligned}
\sum_{i=0}^{2^n-1}\zeta^{-i}\theta_i(g)=
\begin{cases}
2^n&\text{ if }g=x,\\
0&\text{ if }g\neq x,
\end{cases}\end{aligned}$$ for all $g\in C_{2^n}$. So $\frac{1}{2^n}\sum_{i=0}^{2^n-1}\zeta^{-i}\theta_i\in{\rm CF}(C_{2^n},\mathcal{O}C_{2^n},\mathcal{O})$. Now consider a perfect isometry induced by $\sigma$, a permutation of $\{0,1,\dots,2^n-1\}$. Then we must have $\frac{1}{2^n}\sum_{i=0}^{2^n-1}\zeta^{-i}\theta_{\sigma(i)}\in{\rm CF}(C_{2^n},\mathcal{O}C_{2^n}\mathcal{O})$. So $$\begin{aligned}
\sum_{i=0}^{2^n-1}\zeta^{-i}\theta_{\sigma(i)}(g)\in2^n\mathcal{O},\end{aligned}$$ for all $g\in C_{2^n}$. Therefore, by Lemma \[lem:roots\_in\_o\], we have that for each $g\in C_{2^n}$ that this sum is either zero or all the $\zeta^{-i}\theta_{\sigma(i)}(x)$’s are equal. Certainly they can’t all be zero, as we have a non-zero linear combination of characters. Therefore there exists $y\in C_{2^n}$ such that $$\begin{aligned}
\theta_{\sigma(0)}(y)=\zeta^{-1}\theta_{\sigma(1)}(y)=\dots=\zeta^{-(2^n-1)}\theta_{\sigma(2^n-1)}(y).\end{aligned}$$ Certainly $y$ must generate $C_{2^n}$ as it takes $2^n$ distinct values on the elements of $C_{2^n}$. Define $0\leq j,l<2^n$ by $x=y^j$ and $\theta_{\sigma(0)}(y)=\zeta^l$ and note that $j$ must be odd. Then $$\begin{aligned}
\theta^j_{m+l}(x)=\zeta^{j(m+l)}=[\zeta^m\theta_{\sigma(0)}(y)]^j=[\theta_{\sigma(m)}(y)]^j=\theta_{\sigma(m)}(y^j)=\theta_{\sigma(m)}(x),\end{aligned}$$ for all $0\leq m<2^n$. Therefore, $\theta_{\sigma(m)}=\theta^j_{m+l}$. Finally we note that the isometry $\theta_m\mapsto\theta_{m+l}^j$ is induced by the $\mathcal{O}$-algebra automorphism $x\mapsto\zeta^{-l}x^{\frac{1}{j}}$ and so is indeed a perfect isometry and the proof is complete.
For the following theorem we adopt the notation of Propositions \[prop:self\_A4\] and \[prop:self\_cyclic\].
\[thm:C2nA4\] Every perfect self-isometry of $\mathcal{O}(C_{2^n}\times A_4)$ is of the form $$\begin{aligned}
(I_{j,l,1},I_{\sigma,\epsilon}):(\mathbb{Z}{\rm Irr}(\mathcal{O}C_{2^n})\otimes_{\mathbb{Z}}\mathbb{Z}{\rm Irr}(\mathcal{O}A_4))\to(\mathbb{Z}{\rm Irr}(\mathcal{O}C_{2^n})\otimes_{\mathbb{Z}}\mathbb{Z}{\rm Irr}(\mathcal{O}A_4)),\end{aligned}$$ where $\sigma\in S_4$, $\epsilon\in\{\pm1\}$ and $0\leq j,l<2^n$ with $j$ odd.
The projective indecomposable characters are $$\begin{aligned}
{\rm prj}(\mathcal{O}(C_{2^n}\times A_4))=\{\chi_{P_1},\chi_{P_2},\chi_{P_3}\},\text{ where }\chi_{P_j}=\left(\sum_{i=0}^{2^n-1}\theta_i\right)\otimes\left(\chi_j+\chi_4\right).\end{aligned}$$ Let $I$ be a perfect self-isometry of $\mathcal{O}(C_{2^n}\times A_4)$. By counting constituents we see that $$\begin{aligned}
\label{align:im}
I(\chi_{P_l})=\pm\chi_{P_1},\pm\chi_{P_2},\pm\chi_{P_3},\pm(\chi_{P_1}-\chi_{P_2}),\pm(\chi_{P_1}-\chi_{P_3})\text{ or }\pm(\chi_{P_2}-\chi_{P_3}),\end{aligned}$$ for $1\leq l\leq 3$. Consider the set $$\begin{aligned}
X_m:=\left\{j\big{|}\left\langle\theta_l\otimes\chi_j,I\left(\left(\sum_{i=0}^{2^n-1}\theta_i\right)\otimes\chi_m\right)\right\rangle\neq0,\text{ for some }l\right\},\end{aligned}$$ for $1\leq m\leq 4$. By (\[align:im\]) we have shown that $|X_m|=1$ or $2$ for every $1\leq m\leq 4$. If $|X_1|=2$, then by considering (\[align:im\]) for $l=1$ we see that $X_4=X_1$. Similarly by considering $I(\chi_{P_2})$, we get that $X_2=X_4$. This is now a contradiction as then $$\begin{aligned}
I\left(\left(\sum_{i=0}^{2^n-1}\theta_i\right)\otimes\left(\chi_1+\chi_2+\chi_4\right)\right)\end{aligned}$$ has at most $2.2^n$ constituents with non-zero multiplicity. Therefore $|X_1|=1$ and so by considering $I(\chi_{P_1})$ we get that $|X_4|=1$ and then by considering $I(\chi_{P_2})$ and $I(\chi_{P_3})$ we get that $|X_2|=|X_3|=1$. Moreover, $X_1,X_2,X_3,X_4$ must all be disjoint. By composing $I$ with the perfect isometry $(I_{1,1,1},I_{\sigma,1})$, for some appropriately chosen $\sigma\in S_4$, we may assume $X_m=\{m\}$ for all $1\leq m\leq4$. Therefore $I(\chi_{P_l})=\pm\chi_{P_l}$ for $1\leq l\leq3$ and by considering $$\begin{aligned}
I\left(\left(\sum_{i=0}^{2^n-1}\theta_i\right)\otimes\chi_4\right),\end{aligned}$$ we see that in fact all these signs are the same and we may assume, possibly by composing $I$ with $(I_{1,1,1},I_{1,\epsilon})$, that $$\begin{aligned}
I\left(\left(\sum_{i=0}^{2^n-1}\theta_i\right)\otimes\chi_m\right)=\left(\sum_{i=0}^{2^n-1}\theta_i\right)\otimes\chi_m,\end{aligned}$$ for $1\leq m\leq 4$. Next we note that $$\begin{aligned}
\frac{1}{12}\theta_j\otimes(\chi_1+\chi_2+\chi_3+9\chi_4)\in{\rm CF}(C_{2^n}\times A_4,\mathcal{O}(C_{2^n}\times A_4),\mathcal{O}),\end{aligned}$$ for $0\leq j<2^n$. As $3$ is invertible in $\mathcal{O}$, this implies $$\begin{aligned}
\theta_j\otimes\left(\sum_{m=1}^4\chi_m\right)\in4{\rm CF}(C_{2^n}\times A_4,\mathcal{O}(C_{2^n}\times A_4),\mathcal{O}),\end{aligned}$$ and so $$\begin{aligned}
\label{align:im2}
I\left(\theta_j\otimes\left(\sum_{m=1}^4\chi_m\right)\right)\in4{\rm CF}(C_{2^n}\times A_4,\mathcal{O}(C_{2^n}\times A_4),\mathcal{O}),\end{aligned}$$ for $0\leq j<2^n$. Now set $\theta_{j_m}\otimes\chi_m:=I(\theta_j\otimes\chi_m)$, for $1\leq m\leq4$. Evaluating (\[align:im2\]) at $(x,1)$, $(x,(123))$ and $(x,(132))$ gives $$\begin{aligned}
\zeta^{j_1}+\zeta^{j_2}+\zeta^{j_3}+\zeta^{j_4}&\in4\mathcal{O},\label{zeta1}\\
\zeta^{j_1}+\omega\zeta^{j_2}+\omega^2\zeta^{j_3}&\in4\mathcal{O},\label{zeta2}\\
\zeta^{j_1}+\omega^2\zeta^{j_2}+\omega\zeta^{j_3}&\in4\mathcal{O}\label{zeta3}.\end{aligned}$$ Adding (\[zeta1\]), (\[zeta2\]) and (\[zeta3\]) gives $3\zeta^{j_1}+\zeta^{j_4}\in4\mathcal{O}$. Now Lemma \[lem:roots\_in\_o\] tells us that $\zeta^{j_1}=\zeta^{j_4}$ as certainly $3\zeta^{j_1}+\zeta^{j_4}\neq0$. Therefore, by (\[zeta1\]), $\zeta^{j_2}+\zeta^{j_3}\in2\mathcal{O}$. So again by Lemma \[lem:roots\_in\_o\] $\zeta^{j_2}=\zeta^{j_3}$ as $\zeta^{j_2}=-\zeta^{j_3}$ is prohibited by (\[zeta1\]). Substituting into (\[zeta2\]) gives $\zeta^{j_1}-\zeta^{j_2}\in4\mathcal{O}$. A final use of Lemma \[lem:roots\_in\_o\] tells us that $\zeta^{j_1}=\pm\zeta^{j_2}$ but $2\zeta^{j_1}\notin4\mathcal{O}$ and so we must have $\zeta^{j_1}=\zeta^{j_2}=\zeta^{j_3}=\zeta^{j_4}$. We have shown that we may assume $I$ is of the form $$\begin{aligned}
I(\theta_j\otimes\chi_m)\mapsto\theta_{\sigma(j)}\otimes\chi_m,\end{aligned}$$ for $0\leq j<2^n$ and $1\leq m\leq4$, where $\sigma$ is a permutation of $\{0,1,\dots,2^n-1\}$. In particular the $\mathcal{O}$-algebra automorphism of $Z(\mathcal{O}(C_{2^n}\times A_4)$ induced by $I$ leaves $\mathcal{O}C_{2^n}$ invariant. Using Proposition \[prop:self\_cyclic\] we can compose $I$ with $(I_{s,t,1},I_{1,1})$ for appropriately chosen $s$ and $t$ so that the automorphism induced on $\mathcal{O}C_{2^n}$ is the identity. Therefore $\sigma$ is the identity permutation, $I$ is the identity perfect isometry and the theorem is proved.
\[normaldefectlemma\] Let $B$ be a block of ${\mathcal{O}}G$ for a finite group $G$ with normal defect group $D \cong C_{2^n} \times C_2 \times C_2$ for some $n>1$. Then $B$ is Morita equivalent to ${\mathcal{O}}D$ or ${\mathcal{O}}(C_{2^n} \times A_4)$.
By Lemma \[inertial\_quotient\_lem\] the possible inertial quotients are $1$ and $C_3$. Since in either case the inertial quotient is cyclic the result follows from the main result of [@ku85].
The following appears as [@sa17 Theorem 15], but we include it here for completeness.
\[per\_isom\] Let $G$ be a finite group and $B$ a block of ${\mathcal{O}}G$ with defect group $D \cong C_{2^n} \times C_2 \times C_2$ for $n>1$. Then either $B$ is nilpotent or $l(B)=3$. Let $G'=C_{2^n} \times A_4$ and $B'={\mathcal{O}}G'$. If $B$ is not nilpotent then there is a perfect isometry between $B$ and $B'$.
Write $E$ for the inertial quotient of $B$. If $E=1$, then $B$ is nilpotent and the result holds. Hence we may assume $|E|=3$. We must first show that $l(B)=3$, and we do this by adapting a method used in [@ks13]. We proceed by induction on $n$. Assume that $l(B')=3$ whenever $B'$ is a block with defect group $C_{2^m} \times C_2 \times C_2$ for $m \geq 1$ and inertial index $3$, and observe that this is known to hold for $m=1$ by [@kkl12].
By [@km13] every irreducible character of $B$ has height zero, and so by [@ro92 1.2(ii)] we have $k(B) \leq |D|$. Let $(D,b_D)$ be a maximal $B$-subpair. Since $D$ is abelian, $N_G(D,b_D)$ controls fusion of $B$-subpairs in $(D,b_D)$. If $u \in D$, then let $b_u$ be the unique block of $C_G(u)$ such that $(u,b_u) \in (D,b_D)$. Write $D=\langle x,y_1,y_2 \rangle$, where $y$ has order $2^n$ and $y_1,y_2$ are involutions. Then $\mathcal{X}:=\{ (1,B),(x^i,b_{x^i}),(x^jy_1,b_{x^jy_1}):1 \leq i \leq 2^n-1, 0 \leq j \leq 2^n-1 \}$ form a complete set of $G$-conjugacy class representatives of subsections in $B$. By a well known reformulation of [@na98 5.12] (see exercise 5.7 of [@na98]) we then have $$2^{n+2} \geq k(B) =\sum_{(u,b_u) \in \mathcal{X}} l(b_u).$$ Now since $D$ is abelian, each block $b$ in the above summation may be chosen to have defect group $D$. First let $1 \neq u=x^i$ for some $i$. Then $b_u$ has inertial index $3$. Now $C_G(u)$ has a non-trivial central $2$-subgroup $Z_u \leq \langle x \rangle$. The unique block $\bar{b}_u$ of $C_G(u)/Z_u$ corresponding to $b_u$ has inertial index $3$ and by induction $3=l(\bar{b}_u) = l(b_u)$. Now let $u=x_jy_1$ for some $j$. Then $b_u$ has inertial index $1$ and so is nilpotent, hence $l(b_u)=1$. Substituting, we have $l(B) + 2^{n+2}-3 = k(B) \leq 2^{n+2}$, so $l(B) \leq 3$.
We have a subsection $(u,b_u)$ with defect group $D$ and $l(b_u)=1$. By [@sambale 1.37], the diagonal entries of the contribution matrix of $B$ (with rows labelled by ${\rm Irr}(B)$) are odd squares, and the trace of the contribution matrix is $|D|$. Hence $2^{n+2}$ is a sum of $k(B)$ odd squares, which cannot happen if $k(B)=2^{n+2}-1$ or $k(B)=2^{n+2}-2$. Hence $l(B)=3$.
Let $C$ be the unique block of $N_G(D)$ with $C^G=B$. By [@wa05] there is a perfect isometry between $B$ and $C'$. By Lemma \[normaldefectlemma\] $C$ and $B'$ are Morita equivalent, so there is a perfect isometry between $B'$ and $C$, and we are done.
In the above, the perfect isometry constructed in [@wa05] is additionally compatible with the $*$ construction in [@bp80] and so could be shown to satisfy the hypotheses of Proposition \[prop:index\_p\](iii). However this can also be shown using the machinery of perfect self-isometry groups developed earlier in this section, and this is what we do in the first part of the proof of the following Theorem.
\[index2theorem\] Let $G$, $N$, $B$, $b$ and $D$ be as in Proposition \[prop:grunit\]. Suppose further that $D\cong C_{2^n}\times C_2\times C_2$, for some $n>1$, $D\cap N\cong C_{2^{n-1}}\times C_2\times C_2$ and $b$ is Morita equivalent to the principal block of $\mathcal{O}(C_{2^{n-1}}\times A_4)$ (respectively $\mathcal{O}(C_{2^{n-1}}\times A_5)$). Then $B$ is Morita equivalent to the principal block of $\mathcal{O}(C_{2^n}\times A_4)$ (respectively $\mathcal{O}(C_{2^n}\times A_5)$).
First suppose that $b$ is Morita equivalent to $\mathcal{O}(C_{2^{n-1}}\times A_4)$.
By Proposition \[prop:index\_p\](ii) $l(B) = 3$ and so by Proposition \[per\_isom\] there exists a perfect isometry $$\begin{aligned}
I:\mathbb{Z}{\rm Irr}(B)\to\mathbb{Z}{\rm Irr}(\mathcal{O}(C_{2^n}\times A_4)).\end{aligned}$$ Now $I$ induces an isomorphism of the groups of perfect self-isometries of $B$ and of $\mathcal{O}(C_{2^n}\times A_4)$ via $\alpha \mapsto I \circ \alpha \circ I^{-1}$ for $\alpha$ any perfect self-isometry of $\mathcal{O}(C_{2^n}\times A_4)$, and we denote this isomorphism by $I_{\operatorname{PI}}$. Consider the perfect self-isometry $$\begin{aligned}
J:\mathbb{Z}{\rm Irr}(B)&\to\mathbb{Z}{\rm Irr}(B)\\
\chi&\mapsto\operatorname{sgn}_N^G.\chi,\end{aligned}$$ where $\operatorname{sgn}_N^G$ is the linear character of $G$ with kernel $N$, so for each irreducible character $\theta$ of $b$, $J$ swaps the two extensions of $\theta$ to $G$. We know that $J$ is indeed a perfect isometry as it is induced by the $\mathcal{O}$-algebra automorphism of $\mathcal{O}G$ given by $g\mapsto\operatorname{sgn}_N^G(g)g$ for all $g\in G$. Note that $J$ is a perfect self-isometry of order $2$ and that it induces the trivial $k$-algebra automorphism on $Z(kB)$. Furthermore by Proposition \[prop:index\_p\](i) and (ii) every character in ${\rm prj}(B)$ is fixed under multiplication by $\operatorname{sgn}_N^G$ and so $J$ is the identity on $\mathbb{Z}{\rm prj}(B)$. Therefore $I_{\operatorname{PI}}(J)$ must be of order $2$, induce the identity $k$-algebra automorphism on $Z(k(C_{2^n}\times A_4))$ and be the identity on $\mathbb{Z}{\rm prj}(\mathcal{O}(C_{2^n}\times A_4))$. Adopting the notation of Theorem \[thm:C2nA4\], set $I_{\operatorname{PI}}(J)=(I_{j,l,1},I_{\sigma,\epsilon})$, where $\sigma\in S_4$, $\epsilon\in\{\pm1\}$ and $0\leq j,l<2^n$ with $j$ odd. Then the fact that $I_{\operatorname{PI}}(J)$ is the identity on $\mathbb{Z}{\rm prj}(\mathcal{O}(C_{2^n}\times A_4))$ forces $\sigma$ to be the identity permutation and $\epsilon=1$, the fact that $I_{\operatorname{PI}}(J)$ induces the identity $k$-algebra automorphism on $Z(k(C_{2^n}\times A_4))$ forces $j=1$ and the fact that $I_{\operatorname{PI}}(J)$ has order $2$ forces $l=2^{n-1}$. In other words $I_{\operatorname{PI}}(J)$ is induced by the $\mathcal{O}$-algebra automorphism $$\begin{aligned}
\mathcal{O}(C_{2^n}\times A_4)&\to\mathcal{O}(C_{2^n}\times A_4)\\
x\otimes y&\mapsto -x\otimes y,\end{aligned}$$ for all $y\in\mathcal{O}A_4$, where $x$ is a fixed generator of $C_{2^n}$. We have shown that $$\begin{aligned}
I(\operatorname{sgn}_N^G.\chi)=\operatorname{sgn}_{N'}^{G'}.I(\chi),\end{aligned}$$ for all $\chi\in{\rm Irr}(B)$, where $G':=C_{2^n}\times A_4$, $N':=C_{2^{n-1}}\times A_4$. Therefore $I$ satisfies the hypotheses of Proposition \[prop:index\_p\](iii), where $B':=\mathcal{O}(C_{2^n}\times A_4)$ and $b':=\mathcal{O}(C_{2^{n-1}}\times A_4)$. Let $I_{N,N'}$ be the perfect isometry between $b$ and $b'$ induced by $I$ as in Proposition \[prop:index\_p\] and $I_{\operatorname{Mor}}$ the perfect isometry induced by the Morita equivalence between $b$ and $b'$. Write $I_{N,N'}\circ I_{\operatorname{Mor}}^{-1}=(I_{s,t,1},I_{\tau,\delta})$ in the notation of Theorem \[thm:C2nA4\] applied to $\mathcal{O}(C_{2^{n-1}}\times A_4)$, where $\tau\in S_4$, $\delta\in\{\pm1\}$ and $0\leq s,t<2^{n-1}$ with $s$ odd. By composing $I$ with the perfect self-isometry $(I_{1,1,1},I_{\tau,\delta})^{-1}$ of $B'$ and composing the Morita equivalence $b\sim_{\operatorname{Mor}}b'$ with that induced by the $\mathcal{O}$-algebra automorphism of $b'$ defined by $x\mapsto\zeta^{-t}x^{\frac{1}{s}}$, we may assume that $I_{N,N'}=I_{\operatorname{Mor}}$. Let $\phi_I:Z(B)\to Z(B')$ be the isomorphism of centres from Lemma \[lem:isomcent\] and let $M$ be the $b'$-$b$-bimodule inducing the Morita equivalence $b\sim_{\operatorname{Mor}}b'$. Since $I_{N,N'}=I_{\operatorname{Mor}}$ and by Lemma \[lem:isomcent\](ii) we have that $\phi_I|_{Z(b)}=\phi_{I_{N,N'}}:Z(b)\to Z(b')$ is the isomorphism of centres induced by the Morita equivalence. In other words $$\begin{aligned}
\label{centre:mor}
\phi_I(\alpha)m=m\alpha,\text{ for all }\alpha\in b,m\in M.\end{aligned}$$ Let $a\in B$ be a graded unit as described in Proposition \[prop:grunit\] and set $a':=\phi_I(a)$. Since $\phi_I$ respects the $G/N$ and $G'/N'$-gradings, $a'$ is also a graded unit. We now give $M$ the structure of a module for $$\begin{aligned}
(b'\otimes_{\mathcal{O}}b^{\operatorname{op}})\oplus(a'^{-1}b'\otimes_{\mathcal{O}}(ab)^{\operatorname{op}})\end{aligned}$$ by defining $a'^{-1}.m.a=m$, for all $m\in M$, where \[centre:mor\] ensures that this does indeed define a module. Now by [@mar96 Theorem 3.4] we have proved that $B$ is Morita equivalent to $\mathcal{O}(C_{2^n}\times A_4)$. For the $A_5$ case we note that the principal blocks of $\mathcal{O}A_4$ and $\mathcal{O}A_5$ are perfectly isometric by [@br90 A1.3]. The proof now proceeds exactly as above by replacing the principal block of $\mathcal{O}A_4$ everywhere with that of $\mathcal{O}A_5$ (note that we can replace the principal block of $\mathcal{O}A_4$ with that of $\mathcal{O}A_5$ in Theorem \[thm:C2nA4\]).
Proof of the main theorem and corollaries
=========================================
\[oddindex\] Let $B$ be a block of ${\mathcal{O}}G$ for a finite group $G$ with defect group $D \cong C_{2^n} \times C_2 \times C_2$ for some $n>1$. Let $N \lhd G$ be of odd prime index $w$ and let $b$ be a $G$-stable block of ${\mathcal{O}}N$ covered by $B$, so that $D$ is also a defect group for $b$. If $b$ is not nilpotent, then either $B$ is nilpotent or $B \sim_{\operatorname{Mor}} b$.
By Proposition \[per\_isom\] we have $l(b)=3$ and either $B$ is nilpotent (with $l(B)=1$) or $l(B)=3$. The normal subgroup $G[b]$ of $G$ is defined to be the group of elements of $G$ acting as inner automorphisms on $b \otimes_{{\mathcal{O}}} k$. Let $B'$ be a block of ${\mathcal{O}}G[b]$ covered by $B$. Then $b$ is source algebra equivalent to $B'$, and in particular has isomorphic inertial quotient by [@kkl12 2.2], noting that a source algebra equivalence over $k$ implies one over ${\mathcal{O}}$ by [@pu88 7.8]. Hence we may assume that $G[b]=N$. Then $B$ is the unique block of $G$ covering $B'$ by [@da73 3.5].
Now consider the action of $G$ on the ${\rm IBr}(b)$. If $w>3$, then every $\varphi \in {\rm IBr}(b)$ is fixed and extends to $w$ distinct elements of ${\rm IBr}(G)$. Since $B$ is the unique block of $G$ covering $b$ these all lie in $B$. Hence $l(B)=3w>3$, a contradiction. Hence $w=3$. Then either every element of ${\rm IBr}(b)$ is fixed, in which case $l(B)=3w>3$, a contradiction, or they are permuted in a single orbit, in which case $l(B)=1$ and $B$ is nilpotent.
In the following write ${\rm rk}_p(Q)$ for the rank of a $p$-group $Q$, that is, $p^{{\rm rk}_p(Q)}$ is the size of the largest elementary abelian subgroup of $Q$.
\[thm:main\] Let $G$ be a finite group and $B$ a block of $\mathcal{O}G$ with defect group $D\cong C_{2^n}\times C_2\times C_2$ for $n>1$. Then $B$ is Morita equivalent to the principal block of $\mathcal{O}(C_{2^n}\times C_2\times C_2)$, $\mathcal{O}(C_{2^n}\times A_4)$ or $\mathcal{O}(C_{2^n}\times A_5)$.
Let $B$ be a block of ${\mathcal{O}}G$ for a finite group $G$ with $[G:O_{2'}(Z(G))$ minimised subject to the condition that $B$ has defect group $D \cong C_{2^n}\times C_2\times C_2$ for some $n>1$ and $B$ is not Morita equivalent to the principal block of ${\mathcal{O}}D$, ${\mathcal{O}}(C_{2^n}\times C_2\times C_2)$, $\mathcal{O}(C_{2^n}\times A_4)$ or $\mathcal{O}(C_{2^n}\times A_5)$.
Suppose that $N \lhd G$ and $b$ is a block of ${\mathcal{O}}N$ covered by $B$. Write $I=I_G(b)$ for the stabilizer of $b$ in $G$, and $B_I$ for the Fong-Reynolds correspondent. Now $B_I$ is Morita equivalent to $B$ and they have isomorphic defect groups. We have $O_{2'}(Z(G)) \leq O_{2'}(Z(I))$, and if $I \neq G$, then $[I:O_{2'}(Z(I))]<[G:O_{2'}(Z(G))]$. Hence by minimality $I=G$.
Now suppose that $b$ is nilpotent. Let $b'$ be a block of ${\mathcal{O}}Z(G)N$ covered by $B$ and covering $b$. By the above argument applied to $Z(G)N$ and $b'$, $b'$ is $G$-stable. Note that $b'$ must also be nilpotent. Using the results of [@kp90], as outlined in [@ekks14 Proposition 2.2], $B$ is Morita equivalent to a block $\tilde{B}$ of a central extension $\tilde{L}$ of a finite group $L$ by a $2'$-group such that there is an $M \lhd L$ with $M \cong D \cap (Z(G)N)$, $G/Z(G)N \cong L/M$, and $\tilde{B}$ has defect group isomorphic to $D$. Note that $[\tilde{L}:O_{2'}(Z(\tilde{L}))] \leq |L| = [G:Z(G)N]|D \cap (Z(G)N)| \leq [G:O_{2'}(Z(G))]$ and that equality only occurs when $N \leq Z(G)O_2(G)$. Hence by minimality $N \leq Z(G)O_2(G)$.
We conclude that $B$ is quasiprimitive, that is, every block of every normal subgroup covered by $B$ is $G$-stable, and that if $B$ covers a nilpotent block of a normal subgroup $N$ of $G$, then $N \leq Z(G)O_2(G)$.
We claim that $O^2(G)=G$. Suppose otherwise, and let $N \lhd G$ be a subgroup of index $2$. Let $b$ be the unique block of $N$ covered by $B$. Then by Lemma \[index\_p\_background\_lem\] $B$ is the unique block of $G$ covering $b$ since $G/N$ is a $2$-group, $G=ND$ and $b$ has defect group $D \cap N$. Let $b_D$ be a block of $C_G(D)$ with $(b_D)^G=B$. Since $B$ has inertial quotient $C_3$ and $N_G(D,b_D)$ controls fusion in $D$, the inertial quotient of $b$ is $C_3$ (if it were $1$, then $b$ would be nilpotent and so $G=Z(G)O_2(G)$, a contradiction by Lemma \[normaldefectlemma\]). If $D \cap N \cong C_{2^n} \times C_2$, then ${\rm Aut}(D \cap N)$ is a $2$-group and so $b$ is nilpotent, a contradiction. If $D \cap N \cong (C_2)^3$, then by [@ea16] $b$ is Morita equivalent to the principal block of ${\mathcal{O}}(C_2 \times A_4)$ or ${\mathcal{O}}(C_2 \times A_5)$ and so Theorem \[index2theorem\] gives a contradiction. Otherwise, since $[N:O_{2'}(Z(N))]<[G:O_{2'}(Z(G))]$ by minimality we also have a contradiction by Theorem \[index2theorem\]. Hence $O^2(G)=G$.
Before proceeding we recall the definition and some properties of the generalized Fitting subgroup $F^*(G)$ of a finite group $G$. Details may be found in [@asc00]. A *component* of $G$ is a subnormal quasisimple subgroup of $G$. The components of $G$ commute, and we define the *layer* $E(G)$ of $G$ to be the normal subgroup of $G$ generated by the components. It is a central product of the components. The *Fitting subgroup* $F(G)$ is the largest nilpotent normal subgroup of $G$, and this is the direct product of $O_r(G)$ for all primes $r$ dividing $|G|$. The *generalized Fitting subgroup* $F^*(G)$ is $E(G)F(G)$. A crucial property of $F^*(G)$ is that $C_G(F^*(G)) \leq F^*(G)$, so in particular $G/F^*(G)$ may be viewed as a subgroup of ${\rm Out}(F^*(G))$.
Write $L_1,\ldots,L_t$ for the components of $G$, so $E(G)=L_1\cdots L_t \lhd G$. Note that $G$ permutes the $L_i$. There must be at least one component, since otherwise the block $b'$ of $F^*(G)$ covered by $B$ is nilpotent and so $F^*(G)=Z(G)O_2(G)$. Therefore $D \leq C_G(F^*(G)) \leq F^*(G)=Z(G)O_2(G)$, so that $D \lhd G$, a contradiction by Lemma \[normaldefectlemma\].
We claim that $O_2(G) \leq Z(G)$. Write $N=C_G(O_2(G))$ and $b$ for the unique block of $N$ covered by $B$. Note that $D \leq N \lhd G$. If $O_2(G) \cong C_{2^m}$ for $m \geq 1$ or $O_2(G) \cong C_{2^m} \times C_2$ for $m>1$, then ${\rm Aut}(O_2(G))$ and so $G/N$ is a $2$-group, which forces $N=G$ as $O^2(G)=G$. Suppose that $O_2(G) \cong C_{2^m} \times (C_2)^2$ for some $m \geq 1$. Let $b_D$ be a block of $C_G(D)$ with $(b_D)^{C_G(D)}=b$ (and so $(b_D)^G=B$). Since $N_G(D,b_D)$ controls fusion in $D$ there must be a $G$-stable subgroup of $O_2(G)$ of order $4$. Hence $G/N$ is isomorphic to a subgroup of $S_3$. Since $O^2(G)=G$, we have $[G:N]|3$. Then by Proposition \[oddindex\] and minimality (noting that $Z(G)\leq N$) we again have $G=N$, so $O_2(G) \leq Z(G)$ as claimed.
We have shown that $F^*(G)=E(G)Z(G)$. We next show that $t=1$, that is, $E(G)$ is quasisimple. Write $b^*$ for the unique block of $F^*(G)$ covered by $B$. Then $D \cap F^*(G)$ is a defect group for $b^*$. Hence $(D \cap F^*(G))/O_p(Z(G))$ is a defect group for a block of $F^*(G)/O_p(Z(G))$. Therefore $(D \cap F^*(G))/O_p(Z(G))$ is a radical $2$-subgroup of $F^*(G)/O_p(Z(G))$ (recall that a $p$-subgroup $Q$ of a finite group $H$ is radical if $Q=O_p(N_H(Q))$ and that defect groups are radical $p$-subgroups) and so $(D \cap F^*(G))Z(G)/Z(G)$ is a radical $2$-subgroup of $F^*(G)/Z(G)\cong (L_1Z(G)/Z(G)) \times \cdots \times (L_tZ(G)/Z(G))$. By [@ou95 Lemma 2.2] it follows that $(D \cap F^*(G))Z(G)/Z(G) = D_1 \times \cdots \times D_m$, where $D_i = (D \cap F^*(G))Z(G)/Z(G)) \cap (L_iZ(G)/Z(G))$ (and $D_i$ is a radical $2$-subgroup but not necessarily a defect group). Write $b_i$ for the block of $L_i$ covered by $B$ and $\bar{b}_i$ for the unique block of $L_iO_2(G)/O_2(G)$ corresponding to $b_i$. If ${\rm rk}_p(D_i)=1$ for some $i$, then $\bar{b}_i$ has cyclic defect group and so is nilpotent, hence $b_i$ is also nilpotent by [@wa94] (where the result is stated over $k$, but follows over ${\mathcal{O}}$ immediately), a contradiction. Hence since $\sum_{i=1}^t {\rm rk}_p(D_i) \leq {\rm rk}_p(D) = 3$ we have $t=1$.
Now by the Schreier conjecture $G/F^*(G)$ is solvable. Suppose that $G \neq F^*(G)$. Since $O^2(G)=G$ there is $N \lhd G$ of odd prime index. Let $b$ be the unique block of $N$ covered by $G$. Note that we may assume $Z(G) \leq N$, as otherwise $B$ and $b$ are Morita equivalent by [@kkl12 2.2] and we may replace $G$ and $B$ by $N$ and $b$. Therefore we have $[N:O_{2'}(Z(N))] < [G:O_{2'}(Z(G))]$ and so Proposition \[oddindex\] leads to a contradiction. Hence we may assume $G=F^*(G)$ and so $G=L_1Z(G)$. Further application of Proposition \[oddindex\] and Theorem \[index2theorem\] allows us to assume that $G=L_1$. Applying [@ekks14 6.1], one of the following occurs, both leading to a contradiction, and we are done:
\(i) $B$ is Morita equivalent to a block $C$ of ${\mathcal{O}}H$ for a finite group $H$ with $H=H_0 \times H_1$ such that $H_0$ is abelian with Sylow $2$-subgroup $C_{2^n}$ and the block of $H_1$ covered by $C$ has defect groups $C_2 \times C_2$. In this case it follows from [@li94] that $B$ is Morita equivalent to the principal block of ${\mathcal{O}}D$, ${\mathcal{O}}(C_{2^n} \times A_4)$ or ${\mathcal{O}}(C_{2^n} \times A_5)$, a contradiction.
\(ii) There is a finite group $H$ with $G \lhd H$ and $B$ is covered by a nilpotent block of ${\mathcal{O}}H$. In this case by [@pu11 4.3] $B$ is Morita equivalent to the unique block of $N_G(D)$ with Brauer correspondent $B$, a contradiction by Lemma \[normaldefectlemma\].
\[cor:derived\] Let $G$ be a finite group and $B$ a block of $\mathcal{O}G$ with defect group $D\cong C_{2^n}\times C_2\times C_2$ for $n>1$. Let $b$ be the unique block of $N_G(D)$ with $b^G=B$. Then $B$ and $b$ are derived equivalent.
By [@ri96 3] the principal blocks of ${\mathcal{O}}A_4$ and ${\mathcal{O}}A_5$ are derived equivalent, and so the same is true of ${\mathcal{O}}(C_{2^n} \times A_4)$ and ${\mathcal{O}}(C_{2^n} \times A_5)$. Hence by Theorem \[thm:main\] there are only two derived equivalence classes of blocks with defect group $C_{2^n}\times C_2\times C_2$, and we are done.
\[cor:derived16\] Let $G$ be a finite group and $B$ a block of $\mathcal{O}G$ with defect group $D$ of order dividing $16$. Let $b$ be the unique block of $N_G(D)$ with $b^G=B$. Then $B$ and $b$ are derived equivalent.
If $D$ is elementary abelian, then this is by [@li94], [@ea16] and [@ea17]. If $D \cong C_4 \times C_4$, then see [@ekks14] where it is shown that there are only two Morita equivalence classes. If $D \cong C_4 \times C_2 \times C_2$, then this is Corollary \[cor:derived\]. In all other cases ${\rm Aut}(D)$ is a $2$-group and so all blocks with that defect group are nilpotent.
\[cor:derivedrank3\] Let $G$ be a finite group and $B$ a block of $\mathcal{O}G$ with defect group $D$ of $2$-rank at most three. Let $b$ be the unique block of $N_G(D)$ with $b^G=B$. Then $B$ and $b$ are derived equivalent.
This follows immediately from the above corollaries, [@ekks14 1.1] and [@wzz17].
[99]{}
J. L. Alperin, *Local representation theory*, Cambridge University Press (1986).
M. Aschbacher, *Finite group theory*, Cambridge Studies in Advanced Mathematics [**10**]{}, Cambridge university Press (1986).
M. Aschbacher, R. Kessar and R. Oliver, *Fusion systems in algebra and topolgy*, London Mathematical Society Lecture Note Series **391**, Cambridge University Press (2011).
M. Broué, *Isométries parfaites, types de blocs, catégories dérivées*, Astérisque [**181-182**]{} (1990), 61–92.
M. Broué and L. Puig, *Characters and local structure in $G$-algebras*, J. Algebra [**63**]{} (1980), 306–317.
E. C. Dade, *Block extensions*, Ill. J. Math. [**17**]{} (1973), 198-272.
C. W. Eaton, *Morita equivalence classes of $2$-blocks of defect three*, Proc. AMS [**144**]{} (2016), 1961–1970.
C. W. Eaton, *Morita equivalence classes of blocks with elementary abelian defect groups of order $16$*, preprint (2017).
C. W. Eaton, R. Kessar, B. Külshammer and B. Sambale, [*$2$-blocks with abelian defect groups*]{}, Adv. Math. [**254**]{} (2014), 706-735.
C. Eaton and M. Livesey, *Loewy lengths of blocks with abelian defect groups*, Proceedings of the American Mathematical Society, Series B
W. Feit, *The representation theory of finite groups*, North-Holland Mathematical Library Volume 25 (1982).
R. Kessar, *A remark on Donovan’s conjecture*, Arch. Math (Basel) [**82**]{} (2005), 391–394.
R. Kessar, S. Koshitani and M. Linckelmann, *Conjectures of Alperin and Broué for $2$-blocks with elementary abelian defect groups of order $8$*, J. Reine Angew. Math. **671** (2012), 85–130.
R. Kessar and G. Malle, *On quasi-isolated blocks and Brauer’s height zero conjecture*, Ann. Math. (2) [**178**]{} (2013), 321–384.
S. Koshitani and B. Külshammer, *A splitting theorem for blocks*, Osaka J. Math. **33** (1996), 343–346.
B. Külshammer, *Crossed products and blocks with normal defect groups*, Comm. Alg. [**13**]{} (1985), 147–168.
B. Külshammer and L. Puig, *Extensions of nilpotent blocks*, Invent. Math. **102** (1990), 17–71.
B. Külshammer and B. Sambale, *The $2$-blocks of defect $4$*, Rerpresentation Theory **17** (2013), 226–236.
M. Linckelmann, *The source algebras of blocks with a Klein four defect group*, J. Algebra **167** (1994), 821–854.
A. Marcus, *On equivalences between blocks of group algebras: Reduction to the simple components*, J. Algebra (2) [**184**]{} (1996), 372–396.
G. Navarro, *Characters and blocks of finite groups*, London Mathematical Society Lecture Note Series **250**, Cambridge University Press (1998).
J. B. Olsson and K. Uno, *Dade’s conjecture for symmetric groups*, J. Algebra [**176**]{} (1995), 534–560.
L. Puig, *Nilpotent blocks and their source algebras*, Invent. Math. **93** (1988), 77–116.
L. Puig, *Nilpotent extensions of blocks*, Math. Z. **269** (2011), 115-136.
J. Rickard, *Splendid equivalences: derived categories and permutation modules*, Proc. London Math. Soc. [**72**]{} (1996), 331–358.
G. R. Robinson, *On Brauer’s $k(B)$ problem*, J. Algebra [**147**]{} (1992), 450–455.
B. Sambale, *Blocks of finite groups and their invariants*, Lecture Notes in Mathematics [**2127**]{}, Springer (2014).
B. Sambale, *Cartan matrices and Brauer’s $k(B)$-conjecture IV*, J. Math. Soc. Japan [**69**]{} (2017), 735–754.
A. Watanabe, *On nilpotent blocks of finite groups*, J. Algebra **163** (1994), 128–134.
A. Watanabe, *On perfect isometries for blocks with abelian defect groups and cyclic hyperfocal subgroups*, Kumamoto J. Math. [**18**]{} (2005), 85-92.
C. Wu, K. Zhang and Y. Zhou, *Blocks with defect group $\mathbb{Z}_{2^n} \times \mathbb{Z}_{2^n} \times
\mathbb{Z}_{2^m}$*, preprint (2017).
[^1]: School of Mathematics, University of Manchester, Manchester, M13 9PL, United Kingdom. Email: [email protected]
[^2]: School of Mathematics, University of Manchester, Manchester, M13 9PL, United Kingdom. Email: [email protected]
[^3]: This research was supported by the EPSRC (grant no. EP/M015548/1).
| {
"pile_set_name": "ArXiv"
} |
**Matrix Identities on**
Weighted Partial Motzkin Paths
William Y.C. Chen$^1$, Nelson Y. Li$^2$, Louis W. Shapiro$^3$ and Sherry H. F. Yan$^4$\
\[2mm\] $^{1,2,4}$Center for Combinatorics, LPMC, Nankai University, Tianjin 300071, P.R. China\
\[2mm\] $^3$Department of Mathematics, Howard University, Washington, DC 20059, USA\
\[2mm\] $^[email protected], $^[email protected], $^[email protected], $^[email protected]\
[**Abstract.**]{} We give a combinatorial interpretation of a matrix identity on Catalan numbers and the sequence $(1, 4,
4^2, 4^3, \ldots)$ which has been derived by Shapiro, Woan and Getu by using Riordan arrays. By giving a bijection between weighted partial Motzkin paths with an elevation line and weighted free Motzkin paths, we find a matrix identity on the number of weighted Motzkin paths and the sequence $(1, k, k^2, k^3, \ldots)$ for any $k \geq 2$. By extending this argument to partial Motzkin paths with multiple elevation lines, we give a combinatorial proof of an identity recently obtained by Cameron and Nkwanta. A matrix identity on colored Dyck paths is also given, leading to a matrix identity for the sequence $(1, t^2+t, (t^2+t)^2, \ldots)$.
[Key words]{}: Catalan number, Schröder number, Dyck path, Motzkin path, partial Motzkin path, free Motzkin path, weighted Motzkin path, Riordan array
[AMS Mathematical Subject Classifications]{}: 05A15, 05A19.
[Corresponding Author:]{} William Y. C. Chen, [email protected]
Introduction
============
This paper is motivated the following matrix identity obtained by Shapiro, Woan and Getu [@shapirotc] in their study of the moments of a Catalan triangle [@chapman; @shapiro; @sulanke]: $$\label{eq2.2}
\begin{bmatrix} 1 \\2 & 1\\5 & 4 & 1\\14 & 14 & 6 & 1\\42 & 48 & 27 & 8 & 1\\& & \cdots &&&\ddots \end{bmatrix}
\begin{bmatrix} 1 \\ 2 \\ 3 \\ 4 \\ 5 \\ \vdots
\end{bmatrix} = \begin{bmatrix} 1 \\ 4 \\ 4^2 \\ 4^3 \\ 4^4 \\ \vdots
\end{bmatrix},$$ where the first column of the first matrix is the Catalan number $C_n={1\over n+1} {2n \choose n}$ and $a_{i,j}$ (the entry in the $i$-th row and $j$-th column) is determined by the following recurrence relation for $j\geq 2 $: $$\label{r2.2}
a_{i,j}=a_{i-1, j-1}+2a_{i-1,j}+a_{i-1,j+1} .$$ Another proof of the above identity is given by Woan, Shapiro and Rogers [@woan] while computing the areas of parallelo-polyominos via generating functions.
The first result of this paper is a combinatorial interpretation of the identity (\[eq2.2\]) in terms of Dyck paths.
One main objective of this paper is to give a matrix identity that extends the sequence $(1, 4, 4^2, 4^3, \ldots)$ to $(1, k, k^2,
k^3, \ldots)$ in (\[eq2.2\]). The following matrix identity was proved by Cameron and Nkwanta [@cn] that arose in a study of elements of order $2$ in [*Riordan groups*]{} [@aigner; @shapiroba; @shapirotr; @sprugnoli]: $$\label{eq1.1}
\begin{bmatrix} 1 \\3 & 1\\11 & 6 & 1\\45 & 31 & 9 & 1\\197 & 156 & 60 & 12 & 1 & &\\& & \cdots &&&\ddots \end{bmatrix}
\begin{bmatrix} 1 \\ 3 \\ 7 \\ 15 \\ 31 \\ \vdots
\end{bmatrix} = \begin{bmatrix} 1 \\ 6 \\ 6^2 \\ 6^3 \\ 6^4 \\
\vdots
\end{bmatrix},$$ where the entry $a_{i,j}$ ($i$th row and $j$th column) in the above matrix satisfies the recurrence relation $$\label{r1.1}
a_{i,j}=a_{i-1,j-1}+3a_{i-1,j}+2a_{i-1,j+1}$$ for $j\geq 2$ and the $a_{i,1}$ is the $i$-th [*little Schröder number*]{} $s_i$ (sequence A001003 in [@sloane]), which counts Schröder paths of length $2(i+1)$. A [*Schröder path*]{} is a lattice path starting at (0, 0) and ending at $(2n, 0)$ and using steps $H=(2, 0)$, $U=(1, 1)$ and $D=(1,
-1)$ such that no steps are below the $x$-axis and there are no peaks at level one. Imposing this last peak condition gives us little Schröder numbers while without it we would have the [*large Schröder numbers*]{}.
For $k=3$, we obtain the following matrix identity on Motzkin numbers:
$$\label{3p}
\begin{bmatrix} 1 \\1 & 1\\2 & 2 & 1\\4 & 5 & 3 & 1\\9 & 12 & 9 & 4 & 1\\& & \cdots &&&\ddots \end{bmatrix}
\begin{bmatrix} 1 \\ 2 \\ 3 \\ 4 \\ 5 \\ \vdots
\end{bmatrix} = \begin{bmatrix} 1 \\ 3 \\ 3^2 \\ 3^3 \\ 3^4 \\ \vdots
\end{bmatrix},$$
where the first column is the sequence of Motzkin numbers, and matrix $A=(a_{ij})$ is generated by the following recurrence relation: $$a_{i,j}= a_{i-1, j-1} + a_{i-1, j} + a_{i-1, j+1}.$$ For $k=5$, we find the following matrix identity $$\label{5p}
\begin{bmatrix} 1 \\3 & 1\\10 & 6 & 1\\36 & 29 & 9 & 1\\137 & 132 & 57 & 12 & 1\\& & \cdots &&&\ddots \end{bmatrix}
\begin{bmatrix} 1 \\ 2 \\ 3 \\ 4 \\ 5 \\ \vdots
\end{bmatrix} = \begin{bmatrix} 1 \\ 5 \\ 5^2 \\ 5^3 \\ 5^4 \\ \vdots
\end{bmatrix},$$ where the first column sequence A002212 in [@sloane], which has two interpretations, the number of $3$-Motzkin paths or the number of ways to assemble benzene rings into a tree [@hr]. Recall that a $3$-Motzkin path is a lattice path from $(0,0)$ to $(n-1,0)$ that does not go below the $x$-axis and consists of up steps $U=(1,1)$, down steps $D=(1,-1)$, and three types of horizontal steps $H=(1,0)$. The above matrix $A=(a_{i,j})$ is generated by the first column and the following recurrence relation $$a_{i,j}= a_{i-1,j-1} + 3a_{i-1,j} +
a_{i-1,j+1}.$$ We may prove the above identities (\[3p\]) and (\[5p\]) by using method of Riordan arrays. So the natural question is to find a matrix identity for the sequence $(1, k, k^2, k^3, \ldots)$. We need the combinatorial interpretation of the entries in the matrix in terms of weighted partial Motzkin paths, as given by Cameron and Nkwanta [@cn]. To be precise, a partial Motzkin path, also called a Motzkin path from $(0,0)$ to $(n,k)$ in [@cn], is just a Motzkin path but without the requirement of ending on the x-axis. A weighted partial Motkzin a partial Motzkin path with the weight assignment that the horizontal steps are endowed with a weight $k$ and the down steps are endowed with a weight $t$, where $k$ and $t$ are regarded as positive integers. In this sense, our weighted Motzkin paths can be regarded as further generalization of $k$-Motzkin paths in the sense of $2$-Motzkin paths and $3$-Motkzin paths [@BdLPP; @deutschs; @sloane].
We introduce the notion of weighted free Motzkin paths which is a lattice path consisting of Motzkin steps without the restrictions that it has to end with a point on the $x$-axis and it does not go below the $x$-axis. We then give a bijection between weighted free Motzkin paths and weighted partial Motzkin paths with an elevation line, which leads to a matrix identity involving the number of weighted partial Motzkin paths and the sequence $(1, k,
k^2, \ldots)$. The idea of the elevation operation is also used by Cameron and Nkwanta in their combinatorial proof of the identity (\[eq2.2\]) in a more restricted form. By extending our argument to weighted partial Motzkin paths with multiple elevation lines, we obtain a combinatorial proof of an identity recently derived by Cameron and Nkwanta, in answer to their question.
We also give a generalization of the matrix identity (\[eq2.2\]) and give a combinatorial proof by using colored Dyck paths.
Riordan Arrays
==============
In this section, we give a brief introduction to the notion of Riordan arrays [@shapiroba; @shapirotr; @sprugnoli]. Let us use (\[eq2.2\]) and (\[eq1.1\]) as examples. Start with two generating functions $g(x)=1+g_1x+g_2x^2+\cdots$ and $f(x)=f_1x+f_2x^2+\cdots$ with $f_1\neq 0$. Let $H=(h_{i,j})_{n,j\geq 0}$ be the infinite lower triangular matrix with nonzero entries on the main diagonal, where $h_{i,j}=[x^i](g(x)(f(x)^j)$ for $i\geq j$, namely, $h_{i,j}$ equals the coefficient of $x^i$ in the expansion of the series $f(x)(g(x)^j)$. If an infinite lower triangular matrix $H$ can be constructed in this way from two generating functions $g(x)$ and $f(x)$, then it is called a [*Riordan array*]{} and is denoted by $H=(g(x),f(x))=(g,f)$.
Suppose we multiply the matrix $H=(g,f)$ by a column vector $(a_0,a_1,\cdots)^T$ and get a column vector $(b_0,b_1,\cdots)^T$. Let $A(x)$ and $B(x)$ be the generating functions for the sequences $(a_0,a_1,\cdots)$ and $(b_0,b_1,\cdots)$ respectively. Then the method of Riordan arrays asserts that $$B(x)=g(x)A(f(x)).$$
For the matrix identity (\[eq2.2\]), let $g(x)$ be the generating function for Catalan numbers $(1, 2, 5, 14, \ldots)$: $$g(x) = {1 -2x-
\sqrt{1-4x} \over 2x^2}.$$ Let $f(x)=xg(x)$. From the recurrence relation (\[r2.2\]) one may derive that the generating function for the sequence in the $j$-th $(j\geq 1)$ column in the matrix in (\[eq2.2\]) equals $g(xg)^{j-1}$. Let $H$ be the Riordan array $(g, xg)$. Since the generating function of $(1,2,3,4\cdots)^T$ equals $A(x)=\frac{1}{(1-x)^2}$, it follows that $B(x)=g(x)A(xg(x))=\frac{1}{1-4x}$ equals the generating function for the right hand side of (\[eq2.2\]). Thus we obtain the identity (\[eq2.2\]).
Let us consider the matrix identity (\[eq1.1\]). Let $g(x)$ be the generating function for the little Schröder numbers as given by $$\label{sg}
g(x)={1-3x-\sqrt{1-6x+x^2}\over 4x^2},$$ and let $f(x) =xg(x)$. Note that the generating function for the sequence $(1, 3, 7, 15, \ldots)$ equals $A(x)=\frac{1}{(1-x)(1-2x)}$. From the recurrence relation (\[r1.1\]) one may verify that the matrix in (\[eq1.1\]) is indeed the Riordan array $(g, xg)$. Therefore, the generating function for the right hand side of (\[eq1.1\]) equals $
g(x)A(xg(x))=\frac{1}{1-6x}$, which implies (\[eq1.1\]).
Using the same method, we can verify the matrix identity (\[3p\]) and (\[5p\]). Since we are going to establish a general bijection for weighted Motzkin numbers, here we omit the proofs.
Dyck path interpretation of (\[eq2.2\])
=======================================
In this section, we present a combinatorial interpretation of the matrix identity (\[eq2.2\]) by using Dyck paths. A [*Dyck path*]{} of length $2n$ is a path going from the origin $(0,0)$ to $(2n,0)$ using steps $U=(1,1)$ and down steps $D=(1,-1)$ such that no steps is below the $x$-axis [@De; @stanley]. The number of Dyck paths of length $2n$ equals the Catalan number $C_n$.
For a Dyck path $P$, the points on the $x$-axis except for the initial point are called return points. In this sense, the ending point is always a return point. Formally speaking, a [*composition*]{} of a Dyck path $P$ is sequence of Dyck path $(P_1,
P_2, \ldots, P_j)$ such that $P=P_1 P_2 \cdots P_j$, where $P_1,
P_2, \ldots, P_j$ are Dyck paths. For a composition $(P_1, P_2,
\ldots, P_j)$ of a Dyck path $P$, its length is meant to be the length of $P$ and $j$ is called the number of segments. We may choose certain return points to cut off the Dyck paths into a composition. We use the convention that the ending point is always a cut point. Clearly, a Dyck path with one segment is an ordinary Dyck path.
\[dotheom\] For $j\geq 2$, we have the following recurrence relation $$\label{eq.1}
d_{i,j}=d_{i-1,j-1} +
2d_{i-1,j} + d_{i-1,j+1}.$$
Let $(P_1, P_2, \ldots, P_{j})$ be composition of a Dyck path $P$ of length $2i$. Consider the following cases for $P_1$. Case 1: $P_1=UD$. Then we get a composition length $2(i-1)$ with $j-1$ segments: $(P_2, \ldots, P_j)$. Case 2: $P_1=QUD$ and $Q$ is not empty. Then we get a composition $(Q, P_2, \ldots, P_j)$ of length $2(i-1)$ and $j$ segments. Case 3: $P_1=U Q D$, $Q$ is not empty. We get a composition $(Q, P_2, \ldots, P_j)$ of length $2(i-1)$ with $j$ segments. Case 4: $P_1=Q_1UQ_2D$, where $Q_1$ and $Q_2$ are not empty. Then we get a composition $(Q_1, Q_2,
P_2, \ldots, P_j)$ of length $2(i-1)$ with $j+1$ segments. Adding up the terms in the above cases, we obtain the desired recursion (\[eq.1\]).
From Lemma \[dotheom\] one sees that the entry $a_{i,j}$ in the triangular matrix of the identity (\[eq2.2\]) can be explained as the number compositions of Dyck paths of length $2i$ that contain $j$ segments. We remark that this combinatorial interpretation can also be derived from the generating function of the entries in the $j$-th column in of the matrix in (\[eq2.2\]). The following formula for $a_{i,j}$ has been derived by Cameron and Nkwanta [@cn]: $$a_{i,j}= {j \over i} \, {2i \choose i-j} .$$
Let us rewrite the matrix identity (\[eq2.2\]) as follows $$\label{eq.2}
\sum_{j= 1}^i ja_{i,j}=4^{i-1}.$$ A combinatorial formulation of the above identity is given by Callan [@callan].
We are now ready to give a combinatorial proof of the above identity. Clearly, $4^n$ is the number of sequences of length $n$ on four letters, say, $\{1, 2, 3, 4\}$. The term $ja_{i,j}$ suggests that we should specify a segment in a composition as a distinguished segment. We may use a star $*$ to mark the distinguished segment. We call a composition with a distinguished segment a [*rooted*]{} composition of a Dyck path. Then $ja_{i,j}$ equals the number of rooted compositions of Dyck paths of length $2i$ that contain $j$ segments.
There is a bijection $\phi$ between the set of rooted compositions of Dyck paths of length $2i$ and the set of sequences of length $i-1$ on four letters.
Let $(P_1, P_2, \ldots, P_j)$ be a rooted composition of a Dyck path $P$ of length $2i$. We proceed to construct a sequence of length $i-1$ on the elements $\{ 1, 2,3, 4\}$. We now recursively define a map $\phi$ from rooted compositions of a Dyck path $P$ of length $2i$ to sequences of length $i-1$ on $\{
1, 2, 3, 4\}$. For $i=1$, $P$ is unique, and the sequence is set to be the empty sequence. We now assume that $i>1$. Let $(P_1,
\ldots, P_t^*, \ldots, P_j)$ be a rooted composition of $P$ with $P_t^*$ being the distinguished segment.
We have the following cases.
1. $P_1=UD$ and $t=1$. Then we set $\phi(P)=1\, \phi(P_2^*, P_3, \ldots, P_j)$.
2. $P_1=UD$ and $t\not=1$. Then we set $\phi(P)=2\, \phi(P_2, \ldots, P_t^*\ldots, P_j)$.
3. $P_1=QUD$ and $Q$ is a nonempty Dyck path. Set $\phi(P)=3\,
\phi(Q^*, P_2, \ldots, P_j)$ if $t=1$ and set $\phi(P_1, \ldots, P_j) =3 \, \phi (Q, P_2, \ldots, P_t^*, \ldots, P_j)$ if $t>1$.
4. $P_1=Q_1UQ_2D$, where $Q_1$ and $Q_2$ are nonempty Dyck paths. Then we set $$\phi(P_1, \ldots, P_j)=1\, \phi
(Q_1, Q_2, P_2, \ldots, P_t^*, \ldots, P_j)
\text{ if }
t>1 ,$$ $$\phi(P)=1\, \phi
(Q_1, Q_2^*, P_2, \ldots, P_j) \text{ if } t=1.$$
5. If $P_1=UQD$ and $Q$ is a nonempty Dyck path. Then we set $$\phi(P)=4\, \phi(Q, P_2, \ldots, P_t^*, \ldots,
P_j).$$
In order to show that $\phi$ is a bijection, we construct the reverse map of $\phi$. Let $w=w_1w_2\cdots w_{i-1}$ be a sequence of length $i-1$ on $\{1, 2, 3, 4\}$. If $i=1$, then it corresponds to $UD$. We now assume that $i>1$. Suppose that $w_2w_3\cdots w_{i-1}$ corresponds to a rooted composition $(R_1, R_2, \cdots, R_m)$ of a Dyck path $P$ of length $2(i-1)$ with $R_k$ being the distinguished segment. We proceed to find a rooted composition $(P_1, P_2, \ldots, P_j)$ with $P_t$ being the distinguished segment such that $\phi(P_1,
P_2, \ldots, P_j) = w_1 \phi(R_1, R_2, \ldots, R_m)$.
If $w_1=2$, we have $P_1=UD$ and $(P_2, P_3, \ldots, P_j) =
(R_1, R_2, \ldots, R_m)$. It follows that $t=k+1$ and $j=m+1$. Thus we can recover $(P_1, P_2, \ldots, P_j)$. For the case $w_1=3$, we have $P_1=R_1UD$ and $t=k$. Also, we can recover $(P_2, \cdots, P_j)$ from $(R_2, \ldots, R_m)$. For the case $w_1=4$, we have $t=k$, $P_1=UR_1D$, and $(P_2,
\ldots, P_j)=(R_2, \ldots, R_m)$.
It remains to deal with the situation $w_1=1$, which involves Cases 1 and 4 of the bijection. If $k=1$, then we have $t=1$, $P_1=UD$ and $(P_2, \ldots, P_j)=(R_1,\ldots, R_m)$. If $k=2$, then we have $t=1$, $P_1=R_1UR_2D$ and $(P_2, \ldots, P_j)= (R_3,
\ldots, R_m)$. If $k>2$, then we have $t=k-1$, $P_1=R_1UR_2D$, and $(P_2, \ldots, P_j)=(R_3, \ldots, R_m)$. Thus, we have shown that $\phi$ is a bijection.
An example of the bijection $\phi$ is given in Figure \[dyck\], where normal vertices are drawn with black points, return points that cut off the Dyck paths into segments are drawn with white points, and the distinguished segment is marked with a $*$ on its last return step.
(120,25)
(0,15)[(1,1)[4]{}]{} (4,19)[(1,-1)[3.7]{}]{} (8.3,15.3)[(1,1)[8]{}]{} (16,23)[(1,-1)[8]{}]{} (24,15)[(1,1)[4]{}]{} (28,19)[(1,-1)[3.7]{}]{}(29.5,16.9)[$\ast$]{}(32.3,15.3)[(1,1)[4]{}]{} (36,19)[(1,-1)[3.7]{}]{} (0,15) (4,19) (8,15) (12,19) (16,23) (20,19) (24,15)(28,19) (32,15) (36,19) (40,15)
(43,19)[$\longleftrightarrow$]{} (54,18)[$2$]{}
(58,15)[(1,1)[8]{}]{} (66,23)[(1,-1)[8]{}]{} (74,15)[(1,1)[4]{}]{} (78,19)[(1,-1)[3.7]{}]{}(79.5,16.9)[$\ast$]{}(82.3,15.3)[(1,1)[4]{}]{} (86,19)[(1,-1)[3.7]{}]{} (58,15) (62,19) (66,23) (70,19) (74,15)(78,19) (82,15) (86,19) (90,15)
(93,19)[$\longleftrightarrow$]{}
(0,3)[$23$]{} (5.3,0.3)[(1,1)[8]{}]{} (13,8)[(1,-1)[7.7]{}]{} (18.5,1.9)[$\ast$]{}(21.3,0.3)[(1,1)[4]{}]{} (25,4)[(1,-1)[3.7]{}]{} (5,0) (9,4) (13,8) (17,4) (21,0) (25,4) (29,0)
(32,4)[$\longleftrightarrow$]{}
(43,3)[$234$]{} (48,0)[(1,1)[4]{}]{} (52,4)[(1,-1)[3.7]{}]{} (53.5,1.9)[$\ast$]{}(56.3,0.3)[(1,1)[4]{}]{} (60,4)[(1,-1)[3.7]{}]{} (48,0) (52,4) (56,0) (60,4) (64,0)
(67,4)[$\longleftrightarrow$]{}
(78,3)[$2341$]{} (85,0)[(1,1)[4]{}]{} (89,4)[(1,-1)[3.7]{}]{} (90.5,1.9)[$\ast$]{} (85,0) (89,4) (93,0)
(96,4)[$\longleftrightarrow$]{} (107,3)[$2341$]{}
Weighted Partial Motzkin Paths
===============================
A [*Motzkin path*]{} of length $n$ is a path going from $(0,0)$ to $(n,0)$ consisting of up steps $U=(1,1)$, down steps $D=(1,-1)$ and horizontal steps $H=(1,0)$, which never goes below the $x$-axis. A [*$(k,t)$-Motzkin path*]{} is a Motzkin path such that each horizontal step is weighted by $k$, each down step is weighted by $t$ and each up step is weighted by $1$. The case $k=2, t=1$ gives the $2$-Motzkin paths which have been introduced by Barcucci, del Lungo, Pergola and Pinzani [@BdLPP] and have been studied by Deutsch and Shapiro [@deutschs]. The [*weight* ]{} of a path is the product of the weights of all its steps. Denote by $|P|$ the weight of a path $P$. The [*weight* ]{} of a set of paths is the sum of the total weights of all the paths. For any step, we say that it is at level $k$ if the $y$-coordinate of its end point is at level $k$.
In this section, we aim to establish the following matrix identity on weighted Motzkin numbers.
Let $M=(m_{i,j})_{i,j\geq 1}$ be the lower triangular matrix such that the first column is the sequence of the number of $(k-t-1, t)$-Motzkin paths of length $n$ and $m_{i,j}$ satisfies the following recurrence relation for $j\geq 2$: $$\label{rc.5}
m_{i,j}=m_{i-1,j-1}+(k-t-1)m_{i-1,j}+tm_{i-1,j+1}.$$ Then we have $$\label{eq.5}
(m_{i,j}) \times
\begin{bmatrix} 1 \\ 1+t \\ 1+t+t^2 \\ \vdots\\
\end{bmatrix} = \begin{bmatrix} 1 \\ k \\ k^2 \\ \vdots
\end{bmatrix},$$
It is well known that the number of $2$-Motzkin paths of length $n$ is given by the Catalan number $C_{n+1}$. It follows that the matrix identity (\[eq2.2\]) is a special case of (\[eq.5\]) for $k=4,t=1$. Denote by $f(x)=\sum_{n\geq 0}f_{n}x^n$ the generating function for the number of $(3,2)$-Motzkin paths. Then it is easy to find the functional equation for $f(x)$: $$f(x)=1+3xf(x)+2x^2f^2(x).$$ It follows that $$f(x)={{1-3x-\sqrt{1-6x+x^2}}\over {4x^2}}.$$ From the above generating function, one see that the number of $(3,2)$-Motzkin paths of length $n$ equals the $n$-th little Schröder number. Therefore, the matrix identity (\[eq1.1\]) is a special case of (\[eq.5\]) for $k=6,t=2$.
Let us rewrite matrix identity (\[eq.5\]) in the following form $$\label{mi}
\sum_{j= 1}^i \, m_{i,j}(1+t+\cdots +t^{j-1}) =k^{i-1}.$$
The following combinatorial interpretation of the entries in the matrix in (\[eq.5\]) is due to Cameron and Nkwanta [@cn]. A partial $(k,t)$-Motkzin path is defined as an initial segment of a $(k,t)$-Motkzin path. We say that a partial $(k, t)$-Motzkin path ends at level $j$ if its last step is at level $j$.
Let $m_{i,j}$ be the entries in the matrix in (\[eq.5\]). Then $m_{i,j}$ equals the the number of partial $(k-t-1, t)$-Motzkin paths of length $i-1$ that end at level $j-1$.
[*Proof.*]{} Regarding the first column of the matrix $M$, one sees that a partial $(k-t-1, t)$-Motzkin path that ends at level zero is just a $(k-t-1, t)$-Motzkin path. Let $a_{i,j}$ denote the number of partial $(k-t-1, t)$-Motzkin paths of length $i-1$ ending at level $j-1$. Let $P$ be a partial $(k-t-1,
t)$-Motzkin path of length $i-1$ that ends at level $j-1$ $(j>1)$. By considering the last step of $P$ and its weight, one sees that $a_{i,j}$ satisfies the recurrence relation (\[rc.5\]).
Let $P$ be a partial $(k-t-1, t)$-Motzkin path ending at level $j-1$. We need the notion of an [*elevated* ]{} partial Motzkin path, which has been introduced by Cameron and Nkwanta [@cn] in their combinatorial proof of the following identity which is a reformulation of (\[eq.2\]): $$4^n=
\sum_{k=0}^n \, {(k+1)^2 \over n+1} \, {2n+2\choose n-k}.$$ Let $p$ be an integer with $0\leq p\leq j$. The [*elevation*]{} of $P$ with respect to the horizontal line $y=p$ is defined as follows. For $p=0$, the elevation of $P$ with respect to $y=0$ is just $P$ itself. We now assume $ 0<p\leq j$. Note that there are always up steps of $P$ at levels $j-1$, $j-2$, …, $0$ bearing in mind that an up step is said to be at level $k$ if its initial point is at level $k$. Therefore, for each level from $0$ to $p-1$, one can find a rightmost up step. Note that there are no other steps at the same level to the right of the rightmost up step which is called a [*R-visible*]{} up step with respect to the line $y=p$ in the sense that it can seen far away from the right. The elevation of $P$ with respect to the line $y=p$ is derived from $P$ by changing the R-visible up steps up to level $p-1$ to down steps by elevating their initial points. The line $y=p$ is called an [*elevation line*]{}.
Figure \[mot\] is an illustration of the elevation of a partial Motzkin path with respect to the line $y=2$.
(150,30) (-0.15,0)[(1,1)[4]{}]{} (0.15,0)[(1,1)[4]{}]{}(4,4)[(1,0)[4]{}]{} (8,4)[(1,1)[4]{}]{} (12,8)[(1,-1)[4]{}]{} (16,4)[(1,0)[4]{}]{} (20,4)[(1,1)[8]{}]{} (28,12)[(1,-1)[8]{}]{} (35.85,4)[(1,1)[4]{}]{} (36.15,4)[(1,1)[4]{}]{}(40,8)[(1,1)[8]{}]{}(48,16)[(1,-1)[4]{}]{} (52,12)[(1,0)[4]{}]{} (0,8)(3,0)[20]{}[(1,0)[1]{}]{}
(0,0) (4,4) (8,4) (12,8) (16,4) (20,4) (24,8) (28,12) (32,8) (36,4) (40,8) (44,12) (48,16) (52,12) (56,12)
(63,5)[$\longleftrightarrow$]{}
(79.85,8)[(1,-1)[4]{}]{} (80.15,8)[(1,-1)[4]{}]{}(84,4)[(1,0)[4]{}]{} (88,4)[(1,1)[4]{}]{} (92,8)[(1,-1)[4]{}]{} (96,4)[(1,0)[4]{}]{} (100,4)[(1,1)[8]{}]{} (108,12)[(1,-1)[8]{}]{} (115.85,4)[(1,-1)[4]{}]{} (116.15,4)[(1,-1)[4]{}]{}(120,0)[(1,1)[8]{}]{} (128,8)[(1,-1)[4]{}]{} (132,4)[(1,0)[4]{}]{}
(80,8) (84,4) (88,4) (92,8) (96,4) (100,4) (104,8) (108,12) (112,8) (116,4) (120,0) (124,4) (128,8) (132,4) (136,4)
We now introduce the notion of [*free Motzkin* ]{} paths which are lattice paths starting from $(0,0)$ and using up steps $U=(1,1)$, down steps $D=(1,-1)$ and horizontal steps $H=(1,0)$. Note that there is no restriction so that the paths may go below the $x$-axis. A free $(k,t)$-Motzkin path is a free Motzkin path in which the steps are weighted in the same way as for $(k,t)$-Motzkin paths, namely, an up step has weight one, a horizontal step has weight $k$ and a down step has weight $t$.
For a free Motzkin path $P$ we may analogously define the [*L-visible*]{} down steps as the down steps that are visible from the far left. It is clear that a complete Motzkin path (a partial Motzkin path with ending point on the $x$-axis) has no R-visible up steps. Similarly, a partial Motzkin path has no L-visible down steps.
We have the following summation formula for weighted free Motzkin paths.
The sum of weights of free $(k-t-1, t)$-Motkzin paths of length $i$ equals $k^i$.
The proof of the above lemma is obvious because of the relation $$(1+k-t-1+t)^i=k^i.$$ We are now led to establish a bijection for the identity (\[mi\]).
\[the.1\] There is a bijection between the set of partial $(k-t-1,t)$-Motzkin paths of length $i$ with an elevation line and the set of free $(k-t-1,t)$-Motzkin path of length $i$.
The bijection is just the elevation operation. The reverse map is also easy. For a free Motzkin path, one can identify the L-visible down steps, if any, then change these L-visible down steps to up steps by elevating their end points.
For a partial $(k-t-1, t)$-Motzkin path $P$ with an elevation line $y=p$, suppose that $Q$ is the elevation of $P$ with respect to $y=p$. It is clear that the weight of $Q$ equals $t^p |P|$. If $P$ ends at level $j$, then the possible elevation lines are $y=0$, $y=1, \ldots, y=j-1$. Summing over $j$ to get what we have currently. Thus we arrive at a combinatorial interpretation of the identity (\[mi\]).
As a consequence of Theorem \[the.1\], we obtain the matrix identity (\[eq.5\]). Note that the matrix identity (\[eq2.2\]) is a special case of (\[eq.5\]) for $k=4$ and $t=1$, and the matrix identity (\[eq1.1\]) is also a special case of (\[eq.5\]) for $k=6$ and $t=2$.
An Identity of Cameron and Nkwanta
==================================
In their study of involutions in Riordan groups, Cameron and Nkwanta [@callan] obtained the following identity for $m\geq
0$, and asked for a purely combinatorial proof: $$\binom{n}{m}4^{n-m}=\sum_{k=0}^{n}\frac{k+1}{n+1}\binom{k+m+1}{k-m}\binom{2n+2}{n-k}.$$
It is clear that identity (\[eq.2\]) is a special case for $m=0$. To be consistent with our notation, we may rewrite the above identity in the following form: $$\label{cam}
\binom{i-1}{m}4^{i-1-m}=\sum_{j=1}^{i}\frac{j}{i}\binom{j+m}{2m+1}\binom{2i}{i-j}.$$
We now give a bijective proof (\[cam\]).
We recall that the number of partial $2$-Motzkin paths of length $i-1$ ending at level $j-1$ is given by $a_{i,j}=\frac{j}{i}\binom{2i}{i-j}$. We now consider the set of partial $2$-Motzkin paths with $m$ marked R-visible up steps and $m+1$ elevation lines such that there is exactly one marked R-visible up step between two adjacent elevation lines. We now have a combinatorial interpretation of the summand in (\[cam\]).
The summand in (\[cam\]) counts partial $2$-Motzkin paths of length $i-1$ ending at level $j-1$ with $m$ marked R-visible up steps and $m+1$ elevation lines such that there is exactly one marked step between two adjacent elevation lines.
Let $P$ be a partial $2$-Motzkin path of length $i-1$ ending at level $j-1$. Suppose that there are $m$ marked R-visible up steps with initial points at levels $j_1, j_2, \ldots, j_m$. Let $t_1=j_1$ and $t_{i}=j_{i}-j_{i-1}-1$ with $j_{m+1}=j-1$ for $i\geq 2$. Then one see that the number of ways to choose the $m+1$ elevation lines such that there is exactly $m+1$ marked R-visible up step is equals to $$(t_1+1) (t_2+1) \cdots (t_m+1).$$ Note that the $t_i$’s range over $t_1+t_2+\cdots +t_{m+1}=j-m-1$. Thus, the number of partial $2$-Motzkin paths of length $i-1$ ending at level $j-1$ with the required marked steps and elevation lines equal $$\label{c.1}
\sum_{t_1+t_2+\ldots t_{m+1}=j-1-m}(t_1+1)(t_2+1)\cdots
(t_{m+1}+1).$$ Let $g(x)=\sum_{n\geq 0}(n+1)x^n$. It is clear that $g(x)={1\over
(1-x)^2}$. Hence the summation (\[c.1\]) equals the coefficient of $x^{j-1-m}$ in the expansion of ${1\over (1-x)^{2m+2}}$, that is, the binomial coefficient ${j+m\choose 2m+1}$.
In fact, we may give a combinatorial interpretation of the binomial coefficient ${j+m \choose 2m+1}$ in the above proof. Let $P$ be a partial $2$-Motzkin path of length $i-1$ ending at level $j-1$ with $m$ marked R-visible up steps and $m+1$ elevation lines such that there is exactly one marked up step between two adjacent elevation lines. Suppose that the $k$-th elevation line and the $k$-th marked up step of $P$ are at level $x_k$ and $y_k$, respectively. Such a configuration can be represented as follows: $$t_1\, | \, t_2 \, * \, t_3\, |\, t_4\, * \, t_5\, | \, \cdots \, |
\, t_{2m} \, * \, t_{2m+1} \, |\, t_{2m+2} ,$$ where $t_i$ denotes the numbers of unmarked R-visible up steps. It is clear that we have $t_1+t_2+\ldots+t_{2m+2}=j-1-m$, and the number of solutions of this equation equals the numbers of ways to distribute $j-1-m$ balls into $2m+2$ boxes when a box can have more than one ball. So this number equals the binomial coefficient ${ j+m \choose 2m+1}$.
We are now ready to give a combinatorial proof of the identity of Cameron and Nkwanta. We recall that a $2$-Motzkin path have two kind of horizontal steps, straight steps and wavy steps. We now need to introduce the third kind of horizontal steps – dotted steps. Therefore, the left hand side of (\[cam\]) is the number of free $3$-Motzkin paths with exactly $m$ dotted horizontal steps. We now give the following bijection that leads to a combinatorial interpretation of (\[cam\]).
\[cameron\] There is a bijection between partial $2$-Motzkin paths of length $i$ with $m$ marked R-visible up steps and $m+1$ elevation lines such that there is exactly one marked step between two adjacent elevation lines and free $3$-motzkin paths of length $i$ with exactly $m$ dotted horizontal steps.
Suppose that $P=P_1U^*P_2U^*\ldots P_{m}U^*P_{m+1}$ is a partial $2$-Motzkin path with $m$ marked R-visible up steps and $m+1$ elevation lines, we get a free $3$-motzkin path by changing all the marked up steps to dotted horizontal steps and applying the elevation operation for each $P_k$.
Conversely, given a free $3$-Motzkin path $P=P_1 \dashrightarrow P_2\dashrightarrow \cdots P_{m}
\dashrightarrow P_{m+1}$ with $m$ dotted horizontal steps, where $\dashrightarrow$ denotes a dotted horizontal step, then we can get a partial $2$-motzkin path by changing each dotted horizontal step to a marked up step and the L-visible down steps of each $P_k$ to up steps by elevating their end points.
(138,30) (-0.15,0)[(1,1)[4]{}]{} (0.15,0)[(1,1)[4]{}]{}(4,4)(4.5,5)(5,4) (5,4)(5.5,3)(6,4) (6,4)(6.5,5)(7,4) (7,4)(7.5,3)(8,4) (8,4)[(1,1)[4]{}]{} (12,8)[(1,0)[4]{}]{} (16,8)[(1,-1)[4]{}]{} (20,4)[(1,1)[8]{}]{} (28,12)[(1,-1)[8]{}]{} (36,4)[(1,1)[4]{}]{} (39.85,8)[(1,1)[8]{}]{} (40.15,8)[(1,1)[8]{}]{}(42.5,9)[$*$]{} (40,8)[(1,1)[12]{}]{} (52,20)[(1,0)[4]{}]{} (0,4)(3,0)[20]{}[(1,0)[1]{}]{} (0,16)(3,0)[20]{}[(1,0)[1]{}]{} (0,0) (4,4) (8,4) (12,8) (16,8) (20,4) (24,8) (28,12) (32,8) (36,4) (40,8) (44,12) (48,16) (52,20) (56,20) (63,5)[$\longleftrightarrow$]{} (79.85,8)[(1,-1)[4]{}]{} (80.15,8)[(1,-1)[4]{}]{}(84,4)(84.5,5)(85,4) (85,4)(85.5,3)(86,4) (86,4)(86.5,5)(87,4) (87,4)(87.5,3)(88,4) (88,4)[(1,1)[4]{}]{} (92,8)[(1,0)[4]{}]{} (96,8)[(1,-1)[4]{}]{} (100,4)[(1,1)[8]{}]{} (108,12)[(1,-1)[8]{}]{} (116,4)[(1,1)[4]{}]{} (120,8)(1,0)[5]{}[(1,0)[0.4]{}]{} (123.85,8)[(1,-1)[4]{}]{} (124.15,8)[(1,-1)[4]{}]{}(128,4)[(1,1)[4]{}]{} (132,8)[(1,0)[4]{}]{} (80,8) (84,4) (88,4) (92,8) (96,8) (100,4) (104,8) (108,12) (112,8) (116,4) (120,8) (124,8) (128,4) (132,8) (136,8)
We conclude this section by giving a more general identity. Let $a_{i,j,k}$ be the number of partial $k$-Motzkin paths of length $i-1$ ending at level $j-1$. Then we have $$\label{cameron1}
\binom{i-1}{m}k^{i-1-m}=\sum_{j=1}^{i}a_{i,j,k-2}\binom{j+m}{2m+1}.$$
A Dyck path generalization of (\[eq1.1\])
=========================================
In this section, we give a Dyck path generalization of the matrix identity (\[eq1.1\]) on the little Schröder numbers. A $k$-Dyck path is a Dyck path in which an up step is colored by one of the $k$ colors $\{ 1, 2, \ldots, k\}$ if it not immediately followed by a down step . In this section, we aim to give the following generalization of (\[eq1.1\]).
Let $M=(m_{i,j})$ be a lower triangular matrix with the first column being the weight of $(t^2-t)$-Dyck paths of length $2i$. The other columns of $M$ are given by the following relation: $$\label{eq.6}
m_{i,j}=m_{i-1,j-1}+(t^2-t+1)m_{i-1,j}+(t^2-t)m_{i-1,j+1}.$$ Then we have the following matrix identity $$\label{eq.t}
(m_{i,j}) \times
\begin{bmatrix} 1 \\ t^2-(t-1)^2 \\ \ t^3-(t-1)^3 \\ \vdots
\end{bmatrix} = \begin{bmatrix} 1 \\ t^2+t \\ (t^2+t)^{2} \\
\vdots
\end{bmatrix}.$$
The matrix identity (\[eq1.1\]) is a consequence of (\[eq.t\]) by setting $t=2$. By using generating functions, one can verify that the number of $2$-Dyck paths of length $2n$ equals the number of little Schröder paths of length $n$.
We now proceed to give a combinatorial proof of (\[eqt.t\]). To this end, we need to give a combinatorial interpretation of the entries in the matrix $M$ in (\[eq.t\]). We may define a composition of $k$-Dyck path $P$ as a sequence of $k$-Dyck paths $(P_1, P_2, \ldots, P_j)$ such that $P=P_1P_2\cdots P_j$, where $j$ is called the number of segments.
Let $a_{i,j}$ be the sum of weights of compositions of $(t^2-t)$-Dyck paths of length $2i$ with $j$ segments. Then $a_{i,j}$ satisfies the recurrence relation (\[eq.6\]).
The proof of the above lemma is similar to that of Lemma \[dotheom\]. Let us rewrite (\[eq.t\]) as follows $$\label{eqt.t}
\sum_{j\geq 1}m_{i,j}(t^j-(t-1)^j)=(t^2+t)^{i-1}.$$
In order to deal with $m_{i,j}(t^j-(t-1)^j)$ combinatorially, we introduce a coloring scheme on a composition of a $(t^2-t)$-Dyck path with $j$ segments. Suppose that we have $t$ colors $c_1, c_2,
\ldots, c_t$. If we use these $t$ colors to color the $j$ segments such that the first color $c_1$ must be used, then there are $t^j-(t-1)^j$ ways to accomplish such colorings. We simply call such colorings [*$t$-feasible colorings*]{}.
We can now present a bijection leading to a combinatorial proof of (\[eqt.t\]).
There is a bijection between the set of compositions of $(t^2-t)$-Dyck paths of length $2i$ with a $t$-feasible coloring on the segments and the set of sequences of length $i-1$ on $t^2+t$ letters.
The desired bijection $\sigma$ is constructed as follows. Let $(P_1, P_2, \ldots, P_j)$ be a composition of a $(t^2-t)$-Dyck path $P$ of length $2i$ with a $t$-feasible coloring on the segments. We will use the following alphabet that contains $t^2+t$ letters: $$\label{alphabet}
\{ \alpha_r, \;|\;
1\leq r\leq t\} \cup \{ \beta_s, \;|\; 1\leq s\leq t-1\} \cup \{
\gamma_{k} \; | \; 1 \leq k \leq t^2-t\} \cup \{ \delta\}.$$
For $i=1$, both the composition and the $t$-feasible coloring are unique. We set the corresponding sequence to be empty. For $i\geq 2$, we consider the following cases:
1. If $P_1=UD$, $P_1$ is colored by $c_r$ ($1\leq r\leq t$) and $(P_2, \ldots, P_j)$ still has a $t$-feasible coloring. Then we set $\sigma(P_1, \ldots P_j)=\alpha_r \sigma(P_2, \ldots, P_j)$.
2. If $P_1=UD$, $P_1$ is colored by $c_1$ and $(P_2, \ldots, P_j)$ does not inherit a $t$-feasible coloring. Assume that $P_2$ is colored by $c_{s+1}$ ($1\leq s\leq t-1$). Then we change the color of $P_2$ to $c_1$ and set $\sigma(P_1, \ldots, P_j)=\beta_s \sigma(P_2, \ldots, P_j)$.
3. If $P_1=UDQ$, where $Q$ is not empty. Then we set $\sigma(P_1, \ldots, P_j)=\delta\sigma(Q, P_2, \ldots, P_j)$.
4. If $P_1=UQD$, where $Q$ is not empty and the first up step of $P$ has color $k$ ($1 \leq k \leq
t^2-t$) because the first step of $Q$ is an up step. Then we set $\sigma(P_1, \ldots, P_j)=\gamma_{k}
\sigma(Q, P_2, \ldots, P_j)$.
5. If $P_1=U Q_1 D Q_2$, neither $Q_1$ nor $Q_2$ is empty and the first up step of $P$ has color $c_k$. Since $k$ ranges from $1$ to $t(t-1)$, we may encode a color $c_k$ by a pair of colors $(c_p, c_q)$ where $p$ ranges from $1$ to $t$ and $q$ ranges from $1$ to $t-1$. Moreover, we may use $(c_r, \beta_s)$ to denote a color $c_k$. Then we assign color $c_r$ to $Q_1$, pass the color of $P_1$ to $Q_2$, and set $\sigma(P)=\beta_s\sigma(Q_1, Q_2, P_2, \ldots, P_j)$.
For each case, the resulting path is always a sequence of length $i-1$.
In order to show that $\sigma$ is a bijection, we proceed to construct the inverse map of $\sigma$. Let $S$ be a sequence of length $i-1$ on the alphabet (\[alphabet\]). If $i=1$, then we get the unique Dyck path $UD$ and the unique composition with a $t$-feasible coloring. Note that the up step in the Dyck path $UD$ is not colored. We now assume that $i>1$. It is easy to check that Cases 1, 3, and 4 are reversible. It remains to show that Cases 2 and 5 are reversible. In fact, we only need to ensure that Case 2 and Case 5 can be distinguished from each other. For Case 2, either $j=2$ or $(P_3, \ldots, P_j)$ does not have a $t$-feasible coloring. On the other hand, for Case 5, $(Q_2,
P_2,\ldots, P_j)$ is always nonempty and it has a $t$-feasible coloring. This completes the proof.
We also have a combinatorial interpretation of the matrix identity (\[eq1.1\]) based on little Schröder paths. The idea is similar to the proof given above, so the proof is omitted.
[**Acknowledgments.**]{} This work was supported by the 973 Project on Mathematical Mechanization, the National Science Foundation, the Ministry of Education, and the Ministry of Science and Technology of China. The third author is partially supported by NSF grant HRD 0401697.
[100]{}
E. Barcucci, A. del Lungo, E. Pergola and R. Pinzani, A construction for enumerating $k$-coloured Motzkin paths, Lecture Notes in Computer Science, Vol. 959, Springer, Berlin, 1995, pp. 254-263.
M. Aigner, Catalan-like numbers and determinants, [*J. Combin. Theory, Ser. A*]{}, 87 (1999) 33-51.
D. Callan, A combinatorial interpretation of a Catalan numbers identity, [*Math. Mag.*]{}, 72 (1999) 295-298.
N. Cameron and A. Nkwanta, On some (pseado) involutions in the Riordan group, [*J. Integer Sequences*]{}, 8 (2005), Article 05.3.7.
R. Chapman, Moments of Dyck paths, [*Discrete Math.*]{}, 204 (1999) 113-117.
E. Deutsch, Dyck path enumeration, [*Discrete Math.*]{}, 204 (1999) 167-202.
E. Deutsch and L. Shapiro, A bijection between ordered trees and $2$-Motzkin paths and its many consequences, [*Discrete Math.*]{}, 256 (2002) 655-670.
F. Harary and R. C. Read, The enumeration of tree-like polyhexes, [ *Proc. Edinburgh Math. Soc.*]{}, (2) 17 (1970) 1-13.
L. Shapiro, Bijections and the Riordan group, [*Theoretical Computer Science*]{}, 307 (2003) 403-413.
L. Shapiro, S. Getu, Wen-Jin Woan, and L.C. Woodson, The Riordan group, [*Discrete Appl. Math.*]{}, 34 (1991) 229-239.
L. Shapiro, Wen-Jin Woan, and S. Getu, Runs, slides and moments, [ *SIAM J. Alg. Disc. Math.*]{}, 4 (1983), 459-466.
L. Shapiro, A Catalan triangle, [*Discrete Math.*]{}, 14 (1976), 83-90.
N.J.A. Sloane, S. Plouffe, The Encyclopedia of Integer Sequence, Academic Press, San Diego, 1995, online at [www.research.att.com/126 njas/sequences/]{}.
R. Sprugnoli, Riordan arrays and combinatorial sums, [*Discrete Math.*]{}, 132 (1994) 267-290.
R.P. Stanley, [*Enumerative Combinatorics, Vol. 2*]{}, Cambridge University Press, Cambridge, 1999.
R.A. Sulanke, Moments of generalized Motzkin pahts, [*J. Integer Sequences*]{}, 3 (2000) 00.1.1.
Wen-Jin Woan, L. Shapiro, and D.G. Rogers, The Catalan numbers, the Lebesgue integral, and $4^{n-2}$, [*The Amer. Math. Monthly*]{}, 104 (1997) 927-931.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a symmetric $D\times D$ matrix $M$ over $\{0,1,*\}$, a list $M$-partition of a graph $G$ is a partition of the vertices of $G$ into $D$ parts which are associated with the rows of $M$. The part of each vertex is chosen from a given list in such a way that no edge of $G$ is mapped to a $0$ in $M$ and no non-edge of $G$ is mapped to a $1$ in $M$. Many important graph-theoretic structures can be represented as list $M$-partitions including graph colourings, split graphs and homogeneous sets and pairs, which arise in the proofs of the weak and strong perfect graph conjectures. Thus, there has been quite a bit of work on determining for which matrices $M$ computations involving list $M$-partitions are tractable. This paper focuses on the problem of counting list $M$-partitions, given a graph $G$ and given a list for each vertex of $G$. We identify a certain set of “tractable” matrices $M$. We give an algorithm that counts list $M$-partitions in polynomial time for every (fixed) matrix $M$ in this set. The algorithm relies on data structures such as sparse-dense partitions and subcube decompositions to reduce each problem instance to a sequence of problem instances in which the lists have a certain useful structure that restricts access to portions of $M$ in which the interactions of $0$s and $1$s is controlled. We show how to solve the resulting restricted instances by converting them into particular counting constraint satisfaction problems (${\ensuremath{\mathrm{\#CSP}}}$s) which we show how to solve using a constraint satisfaction technique known as “arc-consistency”. For every matrix $M$ for which our algorithm fails, we show that the problem of counting list $M$-partitions is [$\mathrm{\#P}$]{}-complete. Furthermore, we give an explicit characterisation of the dichotomy theorem — counting list $M$-partitions is tractable (in [$\mathrm{FP}$]{}) if the matrix $M$ has a structure called a derectangularising sequence. If $M$ has no derectangularising sequence, we show that counting list $M$-partitions is [$\mathrm{\#P}$]{}-hard. We show that the meta-problem of determining whether a given matrix has a derectangularising sequence is [$\mathrm{NP}$]{}-complete. Finally, we show that list $M$-partitions can be used to encode cardinality restrictions in $M$-partitions problems and we use this to give a polynomial-time algorithm for counting homogeneous pairs in graphs.'
author:
- 'Andreas Göbel[^1]'
- Leslie Ann Goldberg
- 'Colin McQuillan[^2]'
- David Richerby
- 'Tomoyuki Yamakami[^3]'
bibliography:
- '\\jobname.bib'
title: 'Counting list matrix partitions of graphs[^4]'
---
Introduction
============
A matrix partition of an undirected graph is a partition of its vertices according to a matrix which specifies adjacency and non-adjacency conditions on the vertices, depending on the parts to which they are assigned. For finite sets $D$ and $D'$, the set $\{0,1,*\}^{D\times D'}$ is the set of matrices with rows indexed by $D$ and columns indexed by $D'$ where each $M_{i,j} \in \{0,1,*\}$. For any symmetric matrix $M\in\{0,1,*\}^{D\times D}$, an [*$M$-partition*]{} of an undirected graph $G=(V,E)$ is a function $\sigma\colon V\to D$ such that, for distinct vertices $u$ and $v$,
- $M_{\sigma(u),\sigma(v)}\neq 0$ if $(u,v)\in E$ and
- $M_{\sigma(u),\sigma(v)}\neq 1$ if $(u,v)\not\in E$.
Thus, $M_{i,j}=0$ means that no edges are allowed between vertices in parts $i$ and $j$, $M_{i,j}=1$ means that there must be an edge between every pair of vertices in the two parts and $M_{i,j}=*$ means that any set of edges is allowed between the parts. For entries $M_{i,i}$ on the diagonal of $M$, the conditions only apply to distinct vertices in part $i$. Thus, $M_{i,i}=1$ requires that the vertices in part $i$ form a clique in $G$ and $M_{i,i}=0$ requires that they form an independent set.
For example, if $D=\{i,c\}$, $M_{i,i} = 0$, $M_{c,c}=1$ and $M_{c,i} =
M_{i,c} = *$, i.e., $M=\left(\begin{smallmatrix}0 & *\\ * &
1\end{smallmatrix}\right)$, then an $M$-partition of a graph is a partition of its vertices into an independent set (whose vertices are mapped to $i$) and a clique (whose vertices are mapped to $c$). The independent set and the clique may have arbitrary edges between them. A graph that has such an $M$-partition is known as a split graph [@Golumbic].
As Feder, Hell, Klein and Motwani describe [@FHKM], many important graph-theoretic structures can be represented as $M$-partitions, including graph colourings, split graphs, $(a,b)$-graphs [@Bra96], clique-cross partitions [@EKR], and their generalisations. $M$-partitions also arise as “type partitions” in extremal graph theory [@BT00]. In the special case where $M$ is a $\{0,*\}$-matrix (that is, it has no 1 entries), $M$-partitions of $G$ correspond to homomorphisms from $G$ to the (potentially looped) graph $H$ whose adjacency matrix is obtained from $M$ by turning every $*$ into a 1. Thus, proper $|D|$-colourings of $G$ are exactly $M$-partitions for the matrix $M$ which has 0s on the diagonal and $*$s elsewhere.
To represent more complicated graph-theoretic structures, such as homogeneous sets and their generalisations, which arise in the proofs of the weak and strong perfect graph conjectures [@lovasz; @CRST], it is necessary to generalise $M$-partitions by introducing lists. Details of these applications are given by Feder et al. [@FHKM], who define the notion of a list $M$-partition.
A [*list $M$-partition*]{} is an $M$-partition $\sigma$ that is also required to satisfy constraints on the values of each $\sigma(v)$. Let ${{\mathcal{P}(D)}}$ denote the powerset of $D$. We say that $\sigma$ [*respects*]{} a function $L\colon V(G)\to {{\mathcal{P}(D)}}$ if $\sigma(v)\in
L(v)$ for all $v\in V(G)$. Thus, for each vertex $v$, $L(v)$ serves as a list of allowable parts for $v$ and a *list $M$-partition* of $G$ is an $M$-partition that respects the given list function. We allow empty lists for technical convenience, although there are no $M$-partitions that respect any list function $L$ where $L(v)=\emptyset$ for some vertex $v$.
Feder et al.[@FHKM] study the computational complexity of the following decision problem, which is parameterised by a symmetric matrix $M\in\{0,1,*\}^{D\times D}\!$.
[[List-$M$-partitions]{}]{}.
A pair $(G,L)$ in which $G$ is a graph and $L$ is a function $V(G)\to{{\mathcal{P}(D)}}$.
“Yes”, if $G$ has an $M$-partition that respects $L$; “no”, otherwise.
Note that $M$ is a parameter of the problem rather than an input of the problem. Thus, its size is a constant which does not vary with the input.
A series of papers [@FH; @FHHList; @FHH] described in [@FHKM] presents a complete dichotomy for the special case of homomorphism problems, which are [[List-$M$-partitions]{}]{} problems in which $M$ is a $\{0,*\}$-matrix. In particular, Feder, Hell and Huang [@FHH] show that, for every $\{0,*\}$-matrix $M$ (and symmetrically, for every $\{1,*\}$-matrix $M$), the problem [[List-$M$-partitions]{}]{} is either polynomial-time solvable or [$\mathrm{NP}$]{}-complete.
It is important to note that both of these special cases of [[List-$M$-partitions]{}]{} are constraint satisfaction problems (CSPs) and a famous conjecture of Feder and Vardi [@FV] is that a P versus [$\mathrm{NP}$]{}-complete dichotomy also exists for every CSP. Although general [[List-$M$-partitions]{}]{} problems can also be coded as CSPs with restrictions on the input,[^5] it is not known how to code them without such restrictions. Since the Feder–Vardi conjecture applies only to CSPs with unrestricted inputs, even if proved, it would not necessarily apply to [[List-$M$-partitions]{}]{}.
Given the many applications of [[List-$M$-partitions]{}]{}, it is important to know whether there is a dichotomy for this problem. This is part of a major ongoing research effort which has the goal of understanding the boundaries of tractability by identifying classes of problems, as wide as possible, where dichotomy theorems arise and where the precise boundary between tractability and intractability can be specified.
Significant progress has been made on identifying dichotomies for [[List-$M$-partitions]{}]{}. Feder et al. [@FHKM Theorem 6.1] give a complete dichotomy for the special case in which $M$ is at most $3\times 3$, by showing that [[List-$M$-partitions]{}]{} is polynomial-time solvable or [$\mathrm{NP}$]{}-complete for each such matrix. Later, Feder and Hell studied the [[List-$M$-partitions]{}]{} problem under the name CSP$^*_{1,2}(H)$ and showed [@FHFull Corollary 3.4] that, for every $M$, [[List-$M$-partitions]{}]{} is either [$\mathrm{NP}$]{}-complete, or is solvable in quasi-polynomial time. In the latter case, they showed that [[List-$M$-partitions]{}]{} is solvable in $n^{O(\log n)}$ time, given an $n$-vertex graph. Feder and Hell refer to this result as a “quasi-dichotomy”.
Although the Feder–Vardi conjecture remains open, a complete dichotomy is now known for counting CSPs. In particular, Bulatov [@Bul08] (see also [@DRfull]) has shown that, for every constraint language $\Gamma$, the counting constraint satisfaction problem ${\ensuremath{\mathrm{\#CSP}}}(\Gamma)$ is either polynomial-time solvable, or [$\mathrm{\#P}$]{}-complete. It is natural to ask whether a similar situation arises for counting list $M$-partition problems. We study the following computational problem, which is parameterised by a finite symmetric matrix $M\in\{0,1,*\}^{D\times D}\!$.
[[\#List-$M$-partitions]{}]{}.
A pair $(G,L)$ in which $G$ is a graph and $L$ is a function $V(G)\to{{\mathcal{P}(D)}}$.
The number of $M$-partitions of $G$ that respect $L$.
Hell, Hermann and Nevisi [@HHN] have considered the related problem [[\#$M$-partitions]{}]{} without lists, which can be seen as [[\#List-$M$-partitions]{}]{} restricted to the case that $L(v)=D$ for every vertex $v$. This problem is defined as follows.
[[\#$M$-partitions]{}]{}.
A graph $G$.
The number of $M$-partitions of $G$.
In the problems [[List-$M$-partitions]{}]{}, [[\#List-$M$-partitions]{}]{} and [[\#$M$-partitions]{}]{}, the matrix $M$ is fixed and its size does not vary with the input.
Hell et al. gave a dichotomy for small matrices $M$ (of size at most $3\times 3$). In particular, [@HHN Theorem 10] together with the graph-homomorphism dichotomy of Dyer and Greenhill [@DG] shows that, for every such $M$, [[\#$M$-partitions]{}]{} is either polynomial-time solvable or ${\ensuremath{\mathrm{\#P}}}$-complete. An interesting feature of counting $M$-partitions, identified by Hell et al. is that, unlike the situation for homomorphism-counting problems, there are tractable $M$-partition problems with non-trivial counting algorithms. Indeed the main contribution of the present paper, as described below, is to identify a set of “tractable” matrices $M$ and to give a non-trivial algorithm which solves [[\#List-$M$-partitions]{}]{} for every such $M$. We combine this with a proof that [[\#List-$M$-partitions]{}]{} is ${\ensuremath{\mathrm{\#P}}}$-complete for every other $M$.
Dichotomy theorems for counting list $M$-partitions {#subsec:dichotomy}
---------------------------------------------------
Our main theorem is a general dichotomy for the counting list $M$-partition problem, for matrices $M$ of all sizes. As noted above, since there is no known coding of list $M$-partition problems as CSPs without input restrictions, our theorem is not known to be implied by the dichotomy for [$\mathrm{\#CSP}$]{}.
Recall that [$\mathrm{FP}$]{} is the class of functions computed by polynomial-time deterministic Turing machines. [$\mathrm{\#P}$]{} is the class of functions $f$ for which there is a nondeterministic polynomial-time Turing machine that has exactly $f(X)$ accepting paths for every input $X$; this class can be thought of as the natural analogue of [$\mathrm{NP}$]{} for counting problems. Our main theorem is the following.
\[thm:dichotomy\][ For any symmetric matrix $M\in\{0,1,*\}^{D\times D}\!$, [[\#List-$M$-partitions]{}]{} is either in ${\ensuremath{\mathrm{FP}}}$ or ${\ensuremath{\mathrm{\#P}}}$-complete.]{}
To prove Theorem \[thm:dichotomy\], we investigate the complexity of the more general counting problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}, which has two parameters — a matrix $M\in\{0,1,*\}^{D\times D}$ and a (not necessarily proper) subset [$\mathcal L$]{} of ${{\mathcal{P}(D)}}$. In this problem, we only allow sets in [$\mathcal L$]{} to be used as lists.
[[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}.
A pair $(G,L)$ where $G$ is a graph and $L$ is a function $V(G)\to {\ensuremath{\mathcal L}}$.
The number of $M$-partitions of $G$ that respect $L$.
Note that $M$ and [$\mathcal L$]{} are fixed parameters of [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} — they are not part of the input instance. The problem [[\#List-$M$-partitions]{}]{} is just the special case of [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} where ${\ensuremath{\mathcal L}}= {{\mathcal{P}(D)}}$.
We say that a set ${\ensuremath{\mathcal L}}\subseteq {{\mathcal{P}(D)}}$ is [*subset-closed*]{} if $A\in {\ensuremath{\mathcal L}}$ implies that every subset of $A$ is in [$\mathcal L$]{}. This closure property is referred to as the “inclusive” case in [@FHFull].
\[def:closure\] Given a set ${\ensuremath{\mathcal L}}\subseteq {{\mathcal{P}(D)}}$, we write ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$ for its subset-closure, which is the set $${{\mathscr{S}({\ensuremath{\mathcal L}})}}=\{X \mid \mbox{for some $Y\in {\ensuremath{\mathcal L}}$, $X\subseteq Y$}\}.$$
We prove the following theorem, which immediately implies Theorem \[thm:dichotomy\].
\[thm:fulldichotomy\][Let $M$ be a symmetric matrix in $\{0,1,*\}^{D\times D}$ and let ${\ensuremath{\mathcal L}}\subseteq{{\mathcal{P}(D)}}$ be subset-closed. The problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is either in ${\ensuremath{\mathrm{FP}}}$ or ${\ensuremath{\mathrm{\#P}}}$-complete.]{}
Note that this does not imply a dichotomy for the counting $M$-partitions problem without lists. The problem with no lists corresponds to the case where every vertex of the input graph $G$ is assigned the list $D$, allowing the vertex to be potentially placed in any part. Thus, the problem without lists is equivalent to the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} with ${\ensuremath{\mathcal L}}=\{D\}$, but Theorem \[thm:fulldichotomy\] applies only to the case where [$\mathcal L$]{} is subset-closed.
Polynomial-time algorithms and an explicit dichotomy
----------------------------------------------------
We now introduce the concepts needed to give an explicit criterion for the dichotomy in Theorem \[thm:fulldichotomy\] and to provide polynomial-time algorithms for all tractable cases. We use standard definitions of relations and their arities, compositions and inverses.
For any symmetric $M\in\{0,1,*\}^{D\times D}$ and any sets $X,Y\in{{\mathcal{P}(D)}}$, define the binary relation $$H^M_{X,Y}=\{(i,j)\in X\times Y\mid M_{i,j}=*\}.$$
The intractability condition for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} begins with the following notion of rectangularity, which was introduced by Bulatov and Dalmau [@BD].
A relation $R\subseteq D\times D'$ is [*rectangular*]{} if, for all $i,j\in D$, and $i'\!,j'\in D'\!$, $$(i,i'),(i,j'),(j,i')\in R\implies (j,j')\in R\,.$$
Note that the intersection of two rectangular relations is itself rectangular. However, the composition of two rectangular relations is not necessarily rectangular: for example, $\{(1,1), (1,2),
(3,3)\}\circ \{(1,1), (2,3), (3,1)\} = \{(1,1), (1,3), (3,1)\}$.
Our dichotomy criterion will be based on what we call [$\mathcal L$]{}-$M$-derectangularising sequences. In order to define these, we introduce the notions of pure matrices and $M$-purifying sets.
Given index sets $X$ and $Y$, a matrix $M\in\{0,1,*\}^{X\times Y}$ is [*pure*]{} if it has no $0$s or has no $1$s.
Pure matrices correspond to ordinary graph homomorphism problems. As we noted above, $M$-partitions of $G$ correspond to homomorphisms of $G$ when $G$ is a $\{0,*\}$-matrix. The same is true (by complementation) when $G$ is a $\{1,*\}$-matrix.
For any $M\in\{0,1,*\}^{D\times D}$, a set ${\ensuremath{\mathcal L}}\subseteq {{\mathcal{P}(D)}}$ is [*$M$-purifying*]{} if, for all $X,Y\in{\ensuremath{\mathcal L}}$, the $X$-by-$Y$ submatrix $M|_{X\times Y}$ is pure.
For example, consider the matrix $$M = \left(\begin{matrix}
1 & * & 0 \\
* & 1 & * \\
0 & * & 1
\end{matrix}\right)$$ with rows and columns indexed by $\{0,1,2\}$ in the obvious way. The matrix $M$ is not pure but for ${\ensuremath{\mathcal L}}= \{\{0,1\},
\{2\}\}$, the set ${\ensuremath{\mathcal L}}$ is $M$-purifying and so is the closure ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$.
\[def:derect\] An [*[$\mathcal L$]{}-$M$-derectangularising sequence*]{} of length $k$ is a sequence $D_1,\dots,D_k$ with each $D_i \in{\ensuremath{\mathcal L}}$ such that:
- $\{D_1,\ldots,D_k\}$ is $M$-purifying and
- the relation $H^M_{D_1,D_2} \circ H^M_{D_2, D_3} \circ \dots \circ
H^M_{D_{k-1}, D_k}$ is not rectangular.
If there is an $i\in \{1,\ldots,k\}$ such that $D_i$ is the empty set then the relation $H=H^M_{D_1,D_2} \circ H^M_{D_2, D_3} \circ \dots \circ
H^M_{D_{k-1}, D_k}$ is the empty relation, which is trivially rectangular. If there is an $i$ such that $|D_i|=1$ then $H$ is a Cartesian product, and is therefore rectangular. It follows that $|D_i|\geq 2$ for each $i$ in a derectangularising sequence.
We can now state our explicit dichotomy theorem, which implies Theorem \[thm:fulldichotomy\] and, hence, Theorem \[thm:dichotomy\].
\[thm:explicitdichotomy\][ Let $M$ be a symmetric matrix in $\{0,1,*\}^{D\times D}$ and let ${\ensuremath{\mathcal L}}{}\subseteq{{\mathcal{P}(D)}}$ be subset-closed. If there is an [$\mathcal L$]{}-$M$-derectangularising sequence then the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is ${\ensuremath{\mathrm{\#P}}}$-complete. Otherwise, it is in ${\ensuremath{\mathrm{FP}}}$. ]{}
Sections \[sec:purifiedcsp\], \[sec:arc\] and \[sec:dichotomy\] develop a polynomial-time algorithm which solves the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} whenever there is no [$\mathcal L$]{}-$M$-derectangularising sequence. The algorithm involves several steps.
First, consider the case in which ${\ensuremath{\mathcal L}}$ is subset-closed and $M$-purifying. In this case, Proposition \[prop:purifiediscsp\] presents a polynomial-time transformation from an instance of the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} to an instance of a related counting CSP. Algorithm \[alg:AC\] exploits special properties of the constructed CSP instance so that it can be solved in polynomial time using a CSP technique called arc-consistency. (This is proved in Lemma \[lem:quickarc\].) This provides a solution to the original [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} problem for the $M$-purifying case.
The case in which ${\ensuremath{\mathcal L}}$ is not $M$-purifying is tackled in Section \[sec:dichotomy\]. Section \[sec:DS\] gives algorithms for constructing the relevant data structures, which include a special case of sparse-dense partitions and also subcube decompositions. Algorithm \[alg:purify\] uses these data structures (via Algorithms \[alg:purifystep\], \[alg:Case1\], \[alg:Case2\], \[alg:Case3\] and \[alg:purifytriv\]) to reduce the [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} problem to a sequence of problems [[\#${\ensuremath{\mathcal L}}_i$-$M$-partitions]{}]{} where ${\ensuremath{\mathcal L}}_i$ is $M$-purifying. Finally, the polynomial-time algorithm is presented in Algorithms \[alg:mainpurifying\] and \[alg:main\]. For every ${\ensuremath{\mathcal L}}$ and $M$ where there is no ${\ensuremath{\mathcal L}}$-$M$-derectangularising sequence, either Algorithm \[alg:mainpurifying\] or Algorithm \[alg:main\] defines a polynomial-time function [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} for solving the [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} problem, given an input $(G,L)$. The function [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is not recursive. However, its *definition* is recursive in the sense that the function [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} defined in Algorithm \[alg:main\] calls a function [[\#${\ensuremath{\mathcal L}}_i$-$M$-partitions]{}]{} where ${\ensuremath{\mathcal L}}_i$ is a subset of ${{\mathcal{P}(D)}}$ whose cardinality is smaller than ${\ensuremath{\mathcal L}}$. The function [[\#${\ensuremath{\mathcal L}}_i$-$M$-partitions]{}]{} is, in turn, defined either in Algorithm \[alg:mainpurifying\] or in \[alg:main\].
The proof of Theorem \[thm:explicitdichotomy\] shows that, when Algorithms \[alg:mainpurifying\] and \[alg:main\] fail to solve the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}, the problem is ${\ensuremath{\mathrm{\#P}}}$-complete.
Complexity of the dichotomy criterion
-------------------------------------
Theorem \[thm:explicitdichotomy\] gives a precise criterion under which the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is in ${\ensuremath{\mathrm{FP}}}$ or ${\ensuremath{\mathrm{\#P}}}$-complete, where [$\mathcal L$]{} and $M$ are considered to be fixed parameters. In Section \[sec:meta\], we address the computational problem of determining which is the case, now treating [$\mathcal L$]{} and $M$ as inputs to this “meta-problem”. Dyer and Richerby [@DRfull] studied the corresponding problem for the [$\mathrm{\#CSP}$]{} dichotomy, showing that determining whether a constraint language $\Gamma$ satisfies the criterion for their ${\ensuremath{\mathrm{\#CSP}}}(\Gamma)$ dichotomy is reducible to the graph automorphism problem, which is in [$\mathrm{NP}$]{}. We are interested in the following computational problem, which we show to be [$\mathrm{NP}$]{}-complete.
[[ExistsDerectSeq]{}]{}.
An index set $D$, a symmetric matrix $M$ in $\{0,1,*\}^{D\times D}$ (represented as an array) and a set ${\ensuremath{\mathcal L}}{}\subseteq{{\mathcal{P}(D)}}$ (represented as a list of lists).
“Yes”, if there is an ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$-$M$-derectangularising sequence; “no”, otherwise.
\[thm:meta\][ [[ExistsDerectSeq]{}]{} is [$\mathrm{NP}$]{}-complete under polynomial-time many-one reductions.]{}
Note that, in the definition of the problem [[ExistsDerectSeq]{}]{}, the input ${\ensuremath{\mathcal L}}$ is not necessarily subset-closed. Subset-closedness allows a concise representation of some inputs: for example, ${{\mathcal{P}(D)}}$ has exponential size but it can be represented as ${{\mathscr{S}(\{D\})}}$, so the corresponding input is just ${\ensuremath{\mathcal L}}=\{D\}$. In fact, our proof of Theorem \[thm:meta\] uses a set of lists [$\mathcal L$]{} where $|X|\leq 3$ for all $X\in{\ensuremath{\mathcal L}}$. Since there are at most $|D|^3+1$ such sets, our [$\mathrm{NP}$]{}-completeness proof would still hold if we insisted that the input [$\mathcal L$]{} to [[ExistsDerectSeq]{}]{} must be subset-closed.
Let us return to the original problem [[\#List-$M$-partitions]{}]{}, which is the special case of the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} where ${\ensuremath{\mathcal L}}={{\mathcal{P}(D)}}$. This leads us to be interested in the following computational problem.
[[MatrixHasDerectSeq]{}]{}.
An index set $D$ and a symmetric matrix $M$ in $\{0,1,*\}^{D\times D}$ (represented as an array).
“Yes”, if there is a ${{\mathcal{P}(D)}}$-$M$-derectangularising sequence; “no”, otherwise.
Theorem \[thm:meta\] does not quantify the complexity of [[MatrixHasDerectSeq]{}]{} because its proof relies on a specific choice of [$\mathcal L$]{} which, as we have noted, is not ${{\mathcal{P}(D)}}$. Nevertheless, the proof of Theorem \[thm:meta\] has the following corollary.
\[cor:meta\][ [[MatrixHasDerectSeq]{}]{} is in [$\mathrm{NP}$]{}.]{}
Cardinality constraints
-----------------------
Many combinatorial structures can be represented as $M$-partitions with the addition of cardinality constraints on the parts. For example, it might be required that certain parts be non-empty or, more generally, that they contain at least $k$ vertices for some fixed $k$.
Feder et al. [@FHKM] showed that the problem of determining whether such a structure exists in a given graph can be reduced to a [[List-$M$-partitions]{}]{} problem in which the cardinality constraints are expressed using lists. In Section \[sec:card\], we extend this to counting. We show that any [[\#$M$-partitions]{}]{} problem with additional cardinality constraints of the form, “part $d$ must contain at least $k_d$ vertices” is polynomial-time Turing reducible to [[\#List-$M$-partitions]{}]{}. As a corollary, we show that the “homogeneous pairs” introduced by Chvátal and Sbihi [@CS1987:Bull-free] can be counted in polynomial time. Homogeneous pairs can be expressed as an $M$-partitions problem for a certain $6\times 6$ matrix, with cardinality constraints on the parts.
Preliminaries {#sec:prelim}
=============
For a positive integer $k$, we write $[k]$ to denote the set $\{1,\dots,k\}$. If $\mathcal{S}$ is a set of sets then we use $\bigcap \mathcal{S}$ to denote the intersection of all sets in $\mathcal{S}$. The vertex set of a graph $G$ is denoted $V(G)$ and its edge set is $E(G)$. We write $\{0,1,*\}^{D}$ for the set of all functions $\sigma\colon D\to\{0,1,*\}$ and $\{0,1,*\}^{D\times D'}$ for the set of all matrices $M=(M_{i,j})_{i\in
D,j\in D'}$, where each $M_{i,j}\in\{0,1,*\}$.
We always use the term “$M$-partition” when talking about a partition of the vertices of a graph according to a $\{0,1,*\}$-matrix $M$. When we use the term “partition” without referring to a matrix, we mean it in the conventional sense of partitioning a set $X$ into disjoint subsets $X_1, \dots, X_k$ with $X_1\cup
\dots \cup X_k = X$.
We view computational counting problems as functions mapping strings over input alphabets to natural numbers. Our model of computation is the standard multi-tape Turing machine. We say that a counting problem $P$ is polynomial-time Turing-reducible to another counting problem $Q$ if there is a polynomial-time deterministic oracle Turing machine $M$ such that, on every instance $x$ of $P$, $M$ outputs $P(x)$ by making queries to oracle $Q$. We say that $P$ is polynomial-time Turing-equivalent to $Q$ if each is polynomial-time Turing-reducible to the other. For decision problems (languages), we use the standard many-one reducibility: language $A$ is many-one reducible to language $B$ if there exists a function $f$ that is computable in polynomial time such that $x\in A$ if and only if $f(x)\in B$.
Counting list $M$-partition problems and counting CSPs {#sec:purifiedcsp}
======================================================
Toward the development of our algorithms and the proof of our dichotomy, we study a special case of the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}, in which [$\mathcal L$]{} is $M$-purifying and subset-closed. For such [$\mathcal L$]{} and $M$, we show that the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is polynomial-time Turing-equivalent to a counting constraint satisfaction problem ([$\mathrm{\#CSP}$]{}). To give the equivalence, we introduce the notation needed to specify \#CSPs.
A *constraint language* is a finite set $\Gamma$ of named relations over some set $D$. For such a language, we define the counting problem ${\ensuremath{\mathrm{\#CSP}}}(\Gamma)$ as follows.
${\ensuremath{\mathrm{\#CSP}}}(\Gamma)$.
A set $V$ of variables and a set $C$ of constraints of the form ${\langle (v_1,\dots,v_k),R \rangle}$, where $(v_1,\dots,v_k)\in V^k$ and $R$ is an arity-$k$ relation in $\Gamma$.
The number of assignments $\sigma\colon V\to D$ such that $$\label{eq:satisfying}(\sigma(v_1),\dots,\sigma(v_k))\in
R\text{ for all }{\langle (v_1,\dots,v_k),R \rangle}\in C\,.$$
The tuple of variables $v_1, \dots, v_k$ in a constraint is referred to as the constraint’s *scope*. The assignments $\sigma\colon V\to D$ for which holds are called the [*satisfying assignments*]{} of the instance $(V,C)$. Note that a unary constraint ${\langle v,R \rangle}$ has the same effect as a list: it directly restricts the possible values of the variable $v$. As before, we allow the possibility that $\emptyset\in\Gamma$; any instance that includes a constraint ${\langle (v_1, \dots, v_k), \emptyset \rangle}$ has no satisfying assignments.
\[defgammaprime\] Let $M$ be a symmetric matrix in $\{0,1,*\}^{D\times D}$ and let [$\mathcal L$]{} be a subset-closed $M$-purifying set. Define the constraint language $$\Gamma'_{\!{\ensuremath{\mathcal L}},M} = \{H^M_{X,Y}\mid X,Y\in{\ensuremath{\mathcal L}}\}$$ and let ${\Gamma_{\!{\ensuremath{\mathcal L}}, M}}= \Gamma'_{\!{\ensuremath{\mathcal L}},M} \cup {{\mathcal{P}(D)}}$, where ${{\mathcal{P}(D)}}$ represents the set of all unary relations on $D$.
The unary constraints in ${\Gamma_{\!{\ensuremath{\mathcal L}}, M}}$ will be useful in our study of the complexity of the dichotomy criterion, in Section \[sec:meta\]. First, we define a convenient restriction on instances of ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$.
\[def:simple\] An instance of ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ is *simple* if:
- there is exactly one unary constraint ${\langle v,X_v \rangle}$ for each variable $v\in V\!$,
- there are no binary constraints ${\langle (v,v),R \rangle}$, and
- each pair $u$, $v$ of distinct variables appears in at most one constraint of the form ${\langle (u,v),R \rangle}$ or ${\langle (v,u),R \rangle}$.
\[lemma:simple\] For every instance $(V,C)$ of ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$, there is a simple instance $(V,C')$ such that an assignment $\sigma\colon V\to D$ satisfies $(V,C)$ if and only if it satisfies $(V,C')$. Further, such an instance can be computed in polynomial time.
Observe that the set of binary relations in ${\Gamma_{\!{\ensuremath{\mathcal L}}, M}}$ is closed under intersections: $H^M_{X,Y} \cap H^M_{X'\!,Y'} =
H^M_{X\cap X'\!,Y\cap Y'}$ and this relation is in ${\Gamma_{\!{\ensuremath{\mathcal L}}, M}}$ because [$\mathcal L$]{} is subset-closed. The binary part of ${\Gamma_{\!{\ensuremath{\mathcal L}}, M}}$ is also closed under relational inverse because $M$ is symmetric, so $$\left(H^M_{X,Y}\right)^{-1} = \{(b,a) \mid (a,b)\in H^M_{X,Y}\}
= H^M_{Y,X}\in {\Gamma_{\!{\ensuremath{\mathcal L}}, M}}\,.$$ Since ${{\mathcal{P}(D)}}\subseteq {\Gamma_{\!{\ensuremath{\mathcal L}}, M}}$, the set of unary relations is also closed under intersections.
We construct $C'$ as follows, starting with $C$. Any binary constraint ${\langle (v,v), R \rangle}$ can be replaced by the unary constraint ${\langle v, \{d\mid (d,d)\in R\} \rangle}$. All the binary constraints between distinct variables $u$ and $v$ can be replaced by the single constraint $$\left\langle
(u,v), \bigcap
\{R \mid {\langle (u,v), R \rangle}\in C
\text{ or } {\langle (v,u), R^{-1} \rangle}\in C\}
\right\rangle\,.$$ Let the set of constraints produced so far be $C''\!$. For each variable $v$ in turn, if there are no unary constraints applied to $v$ in $C''\!$, add the constraint ${\langle v, D \rangle}$; otherwise, replace all the unary constraints involving $v$ in $C''$ with the single constraint $$\left\langle
v, \bigcap \{R \mid {\langle v, R \rangle}\in C''\}
\right\rangle\,.$$ $C'$ is the resulting constraint set. The closure properties established above guarantee that $(V,C')$ is a ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ instance. It is clear that it has the same satisfying assignments as $(V,C)$ and that it can be produced in polynomial time.
Our main result connecting the counting list $M$-partitions problem with counting CSPs is the following.
\[prop:purifiediscsp\][For any symmetric $M\in\{0,1,*\}^{D\times
D}$ and any subset-closed, $M$-purifying set [$\mathcal L$]{}, the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is polynomial-time Turing-equivalent to ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$.]{}
Because of its length, we split the proof of the proposition into two lemmas.
For any symmetric $M\in\{0,1,*\}^{D\times D}$ and any subset-closed, $M$-purifying set [$\mathcal L$]{}, ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ is polynomial-time Turing-reducible to [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}.
Consider an input $(V,C)$ to ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$, which we may assume to be simple. Each variable appears in exactly one unary constraint, ${\langle v,X_v \rangle}\in C$. Any variable $v$ that is not used in a binary constraint can take any value in $X_v$ so just introduces a multiplicative factor of $|X_v|$ to the output of the counting CSP. Thus, we will assume without loss of generality that every variable is used in at least one constraint with a relation from $\Gamma'_{\!{\ensuremath{\mathcal L}},M}$ and, by simplicity, there are no constraints of the form ${\langle (v,v),R \rangle}$.
We now define a corresponding instance $(G,L)$ of the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}. The vertices of $G$ are the variables $V$ of the [$\mathrm{\#CSP}$]{} instance. For each variable $v\in V\!$, set $$L(v) = X_v \cap \bigcap\left\{
X \mid
\mbox{for some $u$ and $Y$,
$ {\langle (v,u),H^M_{X,Y} \rangle}\in C$ or
$ {\langle (u,v),H^M_{Y,X} \rangle}\in C$}
\right\}.$$ The edges $E(G)$ of our instance are the unordered pairs $\{u,v\}$ that satisfy one of the following conditions:
- there is a constraint between $u$ and $v$ in $C$ and $M|_{L(u)\times L(v)}$ has a $0$ entry, or
- there is no constraint between $u$ and $v$ in $C$ and $M|_{L(u)\times L(v)}$ has a $1$ entry.
Since every vertex $v$ is used in at least one constraint with a relation $H^M_{X,Y}$ where, by definition, $X$ and $Y$ are in ${\ensuremath{\mathcal L}}$, every set $L(v)$ is a subset of some set $W\in{\ensuremath{\mathcal L}}$. [$\mathcal L$]{} is subset-closed so $L(v)\in{\ensuremath{\mathcal L}}$ for all $v\in V$, as required.
We claim that a function $\sigma\colon V\to D$ is a satisfying assignment of $(V,C)$ if and only if it is an $M$-partition of $G$ that respects $L$. Note that, since [$\mathcal L$]{} is $M$-purifying, no submatrix $M|_{X\times Y}$ ($X,Y\in{\ensuremath{\mathcal L}})$ contains both 0s and 1s.
First, suppose that $\sigma$ is a satisfying assignment of $(V,C)$. For each variable $v$, $\sigma$ satisfies all the constraints ${\langle v,X_v \rangle}$, ${\langle (v,u),H^M_{X,Y} \rangle}$ and ${\langle (u,v),H^M_{Y,X} \rangle}$ containing $v$. Therefore, $\sigma(v)\in X_v$ and $\sigma(v)\in X$ for each binary constraint ${\langle (v,u),H^M_{X,Y} \rangle}$ or ${\langle (u,v),H^M_{Y,X} \rangle}$, so $\sigma$ satisfies all the list requirements.
To show that $\sigma$ is an $M$-partition of $G$, consider any pair of distinct vertices $u,v\in V$. If there is a constraint ${\langle (u,v), H^M_{X,Y} \rangle}\in C$, then $\sigma$ satisfies this constraint so $M_{\sigma(u),\sigma(v)}=*$ and $u$ and $v$ cannot stop $\sigma$ being an $M$-partition. Conversely, suppose there is no constraint between $u$ and $v$ in $C$. If $M|_{L(u)\times L(v)}$ contains a 0, there is no edge $(u,v)\in E(G)$ by construction; otherwise, if $M|_{L(u)\times L(v)}$ contains a 1, there is an edge $(u,v)\in E(G)$ by construction; otherwise, $M_{x,y}=*$ for all $x\in L(u)$, $y\in L(v)$. In all three cases, the assignment to $u$ and $v$ is consistent with $\sigma$ being an $M$-partition.
Conversely, suppose that $\sigma$ is not a satisfying assignment of $(V,C)$. If $\sigma$ does not satisfy some unary constraint ${\langle v,X \rangle}$ then $\sigma(v)\notin L(v)$ so $\sigma$ does not respect [$\mathcal L$]{}. If $\sigma$ does not satisfy some binary constraint ${\langle (u,v), H^M_{X,Y} \rangle}$ where $u$ and $v$ are distinct then, by definition of the relation $H^M_{X,Y}$, $M_{\sigma(u),\sigma(v)}\neq
*$. If $M_{\sigma(u),\sigma(v)}=0$, there is an edge $(u,v)\in E(G)$ by construction, which is forbidden in $M$-partitions; if $M_{\sigma(u),\sigma(v)}=1$, there is no edge $(u,v)\in E(G)$ but this edge is required in $M$-partitions. Hence, $\sigma$ is not an $M$-partition.
For any symmetric $M\in\{0,1,*\}^{D\times D}$ and any subset-closed, $M$-purifying set [$\mathcal L$]{}, the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is polynomial-time Turing-reducible to ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$.
We now essentially reverse the construction of the previous lemma to give a reduction from [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} to ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$. For any instance ($G,L)$ of [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}, we construct a corresponding instance $(V,C)$ of ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ as follows. The set of variables $V$ is $V(G)$. The set of constraints $C$ consists of a constraint ${\langle v,
L(v) \rangle}$ for each vertex $v\in V(G)$ and a constraint ${\langle (u,v),
H^M_{L(u),L(v)} \rangle}$ for every pair of distinct vertices $u$, $v$ such that:
- $(u,v)\in E(G)$ and $M|_{L(u)\times L(v)}$ has a 0 entry, or
- $(u,v)\not\in E(G)$ and $M|_{L(u)\times L(v)}$ has a 1 entry.
We show that a function $\sigma\colon V\to D$ is a satisfying assignment of $(V,C)$ if and only if it is an $M$-partition of $G$ that respects $L$. It is clear that $\sigma$ satisfies the unary constraints if and only if it respects $L$.
If $\sigma$ satisfies $(V,C)$ then consider any pair of distinct vertices $u,v\in V$. If there is a binary constraint involving $u$ and $v$, then $M_{\sigma(u),\sigma(v)} = M_{\sigma(v),\sigma(u)} = *$ so the existence or non-existence of the edge $(u,v)$ of $G$ does not affect whether $\sigma$ is an $M$-partition. If there is no binary constraint involving $u$ and $v$, then either there is an edge $(u,v)\in
E(G)$ and $M_{\sigma(u),\sigma(v)}\neq 0$ or there is no edge $(u,v)$ and $M_{\sigma(u),\sigma(v)}\neq 1$. In all three cases, $\sigma$ maps $u$ and $v$ consistently with it being an $M$-partition.
Conversely, if $\sigma$ does not satisfy $(V,C)$, either it fails to satisfy a unary constraint, in which case it does not respect $L$, or it satisfies all unary constraints (so it respects $L$), but it fails to satisfy a binary constraint ${\langle (u,v),H^M_{L(u),L(v)} \rangle}$. In the latter case, by construction, $M_{\sigma(u),\sigma(v)}\neq *$ so either $M_{\sigma(u),\sigma(v)}=0$ but there is an edge $(u,v)\in E(G)$, or $M_{\sigma(u),\sigma(v)}=1$ and there is no edge $(u,v)\in E(G)$. In either case, $\sigma$ is not an $M$-partition of $G$.
An arc-consistency based algorithm for ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ {#sec:arc}
=================================================================================================================
In the previous section, we showed that a class of [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} problems is equivalent to a certain class of counting CSPs, where the constraint language consists of binary relations and all unary relations over the domain $D$. We now investigate the complexity of such [$\mathrm{\#CSP}$]{}s.
Arc-consistency is a standard solution technique for constraint satisfaction problems [@CSPbook]. It is, essentially, a local search method which initially assumes that each variable may take any value in the domain and iteratively reduces the range of values that can be assigned to each variable, based on the constraints applied to it and the values that can be taken by other variables in the scopes of those constraints.
For any simple ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ instance $(V,C)$, define the vector of [*arc-consistent domains*]{} $(D_v)_{v\in V}$ by the procedure in Algorithm \[alg:ACComp\].
At no point in the execution of the algorithm can any domain $D_v$ increase in size so, for fixed $D$, the running time of the algorithm is at most a polynomial in $|V|+|C|$.
It is clear that, if $(D_v)_{v\in V}$ is the vector of arc-consistent domains for a simple ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ instance $(V,C)$, then every satisfying assignment $\sigma$ for that instance must have $\sigma(v)\in D_v$ for each variable $v$. In particular, if some $D_v=\emptyset$, then the instance is unsatisfiable. (Note, though, that the converse does not hold. If $D=\{0,1\}$ and $R=\{(0,1),(1,0)\}$, the instance with constraints ${\langle x,D \rangle}$, ${\langle y,D \rangle}$, ${\langle z,D \rangle}$, ${\langle (x,y),R \rangle}$, ${\langle (y,z),R \rangle}$ and ${\langle (z,x),R \rangle}$ is unsatisfiable but arc-consistency assigns $D_x = D_y = D_z =
\{0,1\}$.)
The arc-consistent domains computed for a simple instance $(V,C)$ can yield further simplification of the constraint structure, which we refer to as [*factoring*]{}. The factoring applies when the arc-consistent domains restrict a binary relation to a Cartesian product. In this case, the binary relation can be replaced with corresponding unary relations. Algorithm \[alg:factor\] factors a simple instance with respect to a vector $(D_v)_{v\in V}$ of arc-consistent domains, producing a set $F$ of factored constraints.
Recall that there is at most one constraint in $C$ between distinct variables and there are no binary constraints ${\langle (v,v), R \rangle}$ because the instance is simple. Note also that, if $|D_u|\leq 1$ or $|D_v|\leq 1$, then $R\cap (D_u\times D_v)$ is necessarily a Cartesian product. It is easy to see that the result of factoring a simple instance is simple, that Algorithm \[alg:factor\] runs in polynomial time and that the instance $(V,F)$ has the same satisfying assignments as $(V,C)$.
The *constraint graph* of a [$\mathrm{CSP}$]{} instance $(V,C)$ (in any constraint language) is the undirected graph with vertex set $V$ that contains an edge between every pair of distinct variables that appear together in the scope of some constraint.
\[line:choice\]
Algorithm \[alg:AC\] uses arc-consistency to count the satisfying assignments of simple ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ instances. It is straightforward to see that the algorithm terminates, since each recursive call is either on an instance with strictly fewer variables or on one in which at least one variable has had its unary constraint reduced to a singleton and no variable’s unary constraint has increased. For general inputs, the algorithm may take exponential time to run but, in Lemma \[lem:quickarc\] we show that the running time is polynomial for the inputs we are interested in.
We first argue that the algorithm is correct. By Lemma \[lemma:simple\], we may assume that the given instance $(V,C)$ is simple. Every satisfying assignment $\sigma\colon V\to D$ satisfies $\sigma(v)\in D_v$ for all $v\in V$ so restricting our attention to arc-consistent domains does not alter the output. Factoring the constraints also does not change the number of satisfying assignments: it merely replaces some binary constraints with equivalent unary ones. The constraints are factored, so any variable $v$ with $|D_v|=1$ must, in fact, be an isolated vertex in the constraint graph because, as noted above, any binary constraint involving it has been replaced by unary constraints. Therefore, if a component $H_i$ contains a variable $v$ with $|D_v|=1$, that component is the single vertex $v$, which is constrained to take a single value, so the number of satisfying assignments for this component, which we denote $Z_i$, is equal to $1$. (So we have now shown that the if branch in the for loop is correct.) For components that contain more than one variable, it is clear that we can choose one of those variables, $w_i$, and group the set of $M$-partitions $\sigma$ according to the value of $\sigma(w_i)$. (So we have now shown that the else branch is correct.) Because there are no constraints between variables in different components of the constraint graph, the number of satisfying assignments factorises as $\prod_{i=1}^\kappa Z_i$.
For a binary relation $R$, we write $$\begin{aligned}
\pi_1(R) &= \{a \mid (a,b)\in R \text{ for some }b\} \\
\pi_2(R) &= \{b \mid (a,b)\in R \text{ for some }a\}\,.\end{aligned}$$
For the following proof, we will also need the observation of Dyer and Richerby [@DRfull Lemma 1] that any rectangular relation $R\subseteq
\pi_1(R)\times \pi_2(R)$ can be written as $(A_1\times B_1) \cup \dots
\cup (A_\lambda \times B_\lambda)$, where the $A_i$ and $B_i$ partition $\pi_1(R)$ and $\pi_2(R)$, respectively. The subrelations $A_i\times B_i$ are referred to as [*blocks*]{}. A rectangular relation $R\neq \pi_1(R)\times \pi_2(R)$ must have at least two blocks.
\[lem:quickarc\] Suppose that ${\ensuremath{\mathcal L}}$ is subset-closed and $M$-purifying. If there is no [$\mathcal L$]{}-$M$-derectangularising sequence, then Algorithm \[alg:AC\] runs in polynomial time.
We will argue that the number of recursive calls made by the function AC in Algorithm \[alg:AC\] is bounded above by a polynomial in $|V|$. This suffices, since every other step of the procedure is obviously polynomial.
Consider a run of the algorithm on instance $(V,C)$ which, by Lemma \[lemma:simple\], we may assume to be simple. Suppose the run makes a recursive call with input $(V_i,F'_{i,d})$. For each $v\in V_i$, let $D'_{v}$ denote the arc-consistent domain for $v$ that is computed during the recursive call. We will show below that $D'_v\subset D_v$ for every variable $v\in V_i$. This implies that the recursion depth is at most $|D|$. As a crude bound, it follows that the number of recursive calls is at most ${(|V|\cdot |D|)}^{|D|},$ since each recursive call that is made is nested below a sequence of at most $|D|$ previous calls, each of which chose a vertex $v\in V$ and “pinned” it to a domain element $d\in D$ (i.e., introduced the constraint ${\langle v,\{d\} \rangle}$).
Towards showing that the domains of all variables decrease at each recursive call, suppose that we are computing $\mathrm{AC}(V,C)$ and the arc-consistent domains are $(D_v)_{v\in V}$. As observed above, for any component $H_i$ of the constraint graph on which a recursive call is made, we must have $|D_v|>1$ for every $v\in V_i$. Fix such a component and, for each $v\in V_i$, let $D'_v$ be the arc-consistent domain calculated for $v$ in the recursive call on $H_i$. It is clear that $D'_v\subseteq D_v$; we will show that $D'_v
\subset D_v$.
Consider a path $v_1\dots v_\ell$ in $H_i$, where $v_1=w_i$ and $v_\ell=v$. For each $j\in[\ell-1]$, there is exactly one binary constraint in $F_i$ involving $v_j$ and $v_{j+1}$. This is either ${\langle (v_j, v_{j+1}), R_j \rangle}$ or ${\langle (v_{j+1}, v_j), R_j^{-1} \rangle}$ and, without loss of generality, we may assume that it is the former. For $j\in[\ell-1]$, let $R'_j = R_j \cap (D_{v_j} \times D_{v_{j+1}}) =
H^M_{D_{v_j},D_{v_{j+1}}}$. The relation $R'_j$ is pure because $D_{v_j}$ and $D_{v_{j+1}}$ are in the subset-closed set ${\ensuremath{\mathcal L}}$ and, since ${\ensuremath{\mathcal L}}$ is $M$-purifying, so is $\{D_{v_j},D_{v_{j+1}}\}$. These two domains do not form a derectangularising sequence by the hypothesis of the lemma, so $H^M_{D_{v_j},D_{v_{j+1}}}$ is rectangular. If some $R_j=\emptyset$ then $D_{v_j} = D_{v_{j+1}} = \emptyset$ by arc-consistency, contradicting the fact that $|D_v|>1$ for all $v\in
V_i$. If some $R'_j$ has just one block, $R_j\cap (D_{v_j}\times
D_{v_{j+1}})$ is a Cartesian product, contradicting the fact that $F$ is a factored set of constraints. Thus, every $R'_j$ has at least two blocks.
For $j\in[\ell-1]$, let $\Phi_j = R'_1 \circ \dots \circ R'_j$. As above, note that $\{D_{v_1}, \ldots, D_{v_{j+1}}\}$ is $M$-purifying and the sequence $D_{v_1}, \dots, D_{v_{j+1}}$ is not derectangularising, so $\Phi_j$ is rectangular. We will show by induction on $j$ that $\pi_1(\Phi_j) = D_{v_1}$, $\pi_2(\Phi_j) = D_{v_{j+1}}$ and $\Phi_j$ has at least two blocks. Therefore, since the recursive call constrains $\sigma(w_i)$ to be $d$ and $d\in A$ for some block $A\times B\subset \Phi_\ell$, we have $D'_v\subseteq B\subset
D_v$, which is what we set out to prove.
For the base case of the induction, take $j=1$ so $\Phi_1=R'_1$. We showed above that $R'_1$ has at least two blocks and that $R'_1= H^M_{D_{v_1},D_{v_2}}$. By arc-consistency, $\pi_1(R'_1) = D_{v_1}$ and $\pi_2(R'_1) = D_{v_2}$.
For the inductive step, take $j\in [\ell-2]$. Suppose that $\pi_1(\Phi_j)=D_{v_1}$, $\pi_2(\Phi_j)=D_{v_{j+1}}$ and $\Phi_j = \bigcup_{s=1}^\lambda (A_s\times
A'_s)$ has at least two blocks. We have $\Phi_{j+1} = \Phi_j\circ R'_{j+1}$ and $R'_{j+1} = \bigcup_{t=1}^\mu (B_t\times B'_t)$ for some $\mu \geq 2$.
For every $d\in D_{v_1}$, there is a $d'\in D_{v_{j+1}}$ such that $(d,d')\in \Phi_j$ by the inductive hypothesis, and a $d''\in
D_{v_{j+1}}$ such that $(d'\!, d'')\in D_{v_{j+2}}$, by arc-consistency. Therefore, $\pi_1(\Phi_{j+1}) = D_{v_1}$; a similar argument shows that $\pi_2(\Phi_{j+1}) = D_{v_{j+2}}$.
Suppose, towards a contradiction, that $\Phi_{j+1} = D_{v_1}\times
D_{v_{j+2}}$. For this to be the case, we must have $A'_s\cap B_t\neq\emptyset$ for every $s\in\{1,2\}$ and $t\in[\mu]$. Now, let $D^*_{v_{j+1}}=D_{v_{j+1}}\setminus
(A'_2\cap B_2)$ and consider the relation $$R = \{(d_1, d_3) \mid
\mbox{for some $d_2\in D^*_{v_{j+1}} $,
$(d_1, d_2)\in \Phi_j$ and $(d_2, d_3)\in R'_{j+1}$
}\}.$$ Since $A'_1 \subseteq D^*_{v_{j+1}}$ the non-empty sets $A'_1 \cap B_1$ and $A'_1 \cap B_2$ are both subsets of $D^*_{v_{j+1}}$ so $A_1\times B'_1\subseteq R$ and $A_1\times B'_2\subseteq
R$. Similarly, $B_1 \subseteq D^*_{v_{j+1}}$, so $A'_2 \cap B_1 \subseteq D^*_{v_{j+1}}$ so $A_2\times B'_1\subseteq R$. However, $(A_2\times B'_2)\cap R =
\emptyset$, so $R$ is not rectangular. We will now derive a contradiction by showing that $R$ is rectangular. Note that $$R = H^M_{D_{v_1},D_{v_2}} \circ \cdots \circ H^M_{D_{v_{j-1}},D_{v_j}}
\circ
H^M_{D_{v_j},D^*_{v_{j+1}}}
\circ
H^M_{D^*_{v_{j+1}},D_{v_{j+2}}}$$ but this relation is rectangular because the hypothesis of the lemma guarantees that the sequence $$D_{v_1},\ldots,D_{v_{j}},D^*_{v_{j+1}},D_{v_{j+2}}$$ is not an ${\ensuremath{\mathcal L}}$-$M$-derectangularising sequence and all of the elements of this sequence are in ${\ensuremath{\mathcal L}}$, and $\{D_{v_1},\ldots,D_{v_{j}},D^*_{v_{j+1}},D_{v_{j+2}}\}$ is $M$-purifying.
Polynomial-time algorithms and the dichotomy theorem {#sec:dichotomy}
====================================================
Bulatov [@Bul08] showed that every problem of the form ${\ensuremath{\mathrm{\#CSP}}}(\Gamma)$ is either in ${\ensuremath{\mathrm{FP}}}$ or $\nP$-complete. Together with Proposition \[prop:purifiediscsp\], his result immediately shows that a similar dichotomy exists for the special case of the problem [[\#$\mathcal L$-$M$-partitions]{}]{} in which $\mathcal L$ is $M$-purifying and is closed under subsets. Our algorithmic work in Section \[sec:arc\] can be combined with Dyer and Richerby’s explicit dichotomy for ${\ensuremath{\mathrm{\#CSP}}}$ to obtain an explicit dichotomy for this special case of [[\#$\mathcal L$-$M$-partitions]{}]{}. In particular, Lemma \[lem:quickarc\] gives a polynomial-time algorithm for the case in which there is no [$\mathcal L$]{}-$M$-derectangularising sequence. When there is such a sequence, ${\Gamma_{\!{\ensuremath{\mathcal L}}, M}}$ is not “strongly rectangular” in the sense of [@DRfull]. It follows immediately that ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ is [$\mathrm{\#P}$]{}-complete [@DRfull Lemma 24] so [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is also [$\mathrm{\#P}$]{}-complete by Proposition \[prop:purifiediscsp\]. In fact, the dichotomy for this special case does not require the full generality of Dyer and Richerby’s dichotomy. If there is an [$\mathcal L$]{}-$M$-derectangularising sequence then it follows immediately from work of Bulatov and Dalmau [@BD Theorem 2 and Corollary 3] that ${\ensuremath{\mathrm{\#CSP}}}({\Gamma_{\!{\ensuremath{\mathcal L}}, M}})$ is [$\mathrm{\#P}$]{}-complete.
In this section we will move beyond the case in which ${\ensuremath{\mathcal L}}$ is $M$-purifying to provide a full dichotomy for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}. We will use two data structures: *sparse-dense partitions* and a representation of the set of *splits* of a bipartite graph. Similar data structures were used by Hell et al. [@HHN] in their dichotomy for the [[\#$M$-partitions]{}]{} problem for matrices of size at most $3$-by-$3$.
Data Structures {#sec:DS}
---------------
We use two types of graph partition. The first is a special case of a sparse-dense partition [@FHKM] which is also called an $(a,b)$-graph with $a=b=2$.
\[def:bsd\] A bipartite–cobipartite partition of a graph $G$ is a partition $(B,C)$ of $V(G)$ such that $B$ induces a bipartite graph and $C$ induces the complement of a bipartite graph.
\[lem:sparsedense\][@FHKM Theorem 3.1; see also the remarks on $(a,b)$-graphs.] There is a polynomial-time algorithm for finding all bipartite–cobipartite partitions of a graph $G$.
The second decomposition is based on certain sub-hypercubes called subcubes. For any finite set $U\!$, a [*subcube*]{} of $\{0,1\}^U$ is a subset of $\{0,1\}^U$ that is a Cartesian product of the form $\prod_{u\in U} S_u$ where $S_u\in\{\{0\},\{1\},\{0,1\}\}$ for each $u\in U\!$. We can also associate a subcube $\prod_{u\in U} S_u$ with the set of assignments $\sigma\colon U\to \{0,1\}$ such that $\sigma(u)\in S_u$ for all $u\in
U\!$. Subcubes can be represented efficiently by listing the projections $S_u$.
Let $G=(U,U'\!,E)$ be a bipartite graph, where $U$ and $U'$ are disjoint vertex sets, and $E\subseteq U\times U'\!$. A [*subcube decomposition*]{} of $G$ is a list $U_1,\dots,U_k$ of subcubes of $\{0,1\}^U$ and a list $U'_1,\dots, U'_k$ of subcubes of $\{0,1\}^{U'}$ such that the following hold.
- The union $(U_1\times U'_1)\cup \dots \cup (U_k\times U'_k)$ is the set of assignments $\sigma\colon
U\cup U'\to\{0,1\}$ such that: $$\begin{aligned}
& \label{todayone} \mbox{no edge $(u,u')\in E$ has $\sigma(u)=\sigma(u')=0$ and}\\
& \label{todaytwo} \mbox{no pair $(u,u')\in (U\times U')\setminus E$ has $\sigma(u)=
\sigma(u')=1$.}
\end{aligned}$$
- For distinct $i,j\in[k]$, $U_i\times U'_i$ and $U_j\times U'_j$ are disjoint.
- For each $i\in[k]$, either $|U_i|=1$ or $|U'_i|=1$ (or both).
Note that, although we require $U_i\times U'_i$ and $U_j\times U'_j$ to be disjoint for distinct $i,j\in[k]$, we allow $U_i\cap
U_j\neq\emptyset$ as long as $U'_i$ and $U'_j$ are disjoint, and vice-versa. It is even possible that $U_i=U_j$, and indeed this will happen in our constructions below.
\[lem:splittocubes\] A subcube decomposition of a bipartite graph $G=(U,U'\!,E)$ can be computed in polynomial time, with the subcubes represented by their projections.
For a vertex $x$ in a bipartite graph, let $\Gamma(x)$ be its set of neighbours and let ${\overline{\Gamma}}(x)$ be its set of non-neighbours on the other side of the graph. Thus, for $x\in U\!$, ${\overline{\Gamma}}(x) =
U'\setminus \Gamma(x)$ and, for $x\in U'\!$, ${\overline{\Gamma}}(x) =
U\setminus \Gamma(x)$.
Observe that we can write $\{0,1\}^n\setminus
\{0\}^n$ as the disjoint union of $n$ subcubes $\{0\}^{k-1}\times
\{1\}^1\times \{0,1\}^{n-k}$ with $1\leq k\leq n$, and similarly for any other cube minus a single point.
We first deal with two base cases. If $G$ has no edges, then the set of assignments $\sigma\colon U\cup U'\to\{0,1\}$ satisfying (\[todayone\]) and (\[todaytwo\]) is the disjoint union of $$\{0\}^U\times \{0\}^{U'}, \quad
(\{0,1\}^U\setminus\{0\}^U)\times \{0\}^{U'}, \quad \text{and} \quad
\{0\}^U\times(\{0,1\}^{U'}\setminus\{0\}^{U'}).$$ The second and third terms can be decomposed into subcubes as described above to produce the output. Similarly, if $G$ is is a complete bipartite graph, then the set of assignments satisfying (\[todayone\]) and (\[todaytwo\]) is the disjoint union of $$\{1\}^U\times \{1\}^{U'}, \quad
(\{0,1\}^U\setminus\{1\}^U)\times \{1\}^{U'}, \quad \text{and} \quad
\{1\}^U\times(\{0,1\}^{U'}\setminus\{1\}^{U'}).$$
If neither of these cases occurs then there is a vertex $x$ such that neither $\Gamma(x)$ nor ${\overline{\Gamma}}(x)$ is empty. If possible, choose $x\in U$; otherwise, choose $x\in U'\!$. To simplify the description of the algorithm, we assume that $x\in U$; the other case is symmetric. We consider separately the assignments where $\sigma(x)=0$ and those where $\sigma(x)=1$. Note that, for any assignment, if $\sigma(y)=0$ for some vertex $y$, then $\sigma(z)=1$ for all $z\in\Gamma(y)$ and, if $\sigma(y)=1$, then $\sigma(z)=0$ for all $z\in{\overline{\Gamma}}(y)$. Applying this iteratively, setting $\sigma(x)=c$ for $c\in\{0,1\}$ also determines the value of $\sigma$ on some set $S_{x=c}\subseteq
U\cup U'$ of vertices.
Thus, we can compute a subcube decomposition for $G$ recursively. First, compute $S_{x=0}$ and $S_{x=1}$. Then, recursively compute subcube decompositions of $G-S_{x=0}$ (the graph formed from $G$ by deleting the vertices in $S_{x=0}$) and $G-S_{x=1}$. Translate these subcube decompositions into a subcube decomposition of $G$ by extending each subcube $(U_i\times U'_i)$ of $G-S_{x=c}$ to a subcube $(V_i\times V'_i)$ of $G$ whose restriction to $G-S_{x=c}$ is $(U_i\times U'_i)$ and whose restriction to $S_{x=c}$ is an assignment $\sigma$ with $\sigma(x)=c$ (in fact, all assignments that set $x$ to $c$ agree on the set $S_{x=c}$, by construction).
It remains to show that the algorithm runs in polynomial time. The base cases are clearly computable in polynomial time, as are the individual steps in the recursive cases, so we only need to show that the number of recursive calls is polynomially bounded. At the recursive step, we only choose $x\in U'$ when $E(G) = U''\times U'$ for some proper subset $\emptyset\subset U''\subset U$ and, in this case, the two recursive calls are to base cases. Since each recursive call when $x\in U$ splits $U'$ into disjoint subsets, there can be at most $|U'|-1$ such recursive calls, so the total number of recursive calls is linear in $|V(G)|$.
Reduction to a problem with $M$-purifying lists
-----------------------------------------------
Our algorithm for counting list $M$-partitions uses the data structures from Section \[sec:DS\] to reduce problems where [$\mathcal L$]{} is not $M$-purifying to problems where it is (which we already know how to solve from Sections \[sec:purifiedcsp\] and \[sec:arc\]). The algorithm is defined recursively on the set ${\ensuremath{\mathcal L}}$ of allowed lists. The algorithm for parameters ${\ensuremath{\mathcal L}}{}$ and $M$ calls the algorithm for ${\ensuremath{\mathcal L}}_i$ and $M$ where ${\ensuremath{\mathcal L}}_i$ is a subset of ${\ensuremath{\mathcal L}}$. The base case arises when ${\ensuremath{\mathcal L}}_i$ is $M$-purifying.
We will use the following computational problem to reduce [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} to a collection of problems [[\#${\ensuremath{\mathcal L}}'$-$M$-partitions]{}]{} that are, in a sense, disjoint.
[[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{}.
A graph $G$ and a function $L\colon V(G)\to{\ensuremath{\mathcal L}}$.
Functions $L_1,\dots,L_t\colon V(G)\to{\ensuremath{\mathcal L}}$ such that
- for each $i\in[t]$, the set $\{L_i(v) \mid v\in V(G)\}$ is $M$-purifying,
- for each $i\in [t]$ and $v \in V(G)$, $L_i(v) \subseteq L(v)$, and
- each $M$-partition of $G$ that respects $L$ respects exactly one of $L_1,\dots,L_t$.
We will give an algorithm for solving the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} in polynomial time when there is no [$\mathcal L$]{}-$M$-derectangularising sequence of length exactly 2. The following computational problem will be central to the inductive step.
[[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{}.
A graph $G$ and a function $L\colon V(G)\to {\ensuremath{\mathcal L}}$.
Functions $L_1,\dots, L_k\colon
V(G)\to {\ensuremath{\mathcal L}}$ such that
- for each $i\in [k]$ and $v \in V(G)$, $L_i(v) \subseteq L(v)$,
- every $M$-partition of $G$ that respects $L$ respects exactly one of $L_1,\dots,L_k$, and
- for each $i\in[k]$, there is a $W\in{\ensuremath{\mathcal L}}{}$ which is inclusion-maximal in [$\mathcal L$]{} but does not occur in the image of $L_i$.
Note that we can trivially produce a solution to the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{} by letting $L_1, \dots, L_k$ be an enumeration of all possible functions such that all lists $L_i(v)$ have size $1$ and satisfy $L_i(v) \subseteq L(v)$. Such a function $L_i$ corresponds to an assignment of vertices to parts so there is either exactly one $L_i$-respecting $M$-partition or none, which means that every $L$-respecting $M$-partition is $L_i$-respecting for exactly one $i$. However, this solution is exponentially large in $|V(G)|$ and we are interested in solutions that can be produced in polynomial time. Also, if $L(v)=\emptyset$ for some vertex $v$, the algorithm is entitled to output an empty list, since no $M$-partition respects $L$.
The following definition extends rectangularity to $\{0,1,*\}$-matrices and is used in our proof.
A matrix $M\in\{0,1,*\}^{X\times Y}$ is [*$*$-rectangular*]{} if the relation $H^M_{X,Y}$ is rectangular.
Thus, $M$ is $*$-rectangular if and only if $M_{x,y}=M_{x'\!,y}=M_{x,y'}=*$ implies that $M_{x'\!,y'}=*$ for all $x,x'\in X'$ and all $y,y'\in Y''\!$.
We will show in Lemma \[lem:claim\] that the function [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{} from Algorithm \[alg:purifystep\] is a polynomial-time algorithm for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{} whenever [$\mathcal L$]{} is not $M$-purifying and there is no length-2 [$\mathcal L$]{}-$M$-derectangularising sequence. Note that a length-2 [$\mathcal L$]{}-$M$-derectangularising sequence is a pair $X,Y\in{\ensuremath{\mathcal L}}$ such that $M|_{X\times Y}$, $M|_{X\times X}$ and $M|_{Y\times Y}$ are pure and $M|_{X\times Y}$ is not $*$-rectangular. If ${\ensuremath{\mathcal L}}\neq
{{\mathcal{P}(D)}}$, it is possible that a matrix that is not $*$-rectangular has no length-2 [$\mathcal L$]{}-$M$-derectangularising sequence. For example, let $D=\{1,2,3\}$ and ${\ensuremath{\mathcal L}}=
{{\mathcal{P}(\{1,2\})}}$ and let $M_{3,3}=0$ and $M_{i,j}=*$ for every other pair $(i,j)\in D^2\!$. $M$ is not $*$-rectangular but this fact is not witnessed by any submatrix $M|_{X\times Y}$ for $X,Y\in{\ensuremath{\mathcal L}}$.
/\* $v_j\in C_i$\*/
$L_i(v_j) \gets X_0$ $L_i(v_j) \gets X_1$
$L_i(v_j) \gets Y_0$ $L_i(v_j) \gets Y_1$
\[lem:claim\] Let $M$ be a symmetric matrix in $\{0,1,*\}^{D\times D}$ and let ${\ensuremath{\mathcal L}}{}\subseteq{{\mathcal{P}(D)}}$ be subset-closed. If [$\mathcal L$]{} is not $M$-purifying and there is no length-2 [$\mathcal L$]{}-$M$-derectangularising sequence, then Algorithm \[alg:purifystep\] is a polynomial-time algorithm for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{}.
We consider an instance $(G,L)$ of the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{} with $V(G)=\{v_1,\ldots,v_n\}$. If there is a $v_i\in V(G)$ with $L(v_i)=\emptyset$ then no $M$-partition of $G$ respects $L$, so the output is correct. Otherwise, we consider the three cases that can occur in the execution of the algorithm.
#### Case 1.
In this case column $d$ of $M|_{X\times Y}$ contains both a zero and a one. Equivalently, row $d$ of $M|_{Y\times X}$ does. Algorithm \[alg:Case1\] groups the set of $M$-partitions of $G$ that respect $L$, based on the first vertex that is placed in part $d$. For $i\in[n]$, $L_i$ requires that $v_i$ is placed in part $d$ and $v_1, \dots, v_{i-1}$ are not in part $d$; $L_{n+1}$ requires that part $d$ is empty. Thus, no $M$-partition can respect more than one of the $L_i$. Now consider an $L$-respecting $M$-partition $\sigma\colon
V(G)\to D$ and suppose that $i$ is minimal such that $\sigma(v_i)=d$. We claim that $\sigma$ respects $L_i$. We have $\sigma(v_i)=d$, as required. For $j\neq i$, we must have $\sigma(v_j)\in L(v_j)$ since $\sigma$ respects $L$ and we must have $M_{d,\sigma(v_j)}\neq 1$ if $(v_i, v_j)\notin E(G)$ and $M_{d,\sigma(v_j)}\neq 0$ if $(v_i,v_j)\in
E(G)$, since $\sigma$ is an $M$-partition. In addition, by construction, $\sigma(v_j)\neq d$ if $j<i$. Therefore, $\sigma$ respects $L_i$. A similar argument shows that $\sigma$ respects $L_{n+1}$ if $\sigma(v)\neq d$ for all $v\in V(G)$. Hence, any $M$-partition that respects $L$ respects exactly one of the $L_i$.
Finally, we show that, for each $i\in[n+1]$, there is a set $W$ which is inclusion-maximal in [$\mathcal L$]{} and is not in the image of $L_i$. For $i\in [n]$, we cannot have both $a$ and $b$ in $L_i(v_j)$ for any $v_j$, so $X$ is not in the image of $L_i$. $Y$ contains $d$, so $Y$ is not in the image of $L_{n+1}$.
#### Case 2.
In this case, every row of $M|_{X_0\times X}$ contains a 0, while every row of $M|_{X_1\times X}$ fails to contain a zero. Since $M|_{X\times X}$ is not pure, but no row of $M|_{X\times X}$ contains both a zero and a one (since we are not in Case 1), $X_0$ and $X_1$ are non-empty. Note that $M|_{X_0\times X_0}$ and $M|_{X_1\times X_1}$ are both pure, while every entry of $M|_{X_0\times X_1}$ is a $*$.
If $V_X=\emptyset$ then $X$ is an inclusion-maximal member of [$\mathcal L$]{} that is not in the image of $L$, so the output of Algorithm \[alg:Case2\] is correct. Otherwise, $(B_1,C_1),\dots,(B_k,C_k)$ is the list containing all partitions $(B,C)$ of $V_X$ such that $B$ induces a bipartite graph in $G$ and $C$ induces the complement of a bipartite graph. The algorithm returns $L_1,\ldots, L_k$. $X$ is not in the image of any $L_i$ so, to show that $\{L_1, \dots, L_k\}$ is a correct output for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{}, we just need to show that every $M$-partition of $G$ that respects $L$ respects exactly one of $L_1,\dots,L_k$. For $i\neq i'$, $(B_i,C_i)\neq (B_{i'},C_{i'})$ so there is at least one vertex $v_j$ such that $L_i(v_j)=X_0$ and $L_{i'}(v_j)=X_1$ or vice-versa. Since $X_0$ and $X_1$ are disjoint, no $M$-partition can simultaneously respect $L_i$ and $L_{i'}$. It remains to show that every $M$-partition respects at least one of $L_1, \dots, L_k$. To do this, we deduce two structural properties of $M|_{X\times X}$.
First, we show that $M|_{X\times X}$ has no $*$ on its diagonal. Suppose towards a contradiction that $M_{d,d}=*$ for some $d\in X$. If $d\in X_0$, then, for each $d'\in X_1$, $M_{d,d'}=M_{d'\!,d}=*$ because, as noted above, every entry of $M|_{X_0\times X_1}$ is a $*$. Therefore, the $2\times 2$ matrix $M'=M|_{\{d,d'\}\times \{d,d'\}}$ contains at least three $*$s so it is pure. $\{d,d'\}
\subseteq
X\in
{\ensuremath{\mathcal L}}$ so, by the hypothesis of the lemma, the length-2 sequence $\{d,d'\},\{d,d'\}$ is not [$\mathcal L$]{}-$M$-derectangularising, so $M'$ must be $*$-rectangular, so $M_{d'\!,d'}=*$ for all $d'\in X_1$. Similarly, if $M_{d'\!,d'}=*$ for some $d'\in X_1$, then $M_{d,d}=*$ for all $d\in X_0$. Therefore, if $M|_{X\times X}$ has a $*$ on its diagonal, every entry on the diagonal is $*$. But $M$ contains a 0, say $M_{i,j}=0$ with $i,j\in X_0$. For any $k\in X_1$, $$M|_{\{i,j\}\times \{j,k\}} = \begin{pmatrix} 0 & * \\ * & * \end{pmatrix},$$ so the length-2 sequence $\{i,j\}, \{j,k\}$ is [$\mathcal L$]{}-$M$-derectangularising, contradicting the hypothesis of the lemma (note that $\{i,j\},\{j,k\}\subseteq X\in{\ensuremath{\mathcal L}}$).
Second, we show that there is no sequence $d_1,\dots,d_\ell\in X_0$ of odd length such that $$M_{d_1,d_2}=M_{d_2,d_3}=\dots=M_{d_{\ell-1},d_\ell}=M_{d_\ell,d_1}=*\,.$$ Suppose for a contradiction that such a sequence exists. Note that $M|_{X_0\times X_0}$ is $*$-rectangular since $X_0,X_0$ is not an [$\mathcal L$]{}-$M$-derectangularising sequence and $M|_{X_0\times X_0}$ is pure since Case 1 does not apply. We will show by induction that for every non-negative integer $\kappa \leq (\ell-3)/2$, $M_{d_1,d_{\ell-2\kappa-2}}=*$. This gives a contradiction by taking $\kappa=(\ell-3)/2$ since $M_{d_1,d_1}=*$ and we have already shown that $M|_{X_0\times X_0}$ has no $*$ on its diagonal. For every $\kappa$, the argument follows by considering the matrix $M_\kappa = M|_{\{d_1,d_{\ell-2\kappa-1}\} \times \{d_{\ell-2\kappa-2},d_{\ell-2\kappa}\}}$. The definition of the sequence $d_1,\ldots,d_\ell$ together with the symmetry of $M$ guarantees that both entries in row $d_{\ell-2\kappa-1}$ of $M_\kappa$ are equal to $*$. It is also true that $M_{d_1,d_{\ell-2\kappa}}=*$: If $\kappa=0$ then this follows from the definition of the sequence; otherwise it follows by induction. The fact that $M_{d_1,d_{\ell-2\kappa-2}}=*$ then follows by $*$-rectangularity.
This second structural property implies that, for any $M|_{X\times X}$-partition of $G[V_X]$, the graph induced by vertices assigned to $X_0$ has no odd cycles, and is therefore bipartite. Similarly, the vertices assigned to $X_1$ induce the complement of a bipartite graph. Therefore, any $M$-partition of $G$ that respects $L$ must respect at least one of the $L_1, \dots, L_k$, so it respects exactly one of them, as required.
#### Case 3.
Since Cases 1 and 2 do not apply and [$\mathcal L$]{} is not $M$-purifying, there are distinct $X,Y\in
{\ensuremath{\mathcal L}}$ such that $X$ and $Y$ are inclusion-maximal in [$\mathcal L$]{} and $M|_{X\times Y}$ is not pure. As in the previous case, the sets $X_0$, $X_1$, $Y_0$ and $Y_1$ are all non-empty.
If either $V_X$ or $V_Y$ is empty then either $X$ or $Y$ is an inclusion-maximal set in [$\mathcal L$]{} that is not in the image of $L$ so the output of Algorithm \[alg:Case3\] is correct. Otherwise, $(U_1,U'_1),\dots,(U_k,U'_k)$ is a subcube decomposition of the bipartite subgraph $(V_X,V_Y,E)$. The $U_i$s are subcubes of $\{0,1\}^{V_X}$ and the $U'_i$s are subcubes of $\{0,1\}^{V_Y}$. The algorithm returns $L_1,\ldots,L_k$.
Note that if $|U'_i|=1$ then $Y$ is not in the image of $L_i$. Similarly, if $|U'_i|>1$ but $|U_i|=1$ then $X$ is not in the image of $L_i$. The definition of subcube decompositions guarantees that, for every $i$, at least one of these is the case. To show this definition of $L_1,\ldots,L_k$ is a correct output for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{}, we must show that any $M$-partition of $G$ that respects $L$ also respects exactly one $L_i$. Since the sets in $\{U_i \times U'_i \mid i\in[k]\}$ are disjoint subsets of $\{0,1\}^{V_X\cup V_Y}$, any $M$-partition of $G$ that respects $L$ respects at most one $L_i$ so it remains to show that every $M$-partition of $G$ respects at least one $L_i$. To do this, we deduce two structural properties of $M|_{X\times Y}$.
First, we show that every entry of $M|_{X_0\times Y_0}$ is $0$. The definition of $X_0$ guarantees that every row of $M|_{X_0\times Y_0}$ contains a $0$. Since Case 1 does not apply, and $M$ is symmetric, every entry of $M|_{X_0\times Y_0}$ is either $0$ or $*$. Suppose for a contradiction that $M_{i,j}=*$ for some $(i,j)\in
X_0\times Y_0$. Pick $i'\in X_1$. For any $j'\in Y_0\setminus\{j\}$ we have $M_{i,j}=M_{i'\!,j}=M_{i'\!,j'}=*$, so by $*$-rectangularity of $M|_{X\times Y_0}$ we have $M_{i,j'}=*$. Thus, every entry of $M|_{\{i\}\times Y_0}$ is $*$, so there is a $*$ in every $Y_0$-indexed column of $M$. By the same argument, swapping the roles of $X$ and $Y$, every entry in $M|_{X_0\times Y_0}$ is $*$, contradicting the fact that $M|_{X\times Y}$ contains a $0$ since $M|_{X\times Y}$ is not pure.
Second, a similar argument shows that every entry of $M|_{X_1\times Y_1}$ is $1$.
Thus for all $M$-partitions $\sigma$ of $G$ respecting $L$, for all $x\in V_X$ and $y\in V_Y$, if $(x,y)\in E$ then $(\sigma(x),\sigma(y))\notin X_0\times Y_0$ while if $(x,y)\notin E$ then $(\sigma(x),\sigma(y))\notin X_1\times Y_1$. Using the definition of subcube decompositions, this shows that any $M$-partition of $G$ respecting $L$ respects some $L_i$.
We can now give an algorithm for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{}. The algorithm consists of the function [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{}, which is defined in Algorithm \[alg:purifytriv\] for the trivial case in which ${\ensuremath{\mathcal L}}$ is $M$-purifying and in Algorithm \[alg:purify\] for the case in which it is not. Note that for any fixed ${\ensuremath{\mathcal L}}$ and $M$ the algorithm is defined either in Algorithm \[alg:purifytriv\] or in Algorithm \[alg:purify\] and the function [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} is not recursive. However, the *definition* is recursive, so the function [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} defined in Algorithm \[alg:purify\] does make a call to a function [[\#${\ensuremath{\mathcal L}}_i$-$M$-purify]{}]{} for some ${\ensuremath{\mathcal L}}_i$ which is smaller than ${\ensuremath{\mathcal L}}$. The function [[\#${\ensuremath{\mathcal L}}_i$-$M$-purify]{}]{} is in turn defined in Algorithm \[alg:purifytriv\] or Algorithm \[alg:purify\].
The correctness of the algorithm follows from the definition of the problem. The following lemma bounds the running time.
\[lem:reducetopure\] Let $M\in\{0,1,*\}^{D\times D}$ be a symmetric matrix and let ${\ensuremath{\mathcal L}}\subseteq{{\mathcal{P}(D)}}$ be subset-closed. If there is no length-$2$ [$\mathcal L$]{}-$M$-derectangularising sequence, then the function [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} as defined in Algorithms \[alg:purifytriv\] and \[alg:purify\] is a polynomial-time algorithm for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{}.
Note that ${\ensuremath{\mathcal L}}$ is a fixed parameter of the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} — it is not part of the input. The proof is by induction on $|{\ensuremath{\mathcal L}}|$. If $|{\ensuremath{\mathcal L}}|=1$ then ${\ensuremath{\mathcal L}}=\{\emptyset\}$ so it is $M$-purifying. In this case, function [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} is defined in Algorithm \[alg:purifytriv\]. It is clear that it is a polynomial-time algorithm for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{}.
For the inductive step suppose that $|{\ensuremath{\mathcal L}}|>1$. If ${\ensuremath{\mathcal L}}$ is $M$-purifying then function [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} is defined in Algorithm \[alg:purifytriv\] and again the result is trivial. Otherwise, function [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} is defined in Algorithm \[alg:purify\]. Note that ${\ensuremath{\mathcal L}}\subseteq{{\mathcal{P}(D)}}$ is subset-closed and there is no length-$2$ [$\mathcal L$]{}-$M$-derectangularising sequence. From this, we can conclude that, for any subset-closed subset ${\ensuremath{\mathcal L}}'$ of ${\ensuremath{\mathcal L}}$, there is no length-$2$ ${\ensuremath{\mathcal L}}'$-$M$-derectangularising sequence. So we can assume by the inductive hypothesis that for all subset-closed ${\ensuremath{\mathcal L}}'\subset {\ensuremath{\mathcal L}}{}$, the function [[\#${\ensuremath{\mathcal L}}'$-$M$-purify]{}]{} runs in polynomial time.
The result now follows from the fact that the function [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{} runs in polynomial time (as guaranteed by Lemma \[lem:claim\]) and from the fact that each ${\ensuremath{\mathcal L}}_i$ is a strict subset of ${\ensuremath{\mathcal L}}$, which follows from the definition of problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify-step]{}]{}. Each $M$-partition that respects $L$ respects exactly one of $L_1,\dots,L_k$ and, hence, it respects exactly one of the list functions that is returned.
Algorithm for [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} and proof of the dichotomy
---------------------------------------------------------------------------------------------
We can now present our algorithm for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}. The algorithm consists of the function [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} which is defined in Algorithm \[alg:mainpurifying\] for the case in which ${\ensuremath{\mathcal L}}$ is $M$-purifying and in Algorithm \[alg:main\] when it is not.
AC$(V,C)$ where AC is the function from Algorithm \[alg:AC\]
\[lem:positive\] Let $M\in\{0,1,*\}^{D\times D}$ be a symmetric matrix and let ${\ensuremath{\mathcal L}}\subseteq{{\mathcal{P}(D)}}$ be subset-closed. If there is no [$\mathcal L$]{}-$M$-derectangularising sequence, then the function [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} as defined in Algorithms \[alg:mainpurifying\] and \[alg:main\] is a polynomial-time algorithm for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}.
If ${\ensuremath{\mathcal L}}$ is $M$-purifying then the function [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is defined in Algorithm \[alg:mainpurifying\]. Proposition \[prop:purifiediscsp\] shows that the reduction in Algorithm \[alg:mainpurifying\] to a CSP instance is correct and takes polynomial time. The CSP instance can be solved by the function AC in Algorithm \[alg:AC\], whose running time is shown to be polynomial in Lemma \[lem:quickarc\].
If ${\ensuremath{\mathcal L}}$ is not $M$-purifying then the function [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is defined in Algorithm \[alg:main\]. Lemma \[lem:reducetopure\] guarantees that the function [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} is a polynomial-time algorithm for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{}. If the list $L_1,\ldots,L_t$ is empty then there is no $M$-partition of $G$ that respects $L$ so it is correct that the function [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} returns $0$. Otherwise, we know from the definition of the problem [[\#${\ensuremath{\mathcal L}}$-$M$-purify]{}]{} that
- functions $L_1,\ldots,L_t$ are from $V(G)$ to ${\ensuremath{\mathcal L}}$,
- for each $i\in [t]$, the set $\{L_i(v) \mid v\in V(G)\}$ is $M$-purifying,
- for each $i\in [t]$ and $v \in V(G)$, $L_i(v) \subseteq L(v)$, and
- each $M$-partition of $G$ that respects $L$ respects exactly one of $L_1,\dots,L_t$.
The desired result is now the sum, over all $i\in[t]$, of the number of $M$-partitions of $G$ that respect $L_i$. Since the list $L_1, \dots, L_t$ is generated in polynomial time, $t$ is bounded by some polynomial in $|V(G)|$.
Now, for each $i\in[t]$, ${\ensuremath{\mathcal L}}_i$ is a subset-closed subset of ${\ensuremath{\mathcal L}}$. Since there is no ${\ensuremath{\mathcal L}}$-$M$-derectangularising sequence, there is also no ${\ensuremath{\mathcal L}}_i$-$M$-derectangularising sequence. Also, ${\ensuremath{\mathcal L}}_i$ is $M$-purifying. Thus, the argument that we gave for the purifying case shows that $Z_i$ is the desired quantity.
We can now combine our results to establish our dichotomy for the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{}.
[thm:explicitdichotomy]{}[ Let $M$ be a symmetric matrix in $\{0,1,*\}^{D\times D}$ and let ${\ensuremath{\mathcal L}}{}\subseteq{{\mathcal{P}(D)}}$ be subset-closed. If there is an [$\mathcal L$]{}-$M$-derectangularising sequence then the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is ${\ensuremath{\mathrm{\#P}}}$-complete. Otherwise, it is in ${\ensuremath{\mathrm{FP}}}$. ]{}
Suppose that there is an [$\mathcal L$]{}-$M$-derectangularising sequence $D_1,\dots,D_k$. Recall (from Definition \[def:closure\]) the definition of the subset-closure ${{\mathscr{S}({\ensuremath{\mathcal L}}'')}}$ of a set ${\ensuremath{\mathcal L}}'' \subseteq {{\mathcal{P}(D)}}$. Let $${\ensuremath{\mathcal L}}'= {{\mathscr{S}(\{D_1,\ldots,D_k\} )}}.$$ Since $\{D_1,\ldots,D_k\}$ is $M$-purifying, so is ${\ensuremath{\mathcal L}}'\!$, which is also subset-closed. It follows that $\Gamma_{\!{\ensuremath{\mathcal L}}'\!,M}$ is well defined (see Definition \[defgammaprime\]) and contains the relations $H_{D_1,D_2}^M, \ldots,H_{D_{k-1},D_k}^M$ (and possibly others). Since $H_{D_1,D_2}^M \circ H_{D_2,D_3}^M \circ \cdots \circ H_{D_{k-1},D_k}^M$ is not rectangular, ${\ensuremath{\mathrm{\#CSP}}}(\Gamma_{\!{\ensuremath{\mathcal L}}'\!,M})$ is ${\ensuremath{\mathrm{\#P}}}$-complete [@BD Theorem 2 and Corollary 3] (see also [@DRfull Lemma 24]). By Proposition \[prop:purifiediscsp\], the problem [[\#${\ensuremath{\mathcal L}}'$-$M$-partitions]{}]{} is ${\ensuremath{\mathrm{\#P}}}$-complete so the more general problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is also ${\ensuremath{\mathrm{\#P}}}$-complete. On the other hand, if there is no ${\ensuremath{\mathcal L}}$-$M$-derectangularising sequence, then the result follows from Lemma \[lem:positive\].
Complexity of the dichotomy criterion {#sec:meta}
=====================================
The dichotomy established in Theorem \[thm:explicitdichotomy\] is that, if there is an [$\mathcal L$]{}-$M$-derectangularising sequence, then the problem [[\#${\ensuremath{\mathcal L}}$-$M$-partitions]{}]{} is ${\ensuremath{\mathrm{\#P}}}$-complete; otherwise, it is in ${\ensuremath{\mathrm{FP}}}$. This section addresses the computational problem of determining which is the case, given [$\mathcal L$]{} and $M$.
The following lemma will allow us to show that the problem [[ExistsDerectSeq]{}]{} (the problem of determining whether there is an ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$-$M$-derectangularising sequence, given [$\mathcal L$]{} and $M$) and the related problem [[MatrixHasDerectSeq]{}]{} (the problem of determining whether there is a ${{\mathcal{P}(D)}}$-$M$-derectangularising sequence, given $M$) are both in [$\mathrm{NP}$]{}. Note that, for this “meta-problem”, [$\mathcal L$]{} and $M$ are the inputs whereas, previously, we have regarded them as fixed parameters.
\[lem:small\_derect\] Let $M\in\{0,1,*\}^{D\times D}$ be symmetric, and let ${\ensuremath{\mathcal L}}\subseteq {{\mathcal{P}(D)}}$ be subset-closed. If there is an [$\mathcal L$]{}-$M$-derectangularising sequence, then there is one of length at most $512(|D|^3+1)$.
Pick an [$\mathcal L$]{}-$M$-derectangularising sequence $D_1,\dots,D_k$ with $k$ minimal; we will show that $k\leq 512(|D|^3+1)$. Define $$R=H^M_{D_1, D_2} \circ H^M_{D_2, D_3} \circ \dots \circ
H^M_{D_{k-1}, D_k}.$$ Note that $R\subseteq D_1\times D_k$. By the definition of derectangularising sequence, there are $a,a'\in D_1$ and $b,b'\in D_k$ such that $(a,b)$, $(a'\!,b)$ and $(a,b')$ are all in $R$ but $(a'\!,b')\not\in R$. So there exist $$(x_1,\dots,x_k),(y_1,\dots,y_k),(z_1,\dots,z_k)\in D_1\times \dots
\times D_k$$ with $(x_1,x_k)=(a,b)$, $(y_1,y_k)=(a'\!,b)$ and $(z_1,z_k)=(a,b')$ such that $M_{x_i,x_{i+1}} = M_{y_i,y_{i+1}} =
M_{z_i,z_{i+1}}=*$ for every $i\in[k-1]$ but, for any $(w_1,\ldots,w_k)\in D_1 \times \dots \times D_k$ with $(w_1,w_k)=(a'\!,b')$, there is an $i\in[k-1]$ such that $M_{w_i,w_{i+1}}\neq *$.
Setting $D'_i=\{x_i,y_i,z_i\}$ for each $i$ gives an [$\mathcal L$]{}-$M$-derectangularising sequence $D'_1,\dots,D'_k$ with $|D'_i|\leq 3$ for each $1\leq i\leq k$. (Note that any submatrix of a pure matrix is pure.) For all $1\leq s <
t\leq k$ define $$R_{s,t}=H^M_{D'_s, D'_{s+1}} \circ H^M_{D'_{s+1}, D'_{s+2}} \circ \dots \circ
H^M_{D'_{t-1}, D'_t}.$$ Since $D'_1,\ldots,D'_k$ is [$\mathcal L$]{}-$M$-derectangularising, $R_{1,k}$ is not rectangular but, by the minimality of $k$, every other $R_{s,t}$ is rectangular. Note also that no $R_{s,t}=\emptyset$ since, if that were the case, we would have $R_{1,k}=\emptyset$, which is rectangular.
Suppose for a contradiction that $k> 512(|D|^3+1)$. There are at most $|D|^3+1$ subsets of $D$ with size at most three, so there are indices $1\leq i_0<i_1<i_2<\dots<i_{512}\leq k$ such that $D'_{i_0}=\dots=D'_{i_{512}}$. There are at most $2^{|D'_{i_0}|^2}-1 \leq 2^9-1=511$ non-empty binary relations on $D'_{i_0}$, so $R_{i_0,i_m}=R_{i_0,i_n}$ for some $1\leq
m<n\leq 512$. Since $R_{1,k}$ is not rectangular, $$R_{1,k}= R_{1,i_0} \circ R_{i_0,i_n} \circ R_{i_n,k}= R_{1,i_0} \circ R_{i_0,i_m} \circ R_{i_n,k}= R_{1,i_m} \circ R_{i_n,k}$$ is not rectangular. Therefore, $D'_1,D'_2,\dots,D'_{i_m},D'_{1+i_n},D'_{2+i_n},\dots,D'_k$ is an $\mathcal
L$-$M$-derectangularising sequence of length less than $k$, contradicting the minimality of $k$.
Now that we have membership in [$\mathrm{NP}$]{}, we can prove completeness.
[thm:meta]{}[ [[ExistsDerectSeq]{}]{} is [$\mathrm{NP}$]{}-complete under polynomial-time many-one reductions.]{}
We first show that [[ExistsDerectSeq]{}]{} is in [$\mathrm{NP}$]{}. Given $D$, $M\in \{0,1,*\}^{D\times D}$ and ${\ensuremath{\mathcal L}}\subseteq {{\mathcal{P}(D)}}$, a non-deterministic polynomial time algorithm for [[ExistsDerectSeq]{}]{}first “guesses” an ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$-$M$-derectangularising sequence $D_1,\ldots,D_k$ with $k\leq 512{(|D|^3+1)}$. Lemma \[lem:small\_derect\] guarantees that such a sequence exists if the output should be “yes”. The algorithm then verifies that each $D_i$ is a subset of a set in ${\ensuremath{\mathcal L}}$, that $\{D_1,\ldots,D_k\}$ is $M$-purifying, and that the relation $H^M_{D_1,D_2} \circ H^M_{D_2, D_3} \circ \dots \circ
H^M_{D_{k-1}, D_k}$ is not rectangular. All of these can be checked in polynomial time without explicitly constructing ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$.
To show that [[ExistsDerectSeq]{}]{} is [$\mathrm{NP}$]{}-hard, we give a polynomial-time reduction from the well-known [$\mathrm{NP}$]{}-hard problem of determining whether a graph $G$ has an independent set of size $k$.
Let $G$ and $k$ be an input to the independent set problem. Let $V(G)= [n]$ and assume without loss of generality that $k\in[n]$. Setting $D=[n]\times[k]\times[3]$, we construct a $D\times
D$ matrix $M$ and a set [$\mathcal L$]{} of lists such that there is an ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$-$M$-derectangularising sequence if and only if $G$ has an independent set of size $k$.
$M$ will be a block matrix, constructed using the following $3\times
3$ symmetric matrices. Note that each is pure, apart from ${\mathrm{Id}}$. $$\begin{gathered}
{M_\mathrm{start}}= \begin{pmatrix}*&*&0 \\ *&*&0 \\ 0&0&*\end{pmatrix} \qquad
{M_\mathrm{end}}= \begin{pmatrix}*&0&0 \\ 0&*&* \\ 0&*&*\end{pmatrix} \qquad
{M_\mathrm{bij}}= \begin{pmatrix}*&0&0 \\ 0&*&0 \\ 0&0&*\end{pmatrix} \\
{\mathbf{0}}= \begin{pmatrix}0&0&0 \\ 0&0&0 \\ 0&0&0\end{pmatrix} \qquad
{\mathrm{Id}}= \begin{pmatrix}1&0&0 \\ 0&1&0 \\ 0&0&1\end{pmatrix}\,.\end{gathered}$$
For $v\in [n]$ and $j\in[k]$, let $D[v,j] = \{(v,j,c) \mid c\in[3]\}$. Below, when we say that $M|_{D[v,j]\times D[v',j']}= N$ for some $3\times 3$ matrix $N$, we mean more specifically that $M_{(v,j,c),(v'\!,j'\!,c')} = N_{c,c'}$ for all $c,c'\in[3]$. $M$ is constructed as follows.
- For all $v\in[n]$, $M|_{D[v,1] \times D[v,1]}= {M_\mathrm{start}}$ and $M|_{D[v,k]\times D[v,k]} =
{M_\mathrm{end}}$.
- For all $v\in[n]$ and all $j\in\{2,\dots,k-1\}$, $M|_{D[{v,j}]\times D[{v,j}]} =
{M_\mathrm{bij}}$.
- If $v\neq v'\!$, $(v,v')\notin E(G)$ and $j<k$, then
- $M|_{D[{v,j}]\times D[{v',j+1}]} = M|_{D[v',j+1]\times D[{v,j}]} = {M_\mathrm{bij}}$ and
- $M|_{D[{v,j}] \times D[{v',j'}]} = M|_{D[{v',j'}]\times D[{v,j}]} = {\mathbf{0}}$ for all $j'>j+1$.
- For all $v,v'\in[n]$ and $j,j'\in[k]$ not covered above, $M|_{D[v,j]\times D[v',j']} = {\mathrm{Id}}$.
To complete the construction, let ${\ensuremath{\mathcal L}}=\{D[v,j]\mid v\in[n], j\in[k]\}$. We will show that $G$ has an independent set of size $k$ if and only if there is an ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$-$M$-derectangularising sequence.
For the forward direction of the proof, suppose that $G$ has an independent set $I = \{v_1, \dots, v_k\}$ of size $k$. We will show that $$D[v_1,1], D[v_1,1], D[v_2,2], D[v_3,3], \dots, D[v_{k-1}, k-1],
D[v_k,k], D[v_k,k]$$ (where the first and last elements are repeated and the others are not) is ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$-$M$-derectangularising. Since there is no edge $(v_i,v_{i'})\in E(G)$ for $i,i'\in[k]$, the matrix $M|_{D[v_i,i]\times D[v_{i'},i']} $ is always one of ${M_\mathrm{start}}$, ${M_\mathrm{end}}$, ${M_\mathrm{bij}}$ and ${\mathbf{0}}$, so it is always pure. Therefore, $\{D[v_1,1], \dots, D[v_k,k]\}$ is $M$-purifying. It remains to show that the relation $$R = H^M_{D[v_1,1],D[v_1,1]} \circ H^M_{D[v_1,1],D[v_2,2]} \circ \dots
\circ H^M_{D[v_{k-1,k-1}],D[v_{k},k]} \circ H^M_{D[v_k,k],D[v_k,k]}$$ is not rectangular.
Consider $i\in[k-1]$. Since $(v_i,v_{i+1})\notin E(G)$, $M|_{ D[{v_i,i}] \times D[{ v_{i+1}, i+1 }]} = {M_\mathrm{bij}}$ so $H^M_{D[v_i,i], D[v_{i+1},i+1]}$ is the bijection that associates $(v_i,i,c)$ with $(v_{i+1},i+1,c)$ for each $c\in[3]$. Therefore, $$H^M_{D[v_1,1],D[v_1,2]} \circ \dots \circ
H^M_{D[v_{k-1},k-1],D[v_k,k]}$$ is the bijection that associates $(v_1,1,c)$ with $(v_k,k,c)$ for each $c\in[3]$. We have $M|_{D[v_1,1]\times D[v_1,1]}
= {M_\mathrm{start}}$ and $M|_{D[v_k,k] \times D[v_k,k]} = {M_\mathrm{end}}$ so $$\begin{aligned}
H^M_{D[v_1,1], D[v_1,1]} &= \{((v_1,1,c),(v_1,1,c')) \mid c,c'\in[2]\}
\cup \{((v_1,1,3),(v_1,1,3))\} \\
H^M_{D[v_k,k], D[v_k,k]} &= \{((v_k,k,1),(v_k,k,1))\}
\cup \{((v_k,k,c),(v_k,k,c')) \mid c,c'\in\{2,3\}\}\,,\end{aligned}$$ and, therefore, $$\begin{aligned}
R = \{((v_1,1,c),(v_k,k,c')) \mid c,c'\in[3]\}
\setminus \{((v_1,1,3),(v_k,k,1))\}\,,\end{aligned}$$ which is not rectangular, as required.
For the reverse direction of the proof, suppose that there is an ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$-$M$-derectangularising sequence $D_1,\ldots,D_m$. The fact that the sequence is derectangularising implies that $|D_i|\geq 2$ for each $i\in[m]$ — see the remarks following Definition \[def:derect\]. Each set in the sequence is a subset of some $D[v,j]$ in ${\ensuremath{\mathcal L}}$ so for every $i\in[m]$ let $v_i$ denote the vertex in $[n]$ and let $j_i$ denote the index in $[k]$ such that $D_i \subseteq D[v_i,j_i]$. Clearly, it is possible to have $(v_i,j_i)=(v_{i'},j_{i'})$ for distinct $i$ and $i'$ in $[m]$.
We will finish the proof by showing that $G$ has a size-$k$ independent set. Let $$R = H^M_{D_1 , D_2} \circ \dots \circ H^M_{D_{m-1}, D_m },$$ which is not rectangular because the sequence is ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$-$M$-derectangularising. Since $\{D_1,\ldots,D_m\}$ is $M$-purifying, and any submatrix of ${\mathrm{Id}}$ with at least two rows and at least two columns is impure, every pair $(i,i')\in [m]^2$ satisfies $M|_{D[v_i,j_i]\times D[v_{i'},j_{i'}]} \neq {\mathrm{Id}}$. This means that we cannot have $(v_i,v_{i'})\in E(G)$ for any pair $(i,i')\in [m]^2$ so the set $I=\{v_1, \dots, v_m\}$ is independent in $G$. It remains to show that $|I|\geq k$.
Observe that, if $v_i=v_{i'}$, we must have $j_i=j_{i'}$ since, otherwise, the construction ensures that $$M|_{D[v_i,j_i]\times D[v_{i'},j_{i'}]}
= M|_{D[v_i,j_i]\times D[v_i,j_{i'}]}
= {\mathrm{Id}},$$ which we already ruled out. Therefore, $|I| \geq |\{j_1, \dots, j_m\}|$.
We must have $|j_i-j_{i+1}|\leq 1$ for each $i\in[m-1]$ as, otherwise, $M|_{D[v_i,j_i]\times D[v_{i+1},j_{i+1}]}
= {\mathbf{0}}$, which implies that $R=\emptyset$, which is rectangular. There must be at least one $i\in
[m-1]$ such that $v_i=v_{i+1}$ and $j_i=j_{i+1}=1$, so $M|_{D[v_i,j_i]\times D[v_{i+1},j_{i+1}]}
= {M_\mathrm{start}}$. If not, $R$ is a composition of relations corresponding to ${M_\mathrm{bij}}$ and ${M_\mathrm{end}}$ and any such relation is either a bijection, or of the form of ${M_\mathrm{end}}$, so it is rectangular. Similarly, there must be at least one $i$ such that $v_i=v_{i+1}$ and $j_i=j_{i+1}=k$, giving $M|_{D[v_i,j_i]\times D[v_{i+1},j_{i+1}]}
= {M_\mathrm{end}}$. Therefore, the sequence $j_1, \dots, j_m$ contains 1 and $k$. Since $|j_i-j_{i+1}|\leq 1$ for all $i\in[m-1]$, it follows that $[k] \subseteq \{j_1,\dots, j_m\}$, so $|I|\geq k$, as required. In fact, $\{j_1,\dots, j_m\} = [k]$ since each $j_i\in [k]$ by construction.
We defined the problem [[ExistsDerectSeq]{}]{} using a concise input representation: ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$ does not need to be written out in full. Instead, the instance is a subset ${\ensuremath{\mathcal L}}$ containing the maximal elements of ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$. For example, when the instance is ${\ensuremath{\mathcal L}}=\{D\}$, we have ${{\mathscr{S}({\ensuremath{\mathcal L}})}}= {{\mathcal{P}(D)}}$. It is important to note that the [$\mathrm{NP}$]{}-completeness of [[ExistsDerectSeq]{}]{} is not an artifact of this concise input coding. The elements of the list ${\ensuremath{\mathcal L}}$ constructed in the NP-hardness proof have length at most three, so the list ${{\mathscr{S}({\ensuremath{\mathcal L}})}}$ could also be constructed explicitly in polynomial time.
Lemma \[lem:small\_derect\] has the following immediate corollary for the complexity of the dichotomy criterion of the general [[\#List-$M$-partitions]{}]{} problem. Recall that, in this version of the meta-problem, the input is just the matrix $M$.
[cor:meta]{}[ [[MatrixHasDerectSeq]{}]{} is in [$\mathrm{NP}$]{}.]{}
Take ${\ensuremath{\mathcal L}}= \{D\}$ in Lemma \[lem:small\_derect\].
Cardinality constraints {#sec:card}
=======================
Finally, we show how lists can be used to implement cardinality constraints of the kind that often appear in counting problems in combinatorics.
Feder, Hell, Klein and Motwani [@FHKM] point out that lists can be used to determine whether there are $M$-partitions that obey simple cardinality constraints. For example, it is natural to require some or all of the parts to be non-empty or, more generally, to contain at least some constant number of vertices. Given a $D\times D$ matrix $M$, we represent such cardinality constraints as a function $C\colon D\to{\mathbb{Z}_{\geq 0}}$. We say that an $M$-partition $\sigma$ of a graph $G$ *satisfies* the constraint if, for each $d\in D$, $|\{v\in
V(G)\mid \sigma(v)=d\}| \geq C(d)$. Given a cardinality constraint $C$, we write $|C| = \sum_{d\in D} C(d)$.
We can determine whether there is an $M$-partition of $G=(V,E)$ that satisfies the cardinality constraint $C$ by making at most ${|V|}^{|C|}$ queries to an oracle for the list $M$-partitions problem, as follows. Let $L_C$ be the set of list functions $L\colon
V\to {{\mathcal{P}(D)}}$ such that:
- for all $v\in V\!$, either $L(v) = D$ or $|L(v)| = 1$, and
- for all $d\in D$, there are exactly $C(d)$ vertices $v$ with $L(v) = \{d\}$.
There are at most ${|V|}^{|C|}$ such list functions and it is clear that $G$ has an $M$-partition satisfying $C$ if, and only if, it has a list $M$-partition that respects at least one $L\in L_C$. The number of queries is polynomial in $|V|$ as long as the cardinality constraint $C$ is independent of $G$.
For counting, the situation is a little more complicated, as we must avoid double-counting. The solution is to count all $M$-partitions of the input graph and subtract off those that fail to satisfy the cardinality constraint. We formally define the problem [[\#$C$-$M$-partitions]{}]{} as follows, parameterized by a $D\times D$ matrix $M$ and a cardinality constraint function $C\colon D\to {\mathbb{Z}_{\geq 0}}$.
[[\#$C$-$M$-partitions]{}]{}.
A graph $G$.
The number of $M$-partitions of $G$ that satisfy $C$.
\[prop:cardinality\] [[\#$C$-$M$-partitions]{}]{} is polynomial-time Turing reducible to [[\#List-$M$-partitions]{}]{}.
Given the cardinality constraint function $C$, let $R = \{d\in
D\mid C(d)>0\}$: that is, $R$ is the set of parts that have a non-trivial cardinality constraint. For any set $P\subseteq R$, say that an $M$-partition $\sigma$ of a graph $G=(V,E)$ *fails on $P$* if $|\{v\in V\mid \sigma(v) = d\}| < C(d)$ for all $d\in P$. That is, if $\sigma$ violates the cardinality constraints on all parts in $P$ (and possibly others, too). Let $\Sigma$ be the set of all $M$-partitions of our given input graph $G$. For $i\in R$, let $A_i = \{\sigma \in \Sigma \mid \mbox{$\sigma$ fails on $\{i\}$}\}$ and let $A=\bigcup_{i\in R} A_i$. By inclusion-exclusion, $$\begin{aligned}
|A| &= -\!\!\sum_{\emptyset \subset P \subseteq R} {(-1)}^{|P|} \left|\bigcap_{i\in P} A_i\right|\\
&= -\!\!\sum_{\emptyset\subset P \subseteq R} {(-1)}^{|P|} \big|\{\sigma \in \Sigma \mid
\mbox{$\sigma$ fails on $P$}\}\big|\,.\end{aligned}$$
We wish to compute $$\begin{aligned}
\big|\{\sigma\in \Sigma\mid \text{$\sigma$ satisfies $C$}\}\big|
&= \big|\Sigma\big| - |A| \\
&= \big|\Sigma\big|
+ \sum_{\emptyset \subset P\subseteq R}
(-1)^{|P|}
\big|\{\sigma\in\Sigma \mid \text{$\sigma$
fails on $P$}\}\big|\,.
\end{aligned}$$
Therefore, it suffices to show that we can use lists to count the $M$-partitions that fail on each non-empty $P\subseteq R$. For such a set $P$, let $L_P$ be the set of list functions $L$ such that
- for all $v\in V$, either $L(v) = D\setminus P$ or $L(v) =
\{p\}$ for some $p\in P$, and
- for all $p\in P$, $\big|\big\{v\in V\mid L(v)=\{p\}\big\}\big| < C(p)$.
Thus, the set of $M$-partitions that respect some $L\in L_P$ is precisely the set of $M$-partitions that fail on $P$. Also, for distinct $L$ and $L'$ in $L_P$, the set of $M$-partitions that respect $L$ is disjoint from the set of $M$-partitions that respect $L'\!$. So we can compute $ \big|\{\sigma\in\Sigma \mid \text{$\sigma$
fails on $P$}\}\big|$ by making $|L_P|$ calls to [[\#List-$M$-partitions]{}]{}, noting that $|L_P|\leq |V|^{|C|}\!$.
As an example of a combinatorial structure that can be represented as an $M$-partition problem with cardinality constraints, consider the *homogeneous pairs* introduced by Chvátal and Sbihi [@CS1987:Bull-free]. A homogeneous pair in a graph $G=(V,E)$ is a partition of $V$ into sets $U$, $W_1$ and $W_2$ such that:
- $|U|\geq 2$;
- $|W_1|\geq 2$ or $|W_2|\geq 2$ (or both);
- for every vertex $v\in U$, $v$ is either adjacent to every vertex in $W_1$ or to none of them; and
- for every vertex $v\in U$, $v$ is either adjacent to every vertex in $W_2$ or to none of them.
Feder et al. [@FHKM] observe that the problem of determining whether a graph has a homogeneous pair can be represented as the problem of determining whether it has an [$M_{\mathrm{hp}}$]{}-partition satisfying certain constraints, where $D = \{1, \dots, 6\}$ and $${\ensuremath{M_{\mathrm{hp}}}}= \begin{pmatrix}
* & * & 1 & 0 & 1 & 0 \\
* & * & 1 & 1 & 0 & 0 \\
1 & 1 & * & * & * & * \\
0 & 1 & * & * & * & * \\
1 & 0 & * & * & * & * \\
0 & 0 & * & * & * & *
\end{pmatrix}.$$
$W_1$ corresponds to the set of vertices mapped to part $1$ (row 1 of ${\ensuremath{M_{\mathrm{hp}}}}$), $W_2$ corresponds to the set of vertices mapped to part $2$ (row 2 of ${\ensuremath{M_{\mathrm{hp}}}}$), and $U$ corresponds to the set of vertices mapped to parts $3$–$6$.
In fact, there is a one-to-one correspondence between the homogeneous pairs of $G$ in which $W_1$ and $W_2$ are non-empty and the ${\ensuremath{M_{\mathrm{hp}}}}$-partitions $\sigma$ of $G$ that satisfy the following additional constraints. For $d\in D$, let $N_\sigma(d) = |\{v\in V(G)\mid \sigma(v)=d\}|$ be the number of vertices that $\sigma$ maps to part $d$. We require that
- $N_\sigma(3) + N_\sigma(4) + N_\sigma(5) + N_\sigma(6)\geq 2$,
- $N_\sigma(1) > 0$ and $N_\sigma(2) > 0$, and
- at least one $N_\sigma(1)$ and $N_\sigma(2)$ is at least $2$.
To see this, consider a homogeneous pair $(U,W_1,W_2)$ in which $W_1$ and $W_2$ are non-empty. Note that there is exactly one ${\ensuremath{M_{\mathrm{hp}}}}$-partition of $G$ in which vertices in $W_1$ are mapped to part $1$ and vertices in $W_2$ are mapped to part $2$ and vertices in $U$ are mapped to parts $3$–$6$. There is exactly one part available to each $v\in U$ since $v$ has an edge or non-edge to $W_1$ (but not both!) ruling out exactly two parts and $v$ has an edge or non-edge to $W_2$ ruling out an additional part. Going the other way, an ${\ensuremath{M_{\mathrm{hp}}}}$-partition that satisfies the constraints includes a homogeneous pair.
Now let $${\ensuremath{M_{\mathrm{hs}}}}= \begin{pmatrix}
* & 0 & 1\\
0 & * & *\\
1 & * & *
\end{pmatrix}.$$ There is a one-to-one correspondence between the homogeneous pairs of $G$ in which $W_2$ is empty and the ${\ensuremath{M_{\mathrm{hs}}}}$-partitions of $G$ that satisfy the following additional constraints.
- At least two vertices are mapped to parts $2$–$3$ (vertices in these parts are in $U$).
- At least two vertices are mapped to part $1$ (vertices in this part are in $W_1$).
Symmetrically, there is also a one-to-one correspondence between the homogeneous pairs of $G$ in which $W_1$ is empty and the ${\ensuremath{M_{\mathrm{hs}}}}$-partitions of $G$ that satisfy the above constraints. (Partitions according to ${\ensuremath{M_{\mathrm{hs}}}}$ correspond to so-called “homogeneous sets” but we do not need the details of these.)
It is known from [@EKR1997:Hom-pair] that, in deterministic polynomial time, it is possible to determine whether a graph contains a homogeneous pair and, if so, to find one. We show that the homogeneous pairs in a graph can also be counted in polynomial time. We start by considering the relevant list-partition counting problems.
\[thm:hompair\] There are polynomial-time algorithms for [[\#List-${\ensuremath{M_{\mathrm{hp}}}}$-partitions]{}]{} and [[\#List-${\ensuremath{M_{\mathrm{hs}}}}$-partitions]{}]{}.
We first show that there is a polynomial-time algorithm for [[\#List-${\ensuremath{M_{\mathrm{hp}}}}$-partitions]{}]{}. The most natural way to do this would be to show that there is no ${{\mathcal{P}(D)}}$-[$M_{\mathrm{hp}}$]{}-derectangularising sequence and then apply Theorem \[thm:explicitdichotomy\]. In theory, we could show that there is no ${{\mathcal{P}(D)}}$-[$M_{\mathrm{hp}}$]{}-derectangularising sequence by brute force since $|D|=6$, but the number of possibilities is too large to make this feasible. Instead, we argue non-constructively.
First, if there is no ${{\mathcal{P}(D)}}$-[$M_{\mathrm{hp}}$]{}-derectangularising sequence, the result follows from Theorem \[thm:explicitdichotomy\].
Conversely, suppose that $D_1, \dots, D_k$ is a ${{\mathcal{P}(D)}}$-[$M_{\mathrm{hp}}$]{}-derectangularising sequence. Let $M$ be the matrix such that $M_{i,j} = 0$ if $({\ensuremath{M_{\mathrm{hp}}}})_{i,j} = 1$ and $M_{i,j}
= ({\ensuremath{M_{\mathrm{hp}}}})_{i,j}$, otherwise. $D_1, \dots, D_k$ is also a ${{\mathcal{P}(D)}}$-$M$-derectangularising sequence, since $H^M_{X,Y} =
H^{{\ensuremath{M_{\mathrm{hp}}}}}_{X,Y}$ for any $X,Y\subseteq D$ and any sequence $D_1,
\dots, D_k$ is $M$-purifying because $M$ is already pure. Therefore, by Theorem \[thm:explicitdichotomy\], counting list $M$-partitions is [$\mathrm{\#P}$]{}-complete.
However, counting the list $M$-partitions of a graph $G$ corresponds to counting list homomorphisms from $G$ to the $6$-vertex graph $H$ whose two components are an edge and a $4$-clique, and which has loops on all six vertices. There is a very straightforward polynomial-time algorithm for this problem (a simple modification of the version without lists in [@DG]). Thus, ${\ensuremath{\mathrm{\#P}}}={\ensuremath{\mathrm{FP}}}$ so, in particular, there is a polynomial-time algorithm for counting list [$M_{\mathrm{hp}}$]{}-partitions.
The proof that there is a polynomial-time algorithm for [[\#List-${\ensuremath{M_{\mathrm{hs}}}}$-partitions]{}]{} is similar.
\[cor:hompair\] There is a polynomial-time algorithm for counting the homogeneous pairs in a graph.
We are given a graph $G=(V,E)$ and we wish to compute the number of homogeneous pairs that it contains. By the one-to-one correspondence given earlier, it suffices to show how to count ${\ensuremath{M_{\mathrm{hp}}}}$-partitions and ${\ensuremath{M_{\mathrm{hs}}}}$-partitions of $G$ satisfying additional constraints. We start with the first of these. Recall the constraints on the ${\ensuremath{M_{\mathrm{hp}}}}$-partitions $\sigma$ that we wish to count:
- $N_\sigma(3) + N_\sigma(4) + N_\sigma(5) + N_\sigma(6)\geq 2$,
- $N_\sigma(1) > 0$ and $N_\sigma(2) > 0$, and
- at least one $N_\sigma(1)$ and $N_\sigma(2)$ is at least $2$.
Define three subsets $\Sigma_1$, $\Sigma_2$ and $\Sigma_{1,2}$ of the set of [$M_{\mathrm{hp}}$]{}-partitions of $G$ that satisfy the constraints. In the definition of each of $\Sigma_1$, $\Sigma_2$ and $\Sigma_{1,2}$, we will require that parts $1$ and $2$ are non-empty and parts $3$–$6$ contain a total of at least two vertices. In $\Sigma_1$, part $1$ must contain at least two vertices; in $\Sigma_2$, part $2$ must contain at least two vertices; in $\Sigma_{1,2}$, both parts $1$ and $2$ must contain at least two vertices. The number of suitable ${\ensuremath{M_{\mathrm{hp}}}}$-partitions of $G$ is $|\Sigma_1| + |\Sigma_2| - |\Sigma_{1,2}|$.
Each of $|\Sigma_1|$, $|\Sigma_2|$ and $|\Sigma_{1,2}|$ can be computed by counting the ${\ensuremath{M_{\mathrm{hp}}}}$-partitions of $G$ that satisfy appropriate cardinality constraints. Parts $1$ and $2$ are trivially dealt with. The requirement that parts $3$–$6$ must contain at least two vertices between them is equivalent to saying that at least one of them must contain at least two vertices or at least two must contain at least one vertex. This can be expressed with a sequence of cardinality constraint functions and using inclusion–exclusion to eliminate double-counting.
Counting constrained ${\ensuremath{M_{\mathrm{hs}}}}$-partitions of $G$ is similar (but simpler).
[^1]: Department of Computer Science, University of Oxford, Wolfson Building, Parks Road, Oxford, OX1 3QD, United Kingdom.
[^2]: Department of Computer Science, Ashton Building, University of Liverpool, Liverpool, L69 3BX, United Kingdom.
[^3]: Department of Information Science, University of Fukui, 3-9-1 Bunkyo, Fukui City, Fukui 910-8507, Japan.
[^4]: A preliminary version of this paper appeared in the proceedings of CCC 2014. The research leading to these results has received funding from the MEXT Grants-in-Aid for Scientific Research and the EPSRC and the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007–2013) ERC grant agreement no. 334828. The paper reflects only the authors’ views and not the views of the ERC or the European Commission. The European Union is not liable for any use that may be made of the information contained therein.
[^5]: For the reader who is familiar with CSPs, it might be useful to see how a [[List-$M$-partitions]{}]{} problem can be coded as a CSP with restrictions on the input. Given a symmetric $M \in \{0,1,*\}^{D\times D}$, let $M_0$ be the relation on $D\times D$ containing all pairs $(i,j) \in D\times D$ for which $M_{i,j} \neq 1$. Let $M_1$ be the relation on $D\times D$ containing all pairs $(i,j)\in D\times D$ for which $M_{i,j} \neq 0$. Then a [[List-$M$-partitions]{}]{} problem with input $G,L$ can be encoded as a CSP whose constraint language includes the binary relations $M_0$ and $M_1$ and also the unary relations corresponding to the sets in the image of $L$. Each vertex $v$ of $G$ is a variable in the CSP instance with the unary constraint $L(v)$. If $(u,v)$ is an edge of $G$ then it is constrained by $M_1$. If it is a non-edge of $G$, it is constrained by $M_0$. Note that the CSP instance satisfies the restriction that every pair of distinct variables has exactly one constraint, which is either $M_0$ or $M_1$. In a general CSP instance, a pair of variables could be constrained by $M_0$ and $M_1$ or one of them, or neither. It is not clear how to code such a general CSP instance as a list partitions problem.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a simplified model of attention which is applicable to feed-forward neural networks and demonstrate that the resulting model can solve the synthetic “addition” and “multiplication” long-term memory problems for sequence lengths which are both longer and more widely varying than the best published results for these tasks.'
author:
- |
Colin Raffel\
LabROSA, Columbia University\
`[email protected]` Daniel P. W. Ellis\
LabROSA, Columbia University\
`[email protected]`
bibliography:
- 'refs.bib'
title: 'Feed-Forward Networks with Attention Can Solve Some Long-Term Memory Problems'
---
Models for Sequential Data
==========================
Many problems in machine learning are best formulated using sequential data and appropriate models for these tasks must be able to capture temporal dependencies in sequences, potentially of arbitrary length. One such class of models are recurrent neural networks (RNNs), which can be considered a learnable function $f$ whose output $h_t = f(x_t, h_{t - 1})$ at time $t$ depends on input $x_t$ and the model’s previous state $h_{t - 1}$. Training of RNNs with backpropagation through time [@werbos1990backpropagation] is hindered by the vanishing and exploding gradient problem [@pascanu2012difficulty; @hochreiter1997long; @bengio1994learning], and as a result RNNs are in practice typically only applied in tasks where sequential dependencies span at most hundreds of time steps. Very long sequences can also make training computationally inefficient due to the fact that RNNs must be evaluated sequentially and cannot be fully parallelized.
Attention
---------
A recently proposed method for easier modeling of long-term dependencies is “attention”. Attention mechanisms allow for a more direct dependence between the state of the model at different points in time. Following the definition from [@bahdanau2014neural], given a model which produces a hidden state $h_t$ at each time step, attention-based models compute a “context” vector $c_t$ as the weighted mean of the state sequence $h$ by $$c_t = \sum_{j = 1}^T \alpha_{tj} h_j$$ where $T$ is the total number of time steps in the input sequence and $\alpha_{tj}$ is a weight computed at each time step $t$ for each state $h_j$. These context vectors are then used to compute a new state sequence $s$, where $s_t$ depends on $s_{t - 1}$, $c_t$ and the model’s output at $t - 1$. The weightings $\alpha_{tj}$ are then computed by $$e_{tj} = a(s_{t - 1}, h_j), \alpha_{tj} = \frac{\exp(e_{tj})}{\sum_{k = 1}^T \exp(e_{tk})}$$ where $a$ is a learned function which can be thought of as computing a scalar importance value for $h_j$ given the value of $h_j$ and the previous state $s_{t - 1}$. This formulation allows the new state sequence $s$ to have more direct access to the entire state sequence $h$. Attention-based RNNs have proven effective in a variety of sequence transduction tasks, including machine translation [@bahdanau2014neural], image captioning [@xu2015show], and speech recognition [@chan2015listen; @bahdanau2015end]. Attention can be seen as analogous to the “soft addressing” mechanisms of the recently proposed Neural Turing Machine [@graves2014neural] and End-To-End Memory Network [@sukhbaatar2015end] models.
Feed-Forward Attention
----------------------
A straightforward simplification to the attention mechanism described above which would allow it to be used to produce a single vector $c$ from an entire sequence could be formulated as follows: $$\label{eq:ffattention}
e_t = a(h_t), \alpha_t = \frac{\exp(e_t)}{\sum_{k = 1}^T \exp(e_k)}, c = \sum_{t = 1}^T \alpha_t h_t$$ As before, $a$ is a learnable function, but it now only depends on $h_t$. In this formulation, attention can be seen as producing a fixed-length embedding $c$ of the input sequence by computing an adaptive weighted average of the state sequence $h$. A schematic of this form of attention is shown in Figure \[fig:schematic\]. [@sonderby2015convolutional] compared the effectiveness of a standard recurrent network to a recurrent network augmented with this simplified version of attention on the task of protein sequence analysis.
![Schematic of our proposed “feed-forward” attention mechanism (cf. [@cho2015introduction] Figure 1). Vectors in the hidden state sequence $h_t$ are fed into the learnable function $a(h_t)$ to produce a probability vector $\alpha$. The vector $c$ is computed as a weighted average of $h_t$, with weighting given by $\alpha$.[]{data-label="fig:schematic"}](schematic.pdf){width=".8\textwidth"}
A consequence of using an attention mechanism is the ability to integrate information over time. It follows that by using this simplified form of attention, a model could handle variable-length sequences even if the calculation of $h_t$ was feed-forward, i.e. $h_t = f(x_t)$. Using a feed-forward $f$ could also result in large efficiency gains as the computation could be completely parallelized. We investigate the capabilities of this “feed-forward attention” model in Section \[sec:experiments\].
We note here that feed-forward models without attention can be used for sequential data when the sequence length $T$ is fixed, but when $T$ varies across sequences, some form of temporal integration is necessary. An obvious straightforward choice, which can be seen as an extreme oversimplification of attention, would be to compute $c$ as the unweighted average of the state sequence $h_t$, i.e. $$\label{eq:unweighted}
c = \frac{1}{T}\sum_{t = 1}^T h_t$$ This form of integration has been used to collapse the temporal dimension of audio [@dieleman2014recommending] and text document [@lei2015molding] sequences. We will also explore the effectiveness of this approach.
Toy Long-Term Memory Problems {#sec:experiments}
=============================
A common way to measure the long-term memory capabilities of a given model is to test it on the synthetic problems originally proposed by [@hochreiter1997long]. In this paper, we will focus on the “addition” and “multiplication” problems; due to space constraints, we refer the reader to [@hochreiter1997long] or [@sutskever2013importance] for their specification. As proposed by [@hochreiter1997long], we define accuracy as the proportion of sequences for which the absolute error between predicted value and the target value was less than .04. Applying our feed-forward model to these tasks is somewhat disingenuous because they are commutative and therefore may be easier to solve with a model which ignores temporal order. However, as we further argue in Section \[sec:discussion\], we believe these tasks provide a useful demonstration of our model’s ability to refer to arbitrary locations in the input sequence when computing its output.
Model Details
-------------
For all experiments, we used the following model: First, the state $h_t$ was computed from the input at each time step $x_t$ by $h_t = \textrm{LReLU}(W_{xh}x_t + b_{xh})$ where $W_{xh} \in \mathbb{R}^{D \times 2}, b_{xh} \in \mathbb{R}^D$ and $\textrm{LReLU}(x) = \max(x, .01x)$ is the “leaky rectifier” nonlinearity, as proposed by [@maas2013rectifier]. We found that this nonlinearity improved early convergence so we used it in all of our models. We tested models where the context vector $c$ was then computed either as in Equation (\[eq:ffattention\]), with $a(h_t) =\tanh(W_{hc}h_t + b_{hc})$ where $W_{hc} \in \mathbb{R}^{1 \times D}, b_{hc} \in \mathbb{R}$, or simply as the unweighted mean of $h$ as in Equation (\[eq:unweighted\]). We then computed an intermediate vector $s = \textrm{LReLU}(W_{cs}c + b_{cs})$ where $W_{cs} \in \mathbb{R}^{D \times D}, b \in \mathbb{R}^D$ from which the output was computed as $y = \textrm{LReLU}(W_{sy}s + b_{sy})$ where $W_{sy} \in \mathbb{R}^{1 \times D}$, $b_{sy} \in \mathbb{R}$. For all experiments, we set $D = 100$.
We used the squared error of the output $y$ against the target value for each sequence as an objective. Parameters were optimized using “adam”, a recently proposed stochastic optimization technique [@kingma2014adam], with the optimization hyperparameters $\beta_1$ and $\beta_2$ set to the values suggested by [@kingma2014adam] (.9 and .999 respectively). All weight matrices were initialized with entries drawn from a Gaussian distribution with a mean of zero and, for a matrix $W \in \mathbb{R}^{M \times N}$, a standard deviation of $1/\sqrt{N}$. All bias vectors were initialized with zeros. We trained on mini-batches of 100 sequences and computed the accuracy on a held-out test set of 1000 sequences every epoch, defined as 1000 parameter updates. We stopped training when either 100% accuracy was attained on the test set, or after 100 epochs. All networks were implemented using Lasagne [@dieleman2015lasagne], which is built on top of Theano [@bastien2012theano; @bergstra2010theano].
---------------------------------- ---- ----- ----- ------ ------ ------- ---- ----- ----- ------ ------ -------
Task
$T_0$ 50 100 500 1000 5000 10000 50 100 500 1000 5000 10000
(r)[2-7]{} (r)[8-13]{} Attention 1 1 1 1 2 3 1 2 4 2 15 6
Unweighted 1 1 1 2 8 17 2 2 8 33
---------------------------------- ---- ----- ----- ------ ------ ------- ---- ----- ----- ------ ------ -------
: Number of epochs required to achieve perfect accuracy, or accuracy after 100 epochs (greyed-out values), for the experiment described in Section \[sec:fixed\].[]{data-label="tab:fixed"}
Fixed-Length Experiment {#sec:fixed}
-----------------------
Traditionally, the sequence lengths tested in each task vary uniformly between $[T_0, 1.1T_0]$ for different values of $T_0$. As $T_0$ increases, the model must be able to handle longer-term dependencies. The largest value of $T_0$ attained using RNNs with different training, regularization, and model structures has varied from a few hundred [@martens2011learning; @sutskever2013importance; @le2015simple; @krueger2015regularizing; @arjovsky2015unitary] to a few thousand [@hochreiter1997long; @jaegar2012long]. We therefore tested our proposed feed-forward attention models for $T_0 \in \{50, 100, 500, 1000, 5000, 10000\}$. The required number of epochs or accuracy after 100 epochs for each task, sequence length, and temporal integration method (adaptively weighted attention or unweighted mean) is shown in Table \[tab:fixed\]. For fair comparison, we report the best result achieved using any learning rate in $\{.0003, .001, .003, .01\}$. From these results, it’s clear that the feed-forward attention model can quickly solve these long-term memory problems for all sequence lengths we tested. Our model is also efficient: Processing one epoch of 100,000 sequences with $T_0 = 10000$ took 254 seconds using an NVIDIA GTX 980 Ti GPU, while processing the same data with a single-layer vanilla RNN with a hidden dimensionality of 100 (resulting in a comparable number of parameters) took 917 seconds on the same hardware. In addition, there is a clear benefit to using the attention mechanism of Equation (\[eq:ffattention\]) instead of a simple unweighted average over time, which only incurs a marginal increase in the number of parameters (10,602 vs. 10,501, or less than 1%).
Variable-length Experiment
--------------------------
Because the range of sequence lengths $[T_0, 1.1T_0]$ is small compared to the range of $T_0$ values we evaluated, we further tested whether it was possible to train a single model which could cope with sequences with highly varying lengths. To our knowledge, such a variant of these tasks has not been studied before. We trained models of the same architecture used in the previous experiment on minibatches of sequences whose lengths were chosen uniformly at random between 50 and 10000 time steps. Using the attention mechanism of Equation (\[eq:ffattention\]), on held-out test sets of 1000 sequences, our model achieved 99.9% accuracy on the addition task and 99.4% on the multiplication task after training for 100 epochs. This suggests that a single feed-forward network with attention can simultaneously handle both short and very long sequences, with a marginal decrease in accuracy. Using an unweighted average over time, we were only able to achieve accuracies of 77.4% and 55.5% on the variable-length addition and multiplication tasks, respectively.
Discussion {#sec:discussion}
----------
A clear limitation of our proposed model is that it will fail on any task where temporal order matters because computing an average over time discards order information. For example, on the two-symbol temporal order task [@hochreiter1997long] where a sequence must be classified in terms of whether two symbols $X$ and $Y$ appear in the order $X, X$; $Y, Y$; $X, Y$; or $Y, X$, our model can differentiate between the $X, X$ and $Y, Y$ cases perfectly but cannot differentiate between the $X, Y$ and $Y, X$ cases at all. Nevertheless, we submit that for some real-world tasks involving sequential data, temporal order is substantially less important than being able to handle very long sequences. For example, in Joachims’ seminal paper on text document categorization [@joachims1998text], he posits that “word stems work well as representation units and that their ordering in a document is of minor importance for many tasks”. In fact, the current state-of-the-art system for document classification still uses order-agnostic sequence integration [@lei2015molding]. We have also shown in parallel work that our proposed feed-forward attention model can be used effectively for pruning large-scale (sub)sequence retrieval searches, even when the sequences are very long and high-dimensional [@raffel2016pruning].
Our experiments explicitly demonstrate that including an attention mechanism can allow a model to refer to specific points in a sequence when computing its output. They also provide an alternate argument for the claim made by [@bahdanau2014neural] that attention helps models handle very long and widely variable-length sequences. We are optimistic that our proposed feed-forward model will prove beneficial in additional real-world problems requiring order-agnostic temporal integration of long sequences. Further investigation is warranted; to facilitate future work, all of the code used in our experiments is available online.[^1]
Acknowledgements
================
We thank Sander Dieleman, Bart van Merri[ë]{}nboer, S[ø]{}ren Kaae S[ø]{}nderby, Brian McFee, and our anonymous reviewers for discussion and feedback.
[^1]: [`https://github.com/craffel/ff-attention/tree/master/toy‘_problems`](https://github.com/craffel/ff-attention/tree/master/toy_problems)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Luis F. Rodríguez, Yolanda Gómez, and Lizette Guzmán'
title: 'Ionization-bounded and Density-bounded Planetary Nebulae'
---
Introduction
============
Planetary nebulae, and in general any type of nebula photoionized by a central star or cluster of stars can be classified as either ionization-bounded or density-bounded. In the first case, the radius of the ionized region is determined by the absorption of the ionizing photons in the inner parts of the nebula. Even if there is more external gas available, it will remain neutral. In contrast, a density-bounded planetary nebula is one in which the ionized region terminates simply because there is no more gas available. Even if there are ionizing photons available, there will be no dense gas to ionize and the photons will escape to the diffuse halos that are known to surround planetary nebulae (Perinotto et al. 2004; Sandin et al. 2008).
When the nebula is visible in the optical, the presence of lines such as \[O I\] $\lambda$6300 (i. e. Bieging et al. 2007), that originate from the region of transition between ionized and neutral hydrogen, suggests that the nebula is ionization-bounded. Also the association of the planetary nebula with neutral (Rodríguez & Moran 1982) or molecular (Treffers et al. 1976) hydrogen or with other molecular species (Thronson 1983) suggests ionization-boundness. However, the situation is not as simple since a nebula can be ionization-bounded in some directions (with respect to the ionizing star) and density-bounded in others and the evidence of a transition from ionized to neutral or molecular gas could be coming from some directions only.
It is relevant to know if a nebulae is ionization-bounded in all directions since in this case all the ionizing photon flux from the central star is trapped inside the nebula. Only in this case an accurate estimate of this ionizing flux can be derived indirectly by observing the nebular emission, for example in recombination lines or free-free emission. Otherwise, we will underestimate the ionizing photon flux. In an important development, Zijlstra et al. (2008) have reported Very Large Array observations taken over 25 years of NGC 7027 at frequencies above 5 GHz, where the free-free emission is optically thin. They find that the flux density is changing at a yearly rate of -0.145$\pm$0.005% and propose that this is caused by a decrease in the number of ionizing photons coming from the central star. However, their conclusion assumes that this planetary nebula is ionization-bounded in all directions. This assumption seems very reasonable for NGC 7027 that is known to be associated with many molecular emissions and to possess large obscuration, suggesting that the ionized region is completely engulfed in neutral gas. However, this assumption should be tested. Finally, Mellema (2004) and Sch[ö]{}nberner et al. (2005) have noted that the apparent expansion velocity in the plane of the sky of planetary nebulae that are ionization-bounded can be larger than the true expansion velocity of the gas since we also have the contribution of the outward motion of the ionization front as the nebula expands. This discrepancy becomes quite large during the optically-thin stage of nebular evolution (Sch[ö]{}nberner et al. 2005). The knowledge of whether or not a nebula is ionization-bounded can be used to approximately correct for this effect.
![If we divide the emission from a nebula in the plane of the sky in $n$ pie slices, each slice will have an opening angle of $2 \pi / n$ radians (see figure) and its apex will coincide with the ionizing star. From the point of view of an observer in the central star, each of these slices will appear as a spherical wedge with a solid angle of $4 \pi / n$ steradians. []{data-label="fig1"}](pieslice.eps)
In this paper we discuss a simple criterion that can be derived from good quality radio continuum images taken at frequencies where the ionized nebula is optically-thin. This analysis provides a necessary, although not sufficient, condition to classify a nebula as ionization-bounded. If the nebula fails the criterion, it can be considered to be density-bounded, at least in some directions (as seen from the central star).
The criterion
=============
We make the following assumptions: i) the central star produces a photoionizing flux that is isotropic as seen from the position of the star, and ii) the nebula is optically-thin in the free-free continuum. The first assumption implies that, if the nebula is ionization-bounded, the rate of photoionizations (and recombinations) produced in the nebula is independent of direction and is a constant for a given solid angle (as seen from the star). Then, the number of free-free photons (that in steady state is proportional to the number of recombinations, e. g. Schraml & Mezger 1969), produced per differential of solid angle as seen from the star is a constant. This conclusion is valid even if the radius of the ionized gas is different in different directions. A larger density in a given direction will cause a smaller ionized radius, but the number of free-free photons produced per unit of volume will be larger, so the number of free-free photons produced per differential of solid angle will remain constant. The second assumption implies that all free-free photons produced escape from the nebula.
We now ask: what implications have these conclusions in the appearance of the nebula in the sky? As seen by the observer, we can split the nebula in $n$ identical “pie” slices centered in the central star (see Fig. 1), each with an opening angle given by $2 \pi / n$ radians. As seen from the central star each of these slices is a spherical wedge with solid angle given by $4 \pi / n$ steradians. Since this solid angle is constant for all wedges, we conclude that for a nebula to be ionization-bounded, the flux density in each of the slices should be constant, within the noise.
This simple test can be implemented by following the next steps: 1) obtain a good quality radio continuum image of a nebula at a frequency where it is known to be optically thin in the free-free emission, 2) define a center for the nebula either using the stellar position or symmetry considerations applied to the nebula, 3) use the task IRING of the NRAO software package AIPS (Astronomical Image Processing System) to determine the flux density in each of $n$ identical slices centered in the position determined in the previous step. The radius of these slices has to be large enough to contain all the nebular emission. Finally, 4) plot these flux densities as a function of the position angle of the slices to analyze their behavior.
Observations
============
The two sets of observations used in this study were obtained in the standard continuum mode and taken from the archive of the VLA of the NRAO[^1]. One set was for NGC 7027 ( = PN G084.9-03.4), and was taken at 43.34 GHz on 2001 November 9 in the D configuration as part of the VLA Calibrator Flux Density project (R. A. Perley & B. Butler 2008, in preparation). The amplitude calibrator was 1331+305, with an adopted flux density of 1.46 Jy. NGC 7027 was used as its own phase calibrator, with a bootstrapped flux density of 4.90$\pm$0.13 Jy. This set was analyzed with the specific goal of testing the hypothesis of ionization-boundness of Zijlstra et al. (2008). The second set was for Hb 4 ( = PN G003.1+02.9), and was taken at 8.46 GHz on 2001 February 18 in the BnA as part of project AK528. The amplitude calibrator was 1331+305, with an adopted flux density of 5.21 Jy. The phase calibrator was 1745$-$290, with a bootstrapped flux density of 0.67$\pm$0.02 Jy.
In Table 1 we present, for each planetary nebula, the position adopted as the center of the nebula (in both cases taken to be equal to the position of the minimum emission in the “bowl” that characterizes the morphology of these objects), and the frequency of the observations.
PN $\alpha$(2000)$^a$ $\delta$(2000)$^a$ $\nu$(GHz)
----------- -------------------- -------------------- ------------
NGC 7027 21 07 01.74 42 14 09.4 43.34
Hb 4 17 41 52.81 -24 42 07.8 8.46
\[tab:1\]
: Observational Parameters
The data were reduced using the standard VLA procedures in the software package AIPS of NRAO and then self-calibrated in phase and amplitude. The images were made with the ROBUST weighting parameter (Briggs 1995) of AIPS set to 0.
Interpretation and Results
==========================
NGC 7027
--------
In Figure 2 we show the 43.34 GHz image of NGC 7027. This planetary nebula is known to be optically-thin above $\sim$5 GHz (Zijlstra et al. 2008), so we expect this 43.34 GHz image to be tracing optically-thin free-free emission. In Figure 3 we plot the flux densities of each slice as a function of their position angle. The slices were made with an angular radius of 10$''$ and an opening of $10^\circ$, with the position angle incrementing by $10^\circ$. This flux density per slice is approximately constant (only small modulations at the $\sim$10% level are present). We conclude that this result is consistent with NGC 7027 being ionization-bounded in all directions, in agreement with the assumption of Zijlstra et al. (2008). This test is relevant because if NGC 7027 were density-bounded in some directions, the observed decrease in flux density could be due to the expansion of the nebula, allowing more ionizing photons to escape in the directions where the nebula is density-bounded and not to a decrease in the number of ionizing photons from the central star. The assumption of NGC 7027 being ionization-bounded is fundamental for the Zijlstra et al. (2008) analysis and our results are consistent with this assumption.
![43.34 GHz contour and greyscale image of the planetary nebula NGC 7027. The contours are -5, -4, 4, 5, 6, 8, 10, 12, 15, 20, 30, 40, 50, 60, 80, 100, and 120 times 3.1 mJy beam$^{-1}$, the rms noise of the image. The greyscale is shown in the top wedge, in mJy beam$^{-1}$. The synthesized beam ($2\rlap.{''}21 \times 1\rlap.{''}80$ with a position angle of $-79^\circ$) is shown in the bottom left corner of the image. The cross marks the position adopted as the center of the nebular emission (see Table 1). []{data-label="fig2"}](N7027GREY.PS)
![Flux density of the image slices of NGC 7027 as a function of their position angle.[]{data-label="fig3"}](ngc7027.eps)
Unfortunately, the constancy in the flux density of the slices is not a sufficient condition to assure ionization-boundness: an ionized ring seen face-on would show this same constancy but it will be density-bounded along the line of sight.
Hb 4
----
![8.46 GHz contour and greyscale image of the planetary nebula Hb 4. The contours are -5, -4, 4, 5, 6, 8, 10, 12, 15, 20, 30, 40, and 50 times 0.10 mJy beam$^{-1}$, the rms noise of the image. The greyscale is shown in the top wedge, in mJy beam$^{-1}$. The synthesized beam ($0\rlap.{''}91 \times 0\rlap.{''}49$ with a position angle of $-53^\circ$) is shown in the bottom left corner of the image. The cross marks the position adopted as the center of the nebular emission (see Table 1). []{data-label="fig4"}](HB4GREY.PS)
![Flux density of the image slices of Hb 4 as a function of their position angle.[]{data-label="fig5"}](hb4.eps)
In Figure 4 we show the 8.46 GHz image of Hb 4. The 1.4 and 5.0 GHz flux densities of this planetary nebula are 158 mJy (Condon & Kaplan 1998) and 170 mJy (Aaquist & Kwok 1990), respectively. This indicates that the nebula is optically-thin above $\sim$1.4 GHz, so we expect the 8.46 GHz image to be tracing optically-thin free-free emission. In Figure 5 we plot the flux densities of each slice as a function of their position angle. The slices were made, as for NGC 7027, with an angular radius of 10$''$, every $10^\circ$. In contrast to NGC 7027, the flux density per slice of Hb 4 shows a strong modulation. This modulation implies that the nebula is density-bounded at least in the NW and SE position angles, where ionizing photons will be escaping to the interstellar medium. We conclude that this result indicates that Hb 4 is not ionization-bounded in all directions. An estimate of the stellar ionizing flux from indicators like H$\alpha$ or free-free emission will then result in an underestimate of the true value.
Conclusions
===========
We present a simple observational test that allows to establish if an ionized nebula is ionization-bounded or not. The condition is necessary but not sufficient. We applied this test to NGC 7027 and Hb 4, finding that the first object is probably ionization-bounded while the second is density-bounded, at least in some directions.
LFR and YG acknowledge the support of DGAPA, UNAM, and of CONACyT (México). This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France.
Aaquist, O. B., & Kwok, S. 1990, , 84, 229
Bieging, J. H., Boley, P. A., Latter, W. B., & Tielens, A. G. G. M. 2008, , 676, 390
Briggs, D. 1995, Ph.D. thesis, New Mexico Inst. of Mining and Technology
Condon, J. J., & Kaplan, D. L. 1998, , 117, 361
Mellema, G. 2004, A&A, 416, 623
Perinotto, M., Sch[ö]{}nberner, D., Steffen, M., & Calonaci, C. 2004, , 414, 993
Rodríguez, L. F., & Moran, J. M. 1982, , 299, 323
Sandin, C., Sch[ö]{}nberner, D., Roth, M. M., Steffen, M., B[ö]{}hm, P., & Monreal-Ibero, A. 2008, , 486, 545
Sch[ö]{}nberner, D., Jacob, R., & Steffen, M. 2005, , 441, 573
Schraml, J., & Mezger, P. G. 1969, , 156, 269
Thronson, H. A., Jr. 1983, , 264, 599
Treffers, R. R., Fink, U., Larson, H. P., & Gautier, T. N., III 1976, , 209, 793
Zijlstra, A. A., van Hoof, P. A. M., & Perley, R. A. 2008, , 681, 1296
[^1]: The National Radio Astronomy Observatory is operated by Associated Universities Inc. under cooperative agreement with the National Science Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Madeline Brandt, DJ Bruce, Taylor Brysiewicz, Robert Krone, Elina Robeva'
bibliography:
- 'ref.bib'
title: 'The degree of $\operatorname{SO}(n)$'
---
Introduction {#introsect}
============
The *special orthogonal group* $\operatorname{SO}(n,\mathbb{R})$ is the group of automorphisms of $\mathbb{R}^n$ which preserve the standard inner product and have determinant equal to one. The complex special orthogonal group is the complexification of the special orthogonal group and can be thought of more explicitly as the group of matrices $$\operatorname{SO}(n):=\operatorname{SO}(n,{\mathbb{C}})=\left\{M\in \text{Mat}_{n,n}({\mathbb{C}}) \; | \; \det M =1, \quad M^{t}M=\text{Id}\right\}.$$ As these conditions are polynomials in the entries of such a matrix, we view $\operatorname{SO}(n)$ as a complex variety.
Recall that the degree of a complex variety $X$ is the generic number of intersection points of $X$ with a linear space of complementary dimension. Problem $4$ on Grassmannians in [@fitness] asks for a formula for the degree of the of $\operatorname{SO}(n)$. Our primary result is the following theorem, which answers this question completely.
[theorem]{}[maintheorem]{} \[degson\] The degree of $\operatorname{SO}(n)$ is given by $$\deg \operatorname{SO}(n)=2^{n-1}\det \left( {2n-2i-2j}\choose{n-2i} \right)_{1\leq i , j \leq \lfloor{\frac n 2}\rfloor}.$$
Our proof of Theorem \[degson\] uses a formula of Kazarnovskij [@kazarnovskii] (see also Theorem \[Kazarnovskij\]) for the degree of the image of a representation of a connected, reductive, algebraic group over an algebraically closed field. By applying this formula to the case of the standard representation of $\operatorname{SO}(n)$ we are able to express the degree of $\operatorname{SO}(n)$ in terms of its root data and other invariants.
In addition to this result, Theorem \[non-intersecting\] provides a combinatorial interpretation of this degree in terms of non-intersecting lattice paths. In contrast to Theorem \[degson\], the combinatorial statement has the immediate benefit of being obviously non-negative.
Let ${\Bbbk}$ be a field of characteristic zero. We can define $\operatorname{SO}(n,{\Bbbk})$ using the same system of equations since they are defined over the prime field ${\mathbb{Q}}$. For ${\Bbbk}$ that is not algebraically closed, the degree of a variety can be defined in terms of the Hilbert series of its coordinate ring. Since the Hilbert series does not depend on the choice of ${\Bbbk}$, the degree does not either. We choose to work over ${\mathbb{C}}$ not only for simplicity, but also so that we may use the above definition of degree.
Our methods are not specific to $\operatorname{SO}(n)$. The same approach can be used to compute the degree of other algebraic groups. For example, toward the end of Section \[sec:degson\] we provide a similar closed formula for the degree of the symplectic group. This formula is also interpreted combinatorially in Section \[NILP\].
In order to verify Theorem \[degson\], as well as explore the structure of $\operatorname{SO}(n)$ in further depth, it is useful to compute this degree explicitly. We were able to do this for small $n$ using symbolic and numerical computations. A comparison of the success of these two approaches, together with our formula from Theorem \[degson\], is illustrated by the following table.
\[fig:SO\]
$\mathbf{n}$ **Symbolic** **Numerical** **Formula**
-------------- -------------- --------------- -------------
2 2 2 2
3 8 8 8
4 40 40 40
5 384 384 384
6 - 4768 4768
7 - 111616 111616
8 - - 3433600
9 - - 196968448
: Degree of $\operatorname{SO}(n)$ computed in various ways
This project started in the spring of 2014, when Benjamin Recht asked the fifth author to describe the geometry of low-rank semidefinite programming (see Section \[SDP\]). In particular, he asked why the augmented Lagrangian algorithm for solving this problem [@BM] almost always recovers the correct optimum despite the existence of multiple local minima. It quickly became clear that to even compute the number of local extrema, one needs to know the degree of the orthogonal group. In Section \[SDP\] we find a formula for the number of critical points of low-rank semidefinite programming (see Theorem \[thm:LRSDP\]).
The rest of this article is organized as follows. In Section \[backgroundsect\] we give the reader a brief introduction to algebraic groups and state Kazarnovskij’s Theorem. Section \[sec:degson\] proves Theorem \[degson\] by applying Kazarnovskij’s Theorem and simplifying the resulting expressions. After simplification, we are left with a determinant of binomial coefficients which can be interpreted combinatorially using the celebrated Gessel-Viennot lemma which we describe in Section \[NILP\]. The relationship between the degree of $\operatorname{SO}(n)$ and the degree of low-rank semidefinite programming is elaborated upon in Section \[SDP\]. Section \[Computational\] contains descriptions of the symbolic and numerical techniques involved in the explicit computation of $\deg \operatorname{SO}(n)$. Finally, in Section \[reality\] we explore questions involving the real points on $\operatorname{SO}(n)$.
Background {#backgroundsect}
==========
In this section we provide the reader with the necessary language to understand the statement of Kazarnovskij’s Theorem (see Theorem \[Kazarnovskij\]), our main tool for determining the degree of $\operatorname{SO}(n)$. We invite those who already are familiar with Lie theory to skip to the statement of Theorem \[Kazarnovskij\] and continue to Section \[sec:degson\] for our main result. We note, that aside from applying Theorem \[Kazarnovskij\], no understanding of the material in this section is necessary for understanding the remainder of the proof of Theorem \[degson\]. A more thorough treatment of the theory of algebraic groups can be found in [@derksen; @fulton; @humphreys].
An *algebraic group* $G$ is a variety equipped with a group structure such that multiplication and inversion are both regular maps on $G$. When the unipotent radical of $G$ is trivial and $G$ is over an algebraically closed field, we say that $G$ is a *reductive group*. Throughout this section, we let $G$ denote a connected reductive algebraic group over an algebraically closed field ${\Bbbk}$. Let ${{\mathbb{G}}_{\mathrm{m}}}$ denote the multiplicative group of ${\Bbbk}$, so as a set, ${{\mathbb{G}}_{\mathrm{m}}}= {\Bbbk}\setminus \{0\}$. Let $T$ denote a fixed maximal torus of $G$. By *maximal torus*, we mean a subgroup of $G$ isomorphic to ${{\mathbb{G}}_{\mathrm{m}}}^r$ and which is maximal with respect to inclusion. The number $r$ is well-defined and is called the *rank* of $G$. After fixing $T$, we define the *Weyl group* of $G$, denoted $W(G)$, to be the quotient of the normalizer of $T$ by its centralizer, $W(G) = N_G(T)/Z_G(T)$. Like $r$, $W(G)$ does not depend on the choice of $T$ up to isomorphism.
\[maxTori\] We can parametrize $\operatorname{SO}(2,{\mathbb{C}})$ by ${{\mathbb{G}}_{\mathrm{m}}}$ via the map $$\mathbf{R}(t) := \frac{1}{2}\begin{pmatrix}
t+t^{-1} & -i(t-t^{-1}) \\
i(t-t^{-1}) & t+t^{-1}
\end{pmatrix},$$ which is in fact a group isomorphism. (Note that $\mathbf{R}(e^{i\theta})$ is the rotation matrix by angle $\theta$.) Therefore $\operatorname{SO}(2)$ has rank 1.
Fix $r \in \mathbb{N}$. Then $$\begin{aligned}
T_{2r}&:=\left\{
\begin{pmatrix}
\mathbf{R}(t_1) & 0 & 0 & \cdots & 0\\
0 & \mathbf{R}(t_2) & 0 & \cdots & 0 \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & 0 & \cdots & \mathbf{R}(t_r)
\end{pmatrix} \Bigg| t_i \in {{\mathbb{G}}_{\mathrm{m}}}\right\}\cong \operatorname{SO}(2)^{r}\subset \operatorname{SO}(2r)\\
T_{2r+1}&:=\left\{
\begin{pmatrix}
\mathbf{R}(t_1) & 0 & 0 & \cdots & 0 & 0\\
0 & \mathbf{R}(t_2) & 0 & \cdots & 0 & 0\\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots\\
0 & 0 & 0 & \cdots & \mathbf{R}(t_r) & 0\\
0 & 0 & 0 & \cdots & 0 & 1
\end{pmatrix}\; \Bigg| \; t_i \in {{\mathbb{G}}_{\mathrm{m}}}\right\}\cong \operatorname{SO}(2)^{r}\subset \operatorname{SO}(2r+1)\end{aligned}$$ are maximal tori of rank $r$ of their respective groups. Therefore, $\text{rank}(\operatorname{SO}(2r))=\text{rank}(\operatorname{SO}(2r+1))=r$ and we see that the rank of $\operatorname{SO}(n)$ depends fundamentally on the parity of $n$.
The *character group* $M(T)$ is the set of algebraic group homomorphisms from $T$ to ${{\mathbb{G}}_{\mathrm{m}}}$, i.e. group homomorphisms defined by polynomial maps, $$M(T):=\operatorname{Hom}_{{\mathbf{AlgGrp}}}(T,{{\mathbb{G}}_{\mathrm{m}}}).$$ Since $T$ is isomorphic to ${{\mathbb{G}}_{\mathrm{m}}}^r$, all such homomorphisms must be of the form $$(t_1,\ldots,t_r) \mapsto t_1^{a_1}\cdots t_r^{a_r}$$ with $a_1,\ldots,a_r$ integers. This character group is isomorphic to ${\mathbb{Z}}^r$ and for this reason it is often called the character lattice. Dual to this is the *group of 1-parameter subgroups* $$N(T):=\operatorname{Hom}_{{\mathbf{AlgGrp}}}({{\mathbb{G}}_{\mathrm{m}}},T),$$ which is also isomorphic to ${\mathbb{Z}}^r$. Indeed, each 1-parameter subgroup is of the form $t \mapsto (t^{b_1},\ldots,t^{b_r})$ for integers $b_1,\ldots,b_r$. There exists a natural bilinear pairing between $N(T)$ and $M(T)$, given by $$\begin{aligned}
M(T)\times N(T)&\rightarrow \operatorname{Hom}_{{\mathbf{AlgGrp}}}({{\mathbb{G}}_{\mathrm{m}}},{{\mathbb{G}}_{\mathrm{m}}})\cong{\mathbb{Z}}\\
\langle \chi, \sigma \rangle &\mapsto \chi \circ \sigma.\end{aligned}$$
Now if $\rho:G\rightarrow{}\operatorname{GL}(V)$ is a representation of $G$ we attach to it special characters called weights. A *weight* of the representation $\rho$ is a character $\chi\in M(T)$ such that the set $$V_\chi:=\bigcap_{s\in T}\ker(\rho(s)-\chi(s)\operatorname{Id}_V)$$ is non-trivial. This condition is equivalent to saying that all of the matrices in $\{\rho(s) \; | \; s\in T\}$ have a simultaneous eigenvector $v\in V$ such that the associated eigenvalue for $\rho(s)$ is $\chi(s)$. We will use $C_V$ to denote the convex hull of the weights of the representation $\rho$.
\[ex:weight\] An example that will be important for us later will be the standard representation coming from the natural embedding $\rho:\operatorname{SO}(n) \to \operatorname{GL}({\mathbb{C}}^n)$. For any $t \in {{\mathbb{G}}_{\mathrm{m}}}$, the matrix $\mathbf{R}(t) \in \operatorname{SO}(2)$ has eigenvectors $e_1 + ie_2$ and $e_1 - ie_2$ with eigenvalues $t$ and $t^{-1}$ respectively. From the explicit description of $T$ in Example \[maxTori\] we see that the eigenvectors of $\rho(t_1,\ldots,t_r)$ are all vectors of the form $e_{2j-1} \pm ie_{2j}$ with $1\leq j \leq r$ and the eigenvalues are $t_1^{\pm 1},\ldots,t_r^{\pm 1}$. These eigenvalues, viewed as characters, are the weights of $\rho$. Additionally when $n = 2r+1$, we have that $e_{2r+1}$ is an eigenvector with eigenvalue 1, corresponding to the trivial character.
Another representation of a matrix group $G \subseteq \operatorname{End}(V)$ is the *adjoint representation*, $\operatorname{Ad}:G \to \operatorname{GL}(\operatorname{End}(V))$, with $\operatorname{Ad}(g)$ the linear map defined by $A \mapsto gAg^{-1}$. The *roots* of $G$ are the weights of the adjoint representation. Given a linear functional $\ell$ on $M(T)$, we define the *positive roots* of $G$ with respect to $\ell$ to be the roots $\chi$ such that $\ell(\chi)>0$. We denote the positive roots of $G$ by $\alpha_1,\ldots,\alpha_l$. For the algebraic groups in this paper, we can choose $\ell$ to be the inner product with the vector $(r,r-1,\ldots,1)$ so that a root of the form $e_j-e_k$ is positive if and only if $j<k$. To each root $\alpha$, we associate a *coroot* $\check \alpha$, defined to be the linear function $\check \alpha (\vec{x}):= \frac{2\langle\vec{x},\alpha\rangle}{\langle\alpha,\alpha\rangle}$ where $\langle , \rangle$ must be $W(G)$-invariant. Throughout this paper, we fix this to be the standard inner product.
We now compute the roots of $\operatorname{SO}(n)$, starting with $n$ even. It can be shown that the simultaneous eigenvectors of $\operatorname{Ad}(s)$ over all $s \in T$ are matrices $A$ with the following structure. These matrices are zero outside a $2\times 2$ block $B$ in rows $2j-1,2j$ and columns $2k-1,2k$ for some $1\leq j,k \leq r$. Furthermore, $B = v_1v_2^T$ with each $v_k$ equal to one of the eigenvectors of $\mathbf{R}(t)$, $e_1 \pm ie_2$. Indeed, suppose $s \in T$ has blocks along the diagonal $\mathbf{R}(t_j)$ with $t_1,\ldots,t_r \in {{\mathbb{G}}_{\mathrm{m}}}$. Then $\operatorname{Ad}(s)(A)$ will also be zero except in the same $2\times 2$ block, and that block will be $$\mathbf{R}(t_j)B\mathbf{R}(t_k)^T = t_j^{\pm 1}t_k^{\pm 1}B,$$ where the signs in the exponents depend on the choices of $v_1$ and $v_2$. Thus the roots of $\operatorname{SO}(2r)$ are the characters of the form $t_j^{\pm 1}t_k^{\pm 1}$ for $1\leq j,k \leq r$.
In the case that $n$ is odd, $A$ has an extra row and column. Consider $A$ with support only in the last column. Then for $s \in T$, $\operatorname{Ad}(s)(A) = sAs^{-1}$ but $s^{-1}$ acts trivially on the left, while $s$ acts on the last column as an element of $\operatorname{GL}({\mathbb{C}}^n)$ as in the standard representation. As in Example \[ex:weight\] we get weights $t_1^{\pm 1},\ldots,t_r^{\pm 1},1$. The same weights appear for $A$ with support in the last row.
Associated to $G$ is a Lie algebra $\mathfrak{g}$, which comes equipped with a Lie bracket $[\ ,\ ]$. A *Cartan subalgebra* $\mathfrak{h}$ is a nilpotent subalgebra of $\mathfrak{g}$ that is self-normalizing, meaning if $[x,y] \in \mathfrak{h}$ for all $x \in \mathfrak{h}$, then $y \in \mathfrak{h}$. Let $S(\mathfrak{h}^*)$ be the ring of polynomial functions on $\mathfrak{h}$. The Weyl group $W(G)$ acts on $\mathfrak{h}$, and this extends to an action of $W(G)$ on $S(\mathfrak{h}^*)$. The space $S(\mathfrak{h}^*)^{W(G)}$ of polynomials which are invariant up to the action of $W(G)$ is generated by $r$ homogeneous polynomials whose degrees, $c_1+1,\ldots,c_r+1$, are uniquely determined. The values $c_1, \ldots, c_r$ are called *Coxeter exponents*.
We are now prepared to state Kazarnovskij’s theorem.
\[Kazarnovskij\] Let $G$ be a connected reductive group of dimension $m$ and rank $r$ over an algebraically closed field. If $\rho:G\rightarrow{}\operatorname{GL}(V)$ is a representation with finite kernel then, $$\deg\overline{\rho\left(G\right)}=\frac{m!}{|W(G)|(c_1!c_2!\cdots c_r!)^2|\ker(\rho)|}\int_{C_V}(\check{\alpha}_{1}\check{\alpha}_2\cdots\check{\alpha}_l)^2dv.$$ where $W(G)$ is the Weyl group, $c_i$ are Coxeter exponents, $C_V$ is the convex hull of the weights, and $\check \alpha_i$ are the coroots.
If $\rho$ is the standard representation for an algebraic group $G$, then it follows that $\deg \overline{\rho(G)}=\deg G$. Therefore, in order to compute $\deg \operatorname{SO}(n)$, all we must do is apply this theorem for the standard representation of $\operatorname{SO}(n)$. The relevant data for this theorem is given in Table \[data\] below for $\operatorname{SO}(n)$ and $\operatorname{Sp}(n)$.
\[data\]
[ l c c c c c c]{} Group& Dimension& Rank & Positive Roots & Weights & $|W(G)|$ & Coxeter Exponents\
\
$\operatorname{SO}(2r)$& ${2r}\choose{2}$ & $r$ &$\{e_i \pm e_j\}_{ i<j }$ & $\{\pm e_i\}$ & $r!2^{r-1}$ & $1,3,\ldots,2r-3,r-1$\
\
$\operatorname{SO}(2r+1)$ & ${2r+1}\choose{2}$& $r$ & $\{e_i \pm e_{j}\}_{i<j} \cup \{e_i\}$ & $\{\pm e_i\}$ & $r! 2^r$ & $1,3, 5, \ldots, 2r-1$\
\
$\operatorname{Sp}(r)$ & ${2r}\choose{2}$ & $r$ &$\{e_i \pm e_j\}_{i<j }\cup\{2e_i\}$ & $\{\pm e_i\}$ & $r! 2^r$ & $1,3, 5, \ldots, 2r-1$\
\
Main Result: The Degree of $\operatorname{SO}(n)$ {#sec:degson}
=================================================
We now prove our main result, Theorem \[degson\]. At the end of this section we use the same method to obtain a formula for the degree of the symplectic group.
We begin by directly applying Theorem \[Kazarnovskij\] to $\operatorname{SO}(2r)$ and $\operatorname{SO}(2r+1)$ to obtain
$$\begin{aligned}
\label{eqn:even-int}
\deg\operatorname{SO}(2r)&=\frac{\displaystyle \binom{2r}{2}!}{\displaystyle r!2^{r-1}(r-1)!^2\prod_{k=1}^{r-1}(2k-1)!^2} \int_{C_V} \left(\prod_{1\leq i<j\leq r}(x_i^2-x_j^2)^2\right)dv,\\
\label{eqn:odd-int}
\deg\operatorname{SO}(2r +1) &= \frac{\displaystyle \binom{2r +1}{2}!}{r!2^r \displaystyle \prod_{k=1}^r(2k-1)!^2} \int_{C_V} \left(\prod_{1\leq i<j\leq r}(x_i^2-x_j^2)^2\prod_{i=1}^r(2x_i)^2\right)dv.\end{aligned}$$
Thus, to compute the degree of $\operatorname{SO}(n)$ it suffices to find formulas for the integrals above. We do this by first expanding the integrand into monomials, and then integrating the result. We use the well-known expression for the determinant of the Vandermonde matrix, $$\prod_{1\leq i < j\leq r}(y_j-y_i)=\sum_{\sigma\in S_r}\left(\operatorname{sgn}(\sigma)\prod_{i=1}^r y_i^{\sigma(i)-1}\right).$$ Substituting $y_i = x_i^2$ and squaring the entire expression yields $$\label{rewrite-integrand}
\prod_{1\leq i<j\leq r}(x_i^2-x_j^2)^2 = \sum_{\sigma,\tau \in S_r}\left(\operatorname{sgn}(\sigma\tau) \prod_{i=1}^r x_i^{2\sigma(i)+2\tau(i)-4}\right).$$ Additionally, we point out that every variable in the integrand is being raised to an even power and $C_V$ is the convex hull of weights, $\{\pm e_i\}$. Because of this symmetry, the integrals over $C_V$ are $2^r$ times the same integrals over $\Delta_r$, the standard $r$-simplex. We have now reduced the computation of this integral to understanding the integral of any monomial over the standard simplex. The following proposition provides a formula for this.
\[integral-monomial\] Let $\Delta_{r}\subset \mathbb{R}^{r}$ be the standard $r$-simplex. If $\mathbf{a}=(a_1,\ldots,a_r)\in \mathbb{Z}_{>0}^r$ then $$\int_{\Delta_r}\mathbf{x}^\mathbf{a}d\mathbf{x}=\int_{\Delta_r}x_1^{a_1}x_{2}^{a_2}\cdots x_{r}^{a_r}dx_1dx_2\cdots dx_r=
\frac{1}{(r+\sum a_i)!}\prod_{i=1}^r a_i!.$$
We can now get expressions for the integrals in (\[eqn:even-int\]) and (\[eqn:odd-int\]) directly by applying (\[rewrite-integrand\]) and Proposition \[integral-monomial\].
\[two-integrals\] Let $I_{even}(r)$ and $I_{odd}(r)$ denote the integrals in (\[eqn:even-int\]) and (\[eqn:odd-int\]) respectively. Then, $$I_{even}(r) =\frac{r!2^r}{\binom{2r}{2}!}\det\left((2i+2j-4)!\right)_{1\leq i,j\leq r}.$$ $$I_{odd}(r) =\frac{r!2^{3r}}{\binom{2r+1}{2}!}\det\left((2i+2j-2)!\right)_{1\leq i,j\leq r}.$$
As mentioned above, we can compute $I_{odd}$ by considering the integrand only over the simplex. This, along with equation gives us that $$\begin{aligned}
I_{odd}(r) &= 2^r \int_{\Delta_r} \prod_{1\leq i<j\leq r} (x_i^2-x_j^2)^2 \prod_{i=1}^r (2x_i)^2 dv\\
&= 2^r \int_{\Delta_r} \left(\sum_{\sigma,\tau \in S_r}\operatorname{sgn}(\sigma\tau) \prod_{i=1}^r x_i^{2\sigma(i)+2\tau(i)-4}\right) \prod_{i=1}^r(2x_i)^2 dv\\
&= 2^{3r} \sum_{\sigma,\tau \in S_r}\operatorname{sgn}(\sigma\tau)\int_{\Delta_r}\prod_{i=1}^r x_i^{2\sigma(i)+2\tau(i)-2} dv.\end{aligned}$$ As the integrand is homogeneous of degree $4\binom{r}{2}+2r$, applying Proposition \[integral-monomial\] and simplifying yields $$\begin{aligned}
I_{odd}(r)
&= \frac{2^{3r}}{\left(4\binom{r}{2}+3r\right)!} \sum_{\sigma,\tau\in S_r} \operatorname{sgn}(\sigma\tau) \prod_{i=1}^r (2\sigma(i)+2\tau(i)-2)!,\end{aligned}$$ which after replacing $i$ with $\sigma^{-1}(i)$ gives us $$\prod_{i=1}^r(2\sigma(i)+2\tau(i)-2)! = \prod_{i=1}^r(2i+2\tau\sigma^{-1}(i)-2)!.$$ Let $\rho = \tau\sigma^{-1}$. Over all pairs $\sigma,\tau \in S_r$, the permutation $\rho$ appears as each permutation in $S_r$ exactly $r!$ times, and $\operatorname{sgn}(\sigma\tau) = \operatorname{sgn}(\rho)$. Therefore, we have that $$\begin{aligned}
I_{odd}(r)
&= \frac{r!2^{3r}}{\left(4\binom{r}{2}+3r\right)!} \sum_{\rho\in S_r} \operatorname{sgn}(\rho) \prod_{i=1}^r (2i+2\rho(i)-2)!\\
&= \frac{r!2^{3r}}{\binom{2r+1}{2}!} \det\left((2i+2j-2)!\right)_{1\leq i,j\leq r}.\end{aligned}$$ The derivation of $I_{even}$ follows precisely the same steps.
Theorem \[degson\] now follows directly from the subsequent simplification.
$$\begin{aligned}
\deg \operatorname{SO}(2r+1) &= \frac{2^{2r}}{(1!3!\cdots(2r-1)!)^2} \det ((2i+2j-2)!)\\
&=\frac{2^{2r}}{(1!2!\cdots (2r-1)!)} \det\left(\frac{(2i+2j-2)!}{(2i-1)!}\right)\\
&=2^{2r}\det\left(\frac{(2i+2j-2)!}{(2i-1)!(2j-1)!}\right)\\
&=2^{2r}\det\left(\binom{2i+2j-2}{2i-1}\right)_{1\leq i,j\leq r}.\end{aligned}$$
Reversing the order of the rows and columns of the final matrix and reindexing produces the formula given in Theorem \[degson\]. Similarly, for the even case, we have
$$\begin{aligned}
\deg \operatorname{SO}(2r) &= \frac{2}{(1!3!\cdots(2r-3)!(r-1)!)^2} \det ((2i+2j-4)!)\\
&= \frac{2 \cdot (2^{r-1})^2}{(1!3!\cdots(2r-3)!2\cdot 4 \cdots (2r-2))^2} \det ((2i+2j-4)!)\\
&= 2^{2r-1} \det \left(\frac{(2i+2j-4)!}{(2i-2)!(2j-2)!}\right)\\
&=2^{2r-1}\det\left(\binom{4r-2i-2j}{2r-2i}\right)_{1\leq i,j\leq r}.\end{aligned}$$
This finishes the proof of Theorem \[degson\].
Since the orthogonal group $\operatorname{O}(n)$ has two components that are isomorphic to $\operatorname{SO}(n)$, we immediately get a formula for the degree of $\operatorname{O}(n)$.
\[degOn\]
The degree of $\operatorname{O}(n)$ is given by $$\deg \operatorname{O}(n)=2^{n}\det \left( {2n-2i-2j}\choose{n-2i} \right)_{1\leq i , j \leq \lfloor{\frac n 2}\rfloor}.$$
Furthermore, as mentioned in the introduction, there is no reason, [*a priori*]{}, that the steps taken in this section are particular to $\operatorname{SO}(n)$. We now apply these methods to find the degree of $\operatorname{Sp}(r)$, the group of (complex) symplectic matrices.
Recall the *symplectic group* over ${\mathbb{C}}$ is defined to be $$\operatorname{Sp}(r):=\operatorname{Sp}(r,{\mathbb{C}}) = \{ M \in \text{Mat}_{2r,2r}({\mathbb{C}}) \ |\ M^T \Omega M = \Omega\},$$ where $$\Omega =
\begin{pmatrix}
0 & I_r \\
-I_r & 0
\end{pmatrix}.$$
\[thm:symplectic\] The degree of $\operatorname{Sp}(r)$ is given by $$\deg \operatorname{Sp}(r) =\det\left({2i + 2j - 2 \choose 2i-1}\right)_{1\leq i,j\leq r}.$$
For $1 \leq r \leq 5$ the values of $\deg \operatorname{Sp}(r)$ are $2,24,1744,769408,2063048448,\ldots$. This was verified using both numerical and symbolic techniques up to $r=3$.
This is an application of Kazarnovskij’s result which is completely analogous to the computation for the special orthogonal group. The integral is the same as the one for $\operatorname{SO}(2r+1)$ up to factors of 2, so it is evaluated in the same way, and then the expression can be simplified $$\begin{aligned}
\deg \left(\text{Sp}(r)\right) &=\frac{(r(2r+1))!}{r! 2^r (1!3!\cdots(2r-1)!)^2}\int_{C_V} \left(\prod_{1\leq i<j\leq r}(x_i-x_j)^2(x_i+x_j)^2\prod_{i=1}^r x_i^2\right)dv \\
&=\frac{1}{(1!3!\cdots(2r-1)!)^2} \det\left((2i+2j-2)!\right)_{1\leq i,j\leq r}\\
&=\det\left({2i + 2j - 2 \choose 2i-1}\right)_{1\leq i,j\leq r}.\end{aligned}$$
We remark that our formula for $\deg \text{Sp}(r)$ is particularly interesting because the determinant in Theorem \[thm:symplectic\] is the same as the determinant in Theorem \[degson\] when $n=2r+1$.
\[cor:sp\] $$\deg \operatorname{SO}(2r+1)=2^{2r}\deg \text{Sp}(r)$$
Sending the $(i,j)$ entry of the matrix in Theorem \[thm:symplectic\] to the ${(r-i+1,r-j+1)}$ entry does not change the determinant and gives us that $$\deg \text{Sp}(r)=\det\left({4r+2-2i-2j \choose 2r+1-2i}\right)_{1\leq i,j\leq r}.$$ When $n=2r+1$, this is the matrix appearing in Theorem \[degson\] and all that is different is the coefficient in front. Accounting for this coefficient finishes the proof.
Non-Intersecting Lattice Paths {#NILP}
==============================
The formulas given in the previous section for the degrees of $\operatorname{SO}(n),\operatorname{O}(n),$ and $\text{Sp}(r)$ can be interpreted as a count of non-intersecting lattice paths via the Gessel-Viennot Lemma [@GV].
\[GV\] Let $A=\{a_1, \ldots,a_r\}$, $B=\{b_1, \ldots,b_r\}$ be collections of lattice points in $\mathbb{Z}^2$. Let $M_{i,j}$ be the number of lattice paths from $a_i$ to $b_j$ using only unit steps in either the North or East direction. If the only way that a system of these lattice paths from $A \to B$ do not cross each other is by sending $a_i\mapsto b_i$, then the determinant of $M$ equals the number of such non-intersecting lattice paths.
The number of lattice paths from $(0,0)$ to $(i,j)$ is the binomial coefficient ${i+j} \choose i$. Since the matrix involved in the formulas for the degrees of $\operatorname{SO}(n),$ $\operatorname{O}(n),$ and $\text{Sp}(r)$ has binomial coefficients as entries, it is natural to search for a interpretation of its determinant via Gessel-Viennot.
\[non-intersecting\] Let $N(n)$ count the number of non-intersecting lattice paths from $A(n):=\{a_i\}_{i=1}^{\lfloor \frac n 2 \rfloor}$ to $B(n):=\{b_j\}_{j=1}^{\lfloor \frac n 2 \rfloor}$ where $a_i=(2i-n,0)$ and $b_j=(0,n-2j)$. Then $$\begin{aligned}
\deg\operatorname{SO}(n) &= 2^{n-1}N(n), \\
\deg\operatorname{O}(n) &= 2^{n}N(n), \\
\deg\text{Sp}(r) &= N(2r+1).\end{aligned}$$
It is enough to prove this theorem for $\operatorname{SO}(n)$ and apply Corollaries \[degOn\] and \[cor:sp\]. Noticing that the matrix appearing in Theorem \[degson\] is the minor of Pascal’s matrix which skips every other row and every other column up to $\lfloor \frac n 2 \rfloor$ shows that we have a correct point configuration for Gessel-Viennot.
Figure \[fig:GVExample\] computes that $N(5)=24$ by explicitly listing all $24$ non-intersecting lattice paths from $A(5)$ to $B(5)$. Then, according to Theorem \[non-intersecting\], we see that $\deg \operatorname{SO}(5)=2^4 \cdot 24=384$, $\deg \operatorname{O}(5)=2^5 \cdot 24=768$, and $\deg \text{Sp}(2)=24$.
Theorem \[non-intersecting\] suggests a relationship between these non-intersecting lattice paths and the degrees of $\operatorname{SO}(n),\operatorname{O}(n),$ and $\text{Sp}(r)$. Such a direct interpretation could be interesting, and so we pose the question:
\[ourQuestion\] Does Theorem \[non-intersecting\] have a deeper combinatorial interpretation?
Because the formula for the degree of the symplectic group has no coefficient in front of the lattice path count in Theorem \[non-intersecting\], studying the combinatorial meaning of the degree of $\text{Sp}(r)$ may be an ideal starting point to tackle Question \[ourQuestion\].
An Application - The Degree of Low Rank Semidefinite Programming {#SDP}
================================================================
In this section we show how knowing the degree of $\operatorname{SO}(n)$ can be used to compute the number of critical points for a certain optimization problem (cf. Theorem \[thm:LRSDP\]).
Consider the standard formulation of *semidefinite programming* $$\begin{aligned}
\label{primalProblem}
&\text{minimize}_{X\in\mathcal S^n} \quad C\bullet X\notag\\
&\text{such that }\quad A_i\bullet X = b_i, i=1,..., m, \quad X\succeq 0.\end{aligned}$$ Here $\mathcal S^n$ is the set of $n\times n$ real symmetric matrices, $b\in\mathbb Q^m$ is a vector, $C, A_1, ..., A_m\in\mathbb Q\mathcal S^n$ are matrices, and $\bullet$ denotes the trace inner product for matrices: $U\bullet V = \text{trace}(UV)$.
Semidefinite programming can be solved in polynomial time in the size $n$ of the unknown matrix $X$ and in the number of constraints $m$. It is a widely used method in practice, and many NP-hard problems possess semidefinite relaxations [@BV2; @GW2]. However, it is often the case that the size $n$ is very large, and solving exactly can be computationally prohibitive. On the other hand, the rank $r$ of the optimal solution $X^*$ is often much smaller than $n$, and in those cases we can solve more rapidly by replacing $X$ by the low rank positive semidefinite matrix $RR^T$, where $R\in\mathbb R^{n\times r}$. This idea and an algorithm to solve the new problem are due to Burer and Monteiro [@BM]. The problem becomes $$\begin{aligned}
\label{newProblem}
&\text{minimize}_{R\in\mathbb R^{n\times r}} \quad C\bullet (RR^T)\notag\\
&\text{such that }\hspace{0.33cm} A_i\bullet (RR^T) = b_i, ~~i=1, ..., m.\end{aligned}$$ The constraint $X\succeq 0$ is now implicit and the number of variables has decreased from $n^2$ to $nr$. However, the objective function and the constraints are no longer linear; instead, they are quadratic and the feasible set is non-convex. In [@BM] Burer and Monteiro propose a fast algorithm for solving . Despite the existence of multiple local minima, in practice this algorithm quickly finds the global minimum. It starts by choosing the rank $r=1$, and increments it until $C - \sum_{i=1}^my_i A_i\succeq 0$, which ensures that we have arrived at the smallest optimal $r$. For each fixed rank $r$, the optimization problem is non-convex, and its appealing behavior still remains to be examined. In Theorem \[thm:LRSDP\] we give a formula for the number of critical points of this optimization problem. We call a [*critical point*]{} of the optimization problem any point $(R, y)$ which satisfies the Lagrange multipliers equations arising from this problem. Here $y$ is a vector of size $m$, and its entries $y_1,\dots, y_m$ are the new dual variables introduced for the $m$ constraints in (see equation ). Before we state our theorem, we need the following definition.
Let $$\psi_i = 2^{i-1}, \quad \psi_{i,j} = \sum_{k=i}^{j-1}\binom{i+j-2}k \text{ when } i < j,$$ and $$\psi_{i_1,\dots, i_r}= \operatorname{Pf}(\psi_{i_k, i_l})_{1\leq k < l\leq r} \text{ if } r \text{ is even},$$ $$\psi_{i_1,\dots, i_r}= \operatorname{Pf}(\psi_{i_k, i_l})_{0\leq k < l\leq r} \text{ if } r \text{ is odd}$$ where $r>2$, $\psi_{0, k } = \psi_{k}$, and $\operatorname{Pf}$ denotes the Pfaffian. Then, define $\delta(m, n, r)$ as $$\delta(m, n, r) = \sum_{I} \psi_I\psi_{I^c},$$ where the sum runs over all strictly increasing subsequences $I = \{i_1, ..., i_{n-r}\}$ of $\{1, ..., n\}$ of length $n - r$ and such that $i_1 + ... + i_{n-r} = m$.
\[thm:LRSDP\] The number of critical points of the low-rank semidefinite programming algorithm is $$2 (\deg\operatorname{SO}(r)) \delta(m, n, r).$$
The number $\delta(m, n, r)$ is called the [*algebraic degree of semidefinite programming*]{} and was originally defined in [@NRS] as the number of critical points of the original semidefinite programming problem for which the matrix $X$ has rank $r$. The final formula for it was computed in [@VR].
*Proof of Theorem \[thm:LRSDP\]:* In order to analyze the optimality conditions for the program (\[newProblem\]) for a fixed $r$, consider the Lagrangian function $$\begin{aligned}
\label{Lagrangian}
L(R, y) = C\bullet (RR^T) - \sum_{i=1}^my_i(A_i\bullet(RR^T) - b_i).\end{aligned}$$ Taking derivatives, we find out that the critical points $(R, y)$ of this optimization problem are given by the Lagrange multipliers equations: $$\begin{aligned}
\label{equations}
\left(C - \sum_{i=1}^my_iA_i\right)RR^T&= 0\\
A_i\bullet(RR^T) &= b_i, i=1,2,...,m.\notag\end{aligned}$$ In addition, those critical points relevant for applications have to be real and have to satisfy $$\begin{aligned}
\label{addEquation}
\left(C - \sum_{i=1}^my_iA_i\right)\succeq 0,\end{aligned}$$ since this is the constraint in the dual to the optimization problem . However, in this article we are primarily concerned with counting all of the critical points. Analogously, in [@NRS] Nie, Ranestad, and Sturmfels show that the critical points of the original semidefinite programming problem satisfy $$\begin{aligned}
\left(C - \sum_{i=1}^my_iA_i\right)X &= 0,\label{optimalSolutions1}\\
A_i\bullet X &= b_i, i=1,...,m.\label{optimalSolutions2}\end{aligned}$$ In addition, the critical points relevant for applications have to satisfy $$\begin{aligned}
\left(C - \sum_{i=1}^my_iA_i\right)\succeq 0 \text{ and } X&\succeq 0,\label{optimalSolutions3}\end{aligned}$$ but these conditions are disregarded and the total number of critical points is counted. Nie, Ranestad, and Sturmfels show that the number of solutions $(X, y)$ to -, for which the rank of $X$ is $r$, equals $\delta(m, n, r)$ (c.f. Definition 1). Comparing our system of equations to the equations -, we see that the fiber of the map $(R, y) \mapsto (RR^T, y)$ above each point $(X, y)$, satisfying -, consists of all points $(R, y')$, satisfying , and such that $y' = y$ and $X = RR^T$. Given $X$ and one matrix $R$ such that $X = RR^T$, all other matrices $S$ such that $(S, y)$ is in the fiber above $(X, y)$ have the form $S = RU$ where $U$ runs over all orthogonal $r\times r$ matrices. In other words, this fiber is isomorphic to a copy of the orthogonal group $O(r)$. Therefore, the number of solutions to is equal to $2(\deg\operatorname{SO}(r))\delta(m, n, r)$.
The number of critical points of low-rank semidefinite programming grows rapidly with the rank $r$, and the appealing behavior of the augmented Lagrangian algorithm [@BM] still needs to be explained. It would be quite interesting and relevant for applications to examine how many of the critical points computed in Theorem 5 and in [@NRS] are real, and moreover, how many of them satisfy the additional linear matrix inequality constraints and respectively. This is a real algebraic problem and would involve counting polynomial system solutions over semialgebraic sets. This question is addressed more in Section \[reality\].
Computational Methods {#Computational}
=====================
Although we have already derived a formula for the degree of $\operatorname{SO}(n)$, it is natural to want to compute this degree explicitly for particular values of $n$. Aside from merely verifying the formula in Theorem \[degson\], the computation of this degree gives us access to other useful data along the way. In our case, this manifests itself as either a Gröbner basis or a witness set for $\operatorname{SO}(n)$. Once computed, either may be used in further computations (such as those done in Section \[reality\] using witness sets). Additionally, $\operatorname{SO}(n)$ serves as a prime example of when numerical algorithms are better suited for computation than other techniques. Even though our computations focus on $\operatorname{SO}(n)$, these methods are useful for studying many other varieties.
In this section, we describe three techniques which can compute the degree of a variety: a Gröbner basis algorithm, polynomial homotopy continuation, and a numerical monodromy algorithm. The first is symbolic and the last two use numerical algebraic geometry. The results of our symbolic and numerical computations for $\deg \operatorname{SO}(n)$ appear in the first two columns of Table \[fig:SO\]. Code for each method is given in the appendix.
Using Gröbner bases, we were able to compute the degree of $\operatorname{SO}(n)$ for $n \leq 5$. The standard algorithm computes a Gröbner basis for the ideal of $\operatorname{SO}(n)$ over $\mathbb{Q}$ and uses this to produce the Hilbert polynomial. However, since the dimension of $\operatorname{SO}(n)$ grows quadratically in $n$, this method quickly becomes computationally infeasible. Computing a Gröbner basis over a finite field can speed up the computation, but this method is still quite slow.
A common numerical algorithm for computing the degree of $\operatorname{SO}(n)$ over $\mathbb{C}$ randomly chooses an affine linear space $\mathcal L$ of complementary dimension and counts the number of complex solutions $S$ to the zero-dimensional system corresponding to $\operatorname{SO}(n) \cap \mathcal L$. This data is contained in the triple $(\operatorname{SO}(n),\mathcal L, S)$ which is called a *witness set* for $\operatorname{SO}(n)$. This is the fundamental data type in numerical algebraic geometry in the sense that the computation of a witness set is often a necessary step for other numerical algorithms. Such techniques include sampling points on the variety at a rapid rate, studying its asymptotic behaviour, computing its monodromy group, or even studying its real locus, as we do in Section \[reality\]. Both numerical algorithms presented below produce a witness set for $\operatorname{SO}(n)$.
*Polynomial homotopy continuation* computes a witness set by solving a system of polynomials describing these points. Briefly, this method begins with a “start” polynomial system that has similar structure to the “target” system we want to solve, but for which the solutions are obvious. The solutions of the start system are quickly tracked through a homotopy towards those of the target system [@SW]. The most basic start system one uses for this technique has a solution count equal to the product of the degrees of the polynomials in the target system. This number is called the *Bézout bound* and for our case is equal to $2^{n(n+1)/2}$ (for $n=6$, this is already $2097152$). The *polyhedral start system*, however, has a solution count equal to the mixed volume of the Newton polytopes of these polynomials. In our case, this count provides no savings as it is equal to the Bézout bound. Because of how many paths need to be tracked with this method, we were only able to compute the degree of $\operatorname{SO}(n)$ up to $n=5$ with this method, just like with Gröbner bases.
The method that proved to be the most efficient takes advantage of the monodromy group of $\operatorname{SO}(n)$. The basic idea is that if we know some point on a linear cut $W=\mathcal{L} \cap \operatorname{SO}(n)$, we can track this solution from the slice $W$ along some path $\gamma$ to another slice $W'$ using homotopy methods. Tracking this solution along a different path $\gamma'$ back to $W$ then induces a permutation $\sigma_{\gamma,\gamma'}$ on the points in $W$. Therefore, applying this action to a point $x_0 \in W$ will likely produce a new point $\sigma_{\gamma,\gamma'}(x_0) \in W$. One iterates this process hoping to populate the witness set associated to $W$. Other than knowing the degree [*a priori*]{}, stopping criteria for this method tend to be heuristic in nature: one can wait until the algorithm fails to produce new points (suggesting there are no new points to be found) or one can compute a *trace test* [@tracetest] which numerically decides whether or not a witness set is complete. This monodromy method has been implemented in the package *monodromySolver* for *Macaulay2* [@M2] and is explained in much more detail in [@duff2016solving].
A major computational result arising from this project was the computation of witness sets for $\operatorname{SO}(6)$ and $\operatorname{SO}(7)$. This was done in $630$ and $42790$ seconds respectively using *monodromySolver*. The algorithm stopped when no new points were found on ten consecutive iterations.
Real Points on $\operatorname{SO}(n)$ {#reality}
=====================================
An interesting question pertaining to $\operatorname{SO}(n)$ is whether or not this variety always admits some witness set consisting of only real points. Since tracking points of one witness set to those of another is computationally inexpensive via homotopy continuation, we use this method to generate experimental data regarding real points on witness sets of $\operatorname{SO}(3),$ $\operatorname{SO}(4),$ and $\operatorname{SO}(5)$.
The number of coefficients needed to produce a linear cut of $\operatorname{SO}(n)$ is $(n^2+1) {{n}\choose 2}$. We randomly choose these coefficients using the *random* function in *Macaulay2* in order to sample linear cuts of $\operatorname{SO}(n)$. We then use homotopy continuation to track solutions of a precomputed witness set to those lying on the randomly chosen linear cut. Finally, we determine how many solutions in the new cut are real by checking whether each solution is within a $0.001$ numerical tolerance of a real point coordinate-wise. One can certify the results using the software *alphaCertify* which implements Smale’s $\alpha$ theory [@alphaCertified]. For the sake of speed, we chose not to certify all of the results, but instead certify at least one witness set achieving the observed maximum of real points (cf. Table \[fig:SO3Real\], Table \[fig:SO4Real\], and Table \[fig:SO5Real\]).
After computing $1398000$, $1004100$, and $48200$ witness sets for $\operatorname{SO}(3),\operatorname{SO}(4),$ and $\operatorname{SO}(5)$ respectively, we have summarized the number of real solutions found in each witness set in the frequency tables and histograms below. Explicit data and code used can be found in [@dataSite]. Note that very rarely, numerical failures occur because the path that homotopy continuation is being performed over is ill-conditioned (for example, almost singular). These occurrences are also tallied below under “fail”.
-------------------- ------------------------------------------------------------------------ -- -- -- -- -- --
\#(Real Solutions) **[Fail]{}&**[0]{} & **[2]{} & **[4]{} & **[6]{}& **[8]{}&**[Total]{}\
Frequency &2& 285676 & 420049 & 549875 & 127699 & 14699&1398000\
**************
-------------------- ------------------------------------------------------------------------ -- -- -- -- -- --
: Number of real points on witness sets of $\operatorname{SO}(3)$[]{data-label="fig:SO3Real"}
-------------------- -------------- ---------------------------------------------------------------- -- -- -- -- --
\#(Real Solutions) [**Fail**]{} **0 & **2& **4& **6& **[8]{}& **[10]{}\
Frequency &51 &183427 &108273 &132143 &156010 &159630 &124843\
************
-------------------- -------------- ---------------------------------------------------------------- -- -- -- -- --
: Number of real points on witness sets of $\operatorname{SO}(4)$[]{data-label="fig:SO4Real"}
[ |c|c|c|c|c| c | c | c | c | c | c | c | c ||c]{} **[12]{}& **[14]{}& **[16]{}& **[18]{}& **[20]{} &**[22]{} &**[24]{}&**[26]{}&**[28]{}&**[30]{}&**[32]{}&$\cdots$&**[40]{}&**[Total]{}\
**************************
76965 &38243 &16150 &5780 &1897 &510 &145 &23 &9 &1 &0 &$\cdots$ &0 &1004100\
-------------------- -------------- ----------------------------------------------------------------------- -- -- -- -- -- -- -- --
\#(Real Solutions) [**Fail**]{} **0 & **2& **4& **6& **[8]{}& **[10]{}& **[12]{}& **[14]{}& **[16]{}\
Frequency &81 &6162 &2628 &2377 &2306 &2275 &2272 &2275 &2383 &2473\
******************
-------------------- -------------- ----------------------------------------------------------------------- -- -- -- -- -- -- -- --
: Number of real points on witness sets of $\operatorname{SO}(5)$[]{data-label="fig:SO5Real"}
[|c|c|c|c|c|c|c|c|c|c|c|c|c|c]{} **[18]{}& **[20]{}& **[22]{}& **[24]{}& **[26]{}& **[28]{}& **[30]{}& **[32]{}& **[34]{}& **[36]{}& **[38]{}& **[40]{}& **[42]{}& **[44]{}\
****************************
2497 &2527 &2504 &2485 &2280 &2009 &1755 &1644 &1331 &1051 &802 &591 &468 &362\
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --
**[46]{}& **[48]{}& **[50]{}& **[52]{}& **[54]{}& **[56]{}& **[58]{}& **[60]{}& **[62]{} &**[64]{} &**[66]{}&**[68]{}&**[70]{}&**[72]{}&**[74]{}&**[76]{}&**[78]{}&$\cdots$&**[384]{}&**[Total]{}\
235 &150 &118 &60 &44 &21 &16 &8 &4 &3 &1 &0 &0 &0 &0 &2 &0 &$\cdots$ &0 &48200\
**************************************
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --
: Number of real points on witness sets of $\operatorname{SO}(5)$[]{data-label="fig:SO5Real"}
\[fig:so4histogram\]
\[fig:so5histogram\]
In each case, we were able to find a witness set which failed to have any real solutions on it. This is unsurprising as $\operatorname{SO}(n)$ is compact over the real numbers. Despite the fact that all witness sets computed for $\operatorname{SO}(4)$ had fewer than $40$ solutions, and for $\operatorname{SO}(5)$, fewer than $384$, there is little evidence suggesting that a non-trivial upper bound for the number of real solutions on a witness set of $\operatorname{SO}(n)$ exists. We end with a conjecture.
For any $n$, $\operatorname{SO}(n)$ admits some real witness set.
Appendix: Macaulay2 Code {#appendix-macaulay2-code .unnumbered}
========================
This section contains code which computes the degree of $\operatorname{SO}(n)$ for various $n$ using Gröbner bases, polynomial homotopy continuation, *MonodromySolver*, and Theorem \[degson\] respectively. They are all done in *Macaulay2*.
First, we compute the degree of $\operatorname{SO}(5)$ using Gröbner bases. The computation is done over the finite field $\mathbb{Z}_{101}$ for $\operatorname{O}(5)$ and the result is halved to give the degree of $\operatorname{SO}(5)$.
n=5
R = ZZ/101[x_(1,1)..x_(n,n)]
M = genericMatrix(R,n,n)
J = minors(1,M*transpose(M)-id_(R^n))
degOn = degree J
degSOn = degOn//2
Computing the degree of $\operatorname{O}(n)$, rather than $\operatorname{SO}(n)$ directly, is useful because it throws out the polynomial of highest degree in the system. This is especially useful in numerical methods since they perform best with polynomials of low degree.
The code below uses the package *NumericalAlgebraicGeometry* to solve the zero dimensional system given by a linear slice of $\operatorname{O}(3)$. The method *solveSystem* employs the standard method of polynomial homotopy continuation. Again, the answer is halved to give $\deg \operatorname{SO}(3)$.
loadPackage "NumericalAlgebraicGeometry"
n = 3
L = toList apply(
(0,0)..(n-1,n-1), (i,j)->"x"|toString i|toString j
)
R = CC[L]
M = genericMatrix(R,n,n)
B = M*(transpose M) - id_(R^n)
polys = flatten for i from 0 to n-1 list(
for j from i to n-1 list B_(i,j)
)
linearSlice = apply(
binomial(n,2), i->random(1,R)-random(CC)
)
S = solveSystem(polys|linearSlice);
degOn = #S
degSOn = degOn//2
Next, we provide code that computes the degree of $\operatorname{SO}(7)$ using the package *MonodromySolver*. We again do not include the determinant condition, but this time we do *not* need to halve the result. This is because our starting point, the identity matrix, lies on $\operatorname{SO}(7)$ and this method only discovers points on the irreducible component corresponding to our starting point. The linear slices are parametrized by the $t$ and $c$ variables which are varied within the function *monodromySolve* to create monodromy loops. The method stops when ten consecutive loops provide no new points. Although it is possible that this stopping criterion is satisfied prematurely, in our case the program stopped at the correct number, serving as a testament to the practicality of the software and also this stopping criterion.
loadPackage "MonodromySolver"
N=7
d=binomial(N,2)
R=CC[c_1..c_d,t_(1,1,1)..t_(d,N,N)][x_(1,1)..x_(N,N)]
M=genericMatrix(R,N,N)
B=M*transpose(M)-id_(R^N)
polys=flatten for j from 0 to N-1 list(
for k from j to N-1 list B_(j,k)
);
linearSlice=for i from 1 to d list(
c_i+sum(
flatten for j from 1 to N list(
for k from 1 to N list t_(i,j,k)*x_(j,k)
)
)
);
G = polySystem join (polys,linearSlice)
x0coords = flatten entries id_(CC^N)
setRandomSeed 0
(p0, x0) := createSeedPair(G,x0coords)
elapsedTime (V,npaths) =
monodromySolve(G,p0,{x0},NumberOfNodes=>2,NumberOfEdges=>4);
--node1: 111616
--node2: 111616
-- 42790.9 seconds elapsed
Finally, for the mathematician wanting to compute the degree of $\operatorname{SO}(n)$ quickest, we give code that evaluates the formula in Theorem \[degson\].
degSO = method()
degSO(ZZ) := N ->(
n := N//2;
M := matrix for i from 1 to n list (
for j from 1 to n list (
binomial(2*N-2*i-2*j,N-2*i)
)
);
2^(N-1)*(det M)
)
This article was initiated during the Apprenticeship Weeks (22 August-2 September 2016), led by Bernd Sturmfels, as part of the Combinatorial Algebraic Geometry Semester at the Fields Institute.
The authors are very grateful to Jan Draisma for his tremendous help with understanding Kazarnovskij’s formula, and to Kristian Ranestad for many helpful discussions. The authors thank Anton Leykin for performing the computation of $\operatorname{SO}(7)$. The first three authors would also like to thank the Max Planck Institute for Mathematics in the Sciences in Leipzig, Germany for their hospitality where some of this article was completed. The motivation for computing the degree of the orthogonal group came from project that started by the fifth author at the suggestion of Benjamin Recht.
The first author was supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE 1106400. The second author was partially supported by the NSF GRFP under Grant No. DGE-1256259 and the Graduate School and the Office of the Vice Chancellor for Research and Graduate Education at the University of Wisconsin-Madison with funding from the Wisconsin Alumni Research Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The property of superadditivity of the quantum relative entropy states that, in a bipartite system ${\ensuremath{\mathcal H}}_{AB}={\ensuremath{\mathcal H}}_A \otimes {\ensuremath{\mathcal H}}_B$, for every density operator $\rho_{AB}$ one has ${D(\rho_{AB}||\sigma_A\otimes \sigma_B)} \ge {D(\rho_A||\sigma_A)}+{D(\rho_B||\sigma_B)}$. In this work, we provide an extension of this inequality for arbitrary density operators $\sigma_{AB}$. More specifically, we prove that $ \alpha (\sigma_{AB})\cdot {D(\rho_{AB}||\sigma_{AB})} \ge {D(\rho_A||\sigma_A)}+{D(\rho_B||\sigma_B)}$ holds for all bipartite states $\rho_{AB}$ and $\sigma_{AB}$, where $\alpha(\sigma_{AB})= 1+2 \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} \, \sigma_{AB} \, \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} - {\ensuremath{\mathds{1}}}_{AB}}_\infty$.'
address:
- 'Instituto de Ciencias Matemáticas (CSIC-UAM-UC3M-UCM), C/ Nicolás Cabrera 13-15, Campus de Cantoblanco, 28049 Madrid, Spain'
- 'QMATH, Department of Mathematical Sciences, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen, Denmark and NBIA, Niels Bohr Institute, University of Copenhagen, Blegdamsvej 17, 2100 Copenhagen, Denmark'
- 'Departamento de Análisis Matemático, Universidad Complutense de Madrid, 28040 Madrid, Spain and Instituto de Ciencias Matemáticas (CSIC-UAM-UC3M-UCM), C/ Nicolás Cabrera 13-15, Campus de Cantoblanco, 28049 Madrid, Spain'
author:
- Ángela Capel
- Angelo Lucia
- 'David Pérez-García'
title: Superadditivity of quantum relative entropy for general states
---
Introduction and notation {#sec-1}
=========================
The *quantum relative entropy* between two density operators $\rho$ and $\sigma$ in a finite dimensional Hilbert space, ${D(\rho||\sigma)}$, is given by $\tr[\rho (\log \rho - \log \sigma)]$ if $\text{supp}(\rho) \subseteq \text{supp}(\sigma) $ and by $+ \infty$ otherwise[^1]. It constitutes a measure of distinguishability between two quantum states and is a fundamental tool in quantum information theory [@libropetz], [@wilde].
The quantum relative entropy is the quantum analogue of the Kullback-Leibler divergence [@kld], the probabilistic relative entropy. Its origin lies in mathematical statistics, where it is used to measure how much two states differ in the sense of statistical distinguishability. The larger the relative entropy of two states is, the more information for discriminating between the hypotheses associated to them can be obtained from an observation.
One of the main properties of quantum relative entropy is *superadditivity*, which states that in a bipartite system ${\ensuremath{\mathcal H}}_{AB}={\ensuremath{\mathcal H}}_A \otimes {\ensuremath{\mathcal H}}_B$ one has: $$\label{superadditivity}
{D(\rho_{AB}||\sigma_A\otimes \sigma_B)} \ge {D(\rho_A||\sigma_A)}+{D(\rho_B||\sigma_B)}$$ for all $\rho_{AB}$, where we use the standard notation $\rho_A=\tr_B[\rho_{AB}]$ and $\tr_B$ is the partial trace.
Since (Proposition \[prop:sigmaprod\]) $${D(\rho_{AB}||\sigma_A\otimes \sigma_B)} -{D(\rho_A||\sigma_A)}-{D(\rho_B||\sigma_B)}= {D(\rho_{AB}||\rho_A\otimes \rho_B)},$$ is equivalent to the fact that the mutual information $I_\rho(A:B):= {D(\rho_{AB}||\rho_A\otimes \rho_B)}$ is always non-negative, a fact that appears ubiquitously in quantum information theory.
In the form , superadditivity of the quantum relative entropy has found applications in e.g. quantum thermodynamics [@gallego], statistical physics [@libropetz Chapter 13] or hypothesis testing [@petz]. Indeed, as proven recently in [@axcharRE] (building on results from [@matsumoto]), the property of superadditivity, along with the properties of *continuity* with respect to the first variable, *monotonicity* and *additivity* (Proposition \[prop:REprop\]), characterizes axiomatically the quantum relative entropy.
The main aim of this work is to provide a quantitative extension of (\[superadditivity\]) for an arbitrary density operator $\sigma_{AB}$. Note that for all $\rho_{AB}$ and $\sigma_{AB}$, as a consequence of monotonicity of the quantum relative entropy for the partial trace, the following holds: $$\label{monotonicity}
2{D(\rho_{AB}||\sigma_{AB})}\ge {D(\rho_A||\sigma_A)} + {D(\rho_B||\sigma_B)}.$$
Therefore we aim to give a constant $\alpha (\sigma_{AB}) \in [1,2]$ at the LHS of (\[superadditivity\]) that measures how far $\sigma_{AB}$ is from $\sigma_A \otimes \sigma_B$.
Following [@clasico] we will consider as $\alpha(\sigma_{AB})-1$ the distance from ${\ensuremath{\mathds{1}}}$ to “$\sigma_{AB}$ [*multiplied by the inverse of $\sigma_A\otimes\sigma_B$*]{}”. In the case in which $\sigma_{AB}$ and $\sigma_A\otimes\sigma_B$ commute there is a unique way to define this: $\sigma_{AB} \,(\sigma_A^{-1}\otimes \sigma_B^{-1})$. In the non-commutative case, however, there are many possible ways to define the multiplication by the inverse. The one we will take in the result below is a symmetric analogue of the commutative case, $(\sigma_A^{-1/2}\otimes \sigma_B^{-1/2}) \, \sigma_{AB} \, (\sigma_A^{-1/2}\otimes \sigma_B^{-1/2})$. Another one that will appear in the proof of this result is the derivative of the matrix logarithm on $\sigma_A\otimes \sigma_B$ evaluated on $\sigma_{AB}$, $\mathcal{T}_{\sigma_A\otimes \sigma_B}(\sigma_{AB})$, whose explicit equivalent expressions shown in [@lieb] and [@sutter] will be presented later.
\[thm:quasifactorizationAB\] For any bipartite states $\rho_{AB},\sigma_{AB}$: $$\label{eq:superadditivity}
(1+2\|H(\sigma_{AB})\|_{\infty}){D(\rho_{AB}||\sigma_{AB})}\ge {D(\rho_A||\sigma_A)} + {D(\rho_B||\sigma_B)},$$ where
$H(\sigma_{AB}) = \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} \, \sigma_{AB} \, \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} - {\ensuremath{\mathds{1}}}_{AB}$,
and ${\ensuremath{\mathds{1}}}_{AB} $ denotes the identity operator in ${\ensuremath{\mathcal H}}_{AB}$.
Note that $H(\sigma_{AB})=0$ if $\sigma_{AB}=\sigma_A\otimes \sigma_B$.
This result constitutes an improvement to (\[monotonicity\]) whenever $ \|H (\sigma_{AB}) \|_{\infty} \leq 1/2 $ (and, hence, $ 1+2\|H (\sigma_{AB}) \|_{\infty} \leq 2 $). Then, it is likely to be relevant for situations where it is natural to assume $\sigma_{AB} \sim \sigma_A \otimes \sigma_B$. This is the case of (quantum) many body systems where such property is expected to hold for spatially separated regions $A, B$ in the Gibbs state above the critical temperature. Indeed, a classical version of Theorem \[thm:quasifactorizationAB\] proven by Cesi [@cesi] and Dai Pra, Paganoni and Posta [@clasico], was the key step to provide the arguably simplest proof of the seminal result of Martinelli and Olivieri [@marti-oliv] connecting the decay of correlations in the Gibbs state of a classical spin model with the mixing time of the associated Glauber dynamics, via a bound on the log-Sobolev constant.
Notation
--------
We consider a finite dimensional Hilbert space ${\mathcal{H}}$. We denote the set of bounded linear operators acting on ${\mathcal{H}}$ by ${\mathcal{B}}= {\mathcal{B}}({\ensuremath{\mathcal H}})$ (whose elements we denote by lowercase Latin letters: $f,g$...), and its subset of Hermitian operators by ${\mathcal{A}}\subseteq {\mathcal{B}}$ (whose elements we call *observables*). The set of positive semidefinite Hermitian operators is denoted by ${\mathcal{A}}^+$. We also denote the set of density operators by ${\mathcal{S}}= \qty{f \in {\mathcal{A}}^+ \, : \, \tr[f]=1} $ (whose elements we also call *states* and denote by lowercase Greek letters: $\sigma, \rho$...).
A linear map $\mathcal{T}: {\mathcal{B}}\rightarrow {\mathcal{B}}$ is called a *superoperator*. We say that a superoperator $\mathcal{T}$ is *positive* if it maps positive operators to positive operators. Moreover, we denote $\mathcal{T}$ as *completely positive* if $\mathcal{T} \otimes {\ensuremath{\mathds{1}}}: \mathcal{B} \otimes \mathcal{M}_n \rightarrow \mathcal{B} \otimes \mathcal{M}_n $ is positive for every $n \in {\mathbb{N}}$, where $ \mathcal{M}_n $ is the space of complex $n \times n$ matrices. We also say that $\mathcal{T}$ is *trace preserving* if $\tr[\mathcal{T}(f)]= \tr[f]$ for all $f \in {\mathcal{B}}$. Finally, if $\mathcal{T}$ verifies all these properties, i.e., is a completely positive and trace preserving map, it is called a *quantum channel* (for more information on this topic, see [@wolf]).
We denote by $\norm{\cdot}_1$ the trace norm $\left( \norm{f}_1 = \tr[\sqrt{f^* f}] \right)$ and by $\norm{\cdot}_\infty$ the operator norm ${\displaystyle}\left( \norm{f}_\infty = \text{sup} \qty{ \norm{f(x)}_{{\ensuremath{\mathcal H}}} : \norm{x}_{{\ensuremath{\mathcal H}}}=1 }\right)$. In the following section, we will make use of this (Hölder) inequality [@bhatia]: $$\norm{f g}_1 \leq \norm{f}_1 \norm{g}_\infty \text{ \phantom{asfs} for every }f, g \in {\mathcal{B}}.$$
In most of the paper, we consider a bipartite finite dimensional Hilbert space ${\mathcal{H}}_{AB}= {\mathcal{H}}_A \otimes {\mathcal{H}}_B$. When this is the case, we use the previous notation placing the subindex ${AB}$ (resp. $A$, $B$) in each of the previous sets to denote that the operators considered act on ${\ensuremath{\mathcal H}}_{AB}$ (resp. ${\ensuremath{\mathcal H}}_A$, ${\ensuremath{\mathcal H}}_B$). There is a natural inclusion of ${\mathcal{A}}_{A}$ in ${\mathcal{A}}_{AB}$ by identifying ${\mathcal{A}}_{A} = {\mathcal{A}}_{A} \otimes {\ensuremath{\mathds{1}}}_{B} $.
Relative entropy
----------------
Let ${\ensuremath{\mathcal H}}$ be a finite dimensional Hilbert space, $f,g\in {\mathcal{A}}^+$, $f$ verifying $\tr[f] \neq 0$. The *quantum relative entropy* of $f$ and $g$ is defined by [@umegaki]: $${D(f||g)} = \frac{1}{\tr[ f]}\tr \left[ f (\log f - \log g) \right].$$
In most of the paper we only consider density matrices (with trace $1$). Let $\rho , \sigma \in {\mathcal{S}}$. In this case, the quantum relative entropy is given by: $${D(\rho||\sigma)} = \tr \left[ {\rho} (\log \rho - \log \sigma) \right].$$
In the following proposition, we collect some well-known properties of the relative entropy, which will be of use in the following section.
\[prop:REprop\]\
Let ${\ensuremath{\mathcal H}}_{AB}$ be a bipartite finite dimensional Hilbert space, ${\ensuremath{\mathcal H}}_{AB} = {\ensuremath{\mathcal H}}_A \otimes {\ensuremath{\mathcal H}}_B$. Let $\rho_{AB}, \sigma_{AB} \in {\mathcal{S}}_{AB}$. The following properties hold:
1. **Non-negativity.** ${D(\rho_{AB}||\sigma_{AB})} \geq 0$ and ${D(\rho_{AB}||\sigma_{AB})}=0 \Leftrightarrow {\rho_{AB}}=\sigma_{AB}$.
2. **Finiteness.** ${D(\rho_{AB}||\sigma_{AB})} < \infty $ if, and only if, $\text{supp}(\rho_{AB}) \subseteq \text{supp}(\sigma_{AB})$, where supp stands for support.
3. **Monotonicity.** ${D(\rho_{AB}||\sigma_{AB})} \geq {D(T(\rho_{AB})||T(\sigma_{AB}))}$ for every quantum channel $T$.
4. **Additivity.** ${D(\rho_A \otimes \rho_B||\sigma_A \otimes \sigma_B)}= {D(\rho_A||\sigma_A)} + {D(\rho_B||\sigma_B)}$.
These properties, especially the property of non-negativity, allow to consider the relative entropy as a measure of separation of two states, even though, technically, it is not a distance (with its usual meaning), since it is not symmetric and lacks a triangle inequality.
Let us prove now the property of superadditivity, whenever $\sigma_{AB}= \sigma_A \otimes \sigma_B$.
\[prop:sigmaprod\]
Let ${\ensuremath{\mathcal H}}_{AB}={\ensuremath{\mathcal H}}_A \otimes {\ensuremath{\mathcal H}}_B$ and $\rho_{AB}, \sigma_{AB} \in {\mathcal{S}}_{AB}$. If $\sigma_{AB}= \sigma_A \otimes \sigma_B$, then
${D(\rho_{AB}||\sigma_{AB})} = I_\rho(A:B) + {D(\rho_A||\sigma_A)}+{D(\rho_B||\sigma_B)}$,
where $I_\rho(A:B)={D(\rho_{AB}||\rho_A \otimes \rho_B)}$ is the mutual information [@shannon].
As a consequence,
${\displaystyle}{D(\rho_{AB}||\sigma_{A}\otimes \sigma_B )} \geq {D(\rho_A||\sigma_A)}+{D(\rho_B||\sigma_B)}$.
Since $\sigma_{AB}= \sigma_A \otimes \sigma_B$, we have $$\begin{aligned}
{D(\rho_{AB}||\sigma_{A} \otimes \sigma_{B})} &=& \tr[\rho_{AB}(\log \rho_{AB} - \log \sigma_{A} \otimes \sigma_{B})]\\
&=& \tr[\rho_{AB} (\log \rho_{AB} - \log \rho_{A} \otimes \rho_{B} +\log \rho_{A} \otimes \rho_{B} - \log \sigma_{A} \otimes \sigma_{B})] \\
&=& {D(\rho_{AB}||\rho_A \otimes \rho_B)} + {D(\rho_A \otimes \rho_B|| \sigma_{A} \otimes \sigma_{B})}\\
&=& I_\rho (A:B) + {D(\rho_A||\sigma_A)} + {D(\rho_B||\sigma_B)}.\end{aligned}$$
Now, since $I_\rho (A:B)$ is a relative entropy, it is greater or equal than zero (property 1 of Proposition \[prop:REprop\]), so
${\displaystyle}{D(\rho_{AB}||\sigma_{A}\otimes \sigma_B )} \geq {D(\rho_A||\sigma_A)}+{D(\rho_B||\sigma_B)}$.
We prove now a lemma for observables (non necessarily of trace $1$) which yields a relation between the relative entropy of two observables and the relative entropy of some dilations of each of them. In particular, it is a useful tool to express the relative entropy of two observables in terms of the relative entropy of their normalizations (i.e., the quotient of each of them by their trace).
\[lemma:rel-entropy-homogeneity\] Let ${\ensuremath{\mathcal H}}$ be a finite dimensional Hilbert space and let $f,g\in {\mathcal{A}}^+$ such that $\tr[f] \neq 0$. For all positive real numbers $a$ and $b$, we have: $${D(af||bg)} = {D(f||g)} + \log \frac{a}{b}.$$
$$\begin{aligned}
{D(af||bg)} &=& \frac{1}{a \tr f} \left(a \tr \left[ f \left( \log a f - \log bg \right) \right] \right) \\
&=& \frac{1}{\tr f} \qty(\tr[ f \log a] + \tr[ f \log f] - \tr[ f \log b] - \tr[f \log g]) \\
&=&\frac{1}{\tr f} (\tr[ f \left( \log f - \log g \right) ] ) + \log a - \log b \\
&=& {D(f||g)} + \log \frac{a}{b},\end{aligned}$$
where, in the first and third equality, we are using the linearity of the trace, and we are denoting $\log a {\ensuremath{\mathds{1}}}$ by $\log a$ for every $a\geq 0$.
Since the relative entropy of two density matrices is non-negative (property 1 of Proposition \[prop:REprop\]), we have the following corollary:
\[lemma:cond-ent\] Let ${\ensuremath{\mathcal H}}$ be a finite dimensional Hilbert space and let $f,g\in {\mathcal{A}}^+$ such that $\tr[f] \neq 0$ and $\tr[g] \neq 0$. Then, the following inequality holds: $$\label{eq:cond-ent}
{D(f||g)} \ge - \log \frac{\tr[ g]}{\tr[ f]}.$$
Since $f/\tr[f]$ and $g / \tr[g]$ are density matrices, we have that
${D(f/\tr[f] \,||\, g / \tr[g])} \ge 0 $,
and we can apply Lemma \[lemma:rel-entropy-homogeneity\]:
${\displaystyle}0 \leq {D(f/\tr[f] \,||\, g / \tr[g])} = {D(f||g)} + \log \frac{\tr[g]}{\tr[f]} $.
Proof of main result {#sec:results}
====================
We divide the proof of Theorem \[thm:quasifactorizationAB\] in four steps.
In the first step, we provide a lower bound for the relative entropy of $\rho_{AB}$ and $\sigma_{AB}$ in terms of ${D(\rho_{A}||\sigma_{A})}$, ${D(\rho_{B}||\sigma_{B})} $ and an error term, which we will further bound in the following steps.
\[step:1\] For density matrices $\rho_{AB}, \sigma_{AB} \in {\mathcal{S}}_{AB}$, it holds that $$\label{eq:step-1}
{D(\rho_{AB}||\sigma_{AB})} \ge {D(\rho_{A}||\sigma_{A})} + {D(\rho_{B}||\sigma_{B})} - \log \tr M,$$ where $ M = \exp \bqty{ \log \sigma_{AB} - \log \sigma_{A} \otimes \sigma_{B} +\log \rho_{A} \otimes \rho_{B} } $ and equality holds (both sides being equal to zero) if $\rho_{AB} =
\sigma_{AB}$.\
Moreover, if $\sigma_{AB} = \sigma_A \otimes \sigma_B$, then $\log \tr M =0$.
It holds that: $$\begin{aligned}
{\displaystyle}{D(\rho_{AB}||\sigma_{AB})} & - & \left[ {D(\rho_{A}||\sigma_{A})} + {D(\rho_{B}||\sigma_{B})} \right]=\\
&=& {D(\rho_{AB}||\sigma_{AB})} - {D(\rho_{A} \otimes \rho_{B}||\sigma_{A} \otimes \sigma_{B})} \\
&=& \tr \left[ {\rho_{AB}} \left( \log {\rho_{AB}} - \underbrace{ \left( \, \log {\sigma_{AB}} - \log \sigma_{A} \otimes \sigma_{B} +\log \rho_{A} \otimes \rho_{B} \, \right) }_{\log M} \right) \right] \\
& =& {D(\rho_{AB}||M)},\end{aligned}$$ where $M$ is defined as in the statement of the step and in the first equality we have used the fourth property of Proposition \[prop:REprop\].
We can now apply Corollary \[lemma:cond-ent\] to obtain that
${D(\rho_{AB}||M)} = \tr[ {\rho_{AB}} (\log {\rho_{AB}} - \log M) ] \ge - \log \tr M$.
It is easy to check, given the definition of $M$, that $M=\sigma_{AB}$ if $\rho_{AB} = \sigma_{AB}$, so both sides are equal to zero in this case.
Also, if $\sigma_{AB}= \sigma_A \otimes \sigma_B$, $M$ is equal to $\rho_A\otimes \rho_B$. In both cases we have $\log \tr M = 0$.
Our target now is to bound the error term, $\log \tr M$, in terms of the relative entropy of $\rho_{AB}$ and $\sigma_{AB}$ times a constant which depends only on $A$, $B$ and $\sigma_{AB}$, and represents how far $\sigma_{AB}$ is from being a tensor product. In the second step of the proof, we will bound this term by the trace of the product of a term which contains this ‘distance’ between $\sigma_{AB}$ and $\sigma_{A} \otimes \sigma_{B}$ and another term which depends on $\rho_{AB}$ and not on $\sigma_{AB}$. However, before that, we need to introduce some concepts and results.
First, we recall the Golden-Thompson inequality, proven independently in [@golden] and [@thompson] (and extended to the infinite dimensional case in [@GTinfdim] and [@GTinfdim2]), which says that for Hermitian operators $f$ and $g$, $$\tr[e^{f+g}] \leq \tr[e^f e^g],$$ where we denote by $e^f$ the exponential of $f$, given by
${\displaystyle}e^f:= \ov{\infty}{\un{k=0}{\sum}} \frac{f^k}{k!} $.
The trivial generalization of the Golden-Thompson inequality to three operators instead of two in the form $\tr[e^{f+g+h}] \leq \tr[e^f e^g e^h]$ is false, as Lieb mentioned in [@lieb]. However, in the same paper, he provides a correct generalization of this inequality for three operators. This result has recently been extended by Sutter et al. in [@sutter] via de so-called multivariate trace inequalities (see also the subsequent paper by Wilde [@multioperator2], where similar inequalities are derived following the statements of [@multioperator1]).
\[thm:Lieb\] Let $f, g$ be positive semidefinite operators, and recall the definition of $\mathcal{T}_g$: $$\mathcal T_g(f) = \int_0^\infty \dd{t} (g+t)^{-1} f (g+t)^{-1} .$$ $\mathcal T_g$ is positive semidefinite if $g$ is. We have that $$\tr[ \exp(-f+g+h)] \le \tr[ e^h \mathcal T_{e^f}(e^g)].$$
This superoperator $\mathcal T_g$ provides a pseudo-inversion of the operator $g$ with respect to the operator where it is evaluated. In particular, if $f$ and $g$ commute, it is exactly the standard inversion, as we can see in the following corollary.
If $f$ and $g$ commute, then $$\mathcal T_g(f) = f \int_0^\infty \dd{t} (g + t)^{-2} = f g^{-1},$$ and therefore $$\tr[ \exp(-f+g+h) ]\le \tr[ e^h e^{-f} e^g ] = \tr[ e^h e^{-f+g}] .$$ This shows that Lieb’s theorem is really a generalization of Golden-Thompson inequality.
We use an alternative definition of this superoperator to obtain a necessary tool for the proof of Step \[step:2\]. In [@sutter Lemma 3.4], Sutter, Berta and Tomamichel prove the following result:
For ${f} $ a positive semidefinite operator and $g $ a Hermitian operator the following holds:
${\displaystyle}\mathcal{T}_g(f)= \int_{-\infty}^\infty dt \, \beta_0 (t) \, g^{\frac{-1-it}{2}} \,f \, g^{\frac{-1+it}{2}} ,$
with
${\displaystyle}\beta_0(t)= \frac{\pi}{2} (\cosh(\pi t)+ 1)^{-1} $.
Using this expression for $\mathcal{T}_{\sigma_A \otimes \sigma_B} (\sigma_{AB})$, we can prove the following result, which is a quantum version of a result used in [@clasico].
\[lemma:remark\] For every operator $O_{A} \in {\mathcal{B}}_A$ and $O_{B} \in {\mathcal{B}}_B$ the following holds:
${\displaystyle}\tr[ L(\sigma_{AB}) \, \sigma_{A} \otimes O_{B} ]= \tr[ L(\sigma_{AB}) \, O_{A} \otimes
\sigma_{B} ]= 0,$
where
${\displaystyle}L(\sigma_{AB}) = \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB}$.
We only prove
$\tr[ L(\sigma_{AB}) \, \sigma_{A} \otimes O_{B} ]=0$,
since the other equality is completely analogous. $$\begin{aligned}
& \tr[ L(\sigma_{AB}) \, \sigma_{A} \otimes O_{B} ]=\\
&\phantom{asdad} = \tr[ \left( \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB}\right) \sigma_{A} \otimes O_{B}]\\
&\phantom{asdad}=\tr[ \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) \sigma_{A} \otimes O_{B}]- \tr[\sigma_{A} \otimes O_{B}] \\
&\phantom{asdad}= \tr[\int_{- \infty}^\infty dt \, \beta_0 (t) \left( \sigma_A \otimes \sigma_B \right)^{\frac{-1-it}{2}} \, \sigma_{AB} \, \left( \sigma_A \otimes \sigma_B \right)^{\frac{-1+it}{2}} \, \sigma_{A} \otimes O_{B}] - \tr[O_B]\\
&\phantom{asdad}= \int_{- \infty}^\infty dt \, \beta_0 (t) \tr[ \sigma_A^{\frac{-1-it}{2}} \otimes \sigma_B^{\frac{-1-it}{2}} \, \sigma_{AB} \, \sigma_A^{\frac{-1+it}{2}} \otimes \sigma_B^{\frac{-1+it}{2}} \, \sigma_{A} \otimes O_{B}]-\tr[O_{B}],\end{aligned}$$ because $\tr[\sigma_A]=1$, the integral commutes with the trace, $\beta_0(t)$ is a scalar for every $t \in {\mathbb{R}}$ and the exponent in the power of a tensor product can be split into both terms.
Now, since the trace is cyclic and using the fact that any operator in ${\ensuremath{\mathcal H}}_B$ commutes with every operator in ${\ensuremath{\mathcal H}}_A$, we have: $$\begin{aligned}
& \tr[ L(\sigma_{AB}) \, \sigma_{A} \otimes O_{B} ]=\\
&\phantom{asdad}= \int_{- \infty}^\infty dt \, \beta_0 (t) \tr[ \sigma_{AB} \, \sigma_A^{\frac{-1+it}{2}} \otimes \sigma_B^{\frac{-1+it}{2}} \, \sigma_{A} \otimes O_{B} \, \sigma_A^{\frac{-1-it}{2}} \otimes \sigma_B^{\frac{-1-it}{2}}]-\tr[O_{B}]\\
&\phantom{asdad}= \int_{- \infty}^\infty dt \, \beta_0 (t) \tr[ \sigma_{AB} \, \left( \sigma_A^{\frac{-1+it}{2}} \sigma_{A} \, \sigma_A^{\frac{-1-it}{2}} \right) \otimes \left( \sigma_B^{\frac{-1+it}{2}} O_{B} \,\sigma_B^{\frac{-1-it}{2}} \right)]-\tr[O_{B}]\\
&\phantom{asdad}= \int_{- \infty}^\infty dt \, \beta_0 (t) \tr[ \sigma_{AB} \, {\ensuremath{\mathds{1}}}_A \otimes \left( \sigma_B^{\frac{-1+it}{2}} O_{B} \,\sigma_B^{\frac{-1-it}{2}} \right)]-\tr[O_{B}]\\
&\phantom{asdad}= \int_{- \infty}^\infty dt \, \beta_0 (t) \tr[ \sigma_{B} \, \sigma_B^{\frac{-1+it}{2}} O_{B} \,\sigma_B^{\frac{-1-it}{2}} ]-\tr[O_{B}]\\
&\phantom{asdad}= \int_{- \infty}^\infty dt \, \beta_0 (t) \tr[ \sigma_B^{\frac{-1-it}{2}} \, \sigma_{B} \, \sigma_B^{\frac{-1+it}{2}} O_{B} ]-\tr[O_{B}]\\
&\phantom{asdad} = \tr[ O_{B}] \int_{- \infty}^\infty dt \, \beta_0 (t) -\tr[O_{B}]\\
&\phantom{asdad}= 0,\end{aligned}$$ where we have used
${\displaystyle}\int_{- \infty}^\infty dt \, \beta_0 (t)=1 $,
and the fact that, for every $f_A \in {\mathcal{B}}_A$ and $g_{AB} \in {\mathcal{S}}_{AB}$, the following holds:
${\displaystyle}\tr[f_A \otimes {\ensuremath{\mathds{1}}}_B \, g_{AB}] = \tr[f_A \, g_A] $.
We are now in position to develop the second step of the proof.
\[step:2\] With the same notation of \[step:1\], we have that $$\log \tr M \le \tr[L(\sigma_{AB}) \left( \rho_{A} - \sigma_A \right) \otimes \left( \rho_{B} - \sigma_B \right) ],$$ where
${\displaystyle}L(\sigma_{AB})= \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB}$.
We apply Lieb’s theorem to the error term of inequality (\[eq:step-1\]): $$\begin{aligned}
\tr M &=&
\tr\left[ \exp( \underbrace{\log \sigma_{AB} }_{g} -
\underbrace{\log \sigma_{A} \otimes \sigma_{B} }_{f} +
\underbrace{\log \rho_{A} \otimes \rho_{B} }_{h} ) \right] \\
&\leq & \tr[ \rho_{A} \otimes \rho_{B} \mathcal T_{\sigma_{A} \otimes \sigma_{B}} ( \sigma_{AB})]\\
&=& \tr[ \rho_{A} \otimes \rho_{B} \underbrace{\left( \mathcal T_{\sigma_{A} \otimes \sigma_{B}} (\sigma_{AB}) -{\ensuremath{\mathds{1}}}_{AB} \right)}_{L(\sigma_{AB})}] + \underbrace{\tr[ \rho_{A} \otimes \rho_{B}]}_1,\end{aligned}$$ where we are adding and substracting $\rho_A \otimes \rho_B$ inside the trace in the last equality.
Now, using the fact $\log(x)\le x-1$, we have
$\log \tr M \leq \tr M - 1 \leq \tr[ L(\sigma_{AB}) \, \rho_{A} \otimes \rho_{B} ]$.
Finally, in virtue of Lemma \[lemma:remark\], it is clear that
$ {\displaystyle}\tr[ L(\sigma_{AB}) \, \rho_{A} \otimes \rho_{B} ]= \tr[L(\sigma_{AB}) \left( \rho_{A} - \sigma_A \right) \otimes \left( \rho_{B} - \sigma_B \right) ] $.
Therefore,
$ \log \tr M \leq \tr[L(\sigma_{AB}) \left( \rho_{A} - \sigma_A \right) \otimes \left( \rho_{B} - \sigma_B \right) ]$.
Notice that if $\sigma_{AB}= \sigma_A \otimes \sigma_B$, then $\mathcal T_{\sigma_{A} \otimes \sigma_{B}} ( \sigma_{AB})= \left( \sigma_{A} \otimes \sigma_{B} \right)^{-1} \sigma_{A} \otimes \sigma_{B}= {\ensuremath{\mathds{1}}}_{AB} $, so $L(\sigma_{AB})=0$.
In the third step of the proof, we need to bound $\tr[L(\sigma_{AB}) \left( \rho_{A} - \sigma_A \right) \otimes \left( \rho_{B} - \sigma_B \right) ]$ in terms of the relative entropy of $\rho_{AB}$ and $\sigma_{AB}$ times a constant depending only on $L(\sigma_{AB})$ (since $L(\sigma_{AB})$ represents how entangled $\sigma_{AB}$ is between the regions $A$ and $B$). The first well-known result we will use in this step is Pinsker’s inequality [@csiszar; @pinsker].
\[thm:pinsker\] For $\rho_{AB}$ and $\sigma_{AB}$ density matrices, it holds that $$\norm{\rho_{AB}-\sigma_{AB}}_1^2 \le 2 {D(\rho_{AB}||\sigma_{AB})}.$$
This result will be of use at the end of the proof to finally obtain the relative entropy in the right-hand side of the desired inequality. However, it is important to notice the different scales of the ${\mathbb{L}}^1$-norm of the difference between $\rho_{AB}$ and $\sigma_{AB}$ and the relative entropy of $\rho_{AB}$ and $\sigma_{AB}$ in Pinsker’s inequality. Since we are interested in obtaining the relative entropy with exponent one, we will need to increase the degree of the term with the trace we already have and from which we will construct an ${\mathbb{L}}^1$-norm (since, for the moment, its degree is one). We will see later that the fact that in $ \tr[L (\sigma_{AB}) \left( \rho_{A} - \sigma_A \right) \otimes \left( \rho_{B} - \sigma_B \right) ]$ we have $ \left( \rho_{A} - \sigma_A \right) \otimes \left( \rho_{B} - \sigma_B \right) $ split into two regions, the multiplicativity of the trace with respect to tensor products and the monotonicity of the relative entropy play a decisive role in the proof.
Another important fact that we notice in the left-hand side of Pinsker’s inequality is that there is a difference between two states (in fact, the ones appearing in the relative entropy). This justifies the use of Lemma \[lemma:remark\] at the end of Step \[step:2\], to obtain something similar to the difference between $\rho_{AB}$ and $\sigma_{AB}$.
We are now ready to prove the third step in the proof of Theorem \[thm:quasifactorizationAB\].
\[step:3\] With the notation of Theorem \[thm:quasifactorizationAB\], $$\tr[L (\sigma_{AB}) \left( \rho_{A} - \sigma_A \right) \otimes \left( \rho_{B} - \sigma_B \right) ]\le 2 \norm{L(\sigma_{AB})}_\infty {D(\rho_{AB}||\sigma_{AB})}.$$
We use the multiplicativity with respect to tensor products of the trace norm and Hölder’s inequality between the trace norm and the operator norm. Thus, $$\begin{aligned}
\tr[ L(\sigma_{AB}) \, (\rho_{A} - \sigma_{A})\otimes(\rho_{B}-\sigma_{B}) ]&\le & \norm{L(\sigma_{AB})}_\infty \norm{ \left( \rho_{A} - \sigma_{A} \right) \otimes \left( \rho_{B}-\sigma_{B} \right) }_1\\
&=& \norm{L(\sigma_{AB})}_\infty \norm{\rho_{A} - \sigma_{A}}_1 \norm{\rho_{B}-\sigma_{B}}_1.\end{aligned}$$
Finally, Pinsker’s inequality (Theorem \[thm:pinsker\]) implies that
${\displaystyle}\norm{\rho_{A} - \sigma_{A}}_1 \leq \sqrt{ 2 {D(\rho_{A}||\sigma_{A})} }, \phantom{sadda} \norm{\rho_{B} - \sigma_{B}}_1 \leq \sqrt{ 2 {D(\rho_{B}||\sigma_{B})} } $.
Therefore,
${\displaystyle}\norm{\rho_{A} - \sigma_{A}}_1 \norm{\rho_{B}-\sigma_{B}}_1\le
2 \sqrt{{D(\rho_{A}||\sigma_{A})} {D(\rho_{B}||\sigma_{B})}} \le 2 {D(\rho_{AB}||\sigma_{AB})},$
where in the last inequality we have used monotonicity of the relative entropy with respect to the partial trace (Proposition \[prop:REprop\]).
If we now put together Steps \[step:1\], \[step:2\] and \[step:3\], we obtain the following expression $$(1 + 2 \norm{L(\sigma_{AB})}_\infty) {D(\rho_{AB}||\sigma_{AB})}\ge {D(\rho_A||\sigma_A)} + {D(\rho_B||\sigma_B)},$$ with
${\displaystyle}L(\sigma_{AB}) = \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB}$.
This inequality already constitutes a quantitative extension of (\[superadditivity\]) for arbitrary density operators $\sigma_{AB}$ in the sense that if $\sigma_{AB}$ is a tensor product between $A$ and $B$, we recover the usual superadditivity, and in general $\norm{L(\sigma_{AB})}_\infty$ measures how far $\sigma_{AB}$ is from $\sigma_A \otimes \sigma_B$.
In the fourth and final step of the proof, we bound $\norm{L(\sigma_{AB})}_\infty$ by
$\norm{ \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} \, \sigma_{AB} \, \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} - {\ensuremath{\mathds{1}}}_{AB}}_\infty$,
a quantity from which the closeness to $0$ whenever $\sigma_{AB}$ is near from being a tensor product is directly deduced. It also has some physical interpretation in quantum many body systems that will be discussed after proving Step \[step:4\].
First, we need to introduce the setting of non-commutative ${\mathbb{L}}^p$ spaces with a $\rho$-weighted norm [@kosaki]. The central property of these non-commutative ${{\mathbb{L}}}^p$ spaces is that they are equipped with a *weighted norm* which, for a full rank state $\rho\in {\mathcal{S}}_{AB}$, is given by
${\displaystyle}\norm{f}_{{\mathbb{L}}^p(\rho)} := \tr[\abs{\rho^{1/2p} f \rho^{1/2p}}^p]^{1/p} \, $ for every $\,f \in {\mathcal{A}}_{AB}$.
Analogously, the $\rho$-*weighted inner product* is given by
${\displaystyle}\left\langle f, g \right\rangle_\rho := \tr[\sqrt{\rho} f \sqrt{\rho} g] \, $ for every $\,f,g \in {\mathcal{A}}_{AB}$.
Some fundamental properties of these spaces are collected in the following proposition.
\[prop:noncomLp\]
Let $\rho \in {\mathcal{S}}_{AB}$. The following properties hold for $\rho$-weighted norms:
1. **Order.** $\forall p,q \in [1, \infty)$, with $p \leq q$, we have $\norm{f}_{{\mathbb{L}}^p(\rho)} \leq \norm{f}_{{\mathbb{L}}^q(\rho)} \, \forall f \in {\mathcal{A}}_{AB}$.
2. **Duality.** $\forall f \in {\mathcal{A}}_{AB}$, we have $\norm{f}_{{\mathbb{L}}^p(\rho)} = \sup \qty{ \left\langle g, f \right\rangle_\rho, g \in {\mathcal{A}}_{AB}, \norm{g}_{{\mathbb{L}}^q(\rho)} \leq 1 }$ for $1/p+1/q=1$.
3. **Operator norm.** $\forall f \in {\mathcal{A}}_{AB}$, we have $\norm{f}_{{\mathbb{L}}^\infty(\rho)} = \norm{f}_\infty$, the usual operator norm.
Another tool we will use in the proof of Step \[step:4\] is the following result.
\[lemma:contract\] Consider $\rho \in {\mathcal{S}}_{AB}$ and let $T$ be a quantum channel verifying $T^*(\rho)=\rho$, where $T^*$ denotes the dual of $T$ with respect to the Hilbert-Schmidt scalar product. Then, $T$ is contractive between ${\mathbb{L}}^1(\rho)$ and ${\mathbb{L}}^1(\rho)$, i.e., the following inequality holds for every $X \in {\mathcal{B}}_{AB}$: $$\norm{T(X)}_{{\mathbb{L}}^1(\rho)} \leq \norm{X}_{{\mathbb{L}}^1(\rho)}.$$
Using the property of duality for the $\rho$-weighted norms of ${\mathbb{L}}^p$-spaces (property 2 of Proposition \[prop:noncomLp\]), we can write: $$\begin{aligned}
{\displaystyle}\norm{T(X)}_{{\mathbb{L}}^1(\rho)} &= \un{\norm{Y}_{{\mathbb{L}}^\infty(\rho)} \leq 1}{\text{sup}} \tr[T(X)\,\rho^{1/2} \, Y \, \rho^{1/2}] \\
&= \un{\norm{Y}_\infty \leq 1}{\text{sup}} \tr[T(X)\,\rho^{1/2} \, Y \, \rho^{1/2}] \\
&= \un{- {\ensuremath{\mathds{1}}}\leq Y \leq {\ensuremath{\mathds{1}}}}{\text{sup}} \tr[T(X)\, \rho^{1/2} \, Y \, \rho^{1/2}],\end{aligned}$$ where in the first step we have used the fact that, for every $\rho \in {\mathcal{S}}_{AB}$, $\norm{\cdot}_{{\mathbb{L}}^\infty(\rho)}$ coincides with the operator norm.
Recalling now that $T^*$ is the dual of $T$ with respect to the Hilbert-Schmidt scalar product, we have: $$\begin{aligned}
{\displaystyle}\tr[T(X)\, \rho^{1/2} \, Y \, \rho^{1/2}] &= \tr[X\, T^*( \rho^{1/2} \, Y \, \rho^{1/2})] \\
&= \tr[X\, \rho^{1/2} \, \rho^{-1/2} \, T^*( \rho^{1/2} \, Y \, \rho^{1/2}) \, \rho^{-1/2} \, \rho^{1/2} ].\end{aligned}$$
Since we are considering the supremum over the observables verifying $-{\ensuremath{\mathds{1}}}\leq Y \leq {\ensuremath{\mathds{1}}}$, if we apply to these inequalitites $T^*(\rho^{1/2} \cdot \rho^{1/2})$, we have $-\rho \leq T^*(\rho^{1/2} \, Y \, \rho^{1/2}) \leq \rho $ (because of the assumption $T^*(\rho)= \rho$).
Hence, if we denote $Z= \rho^{-1/2} \, T^*( \rho^{1/2} \, Y \, \rho^{1/2}) \, \rho^{-1/2}$, it is clear that whenever $ -{\ensuremath{\mathds{1}}}\leq Y \leq {\ensuremath{\mathds{1}}}$, also $-{\ensuremath{\mathds{1}}}\leq Z \leq {\ensuremath{\mathds{1}}}$. Therefore, $$\begin{aligned}
\norm{T(X)}_{{\mathbb{L}}^1(\rho)} &= \un{- {\ensuremath{\mathds{1}}}\leq Y \leq {\ensuremath{\mathds{1}}}}{\text{sup}} \tr[T(X)\, \rho^{1/2} \, Y \, \rho^{1/2}] \\
&= \un{- {\ensuremath{\mathds{1}}}\leq Y \leq {\ensuremath{\mathds{1}}}}{\text{sup}} \tr[X\, \rho^{1/2} \, \rho^{-1/2} \, T^*( \rho^{1/2} \, Y \, \rho^{1/2}) \, \rho^{-1/2} \, \rho^{1/2} ] \\
&\leq \un{- {\ensuremath{\mathds{1}}}\leq Z \leq {\ensuremath{\mathds{1}}}}{\text{sup}} \tr[X\, \rho^{1/2} \, Z \, \rho^{1/2} ] \\
&= \norm{X}_{{\mathbb{L}}^1(\rho)} ,\end{aligned}$$ where the last equality comes again from the property of duality of weighted ${\mathbb{L}}^p$-norms.
In the proof of the previous lemma we have made strong use of the property of duality of ${\mathbb{L}}^p(\rho)$. Indeed, considering the ${\mathbb{L}}^1(\rho)$-norm as dual of the operator norm, has been essential to obtain the desired result. Using similar tools, we can now prove the last step in the proof of Theorem \[thm:quasifactorizationAB\].
\[step:4\] With the notation of the previous steps, we have $$\norm{L(\sigma_{AB})}_\infty \leq \norm{ \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} \, \sigma_{AB} \, \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} - {\ensuremath{\mathds{1}}}_{AB}}_\infty.$$
The strategy we follow in this proof is the opposite to the one used in the previous lemma, i.e., we study now the ${\mathbb{L}}^\infty(\sigma_A \otimes \sigma_B)$-norm as the dual of the ${\mathbb{L}}^1(\sigma_A \otimes \sigma_B)$-norm. Since $\norm{\cdot}_{{\mathbb{L}}^\infty (\rho_{AB})}$ coincides with the usual $\infty$-norm (operator norm) for every $\rho_{AB} \in {\mathcal{S}}_{AB}$, we can write
${\displaystyle}\norm{L(\sigma_{AB})}_\infty = \norm{\mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB}}_{{\mathbb{L}}^\infty(\sigma_A \otimes \sigma_B)}$.
Using the aforementioned property of duality for the $\sigma_A \otimes \sigma_B$-weighted norms of ${\mathbb{L}}^p$-spaces, we have: $$\begin{aligned}
& \norm{\mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB}}_{{\mathbb{L}}^\infty(\sigma_A \otimes \sigma_B)} = \\
& \phantom{asdasddad}=\un{\norm{O_{AB}}_{{\mathbb{L}}^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \left\langle O_{AB} , \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB} \right\rangle_{\sigma_A \otimes \sigma_B} \\
&\phantom{asdasddad} =\un{\norm{O_{AB}}_{{\mathbb{L}}^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \tr[ (\sigma_A \otimes \sigma_B)^{1/2} \, O_{AB}\, (\sigma_A \otimes \sigma_B)^{1/2} \left( \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB} \right)] \\
&\phantom{asdasddad} =\un{\norm{O_{AB}}_{{\mathbb{L}}^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \left( \underbrace{\tr[ \sigma_A^{1/2} \otimes \sigma_B^{1/2} \, O_{AB}\, \sigma_A^{1/2} \otimes \sigma_B^{1/2} \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) ]}_{R} \right. \\
& \left. \phantom{asdasddadaszxczzxdasdasdads} - \underbrace{\tr[\sigma_A^{1/2} \otimes \sigma_B^{1/2} \, O_{AB}\, \sigma_A^{1/2} \otimes \sigma_B^{1/2} ] }_{S} \right).\end{aligned}$$
Let us analyze the terms $R$ and $S$ separately. For $R$, we have: $$\begin{aligned}
R&= \tr[ \sigma_A^{1/2} \otimes \sigma_B^{1/2} \, O_{AB}\, \sigma_A^{1/2} \otimes \sigma_B^{1/2} \mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) ] \\
& = \tr[ (\sigma_A \otimes \sigma_B)^{1/2} \, O_{AB}\, (\sigma_A \otimes \sigma_B)^{1/2} \int_{- \infty}^\infty dt \, \beta_0 (t) \left( \sigma_A \otimes \sigma_B \right)^{\frac{-1-it}{2}} \, \sigma_{AB} \, \left( \sigma_A \otimes \sigma_B \right)^{\frac{-1+it}{2}} ] \\
& = \tr[ O_{AB}\, \int_{- \infty}^\infty dt \, \beta_0 (t) \left( \sigma_A \otimes \sigma_B \right)^{\frac{-it}{2}} \, \sigma_{AB} \, \left( \sigma_A \otimes \sigma_B \right)^{\frac{it}{2}} ] \\
& =\int_{- \infty}^\infty dt \, \beta_0 (t) \tr[O_{AB} \, \left( \sigma_A \otimes \sigma_B \right)^{\frac{-it}{2}} \, \sigma_{AB} \, \left( \sigma_A \otimes \sigma_B \right)^{\frac{it}{2}}] \\
& =\int_{- \infty}^\infty dt \, \beta_0 (t) \tr[\left( \sigma_A \otimes \sigma_B \right)^{\frac{it}{2}} \; O_{AB} \, \left( \sigma_A \otimes \sigma_B \right)^{\frac{-it}{2}} \, \sigma_{AB} ] \\
& = \tr[ \sigma_{AB} \underbrace{ \int_{- \infty}^\infty dt \, \beta_0 (t) \left( \sigma_A \otimes \sigma_B \right)^{\frac{it}{2}} \, O_{AB} \, \left( \sigma_A \otimes \sigma_B \right)^{\frac{-it}{2}}}_{\widetilde{O}_{AB}} ],\end{aligned}$$ where in the third and last equality we have used the fact that the integral and the trace commute, and the fourth equality is due to the cyclicity of the trace. We have also defined:
${\displaystyle}\widetilde{O}_{AB}:= \int_{- \infty}^\infty dt \, \beta_0 (t) \left( \sigma_A \otimes \sigma_B \right)^{\frac{it}{2}} \, O_{AB} \, \left( \sigma_A \otimes \sigma_B \right)^{\frac{-it}{2}}$.
If we were able to express $S$ in terms of $\widetilde{O}_{AB}$, we could simplify the expression that appears in the supremum above. We can do that in the following way:
$$\begin{aligned}
S& = \tr[\sigma_A^{1/2} \otimes \sigma_B^{1/2} \, O_{AB}\, \sigma_A^{1/2} \otimes \sigma_B^{1/2} ] \\
&= \tr[\sigma_A^{1/2} \otimes \sigma_B^{1/2} \, O_{AB}\, \sigma_A^{1/2} \otimes \sigma_B^{1/2} \int_{- \infty}^\infty dt \, \beta_0 (t) ] \\
&= \int_{- \infty}^\infty dt \, \beta_0 (t) \tr[\sigma_A^{1/2} \otimes \sigma_B^{1/2} \, O_{AB}\, \sigma_A^{1/2} \otimes \sigma_B^{1/2} ] \\
&= \int_{- \infty}^\infty dt \, \beta_0 (t) \tr[(\sigma_A\otimes \sigma_B) \, (\sigma_A\otimes \sigma_B)^{\frac{it}{2}} \, O_{AB}\, (\sigma_A\otimes \sigma_B)^{\frac{-it}{2}} ] \\
&= \tr[(\sigma_A\otimes \sigma_B) \int_{- \infty}^\infty dt \, \beta_0 (t) (\sigma_A\otimes \sigma_B)^{\frac{it}{2}} \, O_{AB}\, (\sigma_A\otimes \sigma_B)^{\frac{-it}{2}} ] \\
&= \tr[(\sigma_A\otimes \sigma_B) \, \widetilde{O}_{AB} ],\end{aligned}$$
where we have used again the properties of cyclicity of the trace and commutativity of the integral and the trace.
Placing now the values for $R$ and $S$ that we have just computed in the supremum of the first part of the proof, we have: $$\begin{aligned}
\norm{\mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB}}_{{\mathbb{L}}^\infty(\sigma_A \otimes \sigma_B)} &= \un{\norm{O_{AB}}_{L^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \left( \tr[\sigma_{AB} \, \widetilde{O}_{AB} ] - \tr[\sigma_A \otimes \sigma_B \, \widetilde{O}_{AB} ] \right) \\
&= \un{\norm{O_{AB}}_{L^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \tr[\, \widetilde{O}_{AB} \left( \sigma_{AB} - \sigma_A \otimes \sigma_B \right) ] .\end{aligned}$$
This expression looks much simpler than the one we had before. However, we need to prove that $\norm{\widetilde{O}_{AB}}_{L^1(\sigma_A \otimes \sigma_B)} \leq 1$ in order to see $\widetilde{O}_{AB}$ as one of the terms where the supremum is taken. Indeed, if we consider the map $T: {\mathcal{A}}_{AB} \rightarrow {\mathcal{A}}_{AB}$ given by
${\displaystyle}O_{AB} \mapsto \int_{- \infty}^\infty dt \, \beta_0 (t) (\sigma_A\otimes \sigma_B)^{\frac{it}{2}} \, O_{AB}\, (\sigma_A\otimes \sigma_B)^{\frac{-it}{2}} $,
it is clearly a quantum channel and also verifies $T^*(\sigma_A \otimes \sigma_B)= \sigma_A \otimes \sigma_B$. Hence, in virtue of Lemma \[lemma:contract\], we have
${\displaystyle}\norm{\widetilde{O}_{AB}}_{{\mathbb{L}}^1(\sigma_A \otimes \sigma_B)} \leq \norm{O_{AB}}_{{\mathbb{L}}^1(\sigma_A \otimes \sigma_B)} $,
and, therefore,
${\displaystyle}\un{\norm{O_{AB}}_{L^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \tr[\widetilde{O}_{AB} \left( \sigma_{AB} - \sigma_A \otimes \sigma_B \right) ] \leq \un{\norm{\Omega_{AB}}_{L^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \tr[\Omega_{AB} \left( \sigma_{AB} - \sigma_A \otimes \sigma_B \right) ]. $
In this last supremum over elements of $1$-norm, we can undo the previous transformations in order to obtain again an $\infty$-norm. First, we need to write the term in the supremum as a $\sigma_A \otimes \sigma_B$-product of two terms: $$\begin{aligned}
& \tr[\Omega_{AB} \left( \sigma_{AB} - \sigma_A \otimes \sigma_B \right) ] = \\
& \phantom{asdasdd} = \tr[ (\sigma_A \otimes \sigma_B)^{1/2} \, (\sigma_A \otimes \sigma_B)^{-1/2} \, \sigma_{AB} \, (\sigma_A \otimes \sigma_B)^{-1/2} \, (\sigma_A \otimes \sigma_B)^{1/2} \, \Omega_{AB} ] \\
& \phantom{asdasddas} - \tr[ (\sigma_A \otimes \sigma_B)^{1/2} \, \Omega_{AB} \, (\sigma_A \otimes \sigma_B)^{1/2} ] \\
& \phantom{asdasdd} = \left\langle \Omega_{AB} , (\sigma_A \otimes \sigma_B)^{-1/2} \, \sigma_{AB} \, (\sigma_A \otimes \sigma_B)^{-1/2} \right\rangle_{\sigma_A \otimes \sigma_B} -\left\langle \Omega_{AB} , {\ensuremath{\mathds{1}}}_{AB} \right\rangle_{\sigma_A \otimes \sigma_B}\\
& \phantom{asdasdd} = \left\langle \Omega_{AB} , (\sigma_A \otimes \sigma_B)^{-1/2} \, \sigma_{AB} \, (\sigma_A \otimes \sigma_B)^{-1/2} -{\ensuremath{\mathds{1}}}_{AB} \right\rangle_{\sigma_A \otimes \sigma_B}. \end{aligned}$$
Finally, using again the property of duality for the norms of ${\mathbb{L}}^1(\sigma_A \otimes \sigma_B)$ and $ {\mathbb{L}}^\infty(\sigma_A \otimes \sigma_B)$, we have: $$\begin{aligned}
& \un{\norm{\Omega_{AB}}_{L^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \tr[\Omega_{AB} \left( \sigma_{AB} - \sigma_A \otimes \sigma_B \right) ] \\
& \phantom{asdasdad} = \un{\norm{\Omega_{AB}}_{L^1(\sigma_A \otimes \sigma_B)} \leq 1}{\text{sup}} \left\langle \Omega_{AB} , (\sigma_A \otimes \sigma_B)^{-1/2} \, \sigma_{AB} \, (\sigma_A \otimes \sigma_B)^{-1/2} - {\ensuremath{\mathds{1}}}_{AB} \right\rangle_{\sigma_A \otimes \sigma_B} \\
& \phantom{asdasdad} = \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} \, \sigma_{AB} \, \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} - {\ensuremath{\mathds{1}}}_{AB}}_{{\mathbb{L}}^{\infty}(\sigma_A \otimes \sigma_B)} \\
& \phantom{asdasdad}= \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} \, \sigma_{AB} \, \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} - {\ensuremath{\mathds{1}}}_{AB}}_{\infty},\end{aligned}$$ where we have used again the fact that $\norm{\cdot}_{{\mathbb{L}}^\infty (\rho_{AB})}$ coincides with the usual $\infty$-norm for every $\rho_{AB} \in {\mathcal{S}}_{AB}$.
In conclusion,
${\displaystyle}\norm{\mathcal{T}_{\sigma_{A} \otimes \sigma_{B}} \left(\sigma_{AB} \right) - {\ensuremath{\mathds{1}}}_{AB}}_{\infty} \leq \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} \, \sigma_{AB} \, \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} - {\ensuremath{\mathds{1}}}_{AB}}_{\infty}. $
By putting together Step \[step:1\], Step \[step:2\], Step \[step:3\] and Step \[step:4\], we conclude the proof of Theorem \[thm:quasifactorizationAB\].
This result constitutes an extension of the superadditivity property, i.e., the constant $H(\sigma_{AB})$ that appears in the statement of the main theorem is $0$ when $\sigma_{AB}=\sigma_A \otimes \sigma_B$ and is small whenever $\sigma_{AB} \sim \sigma_A \otimes \sigma_B$. A trivial upper bound can be found with respect to the trace distance as follows, $$\begin{aligned}
& \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} \, \sigma_{AB} \, \sigma_A^{-1/2} \otimes \sigma_B^{-1/2} - {\ensuremath{\mathds{1}}}_{AB}}_{\infty} = \\
& \phantom{asdasdaad} = \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} ( \sigma_{AB} - \sigma_A \otimes \sigma_B) \sigma_A^{-1/2} \otimes \sigma_B^{-1/2}}_\infty \\
& \phantom{asdasdaad} \leq \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} ( \sigma_{AB} - \sigma_A \otimes \sigma_B) \sigma_A^{-1/2} \otimes \sigma_B^{-1/2}}_1 \\
& \phantom{asdasdaad} \leq \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} }_\infty \norm{\sigma_{AB} - \sigma_A \otimes \sigma_B}_1 \norm{\sigma_A^{-1/2} \otimes \sigma_B^{-1/2} }_\infty \\
& \phantom{asdasdaad} \leq \sigma_{\text{min}}^{-2} \, \norm{\sigma_{AB} - \sigma_A \otimes \sigma_B}_1 .\end{aligned}$$
The term $\norm{H(\sigma_{AB})}_\infty$ is also closely related to certain forms of *decay of correlations* of states that have already appeared in quantum many body systems, such as *LTQO (Local Topological Quantum Order)* [@spiros], or the concept of *local indistinguishability* as a strengthened form of *weak clustering* in [@kast-brand].
Let us suppose that $\norm{H(\sigma_{AB})}_\infty \leq \lambda(\ell)$ for a certain small scalar $\lambda(\ell)$ that decays sufficiently fast as a function of the distance $\ell$ between regions $A$ and $B$ in a many body system, and denote by $\left\langle f \right\rangle_\varphi$ the expected value of an observable $f \in {\mathcal{A}}_{AB} $ with respect to a state $\varphi$ (usually the ground or thermal state of the system). Then, for every observable of the form $O_A \otimes O_B \geq 0$, if the reduced density matrix on $AB$ of $\varphi$ is $\sigma_{AB}$, the previous condition can be rewritten as
${\displaystyle}\abs{\left\langle O_A O_B \right\rangle_\varphi - \left\langle O_A \right\rangle_\varphi \left\langle O_B \right\rangle_\varphi } \leq \lambda \left\langle O_A \right\rangle_\varphi \left\langle O_B \right\rangle_\varphi $.
One can now compare this expression with the definition of decay of correlations
${\displaystyle}\abs{\left\langle O_A O_B \right\rangle_\varphi - \left\langle O_A \right\rangle_\varphi \left\langle O_B \right\rangle_\varphi } \leq \lambda(\ell) \norm{O_A}_\infty \norm{O_B}_\infty$,
or LTQO
${\displaystyle}\abs{\left\langle O_A O_B \right\rangle_\varphi - \left\langle O_A \right\rangle_\varphi \left\langle O_B \right\rangle_\varphi } \leq \lambda(\ell) \left\langle O_A \right\rangle_\varphi \norm{O_B}_\infty$.
Conclusion
==========
In this work, we have proven an extension of the property of superadditivity of the quantum relative entropy for general states. Our result constitutes an improvement to the usual lower bound for the relative entropy of two bipartite states, given by the property of monotonicity, in terms of the relative entropies in the two constituent spaces, whenever the second state is near to be a tensor product. Therefore, it might be relevant for situations where this property is expected to hold, such as quantum many body systems, in which it is likely that the Gibbs state satisfies this property in spatially separated systems.
In [@kast-brand], Kastoryano and Brandao proved, for certain Gibbs samplers, the existence of a positive spectral gap for the dissipative dynamics, via a quasi-factorization result of the variance. This provides a bound for the mixing time of the evolution of the semigroup that drives the system to thermalization which is polynomial in the system size. We leave for future work the possibility of using the result of the present paper to obtain a quasi-factorization of the relative entropy in quantum many body systems, which could allow us to prove, under some conditions of decay of correlations on the Gibbs state, the existence of a positive log-Sobolev constant, obtaining an exponential improvement in the bound for the mixing time obtained in [@kast-brand].
Acknowledgment {#acknowledgment .unnumbered}
==============
We are very grateful to D. Sutter and M. Tomamichel, who detected an error in a previous version of the paper. We also thank M. Junge for fruitful discussions. AC and DPG acknowledge support from MINECO (grant MTM2014-54240-P), from Comunidad de Madrid (grant QUITEMAD+- CM, ref. S2013/ICE-2801), and the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 648913). AC is partially supported by a La Caixa-Severo Ochoa grant (ICMAT Severo Ochoa project SEV-2011-0087, MINECO). AL acknowledges financial support from the European Research Council (ERC Grant Agreement no 337603), the Danish Council for Independent Research (Sapere Aude) and VILLUM FONDEN via the QMATH Centre of Excellence (Grant No. 10059). This work has been partially supported by ICMAT Severo Ochoa project SEV-2015-0554 (MINECO).
[30]{}
<span style="font-variant:small-caps;">R. Bhatia</span>, Matrix Analysis, *Springer Science $\&$ Business Media* **169** (1997), doi:10.1007/978-1-4612-0653-8.
<span style="font-variant:small-caps;">M. Breitenbecker and H.R. Gruemm</span>, Note on trace inequalities, *Commun. Math. Phys.* **26** (1972), 276-279, doi:10.1007/BF01645522.
<span style="font-variant:small-caps;">F. Cesi</span>, Quasi-factorization of the entropy and logarithmic Sobolev inequalities for Gibbs random fields, *Probab. Theory Related Fields* **120** (2001), 569-584, doi:10.1007/PL00008792.
<span style="font-variant:small-caps;">I. Csiszár</span>, Information-type measures of difference of probability distributions and indirect observations, *Stud. Sci. Math. Hung.* **2** (1967), 299-318.
<span style="font-variant:small-caps;">P. Dai Pra, A.M. Paganoni and G. Posta</span>, Entropy inequalities for unbounded spin systems, *Ann. Probab.* **30** (2002), 1959-1976, doi:10.1214/aop/1039548378.
<span style="font-variant:small-caps;">F. Dupuis and M.M. Wilde</span>, Swiveled Rényi entropies, *Quantum Inf. Process.* **15**(3) (2016), 1309-1345, doi:10.1007/s11128-015-1211-x.
<span style="font-variant:small-caps;">R. Gallego, J. Eisert and H. Wilming</span>, Thermodynamic work from operational principles, *New J. Phys.* **18** (2016), 103017, doi:10.1088/1367-2630/18/10/103017.
<span style="font-variant:small-caps;">S. Golden</span>, Lower Bounds for the Helmholtz Function, *Phys. Rev., Series II,* **137** (1965), B1127-B1128, doi:10.1103/PhysRev.137.B1127.
<span style="font-variant:small-caps;">F. Hiai and D. Petz</span>, The proper formula for relative entropy and its asymptotics in quantum probability, *Commun. Math. Phys.* **143**(1) (1991), 99-114, doi:10.1007/BF02100287.
<span style="font-variant:small-caps;">M.J. Kastoryano and F.G.S.L. Brandão</span>, Quantum Gibbs Samplers: The Commuting Case, *Commun. Math. Phys.* **344** (2016), 915-957, doi:10.1007/s00220-016-2641-8.
<span style="font-variant:small-caps;">H. Kosaki</span>, Application of the complex interpolation method to a von Neumann algebra: non-commutative $L^p$-spaces, *J. Funct. Anal.* **56** (1984), 29-78, doi:10.1016/0022-1236(84)90025-9.
<span style="font-variant:small-caps;">S. Kullback and R.A. Leibler</span>, On information and sufficiency, *Annals of Math. Stat.* **22**(1) (1951), 79-86, doi:10.1214/aoms/1177729694.
<span style="font-variant:small-caps;">E.H. Lieb</span>, Convex trace functions and the Wigner–Yanase–Dyson conjecture, *Adv. Math.* **11**(3) (1973), 267-288, doi:10.1016/0001-8708(73)90011-X.
<span style="font-variant:small-caps;">F. Martinelli and E. Olivieri</span>, Approach to equilibrium of Glauber dynamics in the one phase region II. The general case, *Commun. Math. Phys.* **161** (1994), 487-514, doi:10.1007/BF02101930.
<span style="font-variant:small-caps;">K. Matsumoto</span>, Reverse Test and Characterization of Quantum Relative Entropy, preprint (2010), arxiv:1010.1030
<span style="font-variant:small-caps;">S. Michalakis and J. Pytel</span>, Stability of frustration free Hamiltonians, *Commun. Math. Phys.* **332**(2) (2013), 277-302, doi:10.1007/s00220-013-1762-6.
<span style="font-variant:small-caps;">M. Ohya and D. Petz</span>, Quantum Entropy and Its Use, *Texts and Monographs in Physics (Springer-Verlag, Berlin)* (1993).
<span style="font-variant:small-caps;">M.S. Pinsker</span>, Information and Information Stability of Random Variables and Processes, *Holden Day* (1964).
<span style="font-variant:small-caps;">M.B. Ruskai</span>, Inequalities for traces on Von Neumann algebras, *Commun. Math. Phys.* **26** (1972), 280-289, doi:10.1007/BF01645523.
<span style="font-variant:small-caps;">C.E. Shannon</span>, A mathematical theory of communication, *Bell Syst. Tech. J.* **27** (1948), 379-423, 623-656, doi:10.1002/j.1538-7305.1948.tb01338.x, 10.1002/j.1538-7305.1948.tb00917.x.
<span style="font-variant:small-caps;">D. Sutter, M. Berta and M. Tomamichel</span>, Multivariate Trace Inequalities, *Commun. Math. Phys.* **352**(1) (2017), 37-58, doi:10.1007/s00220-016-2778-5.
<span style="font-variant:small-caps;">C.J. Thompson</span>, Inequality with Applications in Statistical Mechanics, *J. Math. Phys.* **6** (1965), 1812-1813, doi:10.1063/1.1704727.
<span style="font-variant:small-caps;">H. Umegaki</span>, Conditional expectation in an operator algebra IV. Entropy and information, *Kodai Math. Sem. Rep.* **14** (1962), 59-85, doi:10.2996/kmj/1138844604.
<span style="font-variant:small-caps;">A. Wehrl</span>, General properties of entropy, *Rev. Mod. Phys.* **50**(2) (1978), 221-260, doi:10.1103/RevModPhys.50.221.
<span style="font-variant:small-caps;">M.M. Wilde</span>, Monotonicity of $p$-norms of multiple operators via unitary swivels, preprint (2016), arxiv:1610.01262.
<span style="font-variant:small-caps;">M.M. Wilde</span>, Quantum Information Theory, Second Edition, (2017), *Cambridge University Press*, doi:10.1017/9781316809976.
<span style="font-variant:small-caps;">H. Wilming, R. Gallego and J. Eisert</span>, Axiomatic Characterization of the Quantum Relative Entropy and Free Energy, *Entropy* **19**(6) (2017), 241, doi:10.3390/e19060241.
<span style="font-variant:small-caps;">M.M. Wolf</span>, Quantum Channels and Operations. Guided tour, (2012), https://www-m5.ma.tum.de/foswiki/pub/M5/Allgemeines/MichaelWolf/QChannelLecture.pdf.
[^1]: It can also be defined in infinite dimensions, as well as generalized von Neumann algebras [@libropetz]. However, in this work, for simplicity we will restrict to finite dimensions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Phase field crystal (PFC) theory is extensively used for modelling the phase behaviour, structure, thermodynamics and other related properties of solids. PFC theory can be derived from dynamical density functional theory (DDFT) via a sequence of approximations. Here, we carefully identify all of these approximations and explain the consequences of each. One approximation that is made in standard derivations is to neglect a term of form $\nabla\cdot[n\nabla{{\cal L}}n]$, where $n$ is the scaled density profile and ${{\cal L}}$ is a linear operator. We show that this term makes a significant contribution to the stability of the crystal, and that dropping this term from the theory forces another approximation, that of replacing the logarithmic term from the ideal gas contribution to the free energy with its truncated Taylor expansion, to yield a polynomial in $n$. However, the consequences of doing this are: (i) the presence of an additional spinodal in the phase diagram, so the liquid is predicted first to freeze and then to melt again as the density is increased; and (ii) other periodic structures, such as stripes, are erroneously predicted to be thermodynamic equilibrium structures. In general, ${{\cal L}}$ consists of a non-local convolution involving the pair direct correlation function. A second approximation sometimes made in deriving PFC theory is to replace ${{\cal L}}$ by a gradient expansion involving derivatives. We show that this leads to the possibility of the density going to zero, with its logarithm going to $-\infty$ whilst being balanced by the fourth derivative of the density going to $+\infty$. This subtle singularity leads to solutions failing to exist above a certain value of the average density. We illustrate all of these conclusions with results for a particularly simple model two-dimensional fluid, the generalised exponential model of index 4 (GEM-4), chosen because a DDFT is known to be accurate for this model. The consequences of the subsequent PFC approximations can then be examined. These include the phase diagram being both qualitatively incorrect, in that it has a stripe phase, and quantitatively incorrect (by orders of magnitude) regarding the properties of the crystal phase. Thus, although PFC models are very successful as phenomenological models of crystallisation, we find it impossible to derive the PFC as a theory for the (scaled) density distribution when starting from an accurate DDFT, without introducing spurious artefacts. However, we find that making a simple one-mode approximation for the *logarithm* of the density distribution $\log\rho({{\bm x}})$ (rather than for $\rho({{\bm x}})$), is surprisingly accurate. This approach gives a tantalising hint that accurate PFC-type theories may instead be derived as theories for the field $\log\rho({{\bm x}})$, rather than for the density profile itself.'
author:
- 'Andrew J. Archer'
- 'Daniel J. Ratliff'
- 'Alastair M. Rucklidge'
- Priya Subramanian
title: 'Deriving phase field crystal theory from dynamical density functional theory: consequences of the approximations'
---
Introduction
============
The phase field crystal (PFC) theory for matter is widely used and has been successfully applied to describe a broad range of phenomena, including, for example, grain boundary dynamics [@Elder2002; @Elder2004], crystal nucleation [@Backofen2010; @Toth2010], crystal growth [@Archer2012], glass formation [@Berry2008a], crack propagation [@Elder2004] and many other properties of condensed matter. For more background and examples of situations to which the PFC theory has been applied, see the excellent review [@Emmerich2012]. The PFC theory was originally proposed, in the spirit of ‘regular’ phase field theory (PFT), as a diffuse-interface theory for the time evolution of an order parameter field [@Elder2002]. The equations of PFT are obtained via symmetry, thermodynamic and other arguments and the result is a theory that is widely used in materials science and other disciplines to model the structure of materials. For more background on PFT see for example Ref. [@Boettinger2002] and references therein.
The central and original idea in extending PFT to arrive at PFC theory is to incorporate aspects of the microscopic structure of the material into the model [@Elder2002]. The result is a theory that operates on atomic length scales and diffusive time scales [@Emmerich2012]. By this we mean that PFC theory is a theory for a field that exhibits numerous maxima, each of which is identified as the average location of the atoms (or more generally ‘particles’) in the system. This idea is powerful because, by building into the theory more information about the underlying material structure, it enables the inclusion of much more of the physics coming from particle correlations to be incorporated. Over the years several variants of PFC theory have been developed that are able to describe a range different crystalline (and even quasicrystalline) structures [@Jaatinen2009; @Pisutha-Arnond2013b; @Wu2010; @Barkan2011; @Achim2014; @Subramanian2016; @Jiang2017; @Savitz2018].
Thus, the original PFC [@Elder2002] may be viewed as the simplest partial differential equation model one can conceive of for a conserved order parameter exhibiting peaks arranged with crystalline ordering. It is obtained from a (scaled) free energy ${{\cal F_{\text{$\alpha$}}}}$ that is a functional of the dimensionless order parameter $n$: $${{\cal F_{\text{$\alpha$}}}}[n] = \int \left(\frac{1}{2}n\left((k_s^2+\nabla^2)^2 - r\right)n
+ \frac{1}{4}n^4\right){{{\rm d}}{{\bm x}}},
\label{eq:PFCFalpha}$$ where $n({{\bm x}},t)$ is a field that depends on position in space ${{\bm x}}$ and on time $t$, and $k_s$ is an inverse length scale that determines the lattice spacing of the crystal. The parameter $r$ defines how near the system is to freezing. The time evolution of the conserved field $n$ is given by the dynamics $$\frac{\partial n}{\partial t} = \nabla^2 \left(\frac{{{\delta{\mkern-1mu}{{\cal F_{\text{$\alpha$}}}}}}}{{{\delta{\mkern-1mu}n}}}\right)
= - \nabla^2 \left(rn - (k_s^2+\nabla^2)^2n - n^3\right),
\label{eq:PFCalphadynamics}$$ where $\frac{{{\delta{\mkern-1mu}{{\cal F_{\text{$\alpha$}}}}}}}{{{\delta{\mkern-1mu}n}}}$ is the functional derivative of ${{\cal F_{\text{$\alpha$}}}}$ with respect to $n({{\bm x}})$.
Given the ingredients in the model, it is therefore no surprise that PFC theory is closely related to the Swift–Hohenberg equation [@Swift1977]: $$\frac{\partial n}{\partial t} = - \frac{{{\delta{\mkern-1mu}{{\cal F_{\text{$\alpha$}}}}}}}{{{\delta{\mkern-1mu}n}}}
= rn - (k_s^2+\nabla^2)^2n - n^3,
\label{eq:SwiftHohenberg}$$ which is one of the archetypal equations in pattern formation theory. As one can see above, both the Swift–Hohenberg equation and PFC theory can be expressed as a different type of dynamics based on the same free energy functional [@Emmerich2012]. The Swift–Hohenberg equation (\[eq:SwiftHohenberg\]) is based on an underlying dynamics that seeks to minimise the free energy over time, whilst the PFC dynamics (\[eq:PFCalphadynamics\]), which also decreases the free energy over time, in addition enforces a conservation of the average value of the order parameter in the system. Thus, the PFC equation (\[eq:PFCalphadynamics\]) is sometimes referred to as the conserved Swift–Hohenberg equation [@Thiele2013; @Sagui1994; @Knobloch2015; @Matthews2000; @Emmerich2012].
In the years after PFC theory was originally proposed it was realised that it could be derived from classical dynamical density functional theory (DDFT) [@Elder2007; @Teeffelen2009; @Huang2010a; @Emmerich2012; @Archer2012], via a sequence of several different approximations. Below, we say much more on what these approximations are. DDFT is a theory for the time evolution of the ensemble average one-body (number) density profile $\rho({{\bm x}},t)$, for a non-equilibrium system of interacting classical particles. DDFT is based on equilibrium density functional theory (DFT) [@Evans1979a; @Evans1992; @Hansen2013] and for an equilibrium system, DDFT is equivalent to DFT. DDFT was originally developed as a theory for Brownian particles with over-damped stochastic equations of motion [@Marconi1999; @Marconi2000; @Archer2004; @Archer2004a], but it has also been extended to describe the dynamics of under-damped systems and atomic or molecular systems where the particle dynamics is governed by Newton’s equations of motion [@Archer2006; @Archer2009; @Goddard2012; @Goddard2013; @DuranOlivencia2017; @Schmidt2018]. This body of work shows that when such systems are not too far from equilibrium, then the dynamics predicted by the original DDFT is still often correct in the long-time limit where the particle dynamics is dominated by diffusive processes. This is because DDFT corresponds to a dynamics given by the continuity equation $$\frac{\partial \rho}{\partial t} = -\nabla \cdot \mathbf{j},
\label{eq:continuity}$$ where the current $\mathbf{j}\propto-\nabla \mu({{\bm x}},t)$, with $\mu({{\bm x}},t)$ a local (non-equilibrium) chemical potential [@Marconi1999; @Marconi2000; @Archer2004; @Archer2004a]. Eq. (\[eq:continuity\]) is of course expected since the total number of particles in the system $N=\int \rho({{\bm x}},t){{{\rm d}}{{\bm x}}}$ is a conserved quantity.
Refs. [@Elder2007; @Teeffelen2009; @Huang2010a; @Emmerich2012; @Archer2012] give various different derivations of the PFC model, starting from DFT and/or DDFT. Here, starting from DDFT, we systematically show how all the various different theories are related and we identify and highlight the significance of each of the approximations that are made in the derivation of PFC theory. We show that there is a particular term of the form $\nabla\cdot[n\nabla{{\cal L}}n]$, where ${{\cal L}}$ is a linear operator, that is almost universally neglected because it is ‘of higher order’ [@Huang2010a], but this term is actually important for stabilizing crystalline structures: its contribution is of the same order as some of the terms that are retained. As we explain in detail, neglecting this term essentially forces one to make the Taylor expansion of the ideal gas logarithmic term in the free energy in order to recover something physically reasonable. We show that neglecting this term, as is done in PFC theory, [and the subsequent replacement of the logarithm by its Taylor series,]{} leads to the spurious appearance in the phase diagram of an extra spinodal and also alters the relative stabilities of the crystal state compared to a stripe phase and also other phases, leading in two dimensions (2D) to the stripe phase becoming the global free energy minimum state for certain parameter values. Essentially, all this behaviour originates because the function $\log(1+n)$ has one root, but when it is replaced by a truncated Taylor expansion, the resulting polynomial generally has two roots. Our arguments also directly apply in three dimensions to explain why lamellar phases occur as equilibrium phases in PFC theory. Recall that most PFC theories predict that as one moves in the phase diagram away from the region where there is coexistence between the liquid and the crystal, moving deeper into the crystalline portion of the phase diagram, such stripe/lamellar phases appear as equilibrium structures and are global minima of the free energy [@Emmerich2012]. Of course, particles with isotropic pair interactions generally never ‘freeze’ to form striped phases, unless they have an unusual and special form for the pair potential between the particles [@Imperio2004; @Imperio2006; @Archer2007]. DDFT, taken together with a reliable approximation for the Helmholtz free energy functional of course does not predict such stripe phases for crystallisation from simple liquids.
The linear operator ${{\cal L}}$ has the form of a non-local convolution involving the pair direct correlation function plus another simpler term (see Eq. below). Another approximation that is often made in deriving PFC theories is to approximate ${{\cal L}}$ by a gradient expansion involving derivatives. We show below that if one makes this approximation whilst simultaneously retaining the logarithmic term from the ideal gas free energy, this results in a theory that still predicts reasonably accurately the freezing transition, but as one increases the average density, moving deeper into the region of the phase diagram where the crystal phase occurs, there is a point where $\rho({{\bm x}})\to0$ at the points in space ${{\bm x}}$ between the density peaks, where the density is a minimum. On increasing the average density beyond this point in the phase diagram, there is no solution to the theory. We analyse in detail this singular behaviour. As $\rho({{\bm x}})\to0$ we have $\log\rho({{\bm x}})\to-\infty$, of course. In the equation for the equilibrium density profile this divergence is initially balanced by the term involving the fourth derivative, $\partial^4\rho/\partial x^4\to+\infty$. However, when the average density in the system is increased beyond the value at which this divergence occurs, we find there is no solution.
We illustrate these conclusions by finding the predicted structures and phase diagram for the 2D version of the GEM-4 (Generalised Exponential Model of index 4) [@Mladek2006; @Prestipino2014], chosen because DDFT based on a simple approximation (the so-called random phase approximation (RPA) [@Likos2001]) for the Helmholtz free energy functional can be very accurate for predicting the equilibrium structures formed in this model and also the thermodynamics [@Mladek2007; @Prestipino2014; @Archer2014]. At higher temperatures, the 2D GEM-4 system exhibits just a single fluid phase and at higher densities a single crystal phase. At lower temperatures, where the RPA DDFT is no longer accurate, there is a hexatic phase and multiple crystalline phases as the density is increased [@Prestipino2014]. Here we do not consider this regime, restricting ourselves to the regime where there is just one fluid and one crystal phase, which are predicted accurately by the RPA DDFT. This enables us to investigate the effect of making subsequent approximations to the DDFT, including those made to derive PFC theory. We find that the PFC type theories spuriously predict three additional phases that are in reality not present in the phase diagram (i.e., are not thermodynamically stable). These are (i) a stripe phase, (ii) what we refer to as ‘down hexagons’ (in contrast to the true crystal structure, which we refer to as ‘up hexagons’) and then at even higher densities a melting to form (iii) another uniform liquid phase. We show how the approximations made in deriving the PFC result in these structures being predicted.
The final contribution of this paper is to show that there is a very simple and accurate ansatz one can make for the form of the equilibrium crystal density profile in DDFT (and so also for DFT, of course). The ansatz is $\rho({{\bm x}})=\rho_0e^{\phi({{\bm x}})}$, where $\rho_0$ is a constant and the field $\phi({{\bm x}})$ is approximated by a sinusoid of the form $\phi({{\bm x}})\approx\phi_0+\phi_1e^{i{{\bm k}}\cdot{{\bm x}}}+$complex conjugate (in one dimension), plus other similar terms (in higher dimensions), where $\phi_0$ and $\phi_1$ are constants. The results presented here are for the GEM-4 model and show why this approximation is unexpectedly accurate: the approximation is able to replicate almost exactly the numerical solution to the DDFT problem, from small to arbitrarily large amplitude density variations. We expect this ansatz also to be reliable for other systems. This form of one-mode theory gives a hint for future directions to develop accurate PFC-type theories, since using a one-mode approximation in PFC is often fairly accurate.
This paper is structured as follows: In Sec. \[sec:2\] we present our systematic step-by-step derivation of PFC, starting from DDFT. After each approximation, we carefully state the model, i.e., we give the corresponding free energy functional and also the expression for the chemical potential, which is a quantity that is a constant at all points in space for equilibrium states. In order to keep track of the different orders in which the approximations can be made, we give each model a name, starting with for the original PFC model in Eq. (\[eq:PFCalphadynamics\]) above, and with DDFT-0 for the original formulation of DDFT below. The different DDFT approximations result in five different versions, DDFT-1 to DDFT-5. Similarly, we explain the various different approximations that can be made to each of these, leading to a corresponding PFC theory, which we name to . Note that the criterion we use here for distinguishing between whether we refer to a theory as a DDFT or a PFC is based on whether the free energy which is minimised by the dynamical equations (i.e., the Lyapunov functional) has the logarithmic ideal gas term or not: if it does not have the logarithm, we refer to it as a . [Table \[tab:DDFTvsPFC\] below is there to help the reader navigate the various models and the approximations made in each one.]{} Sec. \[sec:2\] concludes with a summarising discussion. In Sec. \[sec:3\] we present results for the GEM-4 system comparing predictions for the density profiles and thermodynamics of equilibria for two of the different DDFT theories and also two of the PFC theories. In this section we also present the phase diagrams for the GEM-4 system predicted by these different DDFT and PFC theories. By comparing all of these we are able to assess the accuracy of the different theories and the validity of the various approximations. In Sec. \[sec:4\] we discuss the implications of the main two approximations and analyse the singular behaviour displayed by some models. In Sec. \[sec:5\] we introduce the ansatz $\rho({{\bm x}})=\rho_0e^{\phi({{\bm x}})}$ and derive the new one-mode approximation for . We draw our conclusions in Sec. \[sec:6\]. The paper includes two appendices in which we describe the numerical (continuation) methods we use to calculate the density profiles.
Derivation of the Phase Field Crystal model from DDFT {#sec:2}
=====================================================
In this section we progress from the original formulation of DDFT (which we call DDFT-0) through a series of approximations (DDFT-1, …, DDFT-5), as listed in Table \[tab:DDFTvsPFC\]. Our main starting point is DDFT-2. From this point, there are three main approximations that can be made (or not made): (i) the Ramakrishan–Yussouff (RY) or the random phase approximation (RPA) for the free energy; (ii) the gradient expansion of the convolution term; and (iii) the Taylor expansion of the logarithmic term. Making (or not making) the first two of these approximations results in DDFT-3, DDFT-4 and DDFT-5. Then, making the third approximation from DDFT-2 results in , from DDFT-3 results in , and so on up to . The model can be rescaled to recover the original version of , , see Eqs. and . The various models are summarised in Table \[tab:DDFTvsPFC\]. Amongst the models we present below, DDFT-5 is equivalent to the model derived by Huang [@Huang2010a] and advocated by van Teeffelen [@Teeffelen2009] (named PFC1 in that paper), and DDFT-3 and are equivalent to the models named DDFT and PFC2 by van Teeffelen [@Teeffelen2009].
---------------- ----- ----- ----- ----- ----- ------------------------------------------------ --------------------------- -------------------------------------- -------------------
Name Dynamics $Q$, $C$, $R$
\[2pt\] Yes N/A Yes Yes Yes (\[eq:PFCalphadynamics\]) (\[eq:PFCFalpha\]) —
\[4pt\] DDFT-0 (\[eq:DDFT0dynamics\]) (\[eq:separatedF\]) (\[eq:chempotdefn\]) —
\[4pt\] DDFT-1 Yes (\[eq:DDFT1dynamics\]), (\[eq:DDFT1deltaF\]) (\[eq:DDFTF1\]) (\[eq:ChemPot\]), (\[eq:betadFdn1\]) —
\[4pt\] DDFT-2 Yes Yes (\[eq:DDFT2dynamics\]) (\[eq:DDFTF2\]) (\[eq:betadFdn2\]) (\[eq:qandc\])
\[4pt\] DDFT-3 Yes N/A Yes (\[eq:DDFT3dynamics\]) (\[eq:DDFTF3\]) (\[eq:betadFdn3\])
\[4pt\] DDFT-4 Yes N/A Yes (\[eq:DDFT4dynamics\]) as (\[eq:DDFTF2\]) as (\[eq:betadFdn2\]) (\[eq:qandc\])
\[4pt\] DDFT-5 Yes N/A Yes Yes (\[eq:DDFT5dynamics\]) (\[eq:DDFTF5\]) (\[eq:betadFdn5\])
\[4pt\] Yes Yes Yes (\[eq:PFCdynamics\]), (\[eq:PFCbetadynamics\]) (\[eq:DDFTFPFCbeta\]) (\[eq:betadFdnbeta\]) (\[eq:PFCqandc\])
\[4pt\] Yes N/A Yes Yes (\[eq:PFCgammadynamics\]) (\[eq:DDFTFPFCgamma\]) (\[eq:betadFdngamma\])
\[4pt\] Yes N/A Yes Yes as (\[eq:PFCbetadynamics\]) as (\[eq:DDFTFPFCbeta\]) as (\[eq:betadFdnbeta\]) (\[eq:PFCqandc\])
\[4pt\] Yes N/A Yes Yes Yes (\[eq:PFCepsilondynamics\]) as (\[eq:DDFTFPFCgamma\]) as (\[eq:betadFdngamma\])
\[4pt\]
---------------- ----- ----- ----- ----- ----- ------------------------------------------------ --------------------------- -------------------------------------- -------------------
Dynamic Density Functional Theory: DDFT-0
-----------------------------------------
The starting point for all of our derivations is the key DDFT equation [@Marconi1999; @Marconi2000; @Archer2004; @Archer2004a]: $$\frac{\partial \rho}{\partial t} = \nabla \cdot \left[ \beta M(\rho) \nabla \frac{{{\delta{\mkern-1mu}{{\cal F}}}}}{{{\delta{\mkern-2mu}\rho}}} \right],
\label{eq:DDFT0dynamics}$$ where $\beta=(k_BT)^{-1}$ (with $k_B$ being Boltzmann’s constant and $T$ being temperature), $M(\rho)$ is the positive $\rho$-dependent mobility. The Helmholtz free energy ${{\cal F}}[\rho]$ depends on the density profile $\rho({{\bm x}},t)$ integrated over space; hence ${{\cal F}}[\rho]$ depends on time but not on position [@Archer2004]. The expression ${{{\delta{\mkern-1mu}{{\cal F}}}}}/{{{\delta{\mkern-2mu}\rho}}}$ is the functional derivative of ${{\cal F}}$ with respect to $\rho({{\bm x}},t)$, which therefore depends on both time and on position. DDFT usually takes $M(\rho)=D \rho$, i.e., the mobility is proportional to density [@Marconi1999; @Marconi2000; @Archer2004; @Archer2004a], where $D$ is the diffusion coefficient. We henceforth scale time so that $D=1$. With boundary conditions that do not allow material to enter or leave the system, $N=\int\!\rho({{\bm x}}){{{\rm d}}{{\bm x}}}$ (or equivalently, the mean density) is a constant of the motion and is the total number of particles in the system.
With suitable boundary conditions, one can readily show that the Helmholtz free energy decreases monotonically with time: $$\frac{{{\rm d}}{{\cal F}}}{{{{\rm d}}{t}}} = - \int \! \beta M(\rho)
\left| \nabla \frac{{{\delta{\mkern-1mu}{{\cal F}}}}}{{{\delta{\mkern-2mu}\rho}}} \right|^2 {{{\rm d}}{{\bm x}}}\leq 0,
\label{eq:Fdecreases}$$ so (assuming that ${{\cal F}}[\rho]$ is bounded below) the system typically evolves to a (local) minimum of ${{\cal F}}$, which is a dynamically stable equilibrium of (\[eq:DDFT0dynamics\]). Here, ‘dynamically stable’ means that small perturbations away from the equilibrium decay, and ‘equilibrium’ means that ${\partial \rho}/{\partial t} =0$ and ${{{\rm d}}{{\cal F}}}/{{{{\rm d}}{t}}}=0$. Owing to the dynamics being governed by a continuity equation , such perturbations cannot change the mean density. Local minima of ${{\cal F}}$ that are not the global minimum are thermodynamically metastable. The system can also have dynamically unstable equilibria, for which ${{\cal F}}$ is a saddle or maximum. From (\[eq:Fdecreases\]), we see that all equilibria of (\[eq:DDFT0dynamics\]) satisfy $\nabla({{\delta{\mkern-1mu}{{\cal F}}}}/{{\delta{\mkern-2mu}\rho}})=0$, so $$\frac{{{\delta{\mkern-1mu}{{\cal F}}}}}{{{\delta{\mkern-2mu}\rho}}} = \text{constant} = \mu,
\label{eq:chempotdefn}$$ where $\mu$ is the chemical potential of the equilibrium. This is of course the Euler–Lagrange equation for the problem of finding stationary points of the functional ${{\cal F}}[\rho]$, subject to the constraint of fixed mean density. Note however that when evolving (\[eq:DDFT0dynamics\]) forward in time from an arbitrary initial condition, $\mu$ is not necessarily known *a priori*.
The theory can also be cast in terms of the grand potential (also called the Landau free energy) functional [@Evans1979a; @Evans1992; @Hansen2013]: $$\Omega[\rho] = {{\cal F}}[\rho] - \mu N = {{\cal F}}[\rho] - \mu \int \! \rho({{\bm x}}){{{\rm d}}{{\bm x}}}.
\label{eq:omegadefn}$$ From this it follows that the functional derivative of $\Omega$ is $$\frac{{{\delta{\mkern-1mu}\Omega}}}{{{\delta{\mkern-2mu}\rho}}} = \frac{{{\delta{\mkern-1mu}{{\cal F}}}}}{{{\delta{\mkern-2mu}\rho}}} - \mu,
\label{eq:dOmega}$$ and that this is zero at equilibrium: equilibria are extreme values of $\Omega$. Like the Helmholtz free energy, the grand potential decreases monotonically with time, since Eq. is also true if one replaces ${{\cal F}}$ by $\Omega$. Therefore, for two phases to coexist, they must have the same specific grand potential (i.e., the same pressure) and the same chemical potential. Thus, the global minimum of $\Omega$ for a given $\mu$ and $T$ is the thermodynamic equilibrium state of the system [@Evans1979a; @Evans1992; @Hansen2013].
Following the usual approach in DFT, we separate the Helmholtz free energy into three parts: the ‘ideal gas’ contribution, which is proportional to the temperature but takes no account of particle interactions, an excess (${{\cal F_{\text{ex}}}}$) over the ideal gas contribution arising from the particle interactions, and the contribution due to an external potential ${U_{\text{ext}}}({{\bm x}})$, as follows [@Evans1979a; @Evans1992; @Hansen2013]: $${{\cal F}}[\rho] = k_BT \int \rho\left(\log(\Lambda^d\rho) - 1\right){{{\rm d}}{{\bm x}}}+ {{\cal F_{\text{ex}}}}[\rho] + \int\rho{U_{\text{ext}}}{{{\rm d}}{{\bm x}}},
\label{eq:separatedF}$$ where the integral is taken over the volume $V$ in three dimensions ($d=3$) (or the area in 2D, $d=2$) and where $\Lambda$ is the thermal de Broglie wavelength. Since for our purposes here the value of $\Lambda$ is irrelevant (changing $\Lambda$ will shift the values of ${{\cal F}}$ and $\mu$ by constants), we henceforth set $\Lambda=1$. We also consider bulk systems and so we assume that ${U_{\text{ext}}}=0$. With the separation in Eq. (\[eq:separatedF\]), we have $$\beta\frac{{{\delta{\mkern-1mu}{{\cal F}}}}}{{{\delta{\mkern-2mu}\rho}}} = \log\rho +
\beta\frac{{{\delta{\mkern-1mu}{{\cal F_{\text{ex}}}}}}}{{{\delta{\mkern-2mu}\rho}}},
\label{eq:separatedFdrho}$$ which gives $$\beta\nabla\frac{{{\delta{\mkern-1mu}{{\cal F}}}}}{{{\delta{\mkern-2mu}\rho}}}=\frac{1}{\rho}\nabla\rho+\dots,
\label{eq:rhoinverse}$$ where on the right hand side we only explicitly write the contribution from the ideal gas part of the free energy. Inserting this into Eq. (\[eq:DDFT0dynamics\]) with $M=D\rho$ we obtain $$\frac{\partial \rho}{\partial t} = \nabla^2 \rho+\dots,
\label{eq:rhodiffusion}$$ in which the coefficient in front of the term $\nabla^2 \rho$ is $D$, but our choice of time scaling has $D=1$. Note that this term is *linear* in $\rho$, in spite of it originating from a *nonlinear* logarithmic contribution to the free energy.
We refer to the model up to this point as DDFT-0.
Expansion of ${{\cal F_{\text{ex}}}}$: DDFT-1
---------------------------------------------
To proceed, we must have an expression for the excess Helmholtz free energy functional ${{\cal F_{\text{ex}}}}[\rho]$. We use a functional Taylor expansion, which is also that used in all derivations of PFC theory. This gives the free energy functional of the system of interest in terms of properties of a reference system, which are assumed to be known. The reference system that is chosen is a uniform liquid, with constant density $\rho_0$. The density profile of the system of interest may be varying in space and with an average density that may be different from $\rho_0$. The functional Taylor series expansion of the excess free energy can be written in terms of the density difference ${{\Delta{\mkern-1.0mu}\rho}}({{\bm x}},t)=\rho({{\bm x}},t)-\rho_0$ as follows [@Evans1992; @Hansen2013]: $$\begin{split}
{{\cal F_{\text{ex}}}}[\rho] &= {{\cal F_{\text{ex}}}}[\rho_0]
- k_BT \!\! \int \! c^{(1)}({{\bm x}}_1){{\Delta{\mkern-1.0mu}\rho}}({{\bm x}}_1){{{\rm d}}{{\bm x}}}_1 \\
& \quad{}-\frac{k_BT}{2!} \!\! \int \!
c^{(2)}({{\bm x}}_1,{{\bm x}}_2)
{{\Delta{\mkern-1.0mu}\rho}}({{\bm x}}_1)
{{\Delta{\mkern-1.0mu}\rho}}({{\bm x}}_2)
{{{\rm d}}{{\bm x}}}_1{{{\rm d}}{{\bm x}}}_2 \\
& \quad{}-\frac{k_BT}{3!} \!\! \int \!
c^{(3)}({{\bm x}}_1,{{\bm x}}_2,{{\bm x}}_3) \times{}\\
&\qquad\qquad\qquad {{\Delta{\mkern-1.0mu}\rho}}({{\bm x}}_1)
{{\Delta{\mkern-1.0mu}\rho}}({{\bm x}}_2)
{{\Delta{\mkern-1.0mu}\rho}}({{\bm x}}_3)
{{{\rm d}}{{\bm x}}}_1{{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3\\
& \quad{} + \text{similar fourth order term} + \dots.
\end{split}
\label{eq:expandFex}$$ The expressions $c^{(n)}$ in the equation above are proportional to the first and higher functional derivatives of ${{\cal F_{\text{ex}}}}$ with respect to density, all evaluated at $\rho=\rho_0$: $$c^{(n)}({{\bm x}}_1,\dots,{{\bm x}}_n) = -\beta
\frac{\delta^n\!{{\cal F_{\text{ex}}}}}{{{\delta{\mkern-2mu}\rho}}({{\bm x}}_1)\dots{{\delta{\mkern-2mu}\rho}}({{\bm x}}_n)}[\rho_0].
\label{eq:defncn}$$ These functions $c^{(n)}$ are known as [*direct correlation functions*]{} [@Evans1979a; @Evans1992; @Hansen2013], and are related to $n$-point density correlation functions. In the two-point case, $c^{(2)}$ is the pair direct correlation function and is related to the pair correlation function (i.e., the radial distribution function) through the Ornstein–Zernike equation [@Evans1979a; @Evans1992; @Hansen2013]. These direct correlation functions depend on our choice of $\rho_0$ and depend directly on temperature through the linear factor of $\beta$ in the definition and also indirectly via the fact that the correlations in a liquid change with temperature. Note also that $c^{(1)}[\rho_0]$ is a constant when $\rho_0$ is a constant.
For a homogeneous liquid with distant (or periodic) boundaries, these functions depend on displacements but not on absolute position, so (through a slight abuse of notation) we also write $$c^{(n)}({{\bm x}}_1,\dots,{{\bm x}}_n) = c^{(n)}({{\Delta{\mkern-1.0mu}{{\bm x}}}}_2,{{\Delta{\mkern-1.0mu}{{\bm x}}}}_3\dots,{{\Delta{\mkern-1.0mu}{{\bm x}}}}_n),
\label{eq:shiftorigin}$$ where ${{\Delta{\mkern-1.0mu}{{\bm x}}}}_j={{\bm x}}_j-{{\bm x}}_1$ [@Hansen2013]. We also take the liquid to be isotropic.
We are considering density perturbations away from the liquid state, so it is convenient to write $$\label{eq:rho_sub}
\rho({{\bm x}},t)=\rho_0 (1+n({{\bm x}},t)).$$ We do not assume that $n$ is small, but it is often the case that the average of $n({{\bm x}},t)$ over the whole system is small. Note also that $\rho({{\bm x}},t)=\rho_0$ is a stationary solution of (\[eq:DDFT0dynamics\]). Substituting Eq. (\[eq:expandFex\]) into Eq. (\[eq:DDFT0dynamics\]) and writing only the terms up to $c^{(1)}$, we get: $$\frac{\partial n}{\partial t} = \nabla^2 n
- \nabla^2 c^{(1)} - \nabla \cdot \left[ n \nabla c^{(1)} \right]
+ \dots \\
\label{eq:DDFT1dynamicsc1}$$ That the uniform liquid state is an equilibrium of (\[eq:DDFT0dynamics\]) implies that $n=0$ is a solution of equation (\[eq:DDFT1dynamicsc1\]): all terms not written down involve ${{\Delta{\mkern-1.0mu}\rho}}$ and so are zero for the uniform liquid with density $\rho_0$. Recall that $c^{(1)}[\rho_0]$ is a constant, which means terms involving gradients of this can be dropped. [Whilst this constant term does not influence the structure (density profile) both in and out of equilibrium, it does affect the thermodynamics (i.e., free energy value) and so also mechanical properties [@Wang2018b].]{} With this, we can write the equation for the time evolution of $n({{\bm x}},t)$ (up to ${{\cal O}}(c^{(4)})$) as: $$\begin{split}
\frac{\partial n}{\partial t} =& \nabla^2 n
- \rho_0 \nabla^2 \!\!\int\! c^{(2)}({{\bm x}},{{\bm x}}_2)n({{\bm x}}_2){{{\rm d}}{{\bm x}}}_2 \\
&{}- \rho_0 \nabla \cdot \left[ n \nabla
\!\!\int\! c^{(2)}({{\bm x}},{{\bm x}}_2)n({{\bm x}}_2){{{\rm d}}{{\bm x}}}_2
\right]\\
&{}- \frac{\rho_0^2}{2} \nabla^2 \!\!\int\! c^{(3)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3)n({{\bm x}}_2)n({{\bm x}}_3){{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3\\
&{}- \frac{\rho_0^2}{2} \nabla \cdot \left[ n \nabla
\!\!\int\! c^{(3)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3)n({{\bm x}}_2)n({{\bm x}}_3){{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3
\right]\\
&{}- \frac{\rho_0^3}{6} \nabla^2 \!\!\int\! c^{(4)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3,{{\bm x}}_4) \times{}\\
& \qquad\qquad\qquad n({{\bm x}}_2)n({{\bm x}}_3)n({{\bm x}}_4){{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3{{{\rm d}}{{\bm x}}}_4 \\
&{}- \frac{\rho_0^3}{6} \nabla \cdot \bigg[ n \nabla
\!\!\int\! c^{(4)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3,{{\bm x}}_4) \times{}\\
& \qquad\qquad\qquad n({{\bm x}}_2)n({{\bm x}}_3)n({{\bm x}}_4){{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3{{{\rm d}}{{\bm x}}}_4
\bigg] + \dots\\
\end{split}
\label{eq:DDFT1dynamicspretruncate}$$ where we have suppressed writing the time dependence of $n$ throughout and the ${{\bm x}}$ dependence of $n$ when it is not inside an integral. We have written this equation so that the first line is linear in $n$, the next two lines are quadratic in $n$, the fourth and fifth lines are cubic in $n$, and the last line is quartic in $n$.
Since the first line is linear in $n$, and both terms involve a Laplacian, we can write the linearised version of (\[eq:DDFT1dynamicspretruncate\]) in terms of the negative Laplacian of a linear operator ${{\cal L}}$: $$\frac{\partial n}{\partial t} = - \nabla^2 {{\cal L}}n,
\label{eq:DDFTlinear}$$ where $${{\cal L}}n({{\bm x}}) = - n({{\bm x}}) + \rho_0 \!\!\int\! c^{(2)}({{\bm x}},{{\bm x}}_2)n({{\bm x}}_2){{{\rm d}}{{\bm x}}}_2.
\label{eq:defnL}$$ The non-local operator ${{\cal L}}$ is most conveniently considered in terms of its Fourier transform, or equivalently, in terms of how it acts on modes of the form $\exp(i{{\bm k}}\cdot{{\bm x}})$. If $${{\cal L}}e^{i{{\bm k}}\cdot{{\bm x}}} = \sigma({{\bm k}}) e^{i{{\bm k}}\cdot{{\bm x}}},
\label{eq:defnsigma}$$ then $\sigma({{\bm k}})$ is the eigenvalue of ${{\cal L}}$ with eigenfunction $\exp(i{{\bm k}}\cdot{{\bm x}})$. With this, the linear equation (\[eq:DDFTlinear\]) can readily be solved in terms of linear combinations of functions like $\exp\left({k^2\sigma({{\bm k}}) t + i{{\bm k}}\cdot{{\bm x}}}\right)$, where $k^2\sigma({{\bm k}})$ is the growth rate for a mode with wavevector ${{\bm k}}$, and $k=|{{\bm k}}|$. If $\sigma({{\bm k}})$ is negative for all ${{\bm k}}$, all small amplitude density modulations decay to zero, and the liquid state is dynamically stable.
![Illustrative example of the growth rate $k^2\sigma(k)$ as a function of wavenumber $k$. Small amplitude modes with $k^2\sigma(k)<0$ decay exponentially in time, while those with $k^2\sigma(k)>0$ grow exponentially. Throughout we scale lengths so that the maximum growth rate occurs at $k=1$.[]{data-label="fig:growthrate"}](fig_growthrate)
Recall that for a bulk liquid, $c^{(2)}({{\bm x}},{{\bm x}}_2)=c^{(2)}({{\Delta{\mkern-1.0mu}{{\bm x}}}}_2)$, with ${{\Delta{\mkern-1.0mu}{{\bm x}}}}_2={{\bm x}}_2-{{\bm x}}$, and for spherically symmetric (isotropic) particles, $c^{(2)}({{\bm x}},{{\bm x}}_2)=c^{(2)}(|{{\Delta{\mkern-1.0mu}{{\bm x}}}}_2|)$. Therefore, in this case $\sigma({{\bm k}})=\sigma(k)$, i.e., $\sigma$ depends only on the wavenumber $k=|{{\bm k}}|$. The eigenvalue $\sigma(k)$ can be expressed as: $$\begin{split}
\sigma(k) &= - 1 + \rho_0 \!\!\int\! c^{(2)}(|{{\bm x}}_2-{{\bm x}}|) e^{i{{\bm k}}\cdot({{\bm x}}_2-{{\bm x}})}{{{\rm d}}{{\bm x}}}_2\\
&= -1 + \rho_0 {\hat c}^{(2)}(k),
\label{eq:sigmactwo}
\end{split}$$ where ${\hat c}^{(2)}$ is the Fourier transform of $c^{(2)}$. Recall from (\[eq:defncn\]) that $c^{(2)}$ is proportional to $\beta$, so if ${\hat c}^{(2)}$ has any positive Fourier components, decreasing the temperature (increasing $\beta$) can be expected to lead to a range of wavenumbers with positive growth rates, and the liquid being dynamically unstable to modes with wavenumbers centered around the maximum of $k^2\sigma(k)$, see Fig. \[fig:growthrate\]. We have scaled lengths so that the maximum growth rate occurs at wavenumber $k=1$. This argument, of course, assumes that the product $\beta^{-1}c^{(2)}$ is independent of temperature. This is not true in general, but for some systems it is a good approximation (at least over a limited range of temperatures) – see Ref. [@Somerville2018] for a recent discussion on this for a particular colloidal system. Recall too that for an equilibrium liquid the static structure factor $S(k)=[1-\rho_0{\hat c}^{(2)}(k)]^{-1}$. $S(k)$ is proportional to the Fourier transform of the radial distribution function [@Hansen2013]. So, for the stable uniform liquid, we have $\sigma(k)=-1/S(k)$.
We refer to the state point at which the uniform liquid becomes linearly unstable to density modulations with wavenumber $k\neq0$ as the *spinodal point*, in keeping with the terminology of [@Trudu2006]. The more common usage of the term ‘spinodal’ relates to the onset of the zero-wavenumber phase separation instability of liquid–liquid or gas–liquid phase separation [@Hansen2013; @Archer2004]. At the spinodal point, the density and temperature are such that the liquid is dynamically marginally stable, that is, the maximum of $k^2\sigma(k)$ is zero. Therefore, at higher temperatures, small amplitude density modulations decay, and at lower temperatures, small amplitude density modulations grow. For a given fixed value of $\rho_0$, the spinodal temperature is $T_s$, with a corresponding $\beta_s=(k_BT_s)^{-1}$. Similarly, either increasing the density $\rho_0$ of the liquid or increasing the chemical potential $\mu$ can also lead to crossing the spinodal.
With (\[eq:defnL\]), we can eliminate $c^{(2)}$ in favour of ${{\cal L}}$ in (\[eq:DDFT1dynamicspretruncate\]), and obtain (truncating at ${{{\cal O}}}(c^{(4)})$): $$\begin{split}
\frac{\partial n}{\partial t} =& - \nabla^2 \left({{\cal L}}n + \tfrac{1}{2}n^2\right)
- \nabla \cdot \left[ n \nabla {{\cal L}}n \right]\\
&{}- \frac{\rho_0^2}{2} \nabla^2 \!\!\int\! c^{(3)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3)n({{\bm x}}_2)n({{\bm x}}_3){{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3\\
&{}- \frac{\rho_0^2}{2} \nabla \cdot \left[ n \nabla
\!\!\int\! c^{(3)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3)n({{\bm x}}_2)n({{\bm x}}_3){{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3
\right]\\
&{}- \frac{\rho_0^3}{6} \nabla^2 \!\!\int\! c^{(4)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3,{{\bm x}}_4)\times{}\\
&\qquad\qquad\qquad n({{\bm x}}_2)n({{\bm x}}_3)n({{\bm x}}_4){{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3{{{\rm d}}{{\bm x}}}_4\\
&{}- \frac{\rho_0^3}{6} \nabla \cdot \bigg[ n \nabla
\!\!\int\! c^{(4)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3,{{\bm x}}_4) \times{}\\
& \qquad\qquad\qquad n({{\bm x}}_2)n({{\bm x}}_3)n({{\bm x}}_4){{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3{{{\rm d}}{{\bm x}}}_4
\bigg]\\\
\end{split}
\label{eq:DDFT1dynamics}$$ where we have used the result $\nabla \cdot \left[n \nabla n\right] = \frac{1}{2}\nabla^2n^2$. [For an ideal gas, with ${{\cal L}}n=-n$ and $c^{(2)}=c^{(3)}=c^{(4)}=0$, the first line of the equation above reduces to the diffusion equation, $\frac{\partial n}{\partial t}=\nabla^2n$, similar to (\[eq:rhodiffusion\]).]{}
At this point, we have made no approximations beyond expanding the free energy in Eq. (\[eq:expandFex\]) and truncating at ${{{\cal O}}}(c^{(4)})$. We refer to the model at this stage, truncated in this way, as . In the new variables, and incorporating $c^{(2)}$ into ${{\cal L}}$, the Helmholtz free energy ${{\cal F}}$ can be expressed (up to fourth order) in terms of a scaled free energy ${{{\cal F}_{1}}}={{\cal F}}/\rho_0$, where $$\begin{split}
\beta{{{\cal F}_{1}}}[n] &=
\int \! \big([1+n({{\bm x}}_1)] \log [1+n({{\bm x}}_1)] - n({{\bm x}}_1)\big){{{\rm d}}{{\bm x}}}_1\\
& \quad{}- \frac{1}{2} \! \int \!
\left(n^2({{\bm x}}_1) + n({{\bm x}}_1){{\cal L}}n({{\bm x}}_1) \right){{{\rm d}}{{\bm x}}}_1 \\
& \quad{}-\frac{\rho_0^2}{6} \!\! \int \!
c^{(3)}({{\bm x}}_1,{{\bm x}}_2,{{\bm x}}_3) \times{}\\
&\qquad\qquad\quad n ({{\bm x}}_1)
n ({{\bm x}}_2)
n ({{\bm x}}_3)
{{{\rm d}}{{\bm x}}}_1{{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3\\
& \quad{}-\frac{\rho_0^3}{24} \! \int \!
c^{(4)}({{\bm x}}_1,{{\bm x}}_2,{{\bm x}}_3,{{\bm x}}_4) \times{}\\
&\qquad\qquad\quad n ({{\bm x}}_1)
n ({{\bm x}}_2)
n ({{\bm x}}_3)
n ({{\bm x}}_4)
{{{\rm d}}{{\bm x}}}_1{{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3{{{\rm d}}{{\bm x}}}_4,
\end{split}
\label{eq:DDFTF1}$$ and where we have also dropped terms that do not contribute to (\[eq:DDFT1dynamics\]). In these variables, the DDFT that leads to the dynamics (\[eq:DDFT1dynamics\]) is $$\frac{\partial n}{\partial t} = \nabla \cdot \left[ \beta (1+n) \nabla \frac{{{\delta{\mkern-1mu}{{{\cal F}_{1}}}}}}{{{\delta{\mkern-1mu}n}}} \right].
\label{eq:DDFT1deltaF}$$ Note that, because of the $\log(1+n)$ term in (\[eq:DDFTF1\]), $n$ is constrained so that $1+n$ is always non-negative. Also, because of Eq. , we have $$\frac{{{\delta{\mkern-1mu}{{\cal F}}}}}{\delta\rho}=\frac{{{\delta{\mkern-1mu}{{{\cal F}_{1}}}}}}{{{\delta{\mkern-1mu}n}}}.
\label{eq:derivs_eq}$$ Moreover, states that satisfy $$\frac{{{\delta{\mkern-1mu}{{{\cal F}_{1}}}}}}{{{\delta{\mkern-1mu}n}}} = {{\Delta{\mkern-0.5mu}\mu}},
\label{eq:ChemPot}$$ where ${{\Delta{\mkern-0.5mu}\mu}}=\mu-\mu_0$ and where \[see (\[eq:chempotdefn\]), (\[eq:separatedF\]) and (\[eq:expandFex\])\] $$\mu_0=k_BT\log\Lambda^d\rho_0-k_BTc^{(1)}[\rho_0],
\label{eq:mu_rho_0}$$ are equilibrium solutions of (\[eq:DDFT1dynamics\]), or equivalently, extrema of ${{{\cal F}_{1}}}$. Henceforth, we redefine $\mu$ to be $\beta{{\Delta{\mkern-0.5mu}\mu}}/\rho_0$, which is a shifted and rescaled chemical potential. For the free energy in (\[eq:DDFTF1\]), we have $$\begin{split}
\beta\frac{{{\delta{\mkern-1mu}{{{\cal F}_{1}}}}}}{{{\delta{\mkern-1mu}n}}} &=
\log\left(1+n({{\bm x}})\right)
- n({{\bm x}}) - {{\cal L}}n({{\bm x}}) \\
& \quad{}-\frac{\rho_0^2}{2} \!\! \int \!
c^{(3)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3)
n ({{\bm x}}_2)
n ({{\bm x}}_3)
{{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3\\
& \quad{}-\frac{\rho_0^3}{6} \! \int \!
c^{(4)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3,{{\bm x}}_4) \times{}\\
&\qquad\qquad\quad
n ({{\bm x}}_2)
n ({{\bm x}}_3)
n ({{\bm x}}_4)
{{{\rm d}}{{\bm x}}}_2{{{\rm d}}{{\bm x}}}_3{{{\rm d}}{{\bm x}}}_4.
\end{split}
\label{eq:betadFdn1}$$ At equilibrium, this expression (the rescaled chemical potential $\mu$) does not vary in space. The reference liquid $n=0$ has ${{{\cal F}_{1}}}=0$ and $\mu=0$. The zero value for ${{{\cal F}_{1}}}$ arises (in part) from dropping ${{\cal F_{\text{ex}}}}[\rho_0]$ from (\[eq:expandFex\]), while the zero value for the rescaled chemical potential is a consequence of , which is equivalent to dropping $c^{(1)}$ from (\[eq:expandFex\]) and setting $\Lambda=1$ in .
Simplification of $c^{(3)}$ and $c^{(4)}$: DDFT-2
-------------------------------------------------
As the next step, Huang [@Huang2010a] kept only the zero-wavenumber components of $c^{(3)}$ and $c^{(4)}$, or equivalently, they took $$\begin{split}
c^{(3)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3)&=c^{(3)}_0 \delta({{\bm x}}-{{\bm x}}_2)\delta({{\bm x}}-{{\bm x}}_3),\\
c^{(4)}({{\bm x}},{{\bm x}}_2,{{\bm x}}_3,{{\bm x}}_4)&=c^{(4)}_0 \delta({{\bm x}}-{{\bm x}}_2)\delta({{\bm x}}-{{\bm x}}_3)\delta({{\bm x}}-{{\bm x}}_4),
\label{eq:definecs}
\end{split}$$ where $c^{(3)}_0$ and $c^{(4)}_0$ are constants (our sign convention is opposite to that of [@Huang2010a]). This is equivalent to making a local density approximation (LDA) [@Evans1992] for these terms in the free energy. We could in principle include terms involving $c^{(5)}$ and higher as well, treated in the same way: these would contribute a more general function of $n$ in the free energy, treated with the . However, since we are investigating the effect of approximations that have not yet been discussed, we keep as simple a free energy as possible at this point, consistent with truncating at ${{\cal O}}(c^{(4)})$. With this, the free energy in (\[eq:DDFTF1\]) becomes $$\begin{split}
\beta{{{\cal F}_{2}}}[n] &=
\int \! \big([1+n({{\bm x}})] \log [1+n({{\bm x}})] - n({{\bm x}})\big){{{\rm d}}{{\bm x}}}\\
& \quad{}+ \! \int \! \bigg(
- \frac{1}{2} \left(n^2({{\bm x}}) + n{{\cal L}}n({{\bm x}}) \right) \\
& \qquad\qquad{}-\frac{\rho_0^2}{6}c^{(3)}_0 n^3 ({{\bm x}})
-\frac{\rho_0^3}{24}c^{(4)}_0 n^4 ({{\bm x}})
\bigg){{{\rm d}}{{\bm x}}},\\
\end{split}
\label{eq:DDFTF2}$$ and the four terms involving $c^{(3)}$ and $c^{(4)}$ in (\[eq:DDFT1dynamics\]) become $$\begin{split}
&- \frac{\rho_0^2}{2}c^{(3)}_0 \nabla^2 n^2,\quad
- \frac{\rho_0^2}{2}c^{(3)}_0 \nabla \cdot \left[ n \nabla n^2
\right],\\
&- \frac{\rho_0^3}{6}c^{(4)}_0 \nabla^2 n^3
\quad\text{and}\quad
- \frac{\rho_0^3}{6}c^{(4)}_0 \nabla \cdot \left[ n \nabla n^3
\right].
\end{split}
\label{eq:cthreedelta}$$ Using $\nabla \cdot \left[ n \nabla n^2\right]=\frac{2}{3}\nabla^2n^3$ and $\nabla \cdot \left[ n \nabla n^3\right]=\frac{3}{4}\nabla^2n^4$, Huang [@Huang2010a] combined (\[eq:cthreedelta\]) and (\[eq:DDFT1dynamics\]) to get $$\frac{\partial n}{\partial t} = - \nabla^2 \left({{\cal L}}n + Qn^2 + Cn^3 + Rn^4\right)
- \nabla \cdot \left[ n \nabla {{\cal L}}n \right]
\label{eq:DDFT2dynamics}$$ where $$Q = \frac{1}{2} + \frac{\rho_0^2}{2}c^{(3)}_0 ,\quad
C = \frac{\rho_0^2}{3}c^{(3)}_0 + \frac{\rho_0^3}{6}c^{(4)}_0
\quad\text{and}\quad
R = \frac{\rho_0^3}{8}c^{(4)}_0.
\label{eq:qandc}$$ We also have a chemical potential $$\begin{split}
\mu=\beta\frac{{{\delta{\mkern-1mu}{{{\cal F}_{2}}}}}}{{{\delta{\mkern-1mu}n}}} &=
\log\left(1+n({{\bm x}})\right)
- n({{\bm x}}) - {{\cal L}}n({{\bm x}})\\
& \quad\quad{}- \frac{\rho_0^2}{2} c^{(3)}_0 n^2 ({{\bm x}}) - \frac{\rho_0^3}{6} c^{(4)}_0 n^3 ({{\bm x}}),
\end{split}
\label{eq:betadFdn2}$$ which does not vary in space at equilibrium. Up to this point, we refer to the model as .
Here, we retain the $n^4$ term (as did Huang [@Huang2010a]), because otherwise the dynamics in (\[eq:DDFT2dynamics\]) would not be consistent with the free energy (\[eq:DDFTF2\]) and the DDFT dynamics (\[eq:DDFT1deltaF\]) (with ${{\cal F}}_2$ instead of ${{\cal F}}_1$).
The next three models involve making (or not making) two approximations: (i) assuming the Ramakrishan–Yussouff or random phase approximation, which leads to a quadratic excess Helmholtz free energy functional, and (ii) making a gradient expansion of the linear operator ${{\cal L}}$.
Quadratic excess free energy: DDFT-3
------------------------------------
Often, the free energy functional in (\[eq:expandFex\]) is truncated at ${{\cal O}}({{\Delta{\mkern-1.0mu}\rho}}^2)$. This is known as the Ramakrishan–Yussouff (RY) approximation [@Ramakrishnan1979; @Teeffelen2009; @Emmerich2012], which effectively sets $c^{(3)}=c^{(4)}=0$. A mathematically equivalent approximation arises in the treatment of soft purely repulsive particles modelling soft matter, namely the RPA [@Likos2001]. Here, two soft isotropic particles at ${{\bm x}}_1$ and ${{\bm x}}_2$ separated by a distance $x_{12}=|{{\bm x}}_1-{{\bm x}}_2|$ interact through a potential energy $u(x_{12})$, which depends only on the magnitude of the distance and is finite for all values of $x_{12}$. The excess free energy \[c.f. Eq. (\[eq:expandFex\])\] is then $${{\cal F_{\text{ex}}}}[\rho] = \frac{1}{2} \! \int \!
u(|{{\bm x}}_1-{{\bm x}}_2|) \rho({{\bm x}}_1) \rho({{\bm x}}_2) {{{\rm d}}{{\bm x}}}_1{{{\rm d}}{{\bm x}}}_2.
\label{eq:RPAfreeenergy}$$ This amounts to setting $c^{(3)}=c^{(4)}=0$ and $$c^{(2)}({{\bm x}}_1,{{\bm x}}_2) = - \beta u(|{{\bm x}}_1-{{\bm x}}_2|)
\label{eq:RPAc2u}$$ in . The eigenvalues $\sigma(k)$ can thus be related to the Fourier transform of $u$ through (\[eq:sigmactwo\]) [@Likos2001; @Archer2012]: $$\begin{split}
\sigma(k) &= - 1 - \rho_0 \beta \!\!\int\! u(|{{\bm x}}-{{\bm x}}_2|) e^{i{{\bm k}}\cdot({{\bm x}}_2-{{\bm x}})}{{{\rm d}}{{\bm x}}}_2\\
&= - 1 - \rho_0 \beta {\hat u}(k).
\label{eq:sigmauhat}
\end{split}$$ Setting $c^{(3)}=c^{(4)}=0$ implies from (\[eq:qandc\]) that $Q=\frac{1}{2}$, $C=0$ and $R=0$, and results in a free energy $$\begin{split}
\beta{{{\cal F}_{3}}}[n] &=
\int \! \big((1+n({{\bm x}}_1)) \log (1+n({{\bm x}}_1)) - n({{\bm x}}_1)\big){{{\rm d}}{{\bm x}}}_1\\
& \quad{}- \frac{1}{2} \! \int \!
\left(n^2({{\bm x}}_1) + n({{\bm x}}_1){{\cal L}}n({{\bm x}}_1) \right){{{\rm d}}{{\bm x}}}_1.
\end{split}
\label{eq:DDFTF3}$$ With this choice of free energy, the dynamics in (\[eq:DDFT2dynamics\]) becomes: $$\frac{\partial n}{\partial t} = - \nabla^2 \left({{\cal L}}n + \tfrac{1}{2}n^2\right)
- \nabla \cdot \left[ n \nabla {{\cal L}}n \right],
\label{eq:DDFT3dynamics}$$ along with an analogous version of (\[eq:betadFdn2\]), for the chemical potential: $$\mu=\beta\frac{{{\delta{\mkern-1mu}{{{\cal F}_{3}}}}}}{{{\delta{\mkern-1mu}n}}} =
\log\left(1+n({{\bm x}})\right)
- n({{\bm x}}) - {{\cal L}}n({{\bm x}})
\label{eq:betadFdn3}$$ We refer to this model as ; it is equivalent to DDFT-1 with the RY approximation, and to DDFT-0 with ${{\cal F_{\text{ex}}}}$ given by the RPA approximation.
Before moving on to make further approximations, it is worth noting a useful property that DDFT-3 and the subsequent theories derived from it possess. If the pair potential $u(x_{12})$ in Eq. can be written as $u(x_{12})=\epsilon\psi(x_{12})$, where $\epsilon$ is a parameter that controls the overall strength of the potential, then from Eqs. , and we obtain: $$\mu=
\log\left(1+n({{\bm x}})\right)
+\rho_0\beta\epsilon \!\!\int\! \psi(|{{\bm x}}-{{\bm x}}_2|) n({{\bm x}}_2){{{\rm d}}{{\bm x}}}_2.
\label{eq:convol_form}$$ The consequence of this is that for a given $\psi$, the behaviour of the model depends only on the combination of parameters $\rho_0\beta\epsilon$ and the value of $\mu$. If one changes the value of the reference density $\rho_0$ to some other value, then this is entirely equivalent to solving the system with the original reference density $\rho_0$ at a different value of $\beta\epsilon$. We should emphasize that this is only true if $\psi$ does not change with density, which in general is not true, but is approximately the case for some systems.
Gradient expansion of the linear term: DDFT-4
---------------------------------------------
Returning to DDFT-2, Huang [@Huang2010a] (following [@Elder2004]) expanded ${{\cal L}}$ in powers of the gradient operator $\nabla$, replacing ${{\cal L}}$ by the simplest linear operator that allows a positive growth rate for modes with a wavenumber $k_s$. Scaling lengths so that $k_s=1$ results in: $${{{\cal L}_{\text{grad}}}}n = r n - \gamma(1+\nabla^2)^2 n,
\label{eq:cLgraddefn}$$ so $\sigma(k)=r - \gamma(1-k^2)^2$ from (\[eq:defnsigma\]). This approximation is equivalent (within scaling) to a local gradient expansion of (\[eq:defnL\]), expanding the Fourier transform of $c^{(2)}$ about its maximum: $$\rho_0 {\hat c}^{(2)}(k) = 1 + r - \gamma(1-k^2)^2,
\label{eq:defnGradientExpansion}$$ where the function $\rho_0{\hat c}^{(2)}(k)$ and its second derivative evaluated at $k=1$ are $1+r$ and $-8\gamma$, respectively. Here, $r$ is a parameter, notionally increasing with $\beta$ (and with $\rho_0$) and equal to zero at the spinodal point, when $\beta=\beta_s$. This parameter controls the growth rate of waves with wavenumber $1$: effectively, $r$ is the height of the maximum at $k=1$ in the growth rate curve in Fig. \[fig:growthrate\]. The second parameter $\gamma$ can be used to fit the curvature of ${\hat c}^{(2)}(k)$ at $k=1$.
With this gradient expansion, the dynamics is $$\begin{split}
\frac{\partial n}{\partial t} &= - \nabla^2 \left({{{\cal L}_{\text{grad}}}}n + Qn^2 + Cn^3 + Rn^4\right) \\
&\qquad {} - \nabla \cdot \left[ n \nabla {{{\cal L}_{\text{grad}}}}n \right].
\end{split}
\label{eq:DDFT4dynamics}$$ We refer to this model as : ${{{\cal L}_{\text{grad}}}}$ is now a (local) partial differential operator and (\[eq:DDFT4dynamics\]) is a partial differential equation. The free energy and chemical potential can be found from (\[eq:DDFTF2\]) and (\[eq:betadFdn2\]), setting ${{\cal L}}={{{\cal L}_{\text{grad}}}}$. The lower bound $n\geq-1$ is still respected. This model is equivalent to that written down by [@Huang2010a].
Higher powers (or other functions) of the Laplacian can be retained in ${{{\cal L}_{\text{grad}}}}$, to improve the accuracy of the match between the eigenvalues of ${{\cal L}}$ and ${{{\cal L}_{\text{grad}}}}$, as done for example by [@Jaatinen2009; @Pisutha-Arnond2013b], or to introduce additional unstable length scales, as done for example by [@Wu2010; @Barkan2011; @Achim2014; @Subramanian2016] and others. See also Eq. below and the associated discussion.
RY and gradient expansion: DDFT-5
---------------------------------
Finally, we can make both the RY/RPA approximation ($c^{(3)}_0=c^{(4)}=0$) and replace the linear operator ${{\cal L}}$ by ${{{\cal L}_{\text{grad}}}}$ to get the model advocated in Ref. [@Teeffelen2009]. The free energy and evolution equation are $$\begin{split}
\beta{{{\cal F}_{5}}}[n] &=
\int \! \big((1+n({{\bm x}}_1)) \log (1+n({{\bm x}}_1)) - n({{\bm x}}_1)\big){{{\rm d}}{{\bm x}}}_1\\
& \quad{}- \frac{1}{2} \! \int \!
\left(n^2({{\bm x}}_1) + n({{\bm x}}_1){{{\cal L}_{\text{grad}}}}n({{\bm x}}_1) \right){{{\rm d}}{{\bm x}}}_1,
\end{split}
\label{eq:DDFTF5}$$ and $$\frac{\partial n}{\partial t} = - \nabla^2 \left({{{\cal L}_{\text{grad}}}}n + \tfrac{1}{2}n^2\right)
- \nabla \cdot \left[ n \nabla {{{\cal L}_{\text{grad}}}}n \right],
\label{eq:DDFT5dynamics}$$ along with an analogous version of Eq. (\[eq:betadFdn3\]) for the chemical potential: $$\mu=\beta\frac{{{\delta{\mkern-1mu}{{{\cal F}_{5}}}}}}{{{\delta{\mkern-1mu}n}}} =
\log\left(1+n({{\bm x}})\right)
- n({{\bm x}}) - {{{\cal L}_{\text{grad}}}}n({{\bm x}})
\label{eq:betadFdn5}$$ This model is named PFC1 in [@Teeffelen2009], but here we call it DDFT-5 for consistency.
[PFC models]{}
--------------
The final simplification that can be made (or not made) is to discard the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ (or $\nabla\cdot\left[n\nabla{{{\cal L}_{\text{grad}}}}{n}\right]$) term from the dynamical equations for the four DDFT models DDFT-2, …, DDFT-5, resulting in four PFC models , …, . Huang [@Huang2010a] justify making this simplification on the grounds that this term is not truly quadratic in $n$: the presence of ${{\cal L}}{n}$ in the expression means that it is effectively of higher order. However, we show below that this term does in fact make an important contribution to the free energy: at least as important as the $c^{(3)}$ term.
In addition, dropping this term implies significant changes to the DDFT dynamics, the mobility and the nonlinear terms in the free energy. In fact, the $(1+n)$ factor in the mobility in (\[eq:DDFT1deltaF\]), the logarithm in the ideal gas free energy in (\[eq:separatedF\]) and the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term in (\[eq:DDFT1dynamics\]) are inextricably linked. This can be seen in the progression from (\[eq:separatedF\]) to (\[eq:rhodiffusion\]): the functional derivative of the ideal gas term in (\[eq:separatedF\]) (the first term on the right hand side) leads to the $\log\rho$ term in (\[eq:separatedFdrho\]), the gradient of this leads to $\rho^{-1}\nabla\rho$ in (\[eq:rhoinverse\]), and the mobility being $M=D\rho$ cancels the $\rho^{-1}$, leading to a diffusion equation in (\[eq:rhodiffusion\]). [If the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term is dropped from (\[eq:DDFT2dynamics\]), the equation for $n$ becomes of the form $\frac{\partial n}{\partial t} =
\nabla^2\frac{{{\delta{\mkern-1mu}{{\cal G}}}}}{{{\delta{\mkern-1mu}n}}}$ for some functional ${{\cal G}}[n]$. We can see the implications of this by returning to (\[eq:DDFT0dynamics\]) and taking the steps needed to get to this modified version of (\[eq:DDFT2dynamics\]). Clearly the mobility in (\[eq:DDFT0dynamics\]) has been taken to be constant. If we now think of the ideal gas part of the free energy in (\[eq:separatedF\]) and (\[eq:separatedFdrho\]), but with a constant mobility in the dynamical equation, we end up with the ideal gas term contribution to the equation for $\rho$ being the form $$\frac{\partial\rho}{\partial t} = \frac{1}{\rho}\nabla^2\rho -
\frac{1}{\rho^2}\left|\nabla\rho\right|^2$$ instead of the diffusion equation (\[eq:rhodiffusion\]). This unlikely equation can be avoided, and the diffusion equation recovered at leading order, by expanding the logarithm in (\[eq:DDFTF2\]) in a Taylor series. Thus, dropping the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term is equivalent to taking constant mobility and expanding the logarithm.]{}
It is because of these substantial changes that we opt to use the term ‘DDFT’ for all models based on free energies that have the logarithmic ideal gas term, the non-constant mobility and the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term retained. In contrast, we use the term ‘PFC’ for models based on expanding the logarithm, having a constant mobility and the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term dropped. One consequence of expanding the logarithm [up to ${{\cal O}}(n^4)$, as is done in most PFC derivations [@Emmerich2012], is that the ideal gas part of the free energy contributes cubic and quartic (as well as quadratic) terms to the free energy,]{} so going from DDFT-2 to turns out not to be just a matter of dropping the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term.
So, a consistent free energy–dynamics derivation [@Elder2004; @Teeffelen2009] involves going back to DDFT-2 and replacing the logarithm in (\[eq:DDFTF2\]) by: $$(1+n)\log(1+n)={n}+\tfrac{1}{2}n^2-\tfrac{1}{6}n^3+\tfrac{1}{12}n^4,
\label{eq:expandlogarithm}$$ resulting in a free energy $$\begin{split}
\beta{{\cal F_{\text{$\beta$}}}}[n] &=
\int \! \Big( - \tfrac{1}{2}n{{\cal L}}n
- \tfrac{1}{6}n^3
+ \tfrac{1}{12}n^4\\
& \qquad\quad {} - \frac{\rho_0^2}{6} c^{(3)}_0 n^3
- \frac{\rho_0^3}{24} c^{(4)}_0 n^4\Big){{{\rm d}}{{\bm x}}}_1,
\label{eq:DDFTFPFCbeta}
\end{split}$$ where we have suppressed writing the ${{\bm x}}_1$ dependency of $n({{\bm x}}_1)$. Taking the mobility $M(\rho)$ in (\[eq:DDFT0dynamics\]) to be a constant ($M=D\rho_0$) implies (after scaling) $$\frac{\partial n}{\partial t} = \nabla^2 \left[\beta\frac{{{\delta{\mkern-1mu}{{\cal F_{\text{$\beta$}}}}}}}{{{\delta{\mkern-1mu}n}}} \right],
\label{eq:PFCdynamics}$$ similar to (\[eq:PFCalphadynamics\]). This leads to the PFC dynamical equation: $$\frac{\partial n}{\partial t} =
- \nabla^2 \left[
{{\cal L}}{n} + Qn^2 + Cn^3
\right]
\label{eq:PFCbetadynamics}$$ and to a chemical potential $$\mu=\beta\frac{{{\delta{\mkern-1mu}{{\cal F_{\text{$\beta$}}}}}}}{{{\delta{\mkern-1mu}n}}} =
- {{\cal L}}{n} - Qn^2 - Cn^3,
\label{eq:betadFdnbeta}$$ where $Q$ is as in (\[eq:qandc\]) but $C$ is different: $$Q = \frac{1}{2} + \frac{\rho_0^2}{2}c^{(3)}_0
\quad\text{and}\quad
C = - \frac{1}{3} + \frac{\rho_0^3}{6}c^{(4)}_0.
\label{eq:PFCqandc}$$ We refer to this model as , and recall that the factor of $\beta$ in front of ${{\cal F_{\text{$\beta$}}}}$ is the inverse temperature.
The end result here is that (\[eq:PFCbetadynamics\]) is not the same as DDFT-2 (\[eq:DDFT2dynamics\]) with the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term removed: the cubic coefficient $C$ is different and the quartic contribution $Rn^4$ in (\[eq:DDFT2dynamics\]) is absent. For the cubic coefficient, the contribution proportional to $c^{(3)}_0$ in (\[eq:qandc\]) comes from the non-constant mobility, while the $-\frac{1}{3}$ term in (\[eq:PFCqandc\]) comes from expanding the logarithm. The contribution to $C$ proportional to $c^{(4)}_0$ is the same. Moreover, the $\frac{1}{2}$ in $Q$ in (\[eq:qandc\]) and (\[eq:PFCqandc\]), while having the same numerical value, arises for two different reasons: non-constant mobility versus expanding the logarithm. An additional difference between the DDFT and PFC models is that in the PFC models, the constraint that $n\geq-1$ (i.e., $\rho\geq0$) is not enforced.
As in the DDFT derivations, we can now make (or not make) the RY/RPA approximation and the gradient expansion. We consider first the RY/RPA approximation, setting $c^{(3)}=c^{(4)}=0$ in . The free energy is $$\beta{{\cal F_{\text{$\gamma$}}}}[n] =
\int \! \left( - \tfrac{1}{2}n{{\cal L}}n
- \tfrac{1}{6}n^3
+ \tfrac{1}{12}n^4
\right){{{\rm d}}{{\bm x}}}_1,
\label{eq:DDFTFPFCgamma}$$ the dynamics is $$\frac{\partial n}{\partial t} =
- \nabla^2 \left[
{{\cal L}}{n} + \tfrac{1}{2}n^2 - \tfrac{1}{3}n^3
\right]
\label{eq:PFCgammadynamics}$$ and the chemical potential is $$\mu=\beta\frac{{{\delta{\mkern-1mu}{{\cal F_{\text{$\gamma$}}}}}}}{{{\delta{\mkern-1mu}n}}} =
- {{\cal L}}{n} - \tfrac{1}{2}n^2 + \tfrac{1}{3}n^3.
\label{eq:betadFdngamma}$$ We refer to this model as , and it is effectively the same as but with $Q=\frac{1}{2}$ and $C=-\frac{1}{3}$.
Finally, the gradient expansion can be made, replacing ${{\cal L}}$ by ${{{\cal L}_{\text{grad}}}}$ in all expressions in this subsection, resulting in (without RY/RPA) and (with RY/RPA).
We refer to these models collectively as the PFC models, and have chosen the names to distinguish these from the PFC1 and PFC2 models of Ref. [@Teeffelen2009]. The quadratic term in the dynamics ($Qn^2$) can be removed (provided $C\neq0$) by adding a constant to $n({{\bm x}})$, but we choose not to do this as it implies a change to what was meant by $\rho_0$ in the reference liquid. In addition, a negative $C$ can be scaled to $-1$. With these changes, is equivalent to the original model (\[eq:PFCalphadynamics\]) of [@Elder2002; @Elder2004]: $$\frac{\partial n}{\partial t} = - \nabla^2 \left(r n - \gamma(1+\nabla^2)^2n + Qn^2 + Cn^3\right),
\label{eq:PFCepsilondynamics}$$ where we have written out ${{{\cal L}_{\text{grad}}}}$ explicitly, and $Q=\frac{1}{2}$ and $C=-\frac{1}{3}$ (or $Q=0$ and $C=-1$ after scaling and adding a constant to $n$ – returning to the conserved Swift–Hohenberg equation).
The implication of dropping the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term in the dynamics (\[eq:DDFT3dynamics\]) for DDFT-3 is now apparent: without this term, Eq. (\[eq:DDFT3dynamics\]) reduces to (\[eq:PFCgammadynamics\]) but with the cubic term removed. The absence of the cubic term here implies a free energy as in (\[eq:DDFTFPFCgamma\]) that is not bounded below, i.e., a free energy that is non-physical, and so the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term can have the effect of stabilizing patterns. In addition, dropping the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term is consistent only with a theory with a constant mobility.
Summary
-------
To summarise, we have carefully laid out the various approximations made in the progression from the DDFT-0 starting point (\[eq:DDFT0dynamics\]) to the final PFC (\[eq:PFCalphadynamics\],\[eq:PFCgammadynamics\]) written down in [@Elder2002; @Elder2004]. We have largely followed earlier derivations [@Teeffelen2009; @Huang2010a; @Emmerich2012], seeking to clarify the approximations that are made. Along the way, we have identified four intermediate versions of DDFT, listed for clarity in Table \[tab:DDFTvsPFC\]. The change in name from DDFT to PFC could be made at any point in this progression, but we prefer to make the name change at the point where the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term is dropped (along with all the other changes that are implied by this), since removing this term marks a considerable alteration to the free energy expression and to the dynamics.
The PFC model (\[eq:PFCbetadynamics\]) is appealing in its simplicity, and it gives insight into a variety of crystallisation phenomena, but the derivations of the model from DDFT presented here, as well as the derivation from Ref. [@Huang2010a], are both problematic. Just dropping the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term, as done by Huang [@Huang2010a], means that the dynamics is not equivalent to a DDFT with mobility proportional to $\rho$. On the other hand, the alternative is to expand the logarithm up to ${{\cal O}}(n^4)$ in (\[eq:expandlogarithm\]) in order to provide a *nonlinear* stabilizing term ($\frac{1}{12}n^4$) in the free energy (\[eq:DDFTFPFCgamma\]). However, in the original formulation, the logarithm comes from the ideal gas term in (\[eq:separatedF\]), and leads to a *linear* diffusive term in the dynamics. The stabilizing nonlinear terms in (\[eq:DDFT2dynamics\]) are provided by $c^{(3)}_0$, $c^{(4)}_0$ (in DDFT-2) and by the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term – these are all absent in .
Indeed, all these models only make physical sense if their free energies are bounded below. The free energies for DDFT-0 and DDFT-1 are too general to make any comment, but that for DDFT-2 (\[eq:DDFTF2\]) can be discussed. The $(1+n) \log (1+n) - n$ term is bounded below by zero, and the $-\tfrac{1}{2}\int{n}{{\cal L}}{n}{{{\rm d}}{{\bm x}}}$ term is bounded below because the eigenvalues of ${{\cal L}}$ are bounded above: $$- \int \! n({{\bm x}}){{\cal L}}n({{\bm x}}) {{{\rm d}}{{\bm x}}}\geq - {\sigma_{\text{max}}}\int \! n^2({{\bm x}}) {{{\rm d}}{{\bm x}}},$$ where ${\sigma_{\text{max}}}$ is the maximum over $k$ of $\sigma(k)$ (we have in mind a $\sigma(k)$ as in Fig. \[fig:growthrate\]). In any case, this term, along with the other quadratic and cubic terms, is dominated by the quartic in $n$, which is bounded below provided $c^{(4)}_0<0$. If $c^{(4)}_0=0$, then $c^{(3)}_0<0$ will do, recalling that $n\geq-1$. For DDFT-3, with the RY/RPA approximation $c^{(3)}_0=c^{(4)}=0$, the boundedness of the free energy (\[eq:DDFTF3\]) depends on the $n^2 + n{{\cal L}}{n}$ combination. From (\[eq:defnL\]), the relevant term is $$\begin{split}
-\int \! &
\left(n^2({{\bm x}}_1) + n({{\bm x}}_1){{\cal L}}n({{\bm x}}_1) \right){{{\rm d}}{{\bm x}}}_1 = \\
& \qquad - \rho_0 \!\!\int\! n({{\bm x}}_1)c^{(2)}({{\bm x}}_1,{{\bm x}}_2)n({{\bm x}}_2){{{\rm d}}{{\bm x}}}_1{{{\rm d}}{{\bm x}}}_2.
\end{split}$$ In general, this is not bounded below, but it is in certain circumstances. For example, it is if ${\sigma_{\text{max}}}<-1$, and it is if $c^{(2)}({{\bm x}}_1,{{\bm x}}_2)\leq0$ (or $u(|{{\bm x}}_1-{{\bm x}}_2|)\geq0$ for RPA) for all ${{\bm x}}_1$ and ${{\bm x}}_2$, which is the case in the numerical examples below. The PFC models are not constrained to have $n\geq-1$, but ${{\cal F_{\text{$\beta$}}}}$ (\[eq:DDFTFPFCbeta\]) is bounded by the $n^4$ term as long as its coefficient is positive; ${{\cal F_{\text{$\gamma$}}}}$ (\[eq:DDFTFPFCgamma\]) is always bounded below, because the expansion of the logarithm in (\[eq:expandlogarithm\]) was truncated after an even powered term.
Throughout we have made the simplest choices in the approximations, but other authors have made many other choices. For example, the original PFC paper [@Elder2002], as well as later papers [@Elder2007; @Huang2010a; @Robbins2012a; @Alster2017a; @Elder2017a], included a two-component (binary) version of the PFC model. Recently, Wang [@Wang2018c] took a much closer look at $c^{(3)}$ and $c^{(4)}$, expressing these in terms of isotropic tensors and so allowing these functions to introduce bond angle dependence into the free energy. Some choices of $c^{(3)}$ and $c^{(4)}$ lead to nonlinear terms that include gradients, which can affect the selection of the final stable crystal [@Wu2010a]. The gradient expansion approach has been generalised in two ways: (i) higher order terms or rational functions were considered by [@Jaatinen2009; @Pisutha-Arnond2013b] in order to improve the fit between the functional form and the Fourier transform of $c^{(2)}$, and (ii) PFC models with two unstable length scales have been put forward by several authors [@Pisutha-Arnond2013b; @Wu2010; @Barkan2011; @Achim2014; @Subramanian2016; @Jiang2017], since these allow more complex crystals (face-centered cubic, icosahedral quasicrystals, …) to be stabilized. We discuss the model of [@Jaatinen2009] in more detail below. Alternative approaches involving weighted densities are also possible [@Jaatinen2010a].
Comparison of DDFT and PFC {#sec:3}
==========================
We are interested in the effects of the approximations made in going from DDFT to . A full assessment of the validity of the RY/RPA approximation for ${{\cal F_{\text{ex}}}}$, which in itself constitutes a major simplification, is beyond the scope of the present study. The general conclusion on the validity of the RY/RPA approximation is that it depends on the nature of the interactions between the particles; there are examples in the literature where this approximation is reliable and others where it works badly – see for example the discussion in Refs [@Evans2009; @Lowen2009] and references therein.
Here, we consider one particular system where the RPA is accurate and then we focus on the effects of approximating ${{\cal L}}$ by ${{{\cal L}_{\text{grad}}}}$ and of making the suite of other approximations inherent in going from DDFT to PFC: expanding the logarithm, assuming constant mobility and dropping the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term. To this end, we start with DDFT-3 and solve (\[eq:betadFdn3\]), rewritten here as: $$\text{DDFT-3:}\quad
\log\left(1+n({{\bm x}})\right)
- n({{\bm x}}) - {{\cal L}}n({{\bm x}}) = \mu.
\label{eq:betadFdn3_rewritten}$$ The system that we consider is particles interacting via the GEM-4 [@Mladek2006; @Prestipino2014] potential: this is a model for soft-matter particles and in particular for dendrimers and other polymers in suspension, treating the polymers via an effective pair potential between their centers of mass. This potential is soft, i.e., finite for all values of $x_{12}$ [@Mladek2006; @Prestipino2014; @Likos2001; @Likos2006; @Lenz2012], and is $$u(x_{12}) = \epsilon e^{-\left(x_{12}/R\right)^4},
\label{eq:GEM4Potential}$$ where the parameter $\epsilon$ controls the strength of the potential and $R$ controls its spatial range. We consider here the system in 2D [@Prestipino2014; @Archer2014]. As long as the temperature and density are high enough that the particle cores regularly overlap (the regime in which the system freezes), the RPA approximation is known to be rather accurate for the GEM-4 system and gives a good account of the phase diagram and the structure of the liquid and solid phases [@Mladek2007; @Prestipino2014; @Archer2014].
From Eqs. (\[eq:defnL\]), (\[eq:RPAc2u\]) and we obtain the linear operator ${{\cal L}}$: $${{\cal L}}n({{\bm x}}) = - n({{\bm x}}) - \rho_0\beta\epsilon \!\!\int\!
e^{-|{{\bm x}}-{{\bm x}}_2|^4/R^4} n({{\bm x}}_2){{{\rm d}}{{\bm x}}}_2.
\label{eq:defnL_GEM4}$$ Recall from (\[eq:defnsigma\]) that ${{\cal L}}$ has eigenvalues $\sigma(k)$ with eigenfunctions $e^{i{{\bm k}}\cdot{{\bm x}}}$. We can choose the combined parameter $\rho_0\beta\epsilon$ and soft-particle radius $R$ so that the maximum in $\sigma(k)$ occurs at $k=1$ when the system is at the linear stability threshold, i.e., this is a maximum with $\sigma(1)=0$, similar to Fig. \[fig:growthrate\]. In 2D, to satisfy this condition we must have $\rho_0\beta\epsilon=0.2455$ and $R=5.0962$ – see Appendix \[app:LinearGEM4\] for details.
![The eigenvalue $\sigma(k)$ of ${{\cal L}}$ plotted as a function of wavenumber $k$ for the GEM-4 potential (solid line) for $R=5.0962$ and $\rho_0\beta\epsilon=0.2455$, which is at the threshold where the system becomes linearly unstable. This has $\sigma(0)=-18.75$. We also display $\sigma(k)$ from the gradient expansion of ${{\cal L}}$ (dashed line), i.e., a Taylor expansion in Fourier space around $k=1$, which is the PFC relation for ${{{\cal L}_{\text{grad}}}}$, $\sigma(k)=-\gamma(1-k^2)^2$, with $\gamma=4.37$. Recall that the growth rate of Fourier modes with wavenumber $k$ is is $k^2\sigma(k)$. The dotted line labelled ${{{\cal L}_{\text{grad-8}}}}$ (EOF) is the curve for (\[eq:cLgradJaatinen\]), the eighth-order fitting proposed in [@Jaatinen2009]: it nearly coincides with the GEM-4 curve for $k\leq1$.[]{data-label="fig:dispersionGEM4"}](dispersion_relation){width="8.6truecm"}
With this choice of parameters, the eigenvalue $\sigma(k)$ is shown as a solid line in Fig. \[fig:dispersionGEM4\]. The figure also shows (dashed line) the eigenvalue for the gradient expansion of ${{\cal L}}$ around $k=1$: $${{{\cal L}_{\text{grad}}}}n({{\bm x}}) = - \gamma(1+\nabla^2)^2 n({{\bm x}}),
\label{eq:cLgradGEM4}$$ where $\gamma=4.37$ is chosen to match the second derivative $\frac{{{\rm d}}^2\sigma}{{{\rm d}}{k}^2}$ at $k=1$, as done for example in Refs. [@Elder2004; @Wu2007; @Jaatinen2009]. The dotted line in Fig. \[fig:dispersionGEM4\] is the eigenvalue for (\[eq:cLgradJaatinen\]), the eighth-order fitting model proposed in [@Jaatinen2009] and discussed in more detail below.
![(a) Liquid density ($1+{n_{\text{liq}}}$) and (b) specific grand potential ${\Omega_{\text{liq}}}/A$ as a function of the scaled chemical potential $\mu$, for DDFT-3 (solid black line), DDFT-5 (dashed black line), (indistinguishable from DDFT-3) and (dashed magenta line). []{data-label="fig:liquiddensity"}](liquid_density){width="8.6truecm"}
![(a) Liquid density ($1+{n_{\text{liq}}}$) and (b) specific grand potential ${\Omega_{\text{liq}}}/A$ as a function of the scaled chemical potential $\mu$, for DDFT-3 (solid black line), DDFT-5 (dashed black line), (indistinguishable from DDFT-3) and (dashed magenta line). []{data-label="fig:liquiddensity"}](liquid_omega){width="8.6truecm"}
In what follows we compare solutions of (\[eq:betadFdn3\_rewritten\]) for DDFT-3 with solutions of the analogous equations for DDFT-5, and : $$\begin{aligned}
&\text{DDFT-5:}&
\log\left(1+n\right)
- n - {{{\cal L}_{\text{grad}}}}n &= \mu,
\label{eq:betadFdn5_rewritten}\\
&\text{{\hbox{PFC-$\gamma$}}:}&
-\tfrac{1}{2}n^2 + \tfrac{1}{3} n^3 - {{\cal L}}n &= \mu,
\label{eq:betadFdngamma_rewritten}\\
&\text{{\hbox{PFC-$\epsilon$}}:}&
-\tfrac{1}{2}n^2 + \tfrac{1}{3} n^3 - {{{\cal L}_{\text{grad}}}}n &= \mu.
\label{eq:betadFdnepsilon_rewritten}
\end{aligned}$$ See Appendix \[app:continuation\] for details of the pseudo-arclength continuation numerical method we use for solving these equations. Also, in the supplementary material we include a <span style="font-variant:small-caps;">Matlab</span> code for solving DDFT-3. Note that throughout what follows, we refer to the quantity $1+n({{\bm x}})$ as the ‘density’.
Since the DDFT-5, and represent different forms of Taylor expansion around the reference state with density $\rho_0$, there are a variety of ways comparison between solutions can be made. Here, we opt to fix ${{\cal L}}$ and ${{{\cal L}_{\text{grad}}}}$ as in (\[eq:defnL\_GEM4\]) and (\[eq:cLgradGEM4\]) with the specified values of $\rho_0\beta\epsilon$, $R$ and $\gamma$. This implies that at $\mu=0$ the reference state with $n=0$ is at the spinodal point and is marginally unstable to modes with wavenumber $k=1$. We then vary $\mu$ starting from $\mu=0$ and follow the liquid, stripe and hexagonal solutions of (\[eq:betadFdn3\_rewritten\]) and (\[eq:betadFdn5\_rewritten\])–(\[eq:betadFdnepsilon\_rewritten\]) in appropriately sized two-dimensional domains. For a given value of $\mu$ the different solutions have different values for the mean density $1+\bar{n}=1+\frac{1}{A}\!\!\int\! n({{\bm x}}){{{\rm d}}{{\bm x}}}$, where $A$ is the area of the domain. For each state we calculate the specific grand potential: $$\frac{\Omega[n]}{A} = \frac{{{\cal F}}[n]}{A} - \mu (1+\bar{n}),
\label{eq:specificgrandpotential}$$ where ${{\cal F}}$ is ${{{\cal F}_{3}}}$, ${{{\cal F}_{5}}}$, ${{\cal F_{\text{$\gamma$}}}}$ or ${{\cal F_{\text{$\epsilon$}}}}$, as appropriate. We also minimise $\Omega/A$ with respect to the domain size $A$ by applying the approach described in Appendix \[app:continuation\]. For a given value of the chemical potential $\mu$ and the combined parameter $\rho_0\beta\epsilon$, the thermodynamic equilibrium state is that with the minimum value of $\Omega/A$. Note that equilibria with the same $\mu$ do not necessarily have the same value of $\bar n$, which is important when considering which equilibria might result from initial conditions via the dynamics.
to to
The solution corresponding to the uniform density liquid state with $n({{\bm x}})={n_{\text{liq}}}$ can readily be found. In this case we have ${{\cal L}}{{n_{\text{liq}}}}=\sigma(0){n_{\text{liq}}}$, and so we must solve the following algebraic equations for ${n_{\text{liq}}}$: $$\begin{aligned}
&\text{DDFT-3,5:}&
\log\left(1+{n_{\text{liq}}}\right)
- {n_{\text{liq}}}- \sigma(0) {n_{\text{liq}}}&= \mu,
\label{eq:DDFTnliq}\\
&\text{{\hbox{PFC-$\gamma$}},$\epsilon$:}&
-\tfrac{1}{2}{n_{\text{liq}}}^2 + \tfrac{1}{3}{n_{\text{liq}}}^3 - \sigma(0){n_{\text{liq}}}&= \mu,
\label{eq:PFCnliq}
\end{aligned}$$ recalling that the value of $\sigma(0)$ depends on whether or not the gradient expansion is carried out (see Fig. \[fig:dispersionGEM4\]). Finding ${n_{\text{liq}}}$ for a given value of $\mu$ is done easily using Newton’s method, and the resulting ${n_{\text{liq}}}$ and specific ${\Omega_{\text{liq}}}$ are shown in Fig. \[fig:liquiddensity\]. In all cases, we see that ${n_{\text{liq}}}$ is an increasing function of $\mu$, while ${\Omega_{\text{liq}}}/A$ is a decreasing function of $\mu$. The figure shows that the specific grand potential for the liquid state predicted by all four models are similar close to $\mu=0$, but the predicted liquid state densities are rather different away from $\mu=0$. This difference originates from the different values of $\sigma(0)$ ($-18.75$ for DDFT-3 and , in contrast to $-4.37$ for DDFT-5 and ). We see from Fig. \[fig:liquiddensity\] that the density of the liquid is erroneously predicted to increase too rapidly as $\mu$ is increased by the gradient expansion theories (DDFT-5 and ). This is because these get the value of the isothermal compressibility $\chi_T$ to be too large [@Jaatinen2009]. This compressibility is related to $\sigma(0)$ via $\chi_T=-\beta/[\sigma(0)\rho_0(1+{n_{\text{liq}}})]$ [@Hansen2013]: see Eq. and following discussion. Expanding the logarithm makes relatively little difference over this range of densities.
Since crystallisation occurs at higher densities, we expect a transition from the liquid to the crystal to occur as $\mu$ increases. At the spinodal the uniform liquid becomes linearly unstable and the patterned state solution branches bifurcate from the liquid at this point. To find these states, we seek a solution of the form $$n({{\bm x}}) = {n_{\text{liq}}}+ {{\delta{\mkern-1mu}n}}({{\bm x}}),$$ where near the bifurcation point ${{\delta{\mkern-1mu}n}}\ll1$, and ${{\delta{\mkern-1mu}n}}$ is of the form $e^{i{{\bm k}}\cdot{{\bm x}}}$, so that ${{\cal L}}{{\delta{\mkern-1mu}n}}=\sigma(k){{\delta{\mkern-1mu}n}}$. Expanding Eqs. (\[eq:betadFdn3\_rewritten\]) and (\[eq:betadFdn5\_rewritten\]–\[eq:betadFdnepsilon\_rewritten\]) in powers of ${{\delta{\mkern-1mu}n}}$ we find that the ${{\cal O}}(1)$ equations to solve are just those for finding the liquid state density, Eqs. (\[eq:DDFTnliq\])–(\[eq:PFCnliq\]). The ${{\cal O}}({{\delta{\mkern-1mu}n}})$ equations are $$\begin{aligned}
&\text{DDFT-3,5:}&
\left(\frac{1}{1+{n_{\text{liq}}}} - 1 - \sigma(k)\right){{\delta{\mkern-1mu}n}}&= 0,
\label{eq:DDFTdeltan}\\
&\text{{\hbox{PFC-$\gamma$}},$\epsilon$:}&
\left(-{n_{\text{liq}}}+ {n_{\text{liq}}}^2 - \sigma(k)\right){{\delta{\mkern-1mu}n}}&= 0.
\label{eq:PFCdeltan}
\end{aligned}$$ The spinodal point for DDFT-3,5 or for ,$\epsilon$ is where there are solutions of the equation with ${{\delta{\mkern-1mu}n}}\neq0$.
Since we are looking for a change in stability, we take the extreme value of $\sigma(k)$, i.e., $\sigma(k)=0$ (see Fig. \[fig:dispersionGEM4\]). Then, Eq. (\[eq:DDFTdeltan\]) is solved (with ${{\delta{\mkern-1mu}n}}\neq0$) only for ${n_{\text{liq}}}=0$, which leads to $\mu=0$ from (\[eq:DDFTnliq\]). In contrast, Eq. (\[eq:PFCdeltan\]) with $\sigma(k)=0$ has two solutions, ${n_{\text{liq}}}=0$ and ${n_{\text{liq}}}=1$, leading to $\mu=0$ and $\mu=-\tfrac{1}{6}-\sigma(0)$ from (\[eq:PFCnliq\]). The implication of this is that the PFC has two spinodal points: the liquid loses stability at ${n_{\text{liq}}}=0$ as $\mu$ increases through $0$, but it regains stability at ${n_{\text{liq}}}=1$, which gives $\mu=18.58$ for and $\mu=4.20$ for . This prediction that the liquid regains stability for higher $\mu$ is a consequence of expanding the logarithm, or equivalently of Taylor expanding the $1/(1+{n_{\text{liq}}})$ term in (\[eq:DDFTdeltan\]) and is confirmed by direct computation of the crystal solutions below. Of course, this prediction is erroneous, since the simulation results for the GEM-4 system [@Mladek2006; @Prestipino2014] show no sign of a second spinodal point or the associated stable second liquid in the equilibrium system phase diagram.
In Fig. \[fig:DDFT3solutions\] we display examples of the three different types of periodic solutions that can be found for DDFT-3. These are (i) the crystal solution, which we refer to as ‘up hexagons’, which exhibits a triangular array of isolated density maxima surrounded by hexagonal regions where the density is close to zero. There are also (ii) ‘down hexagons’ which are the opposite, with isolated density minima and hexagonal density maxima. Finally, there is (iii) the stripe state. Depending on the state point these solutions are not necessarily linearly stable. Our naming convention to distinguish the two different hexagonal solutions originates in the convection literature [@Bodenschatz2000]. These solutions were initiated at $\mu\approx0$ and then continued numerically (see Appendix \[app:continuation\]) up to $\mu=10$. For DDFT-3 it is possible to go a bit higher in $\mu$, but with increasing $\mu$ (i.e., increasing average density) the peaks in the density profile get narrower and higher and so more and more grid points are required to resolve these peaks correctly. However, as we show below for some of the other models and for different reasons, it is not possible to continue the solutions this far in $\mu$. The domains on which the profiles are calculated have periodic boundary conditions, with 4 wavelengths in each direction (for stripes), or $8\times\tfrac{8}{\sqrt{3}}$ wavelengths (for hexagons). The wavelength is initially equal to $2\pi$ for $\mu=0$ and is then adjusted by up to about 2% in order to minimise the specific grand potential as $\mu$ is varied; i.e., we minimise $\Omega/A$ with respect to variations in the size of the crystal unit cell or, for the stripe phase, we minimise with respect to variations in the spacing between the stripes – see Appendix B for details.
to to to to
In Fig. \[fig:DDFTPFCsummary\] we display a series of plots showing the maximum, minimum and average values of the density profiles $1+n$ for the stripe and hexagonal structures as a function of $\mu$. We also plot the specific grand potential $\Omega/A$ for the different structures. Recall that for a given $\mu$ the thermodynamic equilibrium phase corresponds to the global minimum of $\Omega/A$. The results for DDFT-3 are shown in Fig. \[fig:DDFTPFCsummary\](a–c). The (a) stripes originate in a supercritical pitchfork bifurcation at $\mu=0$, and (b) hexagons originate in a transcritical bifurcation at the same value of $\mu$. The density of the up hexagons ranges from about $2\times10^{-5}$ up to about 50, for $\mu=10$. All of these branches can be continued to larger values of $\mu$.
DDFT-5, in Fig. \[fig:DDFTPFCsummary\](d–f), initially behaves in the same way, but all three branches have their minimum density heading to zero before $\mu$ gets to 10: this happens at $\mu\approx3.37$ for (d) stripes, and for $\mu\approx0.28$ and $\mu\approx2.73$ for (e) the up and down branches of hexagons, respectively. The numerical method cannot continue the branches beyond these points. We argue in Sec. \[sec:4\] that this is not an artefact of the numerical method, rather it is a genuine feature of solutions of Eq. (\[eq:betadFdn5\_rewritten\]) that the density $1+n$ can go to zero. In this limit, $\log(1+n)\rightarrow-\infty$, but this is balanced by a lack of smoothness in $n({{\bm x}})$: the fourth derivative in ${{{\cal L}_{\text{grad}}}}n$ can go to $+\infty$ and so balance the singularity in $\log(1+n)$. Therefore, $\mu\approx0.28$ is the limit of validity of the DDFT-5 model.
The two PFC examples are similar to each other, and it is easier to discuss , in Fig. \[fig:DDFTPFCsummary\](j–l), first. Here, (j) stripes and (k) hexagons bifurcate from the liquid at $\mu=0$, but they rejoin the liquid at $\mu=4.20$ as explained in the discussion following Eqs. (\[eq:DDFTdeltan\]–\[eq:PFCdeltan\]). The maximum and minimum densities for the up and down hexagon cross between the two bifurcations. The behaviour of , in Fig. \[fig:DDFTPFCsummary\](g–i), is similar, though the second bifurcation is at $\mu=18.58$, off the scale of the figure.
to to
Figure \[fig:DDFTPFCsummary\](c,f,i,l) shows that the curves of the specific grand potential $\Omega/A$ as functions of $\mu$ are very close, so in Fig. \[fig:DDFTPFCrelativeOmega\], we plot instead $(\Omega-{\Omega_{\text{liq}}})/A$ versus $\mu$, where ${\Omega_{\text{liq}}}$ is the specific grand potential for the liquid at the same value of $\mu$. In (a) the DDFT-3 case, the up hexagons clearly have the lowest grand potential for $\mu\geq-2.8$ with the uniform liquid being the global minimum for $\mu<-2.8$, and at no point do stripes come anywhere near, as one should expect from the particle simulation results [@Prestipino2014]. For (b) DDFT-5, the two hexagon branches stop before the stripe branch when their minimum densities go to zero (the limit of validity), but otherwise the relative values for the hexagon and stripe grand potentials is qualitatively similar to DDFT-3. For (c,d) the PFC examples, once again it is easiest to discuss first. In Fig. \[fig:DDFTPFCrelativeOmega\](d), the hexagon and stripe branches bifurcate from the liquid at $\mu=0$ and rejoin the liquid at $\mu=4.20$, with stripes having the lowest grand potential for intermediate values of $\mu$, and up or down hexagons being the lowest grand potential state for smaller or larger values of $\mu$. The behaviour of (c) is similar, but stretched to larger values of $\mu$ (off scale).
The insets in the four panels of Fig. \[fig:DDFTPFCrelativeOmega\] display magnifications that show that the behaviour near the spinodal point at $\mu=0$ is qualitatively similar in all four cases: the up hexagons start with $\Omega>{\Omega_{\text{liq}}}$ for negative $\mu$, but the branch changes direction, forming a cusp, close to which is the thermodynamic coexistence point (Maxwell point), where $\Omega={\Omega_{\text{liq}}}$. The down hexagons start with $\Omega<{\Omega_{\text{liq}}}$ for positive $\mu$, and the stripes, also with $\Omega<{\Omega_{\text{liq}}}$ for positive $\mu$, have a value of the grand potential intermediate between the up and down hexagons. We note that the range of $\mu$ over which this behaviour occurs is about a factor of ten smaller in the DDFT-5 and cases as compared to DDFT-3, also with a roughly ten-fold drop in the overall range of values of $(\Omega-{\Omega_{\text{liq}}})/A$.
to to
The observation that the bulk phase behaviour of the system depends only on $\mu$ and the value of $\rho_0\beta\epsilon$ if the pair potential can be written as $u(x_{12})=\epsilon\psi(x_{12})$ – see the discussion around Eq. – is true for the GEM-4 system. As a consequence, having calculated the coexisting densities for a particular value of $\rho_0\beta\epsilon$, the linear stability threshold, these results can be scaled to give the phase diagram in the full average density $1+\bar{n}$ versus dimensionless temperature $k_BT/\epsilon$ plane, which is one of the usual ways the GEM-4 phase diagram is displayed [@Mladek2006; @Mladek2007; @Prestipino2014; @Archer2014; @Archer2016]. The phase diagrams obtained from doing this are in Fig. \[fig:phase\_diagrams\]. In Fig. \[fig:phase\_diagrams\](a) we display the phase diagram obtained from DDFT-3, which is identical (to within the resolution of the calculations) to that previously calculated in Refs. [@Archer2014; @Archer2016]. For example, when $\beta\epsilon=1$, the average densities $\rho_0(1+\bar{n})R^2$ of the coexisting liquid and the crystal are $5.41$ and $5.68$, respectively. As a result of the scaling behaviour, the coexisting densities (binodals) are two straight lines going from the origin and passing through these two points.
In Fig. \[fig:phase\_diagrams\](b) we display the phase diagram obtained for the DDFT-5. The binodals are a little closer to the linear stability threshold line than for DDFT-3, but other than that, it looks similar overall. Note however that the up hexagon branch cannot be continued beyond $\mu\approx0.28$ (where the minimum density goes to zero): this line is indicated as the ‘limit of validity’. Beyond this line, in the bottom right region of the phase diagram, there is no up hexagon solution to the equations, for the reasons discussed in Sec. \[sec:4\].
In Fig. \[fig:phase\_diagrams\](c) we display the phase diagram for and in (d) for . The binodals almost overlie each other, so the predicted difference between the average densities of the liquid and the crystal at coexistence are much smaller than that predicted by DDFT-3 and DDFT-5. Furthermore, on moving to higher average densities or to lower temperatures $k_BT/\epsilon$ one encounters the stripe phase, followed by the down hexagon phase and then finally the uniform liquid becomes stable again. The prediction of the occurrence of these later phases is of course wrong, signifying a breakdown in the accuracy of the PFC theory at even the qualitative level.
Before finishing this section, we note that it is possible to extend the gradient expansion in (\[eq:cLgradGEM4\]) by including higher powers of the Laplacian. For example, Ref. [@Jaatinen2009] proposed an eighth-order fitting (EOF), which in our notation is $${{{\cal L}_{\text{grad-8}}}}n({{\bm x}}) = - \gamma(1+\nabla^2)^2 n({{\bm x}})
- E_B (1+\nabla^2)^4 n({{\bm x}}),
\label{eq:cLgradJaatinen}$$ where $\gamma$ fits the curvature of the dispersion relation as before, and $E_B$ allows the eigenvalue $\sigma(0)$ of ${{\cal L}}$ to be matched as well, i.e., allows the model to match correctly the isothermal compressibility $\chi_T$. An example of the dispersion relation for this operator is shown as a dotted line in Fig. \[fig:dispersionGEM4\]. This EOF version of the theory, with ${{{\cal L}_{\text{grad-8}}}}$ in (\[eq:cLgradJaatinen\]), improves over the standard version, with (\[eq:cLgradGEM4\]), since $\sigma(0)$ for ${{\cal L}}$ and ${{{\cal L}_{\text{grad-8}}}}$ are the same. Therefore, the liquid properties of DDFT-5 match those of DDFT-3, and the liquid properties of match those of , once ${{{\cal L}_{\text{grad}}}}$ is replaced by ${{{\cal L}_{\text{grad-8}}}}$.
However, the drawbacks of the gradient expansion are still present. With ${{{\cal L}_{\text{grad-8}}}}$, the values of $\mu$ at which the DDFT-5 stripe and hexagon densities go to zero are larger, but this undesirable feature is only deferred, not eliminated. The reason is that the singularity in the logarithm is now balanced against an eighth-order derivative. Note too that introducing even higher derivatives does not cure this problem, it just pushes the singularity to higher order. In addition, with ${{{\cal L}_{\text{grad-8}}}}$ the second liquid spinodal in the model is still present, it is just pushed to higher values of $\mu$ (similar to the value for ) and since this second spinodal is present, there is still a range of values of $\mu$ for which stripes have the lowest specific $\Omega$.
Effect of the approximations {#sec:4}
============================
The qualitative change in going from a DDFT to a PFC model (dropping the $\nabla\cdot\left[n\nabla{{\cal L}}{n}\right]$ term, assuming constant mobility and expanding the logarithm) is apparent in the phase diagrams shown in Fig. \[fig:phase\_diagrams\], comparing (a,b) to (c,d). Here we discuss in detail additional effects of the approximations.
Expanding the logarithm
-----------------------
The effect of expanding the logarithm as in Eq. (\[eq:expandlogarithm\]) is very significant. In Sec. \[sec:3\] we demonstrated that this expansion leads to the liquid having a second spinodal point, illustrated in the phase diagrams in Fig. \[fig:phase\_diagrams\](c,d), with the crystal re-melting as the density is increased. The reason for this is that Eq. (\[eq:PFCdeltan\]) can be solved with $n=0$ and $n=1$, while Eq. (\[eq:DDFTdeltan\]) is only solved by $n=0$, with $\sigma(k)=0$ in both cases.
Intimately connected to the existence of this second spinodal is the transition from stable up hexagons at the $\mu=0$ spinodal to stable down hexagons at the higher density spinodal. These connections to the spinodals mean that the free energy of the up hexagons increases again compared to the liquid state free energy as the chemical potential is increased, in order to reconnect to the liquid state at the upper spinodal. An intermediate region of stable stripes is not inevitable, but is evident in both PFC examples in Fig. \[fig:DDFTPFCrelativeOmega\](c,d).
Taking more terms in the expansion in Eq. (\[eq:expandlogarithm\]) does not help. The highest power should be even (otherwise the free energy is not bounded below), and the improved versions of Eq. (\[eq:PFCdeltan\]), which involves the second derivative of Eq. (\[eq:expandlogarithm\]) with respect to $n$, also have $n=0$ and $n=1$ as (the only) real roots, regardless of how many terms are kept in the expansion of the logarithm. [The exception is if only terms up to $n^2$ are kept in (\[eq:expandlogarithm\]); in this case, $c^{(3)}$ and $c^{(4)}$ (if they are non-zero) provide the stabilizing nonlinearities and may also lead to a spurious spinodal.]{}
to to
Gradient expansion of ${{\cal L}}$
----------------------------------
As discussed in Sec. \[sec:3\] above, the results in Fig. \[fig:DDFTPFCsummary\](d,e) suggest that for certain values of $\mu$, the DDFT-5 equation (\[eq:betadFdn5\_rewritten\]) has solutions for the density $1+n$ which go to zero at certain places. In contrast, the density in the PFC models appears to stay away from zero (although there is no reason for it to do so and there would be no singularity if it did), and in DDFT-3, the density minimum gets smaller and smaller as $\mu$ increases, but remains positive, without the sharp cutoff seen in DDFT-5. In this section we argue that the density reaching zero is not an artefact of numerical difficulties, rather it is a feature of the DDFT-5 equation (\[eq:betadFdn5\_rewritten\]). [Here, we focus on singularities in the solution, not on stability. Our discussion in this section is mainly framed in terms of the stripe solution, which is unstable, but the stable up-hexagon branch has similar issues, as is illustrated in Fig. \[fig:DDFTPFCsummary\](e).]{}
Figure \[figApp:DDFT5singularity\](a) shows that even close to the end of the branch of DDFT-5 stripes, the density profile $1+n$ remains smooth, but since its minimum at $x=x_{\rm min}$ is very close to zero ($1+n(x_{\rm min})\approx5\times10^{-7}$), therefore the logarithm $\log(1+n)$ is sharply spiked towards large negative values at $x_{\rm min}$. Writing out the terms in (\[eq:betadFdn5\_rewritten\]) for a density profile only varying in the $x$-direction, we have: $$\log\left(1+n\right)
+ (\gamma-1) n + \gamma n_{xx} + \gamma n_{xxxx} - \mu = 0,
\label{eq:betadFdn5_rewrittenagain}$$ which suggests that the only way to balance a large negative contribution from $\log(1+n)$ is to have a large positive $\gamma{n_{xxxx}}$. Figure \[figApp:DDFT5singularity\](b) shows that these two terms (solid lines with markers at the grid points) do indeed balance each other. The figure also shows that the other terms in (\[eq:betadFdn5\_rewrittenagain\]) are well behaved and that the equation is satisfied at each grid point. Therefore this singularity is not a numerical artefact, but rather a genuine feature of DDFT-5 stripes: the minimum density goes to zero at a certain finite value of $\mu$. In Fig. \[figApp:DDFT5singularity\](c,d) we show for comparison results from DDFT-3: the stripes have sharply peaked density maxima and density minima just as small as in DDFT-5, but all terms in Eq. (\[eq:betadFdn3\_rewritten\]) (Fig. \[figApp:DDFT5singularity\]d) are well behaved.
As $\mu$ is further increased, the minimum of the density in DDFT-5 gets closer to zero, so the logarithm of the density goes further towards $-\infty$ and correspondingly $\gamma{n_{xxxx}}$ goes towards $+\infty$. Fig. \[fig:DDFTPFCsummary\](d) shows that the density minimum gets to zero at a finite value of $\mu$. We have not been able to develop a consistent asymptotic approximation for this limit in the DDFT-5 equation. However, to illustrate that apparently smooth solutions with logarithmic singularities in their fourth derivatives can easily be found, consider for example taking $\gamma=1$ in (\[eq:betadFdn5\_rewrittenagain\]) and taking a density profile that has a quadratic minimum at $x=x_{\rm min}=0$: $$1 + n(x) = Ax^2 + Bx^4 + Cx^4\log(x^2),
\label{eqApp:densityexpansion}$$ where $A$, $B$ and $C$ are constants. For small $x$, the largest of these three terms is $Ax^2$, so $\log(1+n)\approx\log(Ax^2)$, which goes to $-\infty$ as $x\rightarrow0$. The other terms are $n_{xx}\approx2A+{{\cal O}}(x^2,x^2\log(x^2))$ and $n_{xxxx}\approx24C\log(x^2) + 24B + 100C$. Adding these three together requires $1+24C=0$ to cancel the logarithmic singularity at $x=0$, and the remaining terms are constants or go to zero as $x\rightarrow0$.
As can be seen from Fig. \[figApp:DDFT5singularity\], having an adequate resolution for our numerical calculations was a challenge but for different reasons for the different models. In DDFT-3, the density maxima can be sharply peaked while the logarithm of the density is smooth, so inadequate resolution in the density field prompts an increase in the number of grid points (we implement automatic regridding, as discussed in Appendix \[app:continuation\]). In contrast, in DDFT-5, the density field can be smooth but with minima very close to zero, so its logarithm has very sharp negative peaks. In this case, inadequate resolution in the logarithm of the density prompts regridding. The difference is that in DDFT-5, the equation involves derivatives so any problem is magnified, while in DDFT-3, the equation involves convolutions that smooth out any problems.
These arguments indicate that a singular solution to the DDFT-5 equation (\[eq:betadFdn5\_rewritten\]) of the type seen in Fig. \[figApp:DDFT5singularity\] is possible, with the density going to zero, and that this is not a problem of inadequate numerical resolution, but rather a consequence of replacing the convolution in DDFT-3 with derivatives. A full asymptotic theory should result in a prediction for the value of $\mu$ at which the branch terminates. The stripe solutions of the DDFT-3 equation (\[eq:betadFdn3\_rewritten\]) can also have small density but without any singularity in the solution.
One mode approximation for DDFT {#sec:5}
===============================
The data displayed in Fig. \[figApp:DDFT5singularity\](a,c) lead to an interesting observation: in (a) DDFT-5, the logarithm of the density is sharply negatively peaked, while the density is smooth (at least up to its second derivative), slowly varying and resembles a cosine. In contrast, in (c) DDFT-3, the density is sharply peaked, while the logarithm of the density is slowly varying and resembles a cosine. One of the attractions of PFC theory is that it has slowly varying solutions that are well represented by a few Fourier modes [@Elder2002; @Elder2004; @Emmerich2012; @Wu2010], and this carries over to some extent to DDFT-5. Such Fourier representations of the density profiles in DDFT-3 are unsatisfactory, apart from for the unstable solutions very close to the spinodal point, since any solution of reasonable amplitude is sharply peaked.
to
However, the data in Fig. \[figApp:DDFT5singularity\] suggest that representing the *logarithm* of the density with a few Fourier modes should work well in DDFT-3. In this section, we elaborate how such a theory can be developed.
The key is to write Eq. (\[eq:betadFdn3\_rewritten\]) in terms of $\phi({{\bm x}})\equiv\log(1+n({{\bm x}}))$, and approximate $\phi$ by a few Fourier modes. As a first step, we write out ${{\cal L}}$ using Eq. (\[eq:defnL\_GEM4\]) and re-write (\[eq:betadFdn3\_rewritten\]) as: $$\log(1+n({{\bm x}}))
+ \rho_0\beta\epsilon \!\!\int\!
\psi(|{{\bm x}}-{{\bm x}}_2|) n({{\bm x}}_2){{{\rm d}}{{\bm x}}}_2
- \mu = 0,
\label{eq:betadFdn3_rewrittenagain}$$ where $\psi(|{{\bm x}}-{{\bm x}}_2|)=e^{-|{{\bm x}}-{{\bm x}}_2|^4/R^4}$. The convolution term (including the $\rho_0\beta\epsilon$ prefactor) in this equation is $-n({{\bm x}})-{{\cal L}}n({{\bm x}})$. We know the eigenvalues of ${{\cal L}}$: ${{\cal L}}\exp(ikx)=\sigma(k)\exp(ikx)$, which means that the convolution term acting on a Fourier mode $\exp(ikx)$ has eigenvalue $-(1+\sigma(k))$. We also know that for high wavenumbers, the convolution averages to zero, and indeed $\sigma(k)\rightarrow-1$, as can be seen in Fig. \[fig:dispersionGEM4\].
We focus first on stripes, which have Fourier components only at integer wavenumbers, and notice that $1+\sigma(2)$ is already very small (less than $0.01$), because $\hat{\psi}(2)$ is small. This implies that the Fourier components of the convolution term at $k=2$ and higher will be much smaller than the Fourier components at $k=0$ and $k=1$ – regardless of the spectrum of $n$ itself. The other two terms in Eq. (\[eq:betadFdn3\_rewrittenagain\]) are $\mu$, which is constant ($k=0$ only), and $\log(1+n)$, so $\log(1+n)$ can only have significant Fourier components at $k=0$ and $k=1$: there is nothing to balance modes with $|k|\geq2$. This explains why $\log(1+n)$ in the lower left panel of Fig. \[figApp:DDFT5singularity\] is smooth, and it is also why approximating the logarithm of the density by a few Fourier modes must work regardless of the amplitude of the modulations in the density, or how sharply they are spiked, or of the value of $\mu$.
For the stripe phase we write $$\log(1+n(x)) = \phi(x) = \phi_0 + \phi_1 e^{ix} + {\bar \phi}_1 e^{-ix},
\label{eq:logdensityFourier}$$ where $\phi_0$ and $\phi_1$ are constants (real and complex, respectively) that we need to find. This can easily be generalised for hexagons and other periodic phases by adding more modes in . The $k=0$ and $k=1$ components of $\exp\left(\phi_1e^{ix}+\text{c.c.}\right)$, where $\text{c.c.}$ denotes the complex conjugate, can be expressed in terms of integrals, defining two functions $f_0(\phi_1)$ and $f_1(\phi_1)$ given by $$\begin{split}
f_0(\phi_1) &= \frac{1}{2\pi}\int_0^{2\pi} \exp\left(\phi_1e^{ix}+\text{c.c.}\right){{\rm d}}{}x, \\
f_1(\phi_1) &= \frac{1}{2\pi}\int_0^{2\pi} e^{-ix}\exp\left(\phi_1e^{ix}+\text{c.c.}\right){{\rm d}}{}x,
\label{eq:f0f1defn}
\end{split}$$ i.e., $f_0$ is modified Bessel function of the first kind of order zero and $f_1$ is a Fourier transform generalisation of $f_0$. Using these functions, $n(x)=e^{\phi(x)}-1$ can be written in terms of its Fourier components as $$\begin{split}
e^{\phi(x)}-1 &= e^{\phi_0}f_0(\phi_1) - 1\\
&\quad{} + \left(e^{\phi_0}f_1(\phi_1)e^{ix} + \text{c.c.}\right)\\
&\quad{} + \text{modes with $|k|\geq2$.}
\label{eq:densityFourier}
\end{split}$$ The modes with $|k|\geq2$ in (\[eq:densityFourier\]) are large in amplitude, but they are reduced in significance in Eq. (\[eq:betadFdn3\_rewrittenagain\]) by the convolution, as explained above. The action of the convolution on modes with $|k|<2$ can be written in terms of $\sigma(k)$. Retaining only these terms, we are left with the $k=0$ and $k=1$ components of (\[eq:betadFdn3\_rewrittenagain\]): $$\begin{split}
\phi_0 + (1+\sigma(0))\left(1-e^{\phi_0}f_0(\phi_1)\right) - \mu &= 0,\\
\phi_1 - e^{\phi_0}f_1(\phi_1)(1+\sigma(1)) &= 0.
\label{eq:DDFT3onewaveeqns}
\end{split}$$ Notice that the only information remaining from the GEM-4 potential is the values of $\sigma(0)$ and $\sigma(1)$, i.e., the values of $\hat{u}(0)$ and $\hat{u}(1)$. Recall also that if the reference density $\rho_0$ is chosen to be the value at the spinodal, then we have $\sigma(1)=0$. These equations can also be written in terms of the pair potential as $$\begin{split}
\phi_0 - \rho_0\beta\hat{u}(0)\left(1-e^{\phi_0}f_0(\phi_1)\right) - \mu &= 0,\\
\phi_1 + e^{\phi_0}f_1(\phi_1)\rho_0\beta\hat{u}(1) &= 0.
\label{eq:DDFT3onewaveeqns_u}
\end{split}$$ The two equations in (\[eq:DDFT3onewaveeqns\]) can easily be solved for $\phi_0$ and $\mu$ in terms of $\phi_1$, from which the density can be reconstructed. The agreement between this and the full solutions of DDFT-3 is astonishing. Figure. \[figApp:DDFT3onemodeapprox\](a) shows a 1D example at $\mu=28.9$, with the full solution as a black line and the approximate solution as a dashed line. Even though the density varies by two orders of magnitude, the two are almost indistinguishable. There is similar excellent agreement with the branch of stripe solutions (Fig. \[figApp:DDFT3onemodeapprox\]b) in 2D DDFT-3 (recall that the GEM-4 potential has different values of $\sigma(0)$ in 1D and 2D).
For 2D hexagons, the approach is similar, with the two $e^{\pm{ix}}$ terms in Eq. (\[eq:f0f1defn\]) replaced by six similar terms, with wavenumbers ${{\bm k}}$ that are uniformly spaced around a circle of radius 1 in ${{\bm k}}$-space (c.f. Eq. in Appendix B). The agreement in this case (Fig. \[figApp:DDFT3onemodeapprox\]c) is also very good. If one adds a further six modes with $|{{\bm k}}|=\sqrt{3}$, then the agreement is as good as that for stripes. These approximate solutions can easily be continued up to $\mu=100$ without difficulties, where we observe very sharply peaked density maxima and extremely small but non-zero values for the density minimum.
Discussion and conclusions {#sec:6}
==========================
In this paper, starting from DDFT, we have presented a step-by-step derivation of PFC theory, at each stage explaining the consequences of the approximations. The approximations can be listed under three main groupings: (i) making a truncated functional Taylor expansion approximation for the excess Helmholtz free energy, and then making the RPA/RY approximation. This leads to DDFT-3 in our classification. (ii) Neglecting the $\nabla\cdot[n\nabla{{\cal L}}n]$ term, which effectively also forces making a Taylor expansion of the logarithmic ideal gas term and assuming constant mobility. (iii) Replacing the nonlocal convolution in ${{\cal L}}$ with a local gradient expansion. The consequence of (ii) is to introduce a second spinodal into the phase diagram and to significantly alter the relative stabilities of the different periodic states, in particular making striped states to become an equilibrium phase for some state points, which is contrary to the physics. The consequence of making (iii) without first making (ii), is to generate a theory (DDFT-5) that has a no-solution region in the phase diagram, such as that displayed for the GEM-4 model in Fig. \[fig:phase\_diagrams\](b). All these consequences have been illustrated for the GEM-4 system, chosen because DDFT-3 is fairly accurate for this model for temperatures $k_BT/\epsilon>0.1$, allowing us to see the influence of the subsequent approximations.
[Throughout, there is good quantitative agreement between DDFT-3, and the EOF versions of DDFT-5 and (data not shown) *only* for unstable small amplitude solutions close to the spinodal point. The region of quantitative agreement agreement between DDFT-3 and is circled in red in Fig. \[fig:DDFTPFCsummary\](b,h). Beyond this region, the agreement between the four theories is at best qualitative.]{}
Given all these problematic consequences for PFC theory, especially the issues related to Taylor expanding the logarithmic ideal gas term, it raises the question of why then is PFC theory so successful? In our view, there are several reasons for this. The first reason is that PFC theory is qualitatively correct near to the spinodal. Therefore, it can satisfactorily describe the coexistence between the liquid and the crystal phase, which is often an important aspect in applications of the theory. Second, despite the approximations, PFC theory still incorporates some very important physics: (i) the free energy satisfies the correct symmetries, (ii) the dynamical equation gives a time evolution that decreases the free energy monotonically over time and (iii) the current is proportional to the gradient of the chemical potential. These are all important features for describing many phenomena. Also, many of the features that PFC theory is used to describe are generic, and the model parameters can be scaled to fit (for example) iron [@Jaatinen2009] and graphene [@Fan2017], but could equally well be scaled to match other materials, with similar good agreement. This universality (having the correct symmetries etc.) underlines the importance of PFC theory as a powerful model of generic features of crystalization, but it means that PFC theory will in general not be able to predict any unusual (non-generic) behaviour.
Our results in Sec. \[sec:5\] for the GEM-4 system showing that one can derive a very accurate one-mode amplitude equation approximation for the field $\phi({{\bm x}})=\log(\rho({{\bm x}})/\rho_0)$, rather than the density itself, gives a tantalising hint as to how PFC-type theories may more properly be derived and what the order parameter field in PFC theory really represents: Should we consider the PFC order parameter to be a scaled logarithm of the density distribution or some other similar function of the density, rather than being proportional to the density profile itself? On the basis of the work presented here, the answer to this question is ‘probably yes’, but clearly more work is required to fully address this.
Returning to theories for the density profile, in our view it is preferable to retain the logarithmic ideal gas term in the approximation used for the Helmholtz free energy functional, since this is required to have physical (i.e., positive) density profiles, and also because with this the DDFT dynamics leads to the (correct physics) linear diffusion equation term – see Eq. . The difficult consequence of retaining this in the free energy is that one then has to deal with terms in the dynamical equation of the form $\nabla\cdot[n\nabla{{\cal L}}n]$, which makes solving numerically far more difficult. However, this term also contributes to making the crystalline phase more stable than the stripe phase, which also makes it important. The crucial contribution of this term can especially be seen in DDFT-3, since in this version of the theory it is clearly required for stabilizing the crystal structure. [In general, we would advocate using one of DDFT-1, DDFT-2 or DDFT-3 (depending on how ${{\cal F_{\text{ex}}}}$ is treated) over all existing PFC theories, for studying the properties of real materials.]{}
It is worth noting that in all the approximations made here, we only consider those that retain the form of a dynamics that decreases the free energy monotonically over time – see Eq. . This is a feature of both DDFT and PFC theory. In our view this structure is important and should not be broken by any approximations made, i.e., any that might be made in future in attempting to avoid any of the above mentioned issues.
It is also worth noting that, while we have not discussed the consequences of the LDA in going from DDFT-1 to DDFT-2 (see Eq. (\[eq:definecs\])), the polynomial terms in the chemical potential (\[eq:betadFdn2\]) can potentially lead to the same problem of having a second spinodal, even while retaining the logarithm term. This applies to DDFT-4 as well.
In this paper we have largely focussed on making our arguments in two dimensions, in order to keep the presentation as simple and comprehensible as possible. However, we should emphasis that all of our arguments apply for three dimensional (3D) systems. For example, at the higher temperatures we have focussed on here, the 3D GEM-4 model exhibits at equilibrium a single fluid phase and two crystalline phases: the body-centered cubic phase at lower densities and the face-centered cubic structure at higher densities. These are all accurately predicted by the 3D version of DDFT-3 [@Mladek2006]. There are no columnar or lamellar phases, which are the 3D equivalents of the stripe phase. On the other hand, the 3D versions of the PFC theories presented here all predict a lamellar phase at some state points [@Thiele2013], the 3D generalisations of the down hexagons and a second spinodal with the uniform liquid becoming the equilibrium phase at higher densities.
[It would be interesting to explore how the dynamics of defects, the elasticity and the plasticity of crystals differs between DDFT and ]{}, and our results are also relevant to binary systems. In the derivation in Ref. [@Huang2010a] of a PFC theory for binary mixtures, the generalisation of the $\nabla\cdot[n\nabla{{\cal L}}n]$ term is retained until the last moment in the derivation, but then dropped for the same reasons that is is neglected for one-component systems. Given the importance of this term for one-component systems, it is surely also important for stabilizing crystal structures in binary systems. [Note also that when determining mechanical properties such as elastic constants, the terms in the free energy that are linear in $n$ can be important [@Wang2018b]. These have been neglected here throughout since such terms do not contribute to determining density profiles.]{}
The singularity observed for DDFT-5 as the chemical potential $\mu$ (or equivalently, the average density $1+\bar{n}$) is increased was found by continuing equilibrium solutions determined at lower values of $\mu$. One aspect that needs further investigation relates to determining the influence of this when DDFT-5 is solved for state points where the final equilibrium crystal (or stripe) solution for the density profile does not exist. For example, a situation we have in mind is that studied by van Teeffelen [@Teeffelen2009] consisting of a solidification front propagating into the unstable liquid. These authors compared results for this situation between (in our terminology) DDFT-3, DDFT-5 and . Their results are for two-dimensional dipolar colloidal particles. From the DDFT-5 results (PFC1 in their terminology) displayed in Figs. 4 and 5 of Ref. [@Teeffelen2009], it can be seen that they did not consider values of the coupling parameter large enough to encounter any of the singularities; the density profiles stay well away from zero. It would be interesting to quench deeper into the crystal phase to study the evolution of the density distribution towards the singular state. However, the numerics to resolve this accurately would surely be difficult.
One aspect of PFC theory that the derivation from DDFT highlights is that in general one is not free to independently vary the parameters $r$ and $\bar{n}$ in Eqs. and . For example, for the GEM-4 model there are certain values of $r$ that are not generated by the mapping from DDFT. Of course, by changing the pair potential, different combinations of the PFC model parameters can become accessible. We should also recall that although we have illustrated many of our conclusions by considering the soft-core GEM-4 model, PFC theory can be derived for systems of particles with hard potentials since it is the pair direct correlation function $c^{(2)}({{\bm x}}_1,{{\bm x}}_2)$ that enters the theory; this quantity is finite for all values of ${{\bm x}}_1$ and ${{\bm x}}_2$.
As a final point, we mention that our results will also be of interest to the pure mathematics community. DDFT-3 is also referred to as the McKean–Vlasov equation and in this context there are a number of recent interesting rigorous results [@Carrillo2018; @Gomes2019]. Our results for DDFT-5, showing that for a finite value of $\mu$ there is a singularity with the density profile going to zero, may well be of interest to those who study the mathematics of solutions to partial differential equations with compact support – see for example Ref. [@Bernis1992]. For values of $\mu$ beyond the singular point where $1+n({{\bm x}})\to0$ it may be that the solutions become complex. If one were interested to find these solutions, we believe it might require treating $\mu$ as a complex variable. Of course, all of this is out of the realm where the model represents a theory for matter.
This work was supported in part by a L’Or[é]{}al UK and Ireland Fellowship for Women in Science (PS), by the EPSRC under grants EP/P015689/1 (AJA, DR) and , and by the Leverhulme Trust (RF-2018-449/9, AMR). We are grateful for conversations with Tapio Ala-Nissilä, Daniele Avitabile, Ken Elder, Zhi-Feng Huang, Kai Jiang, Edgar Knobloch, Ron Lifshitz, Chris Malcotte, Daniel Read, Uwe Thiele, Steve Tobias, Gyula Tóth, Laurette Tuckerman and Joanna Tumelty. We are grateful also for constructive comments from two anonymous referees.
Linear theory for GEM-4 {#app:LinearGEM4}
=======================
In this appendix, we discuss how we compute the linear theory for the GEM-4 potential in (\[eq:GEM4Potential\]). To be specific, in a two-dimensional periodic domain, the eigenvalue $\sigma(k)$ is defined by ${{\cal L}}e^{i{{\bm k}}\cdot{{\bm x}}} = \sigma(k)
e^{i{{\bm k}}\cdot{{\bm x}}}$ and (\[eq:defnL\_GEM4\]), which can be written as $$\sigma(k) = - 1 - \rho_0\beta\epsilon \!\!\int\!
e^{-|{{\bm x}}-{{\bm x}}_2|^4/R^4} e^{i{{\bm k}}\cdot({{\bm x}}_2-{{\bm x}})}{{{\rm d}}{{\bm x}}}_2,
\label{eqApp:defn_sigmak_GEM4}$$ where the integral is taken over the periodic domain (the GEM-4 potential is replaced by its periodic extension). We set ${{\bm k}}=(k,0)$, we integrate from $[-N\pi/k,N\pi/k]$ in each dimension, and we choose the integer $N$ large enough that the GEM-4 exponential is effectively zero at the boundaries; $N=4$ suffices. We then scale ${{\bm x}}$ by a factor of $k$, replacing ${{\bm x}}_2-{{\bm x}}$ by ${{\bm x}}/k$, so that the integral becomes: $$\sigma(k) = - 1 - \rho_0\beta\epsilon
\!\!\iint_{-N\pi}^{N\pi}\!
e^{-|{{\bm x}}|^4/(k^4R^4)} e^{ix}\frac{{{{\rm d}}{{\bm x}}}}{k^2},
\label{eqApp:defn_sigmak_GEM4_scaled}$$ where $x$ is the first component of ${{\bm x}}$. With this scaling, the limits of the integral do not depend on $k$.
-------------- -------- ----------------------- ---------- ------------- --------------------
Dimension $R$ $\rho_0\beta\epsilon$ $\gamma$ $\sigma(0)$ $E_B$
\[0.5ex\]
\[-1.5ex\] 1 4.5918 1.1629 3.2051 $-10.680$ $\phantom{0}7.475$
2 5.0962 0.2455 4.3692 $-18.752$ $14.383$
3 5.5719 0.0455 5.6889 $-31.305$ $25.616$
-------------- -------- ----------------------- ---------- ------------- --------------------
: Linear theory for the GEM-4 potential in one, two and three dimensions. Solving $\frac{{{\rm d}}\sigma}{{{\rm d}}{k}}(1)=0$ and $\sigma(1)=0$ gives $R$ and $\rho_0\beta\epsilon$, while $\gamma$ and $\sigma(0)$ are computed from $\frac{{{\rm d}}^2\sigma}{{{\rm d}}{k^2}}(1)$ and (\[eqApp:defn\_sigmak\_GEM4\]) respectively. We also give the values of $E_B$ for the EOF in Eq. .[]{data-label="tabApp:GEM4parameters"}
We choose $R$ and $\rho_0\beta\epsilon$ so that $\sigma(1)=0$ and $\frac{{{\rm d}}\sigma}{{{\rm d}}{k}}(1)=0$. The derivative of $\sigma$ with respect to $k$ is $$\frac{{{\rm d}}\sigma}{{{\rm d}}{k}} = - \rho_0\beta\epsilon
\!\!\iint_{-N\pi}^{N\pi}\!
\left(\frac{4|{{\bm x}}|^4 - 2k^4R^4}{k^7R^4}\right)
e^{-|{{\bm x}}|^4/(k^4R^4)} e^{ix}{{{\rm d}}{{\bm x}}}.
\label{eqApp:defn_sigmak_GEM4_derivative}$$ Evaluating this at $k=1$ and removing the constant factor outside the integral gives a function $F(R)$: $$F(R) = \iint_{-N\pi}^{N\pi}\!
\left(\frac{4|{{\bm x}}|^4 - 2R^4}{R^4}\right)
e^{-|{{\bm x}}|^4/R^4} e^{ix}{{{\rm d}}{{\bm x}}}.
\label{eqApp:GEM4_FRs}$$ We solve the equation $F(R)=0$ using Newton’s method to give $R$. We then calculate $\rho_0\beta\epsilon$ by requiring that $\sigma(1)=0$ in (\[eqApp:defn\_sigmak\_GEM4\_scaled\]). Values for $R$ and $\rho_0\beta\epsilon$ in one, two and three dimensions are given in table \[tabApp:GEM4parameters\]. We compute the GEM-4 dispersion relation in Fig. \[fig:dispersionGEM4\] in a similar way, and use the second derivative of (\[eqApp:defn\_sigmak\_GEM4\_scaled\]) with respect to $k$, at $k=1$, to find the $\gamma$ parameter (also given in table \[tabApp:GEM4parameters\]) in ${{{\cal L}_{\text{grad}}}}$. The table also gives $\sigma(0)$, since this is useful for computing properties of the liquid, as well as $E_B$, the coefficient in the eighth-order model of [@Jaatinen2009], given in Eq. (\[eq:cLgradJaatinen\]).
Numerical method: continuation {#app:continuation}
==============================
We use numerical continuation to solve the four equations (\[eq:betadFdn3\_rewritten\]) and (\[eq:betadFdn5\_rewritten\])–(\[eq:betadFdnepsilon\_rewritten\]) for $n({{\bm x}})$ as the parameters vary. Our approach is based on [@Doedel1991] for the pseudo-arclength continuation method, and we use the approach advocated by [@Kelley2003] to solve the large linear systems at each Newton step. The main parameters are the chemical potential $\mu$, the parameters in the linear operators ${{\cal L}}$ and ${{{\cal L}_{\text{grad}}}}$, and the domain size. In this discussion, we focus on $\mu$ as the parameter that is varied.
Pseudo-arclength continuation
-----------------------------
The main idea behind pseudo-arclength continuation is to suppose that we are looking to calculate a branch of solutions $n({{\bm x}})$ depending on the parameter $\mu$. The branch may have folds (as in Fig. \[fig:DDFTPFCsummary\]), and the method should be able to go around these. The method defines a parameter (the arclength $s$) that increases or decreases monotonically along the branch (including its folds) such that both $n({{\bm x}})$ and $\mu$ can be regarded as single-valued functions of $s$. Then the equation to be solved is $${{\cal G}}(n(s),\mu(s)) = 0,
\label{eqApp:equationtosolve}$$ where ${{\cal G}}$ represents the equation we are solving for $n$. Instead of thinking of $n$ as being a function of position ${{\bm x}}$, we represent $n$ as a series of values $n_i$ on $N$ grid points ${{\bm x}}_i$ ($i=1,\dots,N$), so $n$ is now a vector in $\mathbb{R}^N$, and ${{\cal G}}$ is a function from $\mathbb{R}^{N+1}\rightarrow\mathbb{R}^{N}$. Equation (\[eqApp:equationtosolve\]) represents $N$ equations for $N+1$ unknowns, and so it is supplemented by an orthogonality condition, that the next point on a branch should lie in a plane orthogonal to a line connecting the two previous points. It is this that allows the branch following technique to go around folds. If we have two points on the branch $(n(s),\mu(s))$ at $s_0$ and $s_1$, then we take the derivatives of $n$ and $\mu$ with respect to the arclength to be approximately $$\frac{{{\rm d}}{n}}{{{\rm d}}{s}} = S \frac{n(s_1)-n(s_0)}{s_1-s_0},
\quad
\frac{{{\rm d}}\mu}{{{\rm d}}{s}} = S \frac{\mu(s_1)-\mu(s_0)}{s_1-s_0},
\label{eqApp:pseudoarclengthstep}$$ with the scaling factor $S$ chosen so as to satisfy $$\frac{1}{N}\sum_{i=1}^N \left(\frac{{{\rm d}}{n_i}}{{{\rm d}}{s}}\right)^2 +
\left(\frac{{{\rm d}}\mu}{{{\rm d}}{s}}\right)^2 = 1.
\label{eqApp:arclengthnormalization}$$ The $\frac{1}{N}$ prefactor means that the parameterization of the branch by the arclength is essentially independent of the number of grid points.
The method then proceeds in a predictor–corrector fashion. The predictor step, with a target stepsize $\Delta{s}$ provides $(n_2,\mu_2)$: $$n_2 = n(s_1) + \Delta{s}\frac{{{\rm d}}{n}}{{{\rm d}}{s}},
\quad
\mu_2 = \mu(s_1) + \Delta{s}\frac{{{\rm d}}\mu}{{{\rm d}}{s}}.
\label{eqApp:pseudoarclengthpredictor}$$ Then, $(n_2,\mu_2)$ is used as an initial iterate for a Newton solver for equation (\[eqApp:equationtosolve\]), supplemented by the condition that the Newton iterates lie in a plane orthogonal to the line given in (\[eqApp:pseudoarclengthpredictor\]), parameterised by $\Delta{s}$. This means that we are solving ${{\cal H}}(n,\mu)=0$, where ${{\cal H}}$ is $\mathbb{R}^{N+1}\rightarrow\mathbb{R}^{N+1}$, with the first $N$ equations in ${{\cal H}}$ being the same as ${{\cal G}}$, and the last equation being $$(n - n_2) \cdot \frac{{{\rm d}}{n}}{{{\rm d}}{s}} + (\mu - \mu_2) \frac{{{\rm d}}\mu}{{{\rm d}}{s}} = 0.
\label{eqApp:pseudoarcorthogonality}$$ To be precise, we take $${{\cal H}}(n,\mu)=\begin{pmatrix}
\frac{1}{\sqrt{N}} P\cdot {{\cal G}}(n,\mu)\\
\textrm{Eq.~(\ref{eqApp:pseudoarcorthogonality})}
\end{pmatrix},
\label{eqApp:fullsystem}$$ where $P$ is a linear preconditioner for ${{\cal G}}$ (see below). The $\sqrt{N}$ scaling means that the norm $\lVert{{\cal H}}(n,\mu)\rVert$ (the square root of the sum of the squares of its components) is independent of the number of grid points $N$, and it also means that the equations in ${{\cal G}}$ and the orthogonality condition (\[eqApp:pseudoarcorthogonality\]) are given a similar weighting by the Newton method.
Solving ${{\cal H}}(n,\mu)=0$ results in a new point on the branch of solutions, $(n(s_2),\mu(s_2))$, where $s_2$ is given by $$s_2 = s_1 + \frac{1}{N}(n(s_2)-n(s_1))\cdot\frac{{{\rm d}}{n}}{{{\rm d}}{s}}
+ (\mu(s_2)-\mu(s_1))\frac{{{\rm d}}\mu}{{{\rm d}}{s}},
\label{eqApp:newpseudoarc}$$ with $\frac{{{\rm d}}{n}}{{{\rm d}}{s}}$ and $\frac{{{\rm d}}\mu}{{{\rm d}}{s}}$ given by (\[eqApp:pseudoarclengthstep\]). This last equation comes from replacing $n_2$ by $n(s_2)$ and $\mu_2$ by $\mu(s_2)$ in (\[eqApp:pseudoarclengthpredictor\]) and finding a $\Delta{s}=s_2-s_1$ from $1/N$ times the first equation plus the second equation. This is not quite the actual change to the arclength that was achieved in the step, and the approximation is the reason that the method is called the pseudo-arclength method.
Newton’s method
---------------
For Newton’s method, we define $X=(n,\mu)$ and solve ${{\cal H}}(X)=0$. We start with $X_0$ given by the predictor step above in (\[eqApp:pseudoarclengthpredictor\]), and follow [@Kelley2003], at each step solving the linear equation $$\frac{\partial{{\cal H}}}{\partial{X}} \cdot {{\delta{\mkern-1mu}X}}_n = - {{\cal H}}(X_n),
\label{eqApp:linearNewtonEqns}$$ where $\frac{\partial{{\cal H}}}{\partial{X}}$ is the $(N+1)\times(N+1)$ matrix of derivatives of ${{\cal H}}$ with respect to $X$, and then improving our estimate of the root by using ${{\delta{\mkern-1mu}X}}$.
The Newton method proceeds until convergence, defined by $$\lVert{{\cal H}}(X_n)\rVert < N_\text{abs} + N_\text{rel}\lVert{{\cal H}}(X_0)\rVert,$$ where $N_\text{abs}$ and $N_\text{rel}$ are the Newton absolute and relative convergence tolerances respectively, typically $10^{-10}$ and $10^{-8}$. We also monitored the maximum of $|{{\cal H}}(n({{\bm x}}),\mu)|$ across the domain, and this was typically no larger than ten times $N_\text{abs}$, so the equations are well satisfied at each point in space as well as in norm.
The linear equations in (\[eqApp:linearNewtonEqns\]) are solved to find ${{\delta{\mkern-1mu}X}}_n$ using <span style="font-variant:small-caps;">Matlab</span>’s biconjugate gradient stabilized (l) (`bicgstabl`) method. This allows the matrix–vector multiplications to be evaluated using a function (rather than by explicitly computing a large matrix). The method is iterative, and proceeds until $$\left\lVert\frac{\partial{{\cal H}}}{\partial{X}} \cdot {{\delta{\mkern-1mu}X}}_n + {{\cal H}}(X_n)\right\rVert <
L_\text{rel}\lVert{{\cal H}}(X_n)\rVert,$$ where the relative tolerance $L_\text{rel}$ of the linear solver is chosen so as to balance the number of Newton steps against the number of `bicgstabl` iterations. Based on [@Kelley2003], we choose $$L_\text{rel} = 0.1\sqrt{\lVert{{\cal H}}(X_n)\rVert}
+ \frac{N_\text{abs}}{\lVert{{\cal H}}(X_n)\rVert},$$ subject to the constraint that $L_\text{rel}$ should be no larger than $0.1$. The effect of this is that in the initial first or second of the Newton iterations, when $\lVert{{\cal H}}(X_n)\rVert$ is at its largest, the linear solver is not asked to work too hard to solve (\[eqApp:linearNewtonEqns\]), since any reasonably good approximate solution is likely to improve the estimate of the root, and an absolutely perfect solution isn’t going to do much better. In the middle stages of the Newton iterations, when $\lVert{{\cal H}}(X_n)\rVert$ is about $10^{-6}$, the tolerance $L_\text{rel}$ is about $2\times10^{-4}$, which is not good enough for quadratic convergence of Newton’s method, but is good enough to provide two or three orders of magnitude improvement to the quality of the solution at a considerably lower cost. In the final stages of the Newton iterations, $L_\text{rel}=0.1$, good enough for polishing the solution to the tolerance $N_\text{abs}$ while not attempting to solve the linear problem down to round-off error.
Also based on [@Kelley2003], we implement the Armijo rule, which ensures that each Newton step results in an improvement to the solution. The idea is that the solution of the linear equation (\[eqApp:linearNewtonEqns\]) should give the correct direction for improving the solution of ${{\cal H}}(X)=0$, but taking a full step may not actually result in an improvement, so instead we set $$X_{n+1} = X_n + 2^{-j} {{\delta{\mkern-1mu}X}}_n,
\label{eqApp:armijo}$$ where $j=0,1,2,\dots$ is chosen to be the smallest such that $$\lVert{{\cal H}}(X_{n+1})\rVert <
\lVert{{\cal H}}(X_{n})\rVert.
\label{eqApp:armijocondition}$$ In most cases, the first ($j=0$) Armijo step satisfies (\[eqApp:armijocondition\]), equivalent to the normal Newton method, but when the density is close to zero in the DDFT calculations, and small changes in density lead to large changes in its logarithm, the Armijo rule is helpful.
We do not use a preconditioner in the GEM-4 calculations (so $P$ in (\[eqApp:fullsystem\]) is the identity), but in the gradient expansion calculations, a preconditioner is helpful. In Fourier space, ${{{\cal L}_{\text{grad}}}}$ can easily be inverted, so the preconditioner is ${{{\cal L}_{\text{grad}}^{-1}}}$ when the absolute value of the eigenvalue $\sigma(k)$ is greater than $1$, otherwise the preconditioner is the identity. This has the effect of reducing the number of iterations needed to solve (\[eqApp:linearNewtonEqns\]) by a factor of 10 or even 100.
A sample <span style="font-variant:small-caps;">Matlab</span> code to solve Eq. (\[eq:betadFdn3\_rewritten\]) for DDFT-3 by Newton’s method (without the continuation aspect) is given in the supplementary material.
Additional considerations
-------------------------
We start the computation of each branch close to $n=0$ and $\mu=0$ using an approximate solution derived from weakly nonlinear theory. For example, for hexagons in DDFT-3, we take $$n({{\bm x}}) = \frac{\mu}{\sigma(0)}\left(-1 +
e^{i{{\bm k}}_1\cdot{{\bm x}}} + e^{i{{\bm k}}_2\cdot{{\bm x}}} + e^{i{{\bm k}}_3\cdot{{\bm x}}} +
\text{c.c.}\right),
\label{eqApp:weaklynonlinearhexagons}$$ where the initial value of $\mu$ is small, $\sigma(0)$ comes from Table \[tabApp:GEM4parameters\], ${{\bm k}}_1=(1,0)$, ${{\bm k}}_2=(-\frac{1}{2},\frac{\sqrt{3}}{2})$, ${{\bm k}}_3=(-\frac{1}{2},-\frac{\sqrt{3}}{2})$, and $\text{c.c.}$ stands for complex conjugate.
The equations are posed on periodic domains and we use $N_x\times{N_y}$ grid points, depending on the number of wavelengths in the domain and the nature of the solution. The GEM-4 hexagonal calculations require $8\times\frac{8}{\sqrt{3}}$ wavelength domains, with resolutions starting at $80\times48$ Fourier modes close to onset. At larger amplitude, $512\times320$ Fourier modes (or even more) are needed, especially if the density maxima are sharply peaked (in DDFT-3) or if the density minima are very close to zero (in DDFT-5). In order to accommodate the changing needs for resolution along a branch, we monitor whether the solutions are well resolved and implemented automatic regridding, so as to maintain enough grid points to resolve the solution well, regardless of what features emerge as $\mu$ is varied. Typically we require that the amplitudes of the highest-wavenumber Fourier modes be no higher than $10^{-10}$ times the largest Fourier amplitude.
We also implement automatic pseudo-arclength stepsize control: $\Delta{s}$ is increased by a factor of 1.1 (up to a maximum of 0.1) if the Newton method converges quickly (in fewer than 5 iterations), or is decreased by a factor of 2 (down to a minimum of $10^{-6}$) if it converges slowly (more than 8 iterations) or fails.
Finally, we adjust the domain size continuously along each branch so as to minimise the specific grand potential $\Omega/A$. This is done by adding an extra parameter (the domain stretch factor $K$), so that the real domain size is $KL_x\times{K}L_y$ instead of an unstretched $L_x\times{L_y}$. The number of grid points is not altered. Then the real GEM-4 potential is proportional to $\psi(|{{\bm x}}|)=\exp\left(-K^4|{{\bm x}}|^4/R^4\right)$, where ${{\bm x}}$ is the unstretched coordinate on the unaltered grid. Quantities like the mean value of $n$ are the sum of the values of $n$ at each grid point divided by the number of grid points, and so are not affected by the stretch factor. The only parts of $\Omega/A$ that are affected are the convolutions (for ${{\cal L}}$ and the GEM-4 potential) and the spatial derivatives (for ${{{\cal L}_{\text{grad}}}}$). In the case of the GEM-4 potential, when considering the specific grand potential $\Omega_3/A$ arising from (\[eq:DDFTF3\]) for example, and ${{\cal L}}$ given by (\[eq:defnL\_GEM4\]), the integrals in $\Omega_3/A$ are proportional to $K^2$, with additional $K$ dependence coming from the GEM-4 potential itself. Therefore, $$\frac{{{\rm d}}(\Omega_3/A)}{{{\rm d}}{K}} \propto
\int n({{\bm x}}) \left(2 K \psi \otimes n + K^2 \frac{\partial\psi}{\partial K} \otimes n\right)\,{{{\rm d}}{{\bm x}}},$$ where $\otimes$ represents the convolution integral evaluated on the unstretched grid, and the ${{\bm x}}$ integral is also on the unstretched grid. In the case of ${{{\cal L}_{\text{grad}}}}$, Laplacians on the real grid are a factor of $K^{-2}$ times Laplacians on the unstretched grid, so ${{\rm d}}(\Omega/A)/{{\rm d}}{K}$ is evaluated accordingly. In both cases, an extra equation (${{\rm d}}(\Omega/A)/{{\rm d}}{K}=0$) is added to ${{\cal H}}$ in (\[eqApp:fullsystem\]), and this is solved alongside all the other equations. The Jacobian also needs to be evaluated. In practice this made little difference in the cases considered here, and the domain stretch factor largely stayed between $0.98$ and $1.02$.
[68]{}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 [****, ()]{} [****, ()](\doibase
{10.1103/PhysRevE.86.031603}) @noop [****, ()]{} [****, ()](\doibase
10.1080/00018732.2012.737555) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase {10.1103/PhysRevE.87.013313}) [****, ()](\doibase 10.1103/PhysRevE.81.061601) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.112.255501) [****, ()](\doibase
10.1103/PhysRevLett.117.075501) @noop [****, ()]{} [****, ()](\doibase 10.1107/S2052252518001161) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevE.87.042915) [****, ()](\doibase 10.1103/PhysRevE.49.2225) [****, ()](\doibase
10.1146/annurev-conmatphys-031214-014514) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1080/00018737900101365) in @noop [**]{}, (, , ) Chap. , pp. @noop [**]{} (, , ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1063/1.1778374) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevE.76.031501) [****, ()](\doibase 10.1103/PhysRevLett.96.045701) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase {10.1103/PhysRevB.97.144112}) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.97.105701) [****, ()](\doibase
10.1103/PhysRevB.19.2775) [****, ()](\doibase {10.1103/PhysRevE.85.061408}) [****, ()](\doibase 10.1103/PhysRevE.95.022105) [**** ()](\doibase 10.1098/rsta.2017.0211) [****, ()](\doibase {10.1103/PhysRevB.97.180102}) [****, ()](\doibase
{10.1088/0953-8984/22/36/364102}) [****, ()](\doibase {10.1103/PhysRevE.82.061602}) in @noop [**]{}, (, , ) Chap. , pp. in @noop [**]{}, (, , ) Chap. , pp. @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.109.228301) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase {10.1021/acs.nanolett.7b01742}) @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1142/S0218127491000397) @noop [**]{} (, , )
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the present paper we construct normal numbers in base $q$ by concatenating $q$-ary expansions of pseudo polynomials evaluated at the primes. This extends a recent result by Tichy and the author.'
address: ' 1. Université de Lorraine, Institut Elie Cartan de Lorraine, UMR 7502, Vandoeuvre-lès-Nancy, F-54506, France;2. CNRS, Institut Elie Cartan de Lorraine, UMR 7502, Vandoeuvre-lès-Nancy, F-54506, France'
author:
- 'Manfred G. Madritsch'
title: Construction of normal numbers via pseudo polynomial prime sequences
---
Introduction {#sec:introduction}
============
Let $q\geq 2$ be a positive integer. Then every real $\theta\in[0,1)$ admits a unique expansion of the form $$\theta=\sum_{k\geq1}a_kq^k\quad(a_k\in\{0,\ldots,q-1\})$$ called the $q$-ary expansion. We denote by $\mathcal{N}(\theta,d_1\cdots d_\ell,N)$ the number of occurrences of the block $d_1\cdots d_\ell$ amongst the first $N$ digits, *i.e.* $$\mathcal{N}(\theta,d_1\cdots d_\ell,N):=\#\{0\leq i< n\colon
a_{i+1}=d_1,\ldots,a_{i+\ell}=d_\ell\}.$$ Then we call a number normal of order $\ell$ in base $q$ if for each block of length $\ell$ the frequency of occurrences tends to $q^{-\ell}$. As a qualitative measure of the distance of a number from being normal we introduce for integers $N$ and $\ell$ the discrepancy of $\theta$ by $$\mathcal{R}_{N,\ell}(\theta)=\sup_{d_1\ldots
d_\ell}{\left\vert}\frac{\mathcal{N}(\theta,d_1\cdots
d_\ell,N)}{N}-q^{-k}{\right\vert},$$ where the supremum is over all blocks of length $\ell$. Then a number $\theta$ is normal to base $q$ if for each $\ell\geq1$ we have that $\mathcal{R}_{N,\ell}(\theta)=o(1)$ for $N\to\infty$. Furthermore we call a number absolutely normal if it is normal in all bases $q\geq2$.
Borel [@borel1909:les_probabilites_denombrables] used a slightly different, but equivalent (*cf.* Chapter 4 of [@bugeaud2012:distribution_modulo_one]), definition of normality to show that almost all real numbers are normal with respect to the Lebesgue measure. Despite their omnipresence it is not known whether numbers such as $\log2$, $\pi$, $e$ or $\sqrt2$ are normal to any base. The first construction of a normal number is due to Champernowne [@champernowne1933:construction_decimals_normal] who showed that the number $$\begin{aligned}
0.1\,2\,3\,4\,5\,6\,7\,8\,9\,10\,11\,12\,13\,14\,15\,16\,17\,18\,19\,20\dots\end{aligned}$$ is normal in base $10$. The construction of Champernowne laid the base for a class of normal numbers which are of the form $$\begin{gathered}
\sigma_q=\sigma_q(f)=
0.{\left\lfloor}f(1){\right\rfloor}_q{\left\lfloor}f(2){\right\rfloor}_q{\left\lfloor}f(3){\right\rfloor}_q {\left\lfloor}f(4){\right\rfloor}_q {\left\lfloor}f(5){\right\rfloor}_q {\left\lfloor}f(6){\right\rfloor}_q \dots,\end{gathered}$$ where ${\left\lfloor}\cdot{\right\rfloor}_q$ denotes the expansion in base $q$ of the integer part. Davenport and Erd[ő]{}s [@davenport_erdoes1952:note_on_normal] showed that $\sigma(f)$ is normal for $f$ being a polynomial such that $f({\mathbb{N}})\subset{\mathbb{N}}$. This construction was extended by Schiffer [@schiffer1986:discrepancy_normal_numbers] to polynomials with rational coefficients. Furthermore he showed that for these polynomials the discrepancy $\mathcal{R}_{N,\ell}(\sigma(f))\ll (\log
N)^{-1}$ and that this is best possible. These results where extended by Nakai and Shiokawa [@nakai_shiokawa1992:discrepancy_estimates_class] to polynomials having real coefficients. Madritsch, Thuswaldner and Tichy [@madritsch_thuswaldner_tichy2008:normality_numbers_generated] considered transcendental entire functions of bounded logarithmic order. Nakai and Shiokawa [@nakai_shiokawa1990:class_normal_numbers] used pseudo-polynomial functions, *i.e.* these are function of the form $$\begin{gathered}
\label{mani:pseudopoly}
f(x)=\alpha_0 x^{\beta_0}+\alpha_1x^{\beta_1}+\cdots+\alpha_dx^{\beta_d}\end{gathered}$$ with $\alpha_0,\beta_0,\alpha_1,\beta_1,\ldots,\alpha_d,\beta_d\in{\mathbb{R}}$, $\alpha_0>0$, $\beta_0>\beta_1>\cdots>\beta_d>0$ and at least one $\beta_i\not\in{\mathbb{Z}}$. Since we often only need the leading term we write $\alpha=\alpha_0$ and $\beta=\beta_0$ for short. They were also able to show that the discrepancy is $\mathcal{O}((\log N)^{-1})$. We refer the interested reader to the books of Kuipers and Niederreiter [@kuipers_niederreiter1974:uniform_distribution_sequences], Drmota and Tichy [@drmota_tichy1997:sequences_discrepancies_and] or Bugeaud [@bugeaud2012:distribution_modulo_one] for a more complete account on the construction of normal numbers.
The present method of construction by concatenating function values is in strong connection with properties of $q$-additive functions. We call a function $f$ strictly $q$-additive, if $f(0)=0$ and the function operates only on the digits of the $q$-ary representation, i.e., $$f(n)=\sum_{h=0}^\ell f(d_h)\quad\text{ for }\quad n=\sum_{h=0}^\ell d_hq^h.$$ A very simple example of a strictly $q$-additive function is the sum of digits function $s_q$, defined by $$s_q(n)=\sum_{h=0}^\ell d_h\quad\text{ for }\quad n=\sum_{h=0}^\ell d_hq^h.$$
Refining the methods of Nakai and Shiokawa [@nakai_shiokawa1990:class_normal_numbers] the author obtained the following result.
Let $q\geq2$ be an integer and $f$ be a strictly $q$-additive function. If $p$ is a pseudo-polynomial as defined in (\[mani:pseudopoly\]), then there exists $\eta>0$ such that $$\begin{gathered}
\sum_{n\leq N}f\left({\left\lfloor}p(n){\right\rfloor}\right)
=\mu_fN\log_q(p(N))
+NF\left(\log_q(p(N))\right)
+\mathcal{O}\left(N^{1-\eta}\right),\end{gathered}$$ where $$\mu_f=\frac1q\sum_{d=0}^{q-1}f(d)$$ and $F$ is a $1$-periodic function depending only on $f$ and $p$.
In the present paper, however, we are interested in a variant of $\sigma_q(f)$ involving primes. As a first example, Champernowne [@champernowne1933:construction_decimals_normal] conjectured and later Copeland and Erd[ő]{}s [@copeland_erdoes1946:note_on_normal] proved that the number $$\begin{aligned}
0.2\,3\,5\,7\,11\,13\,17\,19\,23\,29\,31\,37\,41\,43\,47\,53\,59\,61\,67\dots\end{aligned}$$ is normal in base $10$. Similar to the construction above we want to consider the number $$\begin{gathered}
\tau_q=\tau_q(f)=0.{\left\lfloor}f(2){\right\rfloor}_q {\left\lfloor}f(3){\right\rfloor}_q {\left\lfloor}f(5){\right\rfloor}_q {\left\lfloor}f(7){\right\rfloor}_q {\left\lfloor}f(11){\right\rfloor}_q {\left\lfloor}f(13){\right\rfloor}_q \dots,\end{gathered}$$ where the arguments of $f$ run through the sequence of primes.
Then the paper of Copeland and Erd[ő]{}s corresponds to the function $f(x)=x$. Nakai and Shiokawa [@nakai_shiokawa1997:normality_numbers_generated] showed that the discrepancy for polynomials having rational coefficients is $\mathcal{O}((\log N)^{-1})$. Furthermore Madritsch, Thuswaldner and Tichy [@madritsch_thuswaldner_tichy2008:normality_numbers_generated] showed, that transcendental entire functions of bounded logarithmic order yield normal numbers. Finally in a recent paper Madritsch and Tichy [@madritsch_tichy2013:construction_normal_numbers] considered pseudo-polynomials of the special form $\alpha x^\beta$ with $\alpha>0$, $\beta>1$ and $\beta\not\in{\mathbb{Z}}$.
The aim of the present paper is to extend this last construction to arbitrary pseudo-polynomials. Our first main result is the following
\[thm:normal\] Let $f$ be a pseudo-polynomial as in (\[mani:pseudopoly\]). Then $$\mathcal{R}_N(\tau_q(f))\ll(\log N)^{-1}.$$
In our second main result we use the connection of this construction of normal numbers with the arithmetic mean of $q$-additive functions as described above. Known results are due to Shiokawa [@shiokawa1974:sum_digits_prime] and Madritsch and Tichy [@madritsch_tichy2013:construction_normal_numbers]. Similar results concerning the moments of the sum of digits function over primes have been established by Kátai [@katai1977:sum_digits_primes].
Let $\pi(x)$ stand for the number of primes less than or equal to $x$. Then adapting these ideas to our method we obtain the following
\[thm:summatoryfun\] Let $f$ be a pseudo-polynomial as in (\[mani:pseudopoly\]). Then $$\sum_{p\leq P}s_q({\left\lfloor}f(p){\right\rfloor})=\frac{q-1}2\pi(P)\log_qP^\beta+\mathcal{O}(\pi(P)),$$ where the sum runs over the primes and the implicit $\mathcal{O}$-constant may depend on $q$ and $\beta$.
With simple modifications Theorem \[thm:summatoryfun\] can be extended to completely $q$-additive functions replacing $s_q$.
The proof of the two theorems is divided into four parts. In the following section we rewrite both statements in order to obtain as a common base the central theorem – Theorem \[mani:centralthm\]. In Section \[sec:proof-prop-refm1\] we start with the proof of this central theorem by using an indicator function and its Fourier series. These series contain exponential sums which we treat by different methods (with respect to the position in the expansion) in Section \[sec:expon-sum-estim\]. Finally, in Section \[sec:proof-prop-refm2\] we put the estimates together in order to proof the central theorem and therefore our two statements.
Preliminaries {#sec:preliminaries}
=============
Throughout the rest $p$ will always denote a prime. The implicit constant of $\ll$ and $\mathcal{O}$ may depend on the pseudo-polynomial $f$ and on the parameter $\varepsilon>0$. Furthermore we fix a block $d_1\cdots d_\ell$ of length $\ell$ and $N$, the number of digits we consider.
In the first step we want to know in the expansion of which prime the $N$-th digit occurs. This can be seen as the translation from the digital world to the world of blocks. To this end let $\ell(m)$ denote the length of the $q$-ary expansion of an integer $m$. Then we define an integer $P$ by $$\begin{gathered}
\sum_{p\leq P-1}\ell\left(\lfloor f(p)\rfloor\right) <N\leq
\sum_{p\leq P}\ell\left(\lfloor f(p)\rfloor\right),\end{gathered}$$ where the sum runs over all primes. Thus we get the following relation between $N$ and $P$ $$\label{mani:NtoP}
\begin{split}
N&=\sum_{p\leq P}\ell({\left\lfloor}f(p){\right\rfloor})+\mathcal{O}(\pi(P))+\mathcal{O}(\beta \log_q(P))\\
&=\frac{\beta}{\log q}P+\mathcal{O}\left(\frac{P}{\log P}\right).
\end{split}$$ Here we have used the prime number theorem in the form (*cf.* [@tenenbaum1995:introduction_la_theorie Théorème 4.1]) $$\begin{gathered}
\label{pnt}
\pi(x)=\mathrm{Li}\, x+\mathcal{O}\left(\frac x{(\log x)^G}\right),\end{gathered}$$ where $G$ is an arbitrary positive constant and $$\mathrm{Li}\,x=\int_2^x\frac{\mathrm{d}t}{\log t}.$$
Now we show that we may neglect the occurrences of the block $d_1\cdots d_\ell$ between two expansions. We write $\mathcal{N}(f(p))$ for the number of occurrences of this block in the $q$-ary expansion of $\lfloor f(p)\rfloor$. Then implies that $$\begin{gathered}
\label{mani:Ntrunc}
{\left\vert}\mathcal{N}(\tau_q(f);d_1\cdots d_\ell;N)-\sum_{p\leq
P}\mathcal{N}(f(p)){\right\vert}\ll\frac N{\log N}.\end{gathered}$$
In the next step we use the polynomial-like behavior of $f$. In particular, we collect all the values having the same length of expansion. Let $j_0$ be a sufficiently large integer. Then for each integer $j\geq j_0$ there exists a $P_j$ such that $$q^{j-2}\leq f(P_j)<q^{j-1}\leq f(P_j+1)<q^j$$ with $$P_j\asymp q^{\frac j\beta}.$$ Furthermore we set $J$ to be the greatest length of the $q$-ary expansions of $f(p)$ over the primes $p\leq P$, i.e., $$\begin{gathered}
J:=\max_{p\leq P}\ell(\lfloor f(p)\rfloor)=\log_q(f(P))+\mathcal{O}(1)\asymp\log
P.\end{gathered}$$
Now we show that we may suppose that each expansion has the same length (which we reach by adding leading zeroes). For $P_{j-1}<p\leq
P_j$ we may write $f(p)$ in $q$-ary expansion, i.e., $$\begin{gathered}
\label{mani:expansionoffp}
f(p)=b_{j-1}q^{j-1}+b_{j-2}q^{j-2}+\dots+b_{1}q+b_{0}+b_{-1}q^{-1}+\dots.\end{gathered}$$ Then we denote by $\mathcal{N}^*(f(p))$ the number of occurrences of the block $d_1\cdots d_\ell$ in the string $0\cdots0b_{j-1}b_{j-2}\cdots b_1b_0$, where we filled up the expansion with leading zeroes such that it has length $J$. The error of doing so can be estimated by $$\begin{aligned}
0&\leq\sum_{p\leq P}\mathcal{N}^*(f(p))-\sum_{p\leq P}\mathcal{N}(f(p))\\
&\leq\sum_{j=j_0+1}^{J-1}(J-j)\left(\pi(P_{j+1})-\pi(P_{j})\right)+\mathcal{O}(1)\\
&\leq\sum_{j=j_0+2}^{J}\pi(P_{j})+\mathcal{O}(1)\ll\sum_{j=j_0+2}^{J}\frac{q^{j/\beta}}j
\ll\frac P{\log P}\ll\frac N{\log N}.\end{aligned}$$
In the following three sections we will estimate this sum of indicator functions $\mathcal{N}^*$ in order to prove the following theorem.
\[mani:centralthm\] Let $f$ be a pseudo polynomial as in . Then $$\begin{gathered}
\label{mani:centralthm:statement}
\sum_{p\leq
P}\mathcal{N}^*\left({\left\lfloor}f(p){\right\rfloor}\right)=q^{-\ell}\pi(P)\log_qP^\beta+\mathcal{O}\left(\frac{P}{\log
P}\right)\end{gathered}$$
Using this theorem we can simply deduce our two main results.
We insert into and get the desired result.
For this proof we have to rewrite the statement. In particular, we use that the sum of digits function counts the number of $1$s, $2$s, etc. and assigns weights to them, i.e., $$s_q(n)=\sum_{d=0}^{q-1}d\cdot\mathcal{N}(n;d).$$ Thus $$\begin{aligned}
\sum_{p\leq P}s_q({\left\lfloor}p^\beta{\right\rfloor})
&=\sum_{p\leq P}\sum_{d=0}^{q-1}d\cdot\mathcal{N}(p^\beta)
=\sum_{p\leq
P}\sum_{d=0}^{q-1}d\cdot\mathcal{N}^*(p^\beta)+\mathcal{O}\left(\frac{P}{\log
P}\right)\\
&=\frac{q-1}2\pi(P)\log_q(P^\beta)+\mathcal{O}\left(\frac{P}{\log
P}\right)\end{aligned}$$ and the theorem follows.
In the following sections we will prove Theorem \[mani:centralthm\] in several steps. First we use the “method of little glasses” in order to approximate the indicator function by a Fourier series having smooth coefficients. Then we will apply different methods (depending on the position in the expansion) for the estimation of the exponential sums that appear in the Fourier series. Finally we put everything together and get the desired estimate.
Proof of Theorem \[mani:centralthm\], Part I {#sec:proof-prop-refm1}
============================================
We want to ease notation by splitting the pseudo-polynomial $f$ into a polynomial and the rest. Then there exists a unique decomposition of the following form: $$\begin{gathered}
\label{pseudo:poly:split}
f(x)=g(x)+h(x)\end{gathered}$$ where $h\in{\mathbb{R}}[X]$ is a polynomial of degree $k$ (where we set $k=0$ if $h$ is the zero polynomial) and $$g(x)=\sum_{j=1}^r\alpha_jx^{\theta_j}$$ with $r\geq1$, $\alpha_r\neq0$, $\alpha_j$ real, $0<\theta_1<\cdots<\theta_r$ and $\theta_j\not\in{\mathbb{Z}}$ for $1\leq j\leq r$.
Let $\gamma$ and $\rho$ be two parameter which we will frequently use in the sequel. We suppose that $$\begin{gathered}
0<\gamma<\rho<\min\left(\frac1{4(k+1)},\frac{\theta_r}{2}\right).\end{gathered}$$
The aim of this section is to calculate the Fourier transform of $\mathcal{N}^*$. In order to count the occurrences of the block $d_1\cdots d_\ell$ in the $q$-ary expansion of $\lfloor f(p) \rfloor$ ($2\le p \le P$) we define the indicator function $$\begin{aligned}
\mathcal{I}(t)=\begin{cases}
1, &\text{if }\sum_{i=1}^\ell d_iq^{-i}\leq t-\lfloor t\rfloor
<\sum_{i=1}^\ell d_iq^{-i}+q^{-\ell};\\
0, &\text{otherwise;}
\end{cases}\end{aligned}$$ which is a $1$-periodic function. Indeed, we have $$\begin{gathered}
\label{position}
\mathcal{I}(q^{-j}f(p)) = 1 \Longleftrightarrow d_1\cdots d_\ell =
b_{j-1}\cdots b_{j-\ell},\end{gathered}$$ where $f(p)$ has an expansion as in (\[mani:expansionoffp\]). Thus we may write our block counting function as follows $$\begin{gathered}
\label{mani:NthetatoNstar}
\mathcal{N}^*(f(p))=\sum_{j=\ell}^J\mathcal{I}\left(q^{-j}f(p)\right).\end{gathered}$$
In the following we will use Vinogradov’s “method of little glasses” (*cf.* [@vinogradov2004:method_trigonometrical_sums]). We want to approximate $\mathcal{I}$ from above and from below by two $1$-periodic functions having small Fourier coefficients. To this end we will use the following
\[vin:lem12\] Let $\alpha$, $\beta$, $\Delta$ be real numbers satisfying $$\begin{gathered}
0<\Delta<\frac12,\quad\Delta\leq\beta-\alpha\leq1-\Delta.\end{gathered}$$ Then there exists a periodic function $\psi(x)$ with period $1$, satisfying
1. $\psi(x)=1$ in the interval $\alpha+\frac12\Delta\leq x
\leq\beta-\frac12\Delta$,
2. $\psi(x)=0$ in the interval $\beta+\frac12\Delta\leq x
\leq1+\alpha-\frac12\Delta$,
3. $0\leq\psi(x)\leq1$ in the remainder of the interval $\alpha-\frac12\Delta\leq x\leq1+\alpha-\frac12\Delta$,
4. $\psi(x)$ has a Fourier series expansion of the form $$\psi(x)=\beta-\alpha+\sum_{\substack{\nu=-\infty\\\nu\neq0}}^\infty
A(\nu) e(\nu x),$$ where $$\begin{gathered}
\label{mani:A}
{\left\vert}A(\nu){\right\vert}\ll \min \left( \frac 1\nu,
\beta-\alpha,\frac{1}{\nu^2\Delta} \right).
\end{gathered}$$
We note that we could have used Vaaler polynomials [@vaaler1985:some_extremal_functions], however, we do not gain anything by doing so as the estimates we get are already best possible. Setting $$\label{mani:abd}
\begin{split}
\delta=P^{-\gamma},\quad
\begin{aligned}
\alpha_-&=\sum_{\lambda=1}^\ell d_\lambda q^{-\lambda}+(2\delta)^{-1},&
\beta_-&=\sum_{\lambda=1}^\ell d_\lambda q^{-\lambda}+q^{-\ell}-(2\delta)^{-1},\\
\alpha_+&=\sum_{\lambda=1}^\ell d_\lambda q^{-\lambda}-(2\delta)^{-1},&
\beta_+&=\sum_{\lambda=1}^\ell d_\lambda q^{-\lambda}+q^{-\ell}+(2\delta)^{-1}.
\end{aligned}
\end{split}$$ and an application of Lemma \[vin:lem12\] with $(\alpha,\beta,\delta)=(\alpha_-,\beta_-,\delta)$ and $(\alpha,\beta,\delta)=(\alpha_+,\beta_+, \delta)$, respectively, provides us with two functions $\mathcal{I}_-$ and $\mathcal{I}_+$. By our choice of $(\alpha_\pm,\beta_\pm,\delta)$ it is immediate that $$\label{uglI}
\mathcal{I}_-(t)\leq\mathcal{I}(t)\leq\mathcal{I}_+(t) \qquad
(t\in\mathbb{R}).$$ Lemma \[vin:lem12\] also implies that these two functions have Fourier expansions $$\begin{aligned}
\label{mani:Ifourier}
\mathcal{I}_\pm(t)=q^{-\ell}\pm P^{-\gamma}+
\sum_{\substack{\nu=-\infty\\\nu\neq0}}^\infty A_\pm(\nu)e(\nu t)\end{aligned}$$ satisfying $$\begin{gathered}
{\left\vert}A_\pm(\nu){\right\vert}\ll\min({\left\vert}\nu{\right\vert}^{-1},P^{\gamma}{\left\vert}\nu{\right\vert}^{-2}).\end{gathered}$$ In a next step we want to replace $\mathcal{I}$ by $\mathcal{I}_+$ in (\[mani:NthetatoNstar\]). For this purpose we observe, using , and that $$\begin{gathered}
{\left\vert}\mathcal{I}(t)-q^{-\ell}{\right\vert}\ll P^{-\gamma} + \sum_{\substack{\nu=-\infty\\\nu\neq0}}^\infty
A_\pm(\nu)e(\nu t).\end{gathered}$$ Thus setting $t=q^{-j}f(p)$ and summing over $p\leq P$ yields $$\begin{gathered}
\label{mani:0.5}
{\left\vert}\sum_{p\leq P}\mathcal{I}(q^{-j}f(p))-\frac{\pi(P)}{q^{\ell}}{\right\vert}\ll\pi(P)P^{-\gamma}+\sum_{\substack{\nu=-\infty\\\nu\neq0}}^\infty
A_{\pm}(\nu)\sum_{p\leq P}e\left(\frac{\nu}{q^j}f(p)\right).\end{gathered}$$
Now we consider the coefficients $A_\pm(\nu)$. Noting one observes that $$\begin{gathered}
A_\pm(\nu)\ll\begin{cases}
\nu^{-1}, &\text{for }{\left\vert}\nu{\right\vert}\leq P^{\gamma};\\
P^{\gamma}\nu^{-2}, &\text{for }{\left\vert}\nu{\right\vert}>P^{\gamma}.
\end{cases}\end{gathered}$$ Estimating all summands with ${\left\vert}\nu{\right\vert}>P^{\gamma}$ trivially we get $$\begin{gathered}
\sum_{\substack{\nu=-\infty\\\nu\neq0}}^\infty
A_\pm(\nu)e\left(\frac{\nu}{q^j}f(p)\right)
\ll\sum_{\nu=1}^{P^{\gamma}}\nu^{-1}e\left(\frac{\nu}{q^j}f(p)\right)+P^{-\gamma}.\end{gathered}$$ Using this in yields $$\begin{gathered}
{\left\vert}\sum_{p\leq P}\mathcal{I}(q^{-j}f(p))-\frac{\pi(P)}{q^{\ell}}{\right\vert}\ll\pi(P)P^{-\gamma}+\sum_{\nu=1}^{P^{\gamma}}
\nu^{-1}S(P,j,\nu),\end{gathered}$$ where we have set $$\begin{gathered}
\label{S_Pjnu}
S(P,j,\nu):=\sum_{p\leq P}e\left(\frac{\nu}{q^j}f(p)\right).\end{gathered}$$
Exponential sum estimates {#sec:expon-sum-estim}
=========================
In the present section we will focus on the estimation of the sum $S(P,j,\nu)$ for different ranges of $j$. Since $j$ describes the position within the $q$-ary expansion of $f(p)$ we will call these ranges the “most significant digits”, the “least significant digits” and the “digits in the middle”, respectively.
Now, if $\theta_r>k\geq0$, *i.e* the leading coefficient of $f$ origins from the pseudo polynomial part $g$, then we consider the two ranges $$1\leq q^j\leq P^{\theta_r-\rho}\quad\text{and}\quad
P^{\theta_r-\rho}<q^j\leq P^{\theta_r}.$$ For the first one we will apply Proposition \[prop:least\_significant\] and for the second one Proposition \[prop:most\_significant\].
On the other hand, if $k>\theta_r>0$, meaning that the leading coefficient of $f$ origins from the polynomial part $h$, then we have an additional part. In particular, in this case we will consider the three ranges $$1\leq q^j\leq P^{\theta_r-\rho},\quad
P^{\theta_r-\rho}<q^j\leq P^{k-1+\rho},\quad\text{and}\quad
P^{k-1+\rho}<q^j\leq P^{k}.$$ We will, similar to above, treat the first and last range by Proposition \[prop:least\_significant\] and Proposition \[prop:most\_significant\], respectively. For the middle range we will apply Proposition \[prop:middle\_digits\]. Since $2\rho<\theta_r$, we note that the middle range is empty if $k=1$.
Since the size of $j$ represents the position of the digit in the expansion (*cf.* ), we will deal in the following subsection with the “most significant digits”, the “least significant digits” and the “digits in the middle”, respectively.
Most significant digits
-----------------------
We start our series of estimates for the exponential sum $S(P,j,\nu)$ for $j$ being in the highest range. In particular, we want to show the following
\[prop:most\_significant\] Suppose that for some $k\geq1$ we have ${\left\vert}f^{(k)}(x){\right\vert}\geq\Lambda$ for any $x$ on $[a,b]$ with $\Lambda>0$. Then $$S(P,j,\nu)\ll\frac1{\log P}\Lambda^{-\frac1k}+\frac{P}{(\log P)^G}.$$
The main idea of the proof is to use Riemann-Stieltjes integration together with
\[ik:lem8.10\] Let $F\colon[a,b]\to{\mathbb{R}}$ and suppose that for some $k\geq1$ we have ${\left\vert}F^{(k)}(x){\right\vert}\geq\Lambda$ for any $x$ on $[a,b]$ with $\Lambda>0$. Then $${\left\vert}\int_a^be(F(x))\mathrm{d}x{\right\vert}\leq k2^k\Lambda^{-1/k}.$$
We rewrite the sum into a Riemann-Stieltjes integral: $$\begin{aligned}
S(P,j,\nu)=\sum_{p\leq P}e\left(\frac{\nu}{q^j}f(p)\right)
=\int_{2}^{P}e\left(\frac{\nu}{q^j}f(t)\right)\mathrm{d}\pi(t)+\mathcal{O}(1).\end{aligned}$$ Then we apply the prime number theorem in the form to gain the usual integral back. Thus $$\begin{aligned}
S(P,j,\nu)
=\int_{P(\log P)^{-G}}^{P}
e\left(\frac{\nu}{q^j}f(t)\right)
\frac{\mathrm{d}t}{\log t}
+\mathcal{O}\left(\frac{P}{(\log P)^G}\right).\end{aligned}$$ Now we use the second mean-value theorem to get $$\label{mani:res:most}
\begin{split}
S(P,j,\nu)\ll\frac1{\log P}\sup_{\xi}
{\left\vert}\int_{P(\log P)^{-G}}^{\xi}e\left(\frac{\nu}{q^j}f(t)\right)\mathrm{d}t{\right\vert}+\frac{P}{(\log P)^G}.
\end{split}$$ Finally an application of Lemma \[ik:lem8.10\] proves the lemma.
Least significant digits
------------------------
Now we turn our attention to the lowest range of $j$. In particular, the goal is the proof of the following
\[prop:least\_significant\] Let $P$ and $\rho$ be positive reals and $f$ be a pseudo-polynomial as in . If $j$ is such that $$\begin{gathered}
\label{mani:gammarange}
1\leq q^j\leq P^{\theta_r-\rho}\end{gathered}$$ holds, then for $1\leq\nu\leq P^\gamma$ there exists $\eta>0$ (depending only on $f$ and $\rho$) such that $$\begin{gathered}
S(P,j,\nu)=(\log P)^8P^{1-\eta}.\end{gathered}$$
Before we launch into the proof we collect some tools that will be necessary in the sequel. A standard idea for estimating exponential sums over the primes is to rewrite them into ordinary exponential sums over the integers having von Mangoldt’s function as weights and then to apply Vaughan’s identity. We denote by $$\Lambda(n)=\begin{cases}
\log p,&\text{if $n=p^k$ for some prime $p$ and an integer $k\geq1$;}\\
0,&\text{otherwise}.
\end{cases}$$ von Mangoldt’s function. For the rewriting process we use the following
\[mr:lem11\] Let $g$ be a function such that ${\left\vert}g(n){\right\vert}\leq 1$ for all integers $n$. Then $${\left\vert}\sum_{p\leq P}g(p){\right\vert}\ll\frac1{\log P}\max_{t\leq P}
{\left\vert}\sum_{n\leq t}\Lambda(n)g(n){\right\vert}+\mathcal{O}(\sqrt{P}).$$
This is Lemma 11 of [@mauduit_rivat2010:sur_un_probleme]. However, the proof is short and we need some piece later.
We start with a summation by parts yielding $$\sum_{p\leq P}g(p)=\frac1{\log P}\sum_{p\leq
x}\log(p)g(p)+\int_2^P\left(\sum_{p\leq
t}\log(p)g(p)\right)\frac{\mathrm{d}t}{t(\log t)^2}.$$
Now we cut the integral at $\sqrt{P}$ and use Chebyshev’s inequality (*cf.* [@tenenbaum1995:introduction_la_theorie Théorème 1.3]) in the form $\sum_{p\leq
t}\log(p)\leq\log(t)\pi(t)\ll t$ for the lower part. Thus $$\begin{aligned}
{\left\vert}\sum_{p\leq P}g(p){\right\vert}&\leq\left(\frac1{\log
P}+\int_{\sqrt{P}}^P\frac{\mathrm{d}t}{t(\log t)^2}\right)
\max_{\sqrt{P}<t\leq P}{\left\vert}\sum_{p\leq
P}\log(p)g(p){\right\vert}+\mathcal{O}(\sqrt{P})\\
&=\frac2{\log P}\max_{\sqrt{P}<t\leq P}{\left\vert}\sum_{p\leq t}\log(p)g(p){\right\vert}+\mathcal{O}(\sqrt{P}).\end{aligned}$$
Finally we again use Chebyshev’s inequality $\pi(t)\ll t/\log(t)$ to obtain $$\begin{gathered}
\label{mani:log_Mangoldt_equivalence}
{\left\vert}\sum_{n\leq t}\Lambda(n)g(n)-\sum_{p\leq t}\log(p)g(p){\right\vert}\leq\sum_{p\leq\sqrt{t}}\log(p)\sum_{a=2}^{{\left\lfloor}\frac{\log(t)}{\log(p)}{\right\rfloor}}1
\leq\pi(\sqrt{t})\log(t)\ll\sqrt{t}.\end{gathered}$$
In the next step we use Vaughan’s identity to subdivide this weighted exponential sum into several sums of Type I and II.
\[bkmst:lem23\] Assume $F(x)$ to be any function defined on the real line, supported on $[P/2, P]$ and bounded by $F_0$. Let further $U,V,Z$ be any parameters satisfying $3
\leq U < V < Z < P$, $Z \geq 4U^2$, $P \geq 64 Z^2 U$, $V^3 \geq 32 P$ and $Z-\frac12\in\mathbb{N}$. Then $$\left| \sum_{P/2< n\leq P} \Lambda(n) F(n) \right| \ll K
\log P + F_0 + L (\log P)^8 ,$$ where $K$ and $L$ are defined by $$\begin{aligned}
K&=\max_M\sum_{m=1}^\infty d_3(m)\left\vert\sum\limits_{Z<n\leq M} F(mn)\right\vert,\\
L&=\sup\sum_{m=1}^\infty d_4(m)\left\vert\sum\limits_{U < n < V} b(n) F(mn)\right\vert,\end{aligned}$$ where the supremum is taken over all arithmetic functions $b(n)$ satisfying $|b(n)| \leq d_3(n).$
After subdividing the weighted exponential sum with Vaughan’s identity we will use the following lemma in order to estimate the occurring exponential sums.
\[bkmst:lem25\] Let $X,k,q\in \mathbb{N}$ with $k,q\geq 0$ and set $K=2^k$ and $Q=
2^q$. Let $h(x)$ be a polynomial of degree $k$ with real coefficients. Let $g(x)$ be a real $(q+k+2)$ times continuously differentiable function on $[X/2 , X]$ such that $\left| f^{(r)}(x)
\right| \asymp F X^{-r}$ $( r = 1, \dots, q+k+2) $. Then, if $F =
o (X^{q+2})$ for $F$ and $X$ large enough, we have $$\left| \sum_{X/2 < x \leq X} e(g(x) + h(x)) \right| \ll X^{1 - \frac{1}{K}} + X \left( \frac{\log^k X}{F} \right)^{\frac{1}{K}} + X \left( \frac{F}{X^{q+2}} \right)^{\frac{1}{(4KQ-2K)}}.$$
Now we have the necessary tools to state the
An application of Lemma \[mr:lem11\] yields $$S(P,j,\nu)\ll\frac1{\log P}\max{\left\vert}\sum_{n\leq
P}\Lambda(n)e\left(\frac{\nu}{q^j}(g(n)+h(n))\right){\right\vert}+P^{\frac12}.$$ We split the inner sum into $\leq \log
P$ sub sums of the form $${\left\vert}\sum\limits_{X< n \leq
2X}\Lambda(n)e\left(\frac{\nu}{q^j}(g(n)+h(n))\right){\right\vert}$$ with $2X \leq P$ and let $S$ be a typical one of them. We may assume that $X \geq P^{1-\rho}$.
Using Vaughan’s identity (Lemma \[bkmst:lem23\]) with $U = \frac{1}{4} X^{1/5}$, $V= 4 X^{1/3}$ and $Z$ the unique number in $\frac12+\mathbb{N}$, which is closest to $\frac{1}{4}
X^{2/5}$, we obtain $$\begin{gathered}
\label{mani:S}
S \ll 1+(\log X)S_1+(\log X)^8S_2,\end{gathered}$$ where $$\begin{aligned}
S_1&=\sum_{x < \frac{2X}{Z}} d_3(x) \sum_{y > Z, \frac{X}{x} < y < \frac{2X}{x}} e\left(\frac{\nu}{q^j}(g(xy)+h(xy))\right)\\
S_2&=\sum_{\frac XV<x\leq\frac{2X}U} d_4(x) \sum_{U < y < V, \frac{X}{x} < y \leq \frac{2X}{x}} b(y) e\left(\frac{\nu}{q^j}(g(xy)+h(xy))\right)\notag\end{aligned}$$
We start with the estimation of $S_1$. Since $d_3(x)\ll
x^{\varepsilon}$ we have for $$\begin{aligned}
{\left\vert}S_1{\right\vert}\ll X^\varepsilon\sum_{x\leq\frac{2X}Z}
{\left\vert}\sum_{\substack{\frac Xx<y\frac{2X}x\\y>Z}}e\left(\frac{\nu}{q^j}(g(xy)+h(xy))\right){\right\vert}.\end{aligned}$$ For estimating the inner sum we fix $x$ and denote $Y=\frac Xx$. Since $\theta_r\not\in{\mathbb{Z}}$ and $\theta_r>k\geq0$, we have that $${\left\vert}\frac{\partial^\ell g(xy)}{\partial y^\ell}{\right\vert}\asymp X^{\theta_r}Y^{-\ell}.$$
Now on the one hand, since $q^j\leq P^{\theta_r-\rho}$, we have $\nu
q^{-j}X^{\theta_r}\gg X^{\rho}$. On the other hand for $\ell\geq5(\lfloor\theta_r\rfloor+1)$ we get $$\frac{\nu}{q^j}X^{\theta_r}Y^{-\ell}\leq P^\gamma X^{\theta_r-\frac25\ell}\ll X^{-\frac12}.$$
Thus an application of Lemma \[bkmst:lem25\] yields the following estimate: $$\label{mani:estim:S1}
\begin{split}
{\left\vert}S_1{\right\vert}&\ll X^{\varepsilon}\sum_{x \leq 2X/Z} Y \left[
Y^{-\frac{1}{K}} + (\log Y)^kX^{-\frac{\rho}{K}} + X^{-\frac{1}{2} \frac{1}{4K \cdot 8L^5 - 2K}} \right] \\
&\ll X^{1+\varepsilon}(\log X)\left(X^{-\rho} + X^{-\frac{1}{64L^5-4} } \right)^{\frac1K},
\end{split}$$ where we have used that $\frac kK<1$ and $\rho<\frac13$.
For the second sum $S_2$ we start by splitting the interval $(
\frac{X}{V} , \frac{2X}{U} ]$ into $\leq \log X$ subintervals of the form $(X_1, 2X_1]$. Thus $$\begin{aligned}
{\left\vert}S_2{\right\vert}&\leq (\log X)X^{\varepsilon}\sum_{X_1<x\leq
2X_1}{\left\vert}\sum_{\substack{U<y<V\\\frac
Xx<y\leq\frac{2X}x}}b(y)e\left(\frac{\nu}{q^j}(g(xy)+h(xy))\right){\right\vert}\end{aligned}$$
Now an application of Cauchy’s inequality together with ${\left\vert}b(y){\right\vert}\ll X^\varepsilon$ yields $$\begin{aligned}
{\left\vert}S_2{\right\vert}^2
&\leq (\log X)^2X^{2\varepsilon}X_1\sum_{X_1<x\leq
2X_1}{\left\vert}\sum_{\substack{U<y<V\\\frac
Xx<y\leq\frac{2X}x}}b(y)e\left(\frac{\nu}{q^j}(g(xy)+h(xy))\right){\right\vert}^2\\
&\ll (\log X)^2X^{4\varepsilon}X_1\\
&\quad\times\left(X_1\frac{X}{X_1}+{\left\vert}\sum_{X_1<x\leq2X_1} \sum_{A < y_1 < y_2 \leq B}e\left(\frac{\nu}{q^j} (g(xy_1)-g(xy_2) + h(xy_1)-h(xy_2))\right) {\right\vert}\right)\end{aligned}$$ where $A = \max \{U, \frac{X}{x} \} $ and $B = \min \{U, \frac{2X}{x} \}$. Changing the order of summation, we get $$\begin{gathered}
|S_2|^2 \ll (\log X)^2X^{4\varepsilon}X_1\\
\times\left(X+
\sum_{A < y_1 < y_2 \leq B}{\left\vert}\sum_{X_1<x\leq2X_1} e\left(\frac{\nu}{q^j} (g(xy_1)-g(xy_2) + h(xy_1)-h(xy_2))\right) {\right\vert}\right)\end{gathered}$$
As above we want to apply Lemma \[bkmst:lem25\]. To this end we fix $y_1$ and $y_2 \ne y_1$. Similarly to above we get that $${\left\vert}\frac{\partial^\ell\left(g(xy_1)-g(xy_2)+h(xy_1)-h(xy_2)\right)}{\partial x^\ell}{\right\vert}\asymp\frac{{\left\vert}y_1-y_2{\right\vert}}{y_1}X^{\theta_r}X_1^{-\ell}.$$ Now, on the one hand we have $\frac{\nu}{q^j}\frac{{\left\vert}y_1-y_2{\right\vert}}{y_1}X^{\theta_r}\gg X^{\rho}$ and on the other hand $$\frac{\nu}{q^j}\frac{{\left\vert}y_1-y_2{\right\vert}}{y_1}X^{\theta_r}X_1^{-\ell}
\ll X^{\gamma+\theta_r}\left(\frac{X}{V}\right)^{-\ell}
\ll X^{\gamma+\theta_r-\frac23\ell}
\ll X^{-\frac12}$$ if $\ell\geq2\lfloor\theta_r\rfloor+3$. Thus again an application of Lemma \[bkmst:lem25\] yields $$\label{mani:estim:S2}
\begin{split}
{\left\vert}S_2{\right\vert}^2 &\ll (\log X)^2X^{4\varepsilon}X_1\left(X +
\sum_{A<y_1<y_2\leq B} X_1\left( X_1^{-\frac{1}{K}} + X^{-\frac{\rho}{K}} + X^{-\frac{1}{2} \frac{1}{4K \cdot 2L^2 - 2K}}\right)\right) \\
&\ll (\log X)^2X^{4\varepsilon}\left(X^{\frac53} + X^{2-\frac{\rho}{K}} + X^{2- \frac{1}{16KL^2 - 4K}}\right).
\end{split}$$
Plugging the two estimates and into proves the proposition.
The digits in the middle
------------------------
Now we are getting more involved in order to consider those $j$ leading to a position between $\theta_r$ and $k$. These sums correspond to the “digits in the middle” in the proof of Theorem \[mani:centralthm\]. We want to prove the following
\[prop:middle\_digits\] Let $P$ and $\rho$ be positive reals and $f$ be a pseudo-polynomial as in . If $2\rho<\theta_r<k$ and $j$ is such that $$\begin{gathered}
\label{mani:middle_range}
P^{\theta_r-\rho}< q^j\leq P^{k-1+\rho}\end{gathered}$$ holds, then for $1\leq\nu\leq P^\gamma$ we have $$\begin{gathered}
S(P,j,\nu)=\sum_{p\leq P}e\left(\frac{\nu f(p)}{q^j}\right)\ll P^{1-\frac{\rho}{4^k}}.\end{gathered}$$
The main idea in this range is to use that the dominant part of $f$ comes from the polynomial $h$. Therefore after getting rid of the function $g$ we will estimate the sum over the polynomial by the following
\[lem:exponential\_sum\_primes\_poly\] Let $h\in{\mathbb{R}}[X]$ be a polynomial of degree $k\geq2$. Suppose $\alpha$ is the leading coefficient of $h$ and that there are integers $a$, $q$ such that $${\left\vert}q\alpha-a{\right\vert}<\frac1q\quad\text{with}\quad
(a,q)=1.$$ Then we have for any $\varepsilon>0$ and $H\leq X$ $$\sum_{X<p\leq X+H}\log(p)e(h(p))\ll
H^{1+\varepsilon}\left(\frac1q+\frac1{H^{\frac12}}+\frac{q}{H^k}\right)^{4^{1-k}}.$$
This is a slight variant of [@harman1981:trigonometric_sums_over Theorem 1], where we sum over an interval of the form $]X,X+H]$ instead of one of the form $]0,X]$.
Now we have enough tools to state the
As in the Proof of Proposition \[prop:least\_significant\] we start by an application of Lemma \[mr:lem11\] yielding $$S(P,j,\nu)\ll\frac1{\log P}\max{\left\vert}\sum_{n\leq
P}\Lambda(n)e\left(\frac{\nu}{q^j}(g(n)+h(n))\right){\right\vert}+P^{\frac12}.$$
We split the inner sum into $\leq \log P$ sub sums of the form $$S:=\sum_{X<n\leq X+H}\Lambda(n)e\left(\frac{\nu}{q^j}(g(n)+h(n))\right)$$ with $P^{1-2\rho}\leq X\leq P$ and $$H=\min\left(P^{1-\theta_r}{\left\vert}\nu{\right\vert}^{-1}q^j,X\right).$$ Now we want to separate the function parts $g$ and $h$. Therefore we define two functions $T$ and $\varphi$ by $$T(x)=\sum_{X< n\leq
X+x}\Lambda(n)e\left(\frac{\nu}{q^j}h(n)\right)
\quad\text{and}\quad
\varphi(x):=e\left(\frac{\nu}{q^j}g\left(X+x\right)\right)$$ Then an application of summation by parts yields $$\label{mani:eq_1}
\begin{split}
\sum_{X< n\leq X+H}\Lambda(n)
e\left(\frac{\nu}{q^j}(g(n)+h(n))\right)
&=\sum_{n=1}^{H} \varphi(n)(T(n)-T(n-1))\\
&=\sum_{n=1}^{H}T(n)\left(\varphi(n)-\varphi(n+1)\right)+\varphi(H-1)T(H)\\
&\ll{\left\vert}T(H){\right\vert}+\sum_{n=1}^{H-1}\left|\varphi(n)-\varphi(n+1)\right|{\left\vert}T(n){\right\vert}\end{split}$$
Let $\alpha_k$ be the leading coefficient of $P$. Then by Diophantine approximation there always exists a rational $a/b$ with $b>0$, $(a,b)=1$, $$1\leq b\leq H^{k-\rho}
\quad\text{and}\quad
{\left\vert}\frac{\nu\alpha_k}{q^j}-\frac ab{\right\vert}\leq
\frac{H^{\rho-k}}{b}.$$ We distinguish three cases according to the size of $b$.
- **Case 1.** $H^\rho<b$. In this case we may apply Lemma \[lem:exponential\_sum\_primes\_poly\] together with to get $$T(h)\ll
H^{1-\frac{\rho}{4^{k-1}}+\varepsilon}.$$
- **Case 2.** $2\leq b<H^\rho$. In this case we get that $${\left\vert}\frac{\nu\alpha_k}{q^j}{\right\vert}\geq{\left\vert}\frac
ab{\right\vert}-\frac1{b^2}\geq\frac1{2b}\geq\frac12H^{-\rho}\geq\frac12P^{-\rho}.$$ Since $2\rho<\theta_r$, this contradicts our lower bound $q^j\geq P^{\theta_r-\rho}$.
- **Case 3.** $b=1$. This case requires a further distinction according to whether $a=0$ or not.
- **Case 3.1.** ${\left\vert}\frac{\nu\alpha_k}{q^j}{\right\vert}\geq\frac12$. It follows that $$q^j\leq2{\left\vert}\nu\alpha_k{\right\vert}$$ again contradicting our lower bound $q^j\geq P^{\theta_r-\rho}$.
- **Case 3.2.** ${\left\vert}\frac{\nu\alpha_k}{q^j}{\right\vert}<\frac12$. This implies that $a=0$ which yields $$\begin{gathered}
\label{case3.2}
q^j\geq{\left\vert}\nu\alpha_k{\right\vert}H^{k-\rho}.
\end{gathered}$$ We distinguish two further cases according to whether $P^{1-\theta_r}{\left\vert}\nu{\right\vert}^{-1}q^j\leq X$ or not.
- **Case 3.2.1** $P^{1-\theta_r}{\left\vert}\nu{\right\vert}^{-1}q^j\leq X$. This implies that $q^j\leq P^{\theta_r}{\left\vert}\nu{\right\vert}$ and $$H=P^{1-\theta_r}{\left\vert}\nu{\right\vert}^{-1}q^j\geq
P^{1-\rho}{\left\vert}\nu{\right\vert}^{-1}\geq P^{1-2\rho}.$$ Plugging these estimates into gives $$P^{\theta_r}\geq
{\left\vert}\alpha_k{\right\vert}P^{(1-2\rho)(k-\rho)}.$$ However, since $4(k+1)\rho<1$, we have $$(1-2\rho)(k-\rho)>k-1+2\rho\geq\theta_r$$ yielding a contradiction.
- **Case 3.2.2** $P^{1-\theta_r}{\left\vert}\nu{\right\vert}^{-1}q^j>X$. Then $H=X\geq
P^{1-2\rho}$ and becomes $$P^{k-1+\rho}\geq{\left\vert}\nu\alpha_k{\right\vert}P^{(1-2\rho)(k-\rho)}$$ yielding a similar contradiction as in **Case 3.2.1**.
Therefore **Case 1** is the only possible and we may always apply Lemma \[lem:exponential\_sum\_primes\_poly\] together with . Plugging this into yields $$\begin{aligned}
\sum_{X< n\leq X+H}\Lambda(n)
e\left(\frac{\nu}{q^j}(g(n)+h(n))\right)
&\ll H^{1-\frac{\rho}{4^{k-1}}+\varepsilon}\left(1+\sum_{X< n\leq X+H}\left|\varphi(n)-\varphi(n+1)\right|\right)\end{aligned}$$
Now by our choice of $H$ together with an application of the mean value theorem we have that $$\sum_{X\leq n\leq X+H}{\left\vert}\varphi(n)-\varphi(n+1){\right\vert}\ll H\frac{\nu}{q^j}P^{\theta-1}\ll 1.$$ Thus $$\begin{aligned}
\sum_{X\leq n\leq X+H} \Lambda(n)
e\left(\frac{\nu}{q^j}(g(n)+h(n))\right)
\ll H^{1-\frac{\rho}{4^{k-1}}+\varepsilon}.\end{aligned}$$
Proof of Theorem \[mani:centralthm\], Part II {#sec:proof-prop-refm2}
=============================================
Now we use all the tools from the section above in order to estimate
$$\begin{gathered}
\label{distance_from_mean}
\sum_{j=\ell}^J{\left\vert}\sum_{p\leq P}\mathcal{I}(q^{-j}f(p))-\frac{\pi(P)}{q^{\ell}}{\right\vert}\ll\pi(P)H^{-1}J+\sum_{\nu=1}^{H}
\nu^{-1}\sum_{j=\ell}^JS(P,j,\nu).\end{gathered}$$
As indicated in the section above, we split the sum over $j$ into two or three parts according to whether $\theta_r>k$ or not. In any case an application of Proposition \[prop:least\_significant\] yields for the least significant digits that $$\begin{gathered}
\label{estimate:least}
\sum_{1\leq \nu\leq P^\gamma}\nu^{-1}\sum_{1\leq q^{j}\leq
P^{\theta_r-\rho}} S(P,j,\nu)
\ll (\log P)^9JP^{1-\eta}.\end{gathered}$$
Now let us suppose that $\theta_r>k$. Then an application of Proposition \[prop:most\_significant\] yields $$\label{estimate:most_non_integer}
\begin{split}
\sum_{1\leq \nu\leq P^\gamma}\nu^{-1}&\sum_{P^{\theta_r-\rho}< q^{j}\leq
P^{\theta_r}}S(P,j,\nu)\\
&\ll \sum_{1\leq \nu\leq P^\gamma}\nu^{-1}\sum_{P^{\theta_r-\rho}< q^{j}\leq
P^{\theta_r}}\frac1{\log
P}\left(\frac{\nu}{q^j}\right)^{-\frac1{{\left\lfloor}\theta_r{\right\rfloor}}}+\frac{P}{(\log P)^{G-2}}\\
&\ll \frac{P}{\log P}.
\end{split}$$
Plugging the estimates and into we get that $$\sum_{j=\ell}^J{\left\vert}\sum_{p\leq P}\mathcal{I}(q^{-j}f(p))-\frac{\pi(P)}{q^{\ell}}{\right\vert}\ll\frac{P}{\log P},$$ which together with proves Theorem \[mani:centralthm\] in the case that $\theta_r>k$.
On the other side if $\theta_r<k$, then we consider the two ranges $$P^{\theta_r-\rho}<q^j\leq P^{k-1+\rho}
\quad\text{and}\quad
P^{k-1+\rho}<q^j\leq P^k.$$ For the “digits in the middle” an application of Proposition \[prop:middle\_digits\] yields $$\label{estimate:middle}
\begin{split}
\sum_{1\leq \nu\leq P^\gamma}\nu^{-1}\sum_{P^{\theta_r-\rho}< q^{j}\leq
P^{k-1+\rho}}S(P,j,\nu)
&\ll \sum_{1\leq \nu\leq P^\gamma}\nu^{-1}\sum_{P^{\theta_r-\rho}< q^{j}\leq
P^{k-1+\rho}}P^{1-\frac{\rho}{4^k}}\\
&\ll(\log P)JP^{1-\frac{\rho}{4^k}}.
\end{split}$$
Finally we consider the most significant digits. By an application of Proposition \[prop:most\_significant\] we have $$\label{estimate:most_integer}
\begin{split}
\sum_{1\leq \nu\leq P^\gamma}\nu^{-1}&\sum_{P^{k-1+\rho}< q^{j}\leq
P^{k}}S(P,j,\nu)\\
&\ll \sum_{1\leq \nu\leq P^\gamma}\nu^{-1}\sum_{P^{k-1+\rho}< q^{j}\leq
P^{k}}\frac1{\log
P}\left(\frac{\nu}{q^j}\right)^{-\frac1k}+\frac{P}{(\log P)^{G-2}}\\
&\ll \frac{P}{\log P}.
\end{split}$$
Plugging the estimates , and into we get that $$\sum_{j=\ell}^J{\left\vert}\sum_{p\leq P}\mathcal{I}(q^{-j}f(p))-\frac{\pi(P)}{q^{\ell}}{\right\vert}\ll\frac{P}{\log P},$$ which together with proves Theorem \[mani:centralthm\] in the case that $\theta_r<k$.
Acknowledgment {#acknowledgment .unnumbered}
==============
The author wants to thank G[é]{}rald Tenenbaum for many fruitful discussions and suggestions in connection with the proof of Proposition \[prop:middle\_digits\].
[10]{}
V. Bergelson, G. Kolesnik, M. Madritsch, Y. Son, and R. Tichy, *Uniform distribution of prime powers and applications to van der corput sets*, Israel Journal of Mathematics (2013), accepted.
E. Borel, *[Les probabilités dénombrables et leurs applications arithmétiques.]{}*, Palermo Rend. **27** (1909), 247–271 (French).
Y. Bugeaud, *Distribution modulo one and [D]{}iophantine approximation*, Cambridge Tracts in Mathematics, vol. 193, Cambridge University Press, Cambridge, 2012.
D. Champernowne, *[The construction of decimals normal in the scale of ten]{}*, J. Lond. Math. Soc. **8** (1933), 254–260 (English).
A. H. Copeland and P. Erd[ő]{}s, *Note on normal numbers*, Bull. Amer. Math. Soc. **52** (1946), 857–860.
H. Davenport and P. Erd[ő]{}s, *Note on normal decimals*, Canadian J. Math. **4** (1952), 58–63.
M. Drmota and R. F. Tichy, *Sequences, discrepancies and applications*, Lecture Notes in Mathematics, vol. 1651, Springer-Verlag, Berlin, 1997.
G. Harman, *Trigonometric sums over primes. [I]{}*, Mathematika **28** (1981), no. 2, 249–254 (1982).
H. Iwaniec and E. Kowalski, *Analytic number theory*, American Mathematical Society Colloquium Publications, vol. 53, American Mathematical Society, Providence, RI, 2004.
I. K[á]{}tai, *On the sum of digits of primes*, Acta Math. Acad. Sci. Hungar. **30** (1977), no. 1–2, 169–173.
L. Kuipers and H. Niederreiter, *Uniform distribution of sequences*, Wiley-Interscience \[John Wiley & Sons\], New York, 1974, Pure and Applied Mathematics.
M. G. Madritsch, J. M. Thuswaldner, and R. F. Tichy, *Normality of numbers generated by the values of entire functions*, J. Number Theory **128** (2008), no. 5, 1127–1145.
M. G. Madritsch and R. F. Tichy, *Construction of normal numbers via generalized prime power sequences*, J. Integer Seq. **16** (2013), no. 2, Article 13.2.12, 17.
M. G. Madritsch, *The summatory function of $q$-additive functions on pseudo-polynomial sequences*, J. Théor. Nombres Bordeaux **24** (2012), 153–171.
C. Mauduit and J. Rivat, *Sur un problème de [G]{}elfond: la somme des chiffres des nombres premiers*, Ann. of Math. (2) **171** (2010), no. 3, 1591–1646.
Y. Nakai and I. Shiokawa, *A class of normal numbers*, Japan. J. Math. (N.S.) **16** (1990), no. 1, 17–29.
[to3em]{}, *Discrepancy estimates for a class of normal numbers*, Acta Arith. **62** (1992), no. 3, 271–284.
[to3em]{}, *Normality of numbers generated by the values of polynomials at primes*, Acta Arith. **81** (1997), no. 4, 345–356.
J. Schiffer, *Discrepancy of normal numbers*, Acta Arith. **47** (1986), no. 2, 175–186.
I. Shiokawa, *On the sum of digits of prime numbers*, Proc. Japan Acad. **50** (1974), 551–554.
G. Tenenbaum, *Introduction à la théorie analytique et probabiliste des nombres*, second ed., Cours Spécialisés \[Specialized Courses\], vol. 1, Société Mathématique de France, Paris, 1995.
J. D. Vaaler, *Some extremal functions in [F]{}ourier analysis*, Bull. Amer. Math. Soc. (N.S.) **12** (1985), no. 2, 183–216.
I. M. Vinogradov, *The method of trigonometrical sums in the theory of numbers*, Dover Publications Inc., Mineola, NY, 2004, Translated from the Russian, revised and annotated by K. F. Roth and Anne Davenport, Reprint of the 1954 translation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cyclic codes are an important class of linear codes, whose weight distribution have been extensively studied. So far, most of previous results obtained were for cyclic codes with no more than three zeros. Recently, [@Y-X-D12] constructed a class of cyclic codes with arbitrary number of zeros, and computed the weight distributions for several cases. In this paper, we determine the weight distribution for a new family of such codes. This is achieved by certain new methods, such as the theory of Jacobi sums over finite fields and subtle treatment of some complicated combinatorial identities.'
author:
- 'Jing Yang,[^1] Lingli Xia, [^2] Maosheng Xiong [^3]'
title: Weight Distributions of a Class of Cyclic Codes with Arbitrary Number of Zeros II
---
Cyclic codes, weight distribution, Gaussian periods, Jacobi sums.
Introduction {#sec-into}
============
A linear code $\mathcal{C}$ over the finite field $\mathbb{F}_q$ of length $n$ is a subspace of $\mathbb{F}_{q}^n$. It is called *cyclic* if it also satisfies that any $(c_0,c_1,\cdots ,c_{n-1})\in \mathcal{C}$ implies $(c_{n-1},c_0,\cdots,c_{n-2})\in \mathcal{C}$. By the one-to-one correspondence $$\begin{array}{cccl}
\sigma:& \mathcal{C}&\rightarrow &R:=\mathbb{F}_{q}[x]/(x^n-1)\\
&(c_0,c_1,\cdots ,c_{n-1})&\mapsto&c_0+c_1x+\cdots +c_{n-1}x^{n-1},
\end{array}$$ each cyclic code $\mathcal{C}$ is equivalent to an ideal of $R$. Since $R$ is a principal ideal ring, there exists a unique monic polynomial $g(x)$ with least degree such that $\sigma(\mathcal{C})=g(x)R$ and $g(x)\mid (x^n-1)$. The $g(x)$ is called the *generator polynomial* of $\mathcal{C}$ and $h(x):=(x^n-1)/g(x)$ is called the *parity-check polynomial* of $\mathcal{C}$. The cyclic code $\mathcal{C}$ is called irreducible (resp. reducible) if $h(x)$ is irreducible (resp. reducible) over $\mathbb{F}_q$. For $\mathcal{C}$ reducible, we say that $\mathcal{C}$ *has $t$ ($\ge 2)$ zeros* if $h(x)$ has $t$ irreducible factors over $\mathbb{F}_{q}$. (In the literature some authors call $\mathcal{C}$ as “the dual of a cyclic code with $t$ zeros” instead.)
Denote by $A_i$ the number of codewords of $\mathcal{C}$ with Hamming weight $i$. The [*weight enumerator*]{} of $\mathcal{C}$ with length $n$ is a polynomial in $\mathbb{Z}[Y]$ defined by $$A_0+A_1Y+A_2Y^2+ \cdots + A_nY^n.$$ The sequence $(A_0,A_1,\cdots ,A_n)$ is called the *weight distribution* of $\mathcal{C}$. The study of weight distribution of a linear code is important in both theory and application, since the weight distribution of a code gives the minimum distance and thus the error correcting capability of the code, and the weight distribution of a code allows the computation of the probability of error detection and correction with respect to some algorithms [@Klov]. Moreover, the weight distribution is always related to interesting and challenging problems in number theory ([@cal; @Schroof]).
For irreducible cyclic codes, an identity due to McEliece [@McE74] shows that the weights of the codes can be expressed via Gauss sums. Because Gauss sums in general are extremely difficult to evaluate, the weight distribution of irreducible cyclic codes is still quite difficult to obtain, however, extensive studies have been carried out with much success by various number theoretic techniques ([@AL06; @BM72; @BM73; @fit; @McE74; @McE72; @Rao10; @van; @wol]). In particular nice characterizations were given in [@D-Y12; @Vega1; @Vega2] for irreducible cyclic codes with exactly one nonzero weight; necessary and sufficient conditions were provided and conjectures were also raised by Schmidt and White [@schmidt] for irreducible cyclic codes with at most two nonzero weights. Interested readers may consult the survey paper [@D-Y12] for more updated information on the weight distribution of irreducible cyclic codes.
For reducible cyclic codes, it has been known that the determination of weight distribution involves the evaluation of exponential sums. This may be even more difficult in general. For many special families of reducible cyclic codes where neat expressions are available, various delicate techniques from number theory and algebraic combinatorics have been developed and utilized, and for some of such families, the weight distribution can been obtained (see for example [@Ding2; @FL08; @Feng12; @F-M12; @holl; @luo2; @luo3; @luo4; @Ding1; @M09; @Mois09; @Vega12; @Tang12; @Xiong1; @Xiong2; @Xiong3; @zeng]). However, to our best knowledge, most of these literature works focus on reducible cyclic codes with two or three zeros. The exponential sums which have been explicitly evaluated seem to share a common feature that they attain only a few distinct values. For reducible cyclic codes with more than three zeros, not much is known. In a beautiful work [@gegeng], the authors obtained the weight distribution of a class of cyclic codes with arbitrary number of zeros. Their work built upon an unexpected connection between the corresponding exponential sums and the spectra of Hermitian forms graphs which were known in the literature. In another recent work [@Y-X-D12] a general family of reducible cyclic codes with arbitrary number of zeros were constructed and under certain conditions the weight distribution was also obtained. The purpose of this paper is to explore the construction of [@Y-X-D12] much further and to determine the weight distribution for another new family of reducible cyclic codes with arbitrary number of zeros. Compared with [@Y-X-D12], we achieve our goal by more advanced theory of Jacobi sums and by more subtle treatment of some complicated combinatorial identities.
The rest of the paper is organized as follows. The codes we consider will be introduced in Section \[sec-II\], so are the main results (Theorems 1, 2 and 3). Section \[sec-pre\] introduces some mathematical tools such as cyclotomy, Gaussian periods and general Jacobi sums that will be needed later. In Sections \[sec-main\] and \[sec-mainII\] we prove our main theorems. To streamline the proofs of Theorems \[thm-e1\], \[thm-e2\] and \[thm-e3\] we have left out the proof of a complicated combinatorial identity to Section \[sec-app\]. Section \[sec-conclusion\] concludes this paper.
Weight Distribution of Code $\mathcal{C}_{(a_1,\cdots,a_t)}$ {#sec-II}
============================================================
We first fix some notation. Let $p$ be a prime, $q=p^s$, $r=q^m$ for some integers $s,m\geqslant 1$. Let $\mathbb{F}_r$ be a finite field of order $r$ and $\gamma$ be a generator of the multiplicative group $\mathbb{F}_{r}^*:=\mathbb{F}_r \setminus \{0\}$. For any $t \ge 2$, the family of reducible cyclic codes $\mathcal{C}_{(a_1,\cdots,a_t)}$ with $t$ zeros were introduced in [@Y-X-D12] as follows.
*For any $e \geqslant t \geqslant 2$, assume that*
- *$a \not \equiv 0 \pmod{r-1} \mbox{ and } e|(r-1)$;*
- *$a_i \equiv a+\frac{r-1}{e}\Delta_i \pmod{r-1},\, 1\leqslant i \leqslant t$, where $\Delta_i \not \equiv \Delta_j \pmod{e}$ for any $ i \ne j$ and\
$\gcd(\Delta_2-\Delta_1,\ldots,\Delta_t-\Delta_1,e)=1$;*
- *$
\deg h_{a_1}(x)=\cdots=\deg h_{a_t}(x)=m, \mbox{ and } h_{a_i}(x) \neq h_{a_j}(x)$ for any $1\leqslant i\neq j\leqslant t$, where $h_{a_i}(x)$ is the minimal polynomial of $\gamma^{-a_i}$ over $\mathbb{F}_q$*;
- *$N=\gcd \left(\frac{r-1}{q-1},a e\right)$*;
- *$\delta=\gcd(r-1,a_1,a_2,\cdots ,a_{t}),\ n=\frac{r-1}{\delta}$*.
*The cyclic code $\mathcal{C}_{(a_1,\cdots,a_t)}$ with $t$ zeros $\gamma^{-a_1},\cdots,\gamma^{-a_t}$ is given by* $$\label{def}
\begin{array}{l}
\mathcal{C}_{(a_1,\cdots,a_t)}=\left\{ c(x_1,x_2,\cdots,x_{t})=\left({{\mathrm{Tr}}}_{r/q}\left(\sum_{j=1}^t x_j \gamma^{a_ji} \right)\right)_{i=0}^{n-1}~:~\forall \, x_1,\cdots,x_{t}\in\mathbb{F}_{r} \right\},
\end{array}$$ *where ${{\mathrm{Tr}}}_{r/q}$ denotes the trace map from $\mathbb{F}_{r}$ to $\mathbb{F}_{q}$.*
It shall be noted that Condition iii) can be easily verified, for example, it holds if $\frac{r-1}{q^\ell - 1}\nmid N$ for any proper factor $\ell$ of $m$ (i.e. $\ell \mid m$ and $\ell<m$, see [@Y-X-D12 Lemma 6]). In particular this is always the case if $N=2$, which is our interest in the paper.
Delsarte’s Theorem [@Delsarte] states that $\mathcal{C}_{(a_1,\cdots,a_t)}$ is an $[n,tm]$ cyclic code over $\mathbb{F}_{q}$ with parity-check polynomial $h(x)=h_{a_1}(x)\cdots h_{a_{t}}(x)$. This class of codes $\mathcal{C}_{(a_1,\cdots,a_t)}$ contain many interesting cyclic codes as special cases which have been extensively studied in the literature ([@Ding1; @Ding2; @F-M12; @Tang12; @Xiong1; @Xiong2; @Xiong3]), all of which focus on the case $t=2$. For any $t \ge 3$, in [@Y-X-D12] we obtain the weight distribution of $\mathcal{C}_{(a_1,\cdots,a_t)}$ under either of the following conditions:
- for any $t,e \ge 2$ when $N=1$; or
- for any $t=e \ge 2$ with $N=1,2,3$; or with $N=(p^j+1)/k$ for some positive integers $j,k$; or with $N$ being a prime number such that $N \equiv 3\pmod{4},{\left(\frac{p}{N}\right)}=1$ (here ${\left(\frac{*}{*}\right)}$ denotes the Legendre symbol).
In this paper we obtain the weight distribution of $\mathcal{C}_{(a_1,\cdots,a_t)}$ for any $t \ge 2$ such that $t=e-1$ and $N=2$. Note that under these conditions, it is necessary that $q$ is odd, $m$ is even and $2|ae$. Our main results are stated as follows.
\[thm-e1\] For $N=2$ and any $t=e-1 \geqslant 2$, we further assume that $$\begin{aligned}
\label{1:assumption}
e|(q^{m/2}-1) \, \mbox{ \emph{and} } \, 2|a.\end{aligned}$$ Then $\mathcal{C}_{(a_1,\ldots,a_t)}$ is an $[n,tm]$ cyclic code over ${{\mathbb F}}_q$ with the minimal Hamming distance $d=\frac{2(q-1)(r-\sqrt{r})}{(t+1)q\delta}$. It has (at most) $\frac{1}{2}(t^2+5t-2)$ nonzero distinct weights.
- If $q \equiv 1 \pmod{4}$, then the weight distribution is listed in Table \[Table1\].
- If $q \equiv 3 \pmod{4}$, then the weight distribution is listed in Table \[Table2\].
-------------------------------------------------------------------------------------------------------------------------------------------------------------------
Weight Frequency $\quad (\forall \, 1 \leqslant k \leqslant t, 0 \leqslant u\leqslant k+1)$
------------------------------------------------------------------------- -----------------------------------------------------------------------------------------
$0$ once
$\frac{q-1}{(t+1) q \delta}\cdot \bigg\{(k+1)r-(k+1-2u)\sqrt{r}\bigg\}$ $\frac{(r-1)}{r 2^{k+2}}\cdot \binom{t+1}{k+1}\binom{k+1}{u}\cdot \bigg\{2(r-1)^k-$
$(-1)^k\left\{
(1+\sqrt{r})^u(1-\sqrt{r})^{k+1-u}+(1-\sqrt{r})^u(1+\sqrt{r})^{k+1-u}\right\}\bigg\}$
-------------------------------------------------------------------------------------------------------------------------------------------------------------------
: The weight distribution of $\mathcal{C}$ when $N=2$ and $t=e-1\geqslant 2$: Case (i).[]{data-label="Table1"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Weight Frequency $\quad (\forall \, 1 \leqslant k \leqslant t, 0 \leqslant u\leqslant k+1)$
----------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------
$0$ once
$\frac{q-1}{(t+1) q \delta}\cdot \bigg\{(k+1)r-(-1)^{m/2}(k+1-2u)\sqrt{r}\bigg\}$ $\frac{(r-1)}{r 2^{k+2}}\cdot \binom{t+1}{k+1}\binom{k+1}{u}\cdot \bigg\{2(r-1)^k-$
$(-1)^k\left\{
(1+\sqrt{r})^u(1-\sqrt{r})^{k+1-u}+(1-\sqrt{r})^u(1+\sqrt{r})^{k+1-u}\right\}\bigg\}$
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: The weight distribution of $\mathcal{C}$ when $N=2$ and $t=e-1\geqslant 2$: Case (ii).[]{data-label="Table2"}
We remark that if $N=2$ and $t=e-1$ is even, then the condition (\[1:assumption\]) will always be satisfied, so this settles the case completely. In particular the special case $N=2,e=3,t=2$ was already studied in [@Tang12]. When $N=2$ and $t=e-1$ is odd, there are two cases: if the condition (\[1:assumption\]) is satisfied, this is again settled by Theorem \[thm-e1\]; on the other hand, if the condition (\[1:assumption\]) is not satisfied, in principle the weight distribution can still be obtained. However, the formulas become quite complicated. To illustrate that, we first present the weight distribution for the simple case $t=3$ in Theorem \[thm-e2\], and then give a computational formula for the general case in Theorem \[thm-e3\].
\[thm-e2\] For $N=2$ and $t=e-1 =3$. Then $\mathcal{C}_{(a_1,a_2,a_3)}$ is an $[n,3m]$ cyclic code over ${{\mathbb F}}_q$ with the minimal Hamming distance $d=\frac{(q-1)(r-\sqrt{r})}{2q\delta}$. It has (at most) $12$ nonzero weights.
- If $2|a$, then its weight distribution is listed in Table \[Table3\] (or Table \[Table1\] with $t=3$).
- If $2 \nmid a$, then its weight distribution is listed in Table \[Table4\].
Weight Frequency
---------------------------------------- ----------------------------
0 once
$\frac{q-1}{2\delta q}(r+\sqrt{r})$ $3(r-1)$ times
$\frac{q-1}{2\delta q}(r-\sqrt{r})$ $3(r-1)$ times
$\frac{3(q-1)}{4\delta q}(r+\sqrt{r})$ $(r-1)(r-5)/2$ times
$\frac{3(q-1)}{4\delta q}(r-\sqrt{r})$ $(r-1)(r-5)/2$ times
$\frac{(q-1)}{4\delta q}(3r+\sqrt{r})$ $3(r-1)^2/2$ times
$\frac{(q-1)}{4\delta q}(3r-\sqrt{r})$ $3(r-1)^2/2$ times
$\frac{(q-1)}{\delta q}(r+\sqrt{r})$ $(r-1)(r^2-2r+9)/16$ times
$\frac{(q-1)}{\delta q}(r-\sqrt{r})$ $(r-1)(r^2-2r+9)/16$ times
$\frac{(q-1)}{2\delta q}(2r+\sqrt{r})$ $(r-1)(r^2-4r+3)/4$ times
$\frac{(q-1)}{2\delta q}(2r-\sqrt{r})$ $(r-1)(r^2-4r+3)/4$ times
$\frac{(q-1)}{\delta q}r$ $3(r-1)^3/8$ times
: The weight distribution of $\mathcal{C}_{(a_1,a_2,a_3)}$ when $t=e-1=3,N=2$ and $2\mid a$.[]{data-label="Table3"}
Weight Frequency
---------------------------------------- -----------------------------
0 once
$\frac{q-1}{2\delta q}(r+\sqrt{r})$ $(r-1)$ times
$\frac{q-1}{2\delta q}(r-\sqrt{r})$ $(r-1)$ times
$\frac{q-1}{2\delta q}r$ $4(r-1)$ times
$\frac{3(q-1)}{4\delta q}(r+\sqrt{r})$ $(r-1)^2/2$ times
$\frac{3(q-1)}{4\delta q}(r-\sqrt{r})$ $(r-1)^2/2$ times
$\frac{(q-1)}{4\delta q}(3r+\sqrt{r})$ $(r-1)(3r-7)/2$ times
$\frac{(q-1)}{4\delta q}(3r-\sqrt{r})$ $(r-1)(3r-7)/2$ times
$\frac{(q-1)}{\delta q}(r+\sqrt{r})$ $(r-1)^3/16$ times
$\frac{(q-1)}{\delta q}(r-\sqrt{r})$ $(r-1)^3/16$ times
$\frac{(q-1)}{2\delta q}(2r+\sqrt{r})$ $(r-1)(r^2-4r+3)/4$ times
$\frac{(q-1)}{2\delta q}(2r-\sqrt{r})$ $(r-1)(r^2-4r+3)/4$ times
$\frac{(q-1)}{\delta q}r$ $(r-1)(3r^2-6r+11)/8$ times
: The weight distribution of $\mathcal{C}_{(a_1,a_2,a_3)}$ when $t=e-1=3,N=2$ and $2\nmid a$.[]{data-label="Table4"}
We now consider the general case for $N=2$ and $t=e-1$. Denote $g:=\gamma^a$ and $\beta:=\gamma^{(r-1)/e}$. And let $A$ be the Vandermonde matrix of size $(t+1)\times (t+1)$, given by $$\label{matrix}A:=\left(
\begin{array}{llll}
1&1&\cdots&1 \\
1&\beta&\cdots&\beta^{e-1} \\
1&\beta^2&\cdots&\beta^{2(e-1)} \\
\vdots &\vdots &&\vdots\\
1&\beta^{e-1}&\cdots&\beta^{(e-1)^2}
\end{array}
\right),$$ Take $B$ be the $(t+1) \times t$-matrix whose columns consist of the $\{\triangle_1+1,\ldots,\triangle_t+1\} \pmod{e}$ columns of $A$, where $\triangle_i$ are the basic parameters of $\mathcal{C}_{(a_1,\cdots,a_t)}$. Let $$\label{equ-xtoy}
(y_0,\cdots ,y_{t})^T=B(x_1,\cdots ,x_t)^T.$$ Since rank$B=t$ (see also [@Y-X-D12 Lemma 18]), this gives a one-to-one correspondence between $(y_1,\ldots,y_t)$ and $(x_1,\ldots,x_t)$, and there exist some $0 \ne \lambda_i \in \mathbb{F}_{q}$ for $1\leqslant i\leqslant t$ such that $$\label{equ-y0}
y_0+\sum\limits_{i=1}^t \lambda_iy_i=0$$ We note that $\{\lambda_i\}_{i=1}^t$ depend only on the parameters $\{\triangle_i \pmod{e}\}_{i=1}^t$ and $\beta$. We further define $$\label{equ-l1l2}
\begin{array}{l}
l_0=\#\{i\mid\lambda_ig^i \mbox{ is a square in } \mathbb{F}_r,1\leqslant i\leqslant t\};\\
l_1=\#\{i\mid\lambda_ig^i \mbox{ is a nonsquare in } \mathbb{F}_r,1\leqslant i\leqslant t\}.
\end{array}$$ Next, we extend the definition of binomial coefficient to all integers such that $$\binom{n}{i}=0,\ \mbox{for $i<0$ and $i>n$}.$$
With such preparations, we give our main result for the general case as follows.
\[thm-e3\] In the case of $N=2$ and $t=e-1\geqslant 2$, the $\mathcal{C}_{(a_1,\ldots,a_t)}$ is an $[n,tm]$ cyclic code over ${{\mathbb F}}_q$ with the minimal Hamming distance $d=\frac{2(q-1)(r-\sqrt{r})}{(t+1)q\delta}$, and the Hamming weight of its codewords takes the value 0 once and the value $$\begin{aligned}
\frac{(q-1)}{(t+1)q \delta}\left[k(r-1)-2u\eta_0^{(2,r)}-2(k-u)\eta_1^{(2,r)}\right], \end{aligned}$$ for any $2 \le k \le t+1$ and $0 \leqslant u \le k$, with the frequency $$\begin{aligned}
\sum_{k_0=0}^k\sum_{u_0=0}^u\binom{l_0+1}{k_0}\binom{l_1}{k-k_0}
\binom{k_0}{u_0}\binom{k-k_0}{u-u_0}\Omega_{{\scriptsize \underbrace{0 \cdots 0}_{2u_0+k-k_0-u,} \underbrace{1\cdots 1}_{k_0+u-2u_0}}}. \end{aligned}$$ where $\eta_0^{(2,r)},\eta_1^{(2,r)}$ are given by Lemma \[lem-degree2\], $\Omega_{{\scriptsize \underbrace{0 \cdots 0}_{u} \underbrace{1\cdots 1}_{v}}}$ is determined by Lemma \[thm-NNN3\] and $l_0,l_1$ are defined by (\[equ-l1l2\]).
We remark that Theorem \[thm-e3\] is a general computational formula for the weight distribution of $\mathcal{C}_{(a_1,\cdots,a_t)}$, and the results of Theorem \[thm-e1\] and Theorem \[thm-e2\] can be viewed as its corollaries. However, the frequency formula in Theorem \[thm-e3\] is complicated since it depends on the choice of $\triangle_1,\cdots,\triangle_t$, and there seems no easy way to write them down in a simple closed form as Theorem \[thm-e1\].
In the end of this section, we give serval numerical examples to illustrate the our main theorems.
Let $(q, m, e, t)=(5,2,4,3)$. Then $\frac{r-1}{3}=\frac{5^2-1}{4}=6$. Let $\gamma$ be the generator of ${{\mathbb F}}_{25}^*$ with characteristic polynomial $\gamma^2 + 4\gamma + 2=0$. Let $(\triangle_1,\triangle_2,\triangle_3)=(1,2,3)$.
- For $a=2$ we have $(a_1,a_2,a_3)=(8,14,20)$, $(\delta,n)=(2,12)$ and $$h_{a_1}(x)=x^2 + x + 1, \, h_{a_2}(x)= x^2 + 3x + 4, \, h_{a_3}(x)= x^2 + 4x + 1.$$ The parity-check polynomial of $\mathcal{C}$ is then $
h(x)=x^6 + 3x^5 + 3 x^3+3x+4.
$ The code $\mathcal{C}$ is a $[12,6,4]$-cyclic code over ${{\mathbb F}}_5$ with weight enumerator given by $$1+72Y^4+312 Y^{6}+864 Y^{7}+1740 Y^{8}+3408 Y^{9}+5184 Y^{10}+3168 z^{11}+876 Y^{12}.$$ This also follows from Table III. There are 8 distinct non-zero weights because some of the weights in Table III turn out the same. More precisely, $$\begin{aligned}
\label{1:eqns} \begin{array}{ccc}
\frac{(q-1)}{2 \delta q}(r+\sqrt{r}) =\frac{3(q-1)}{4 \delta q}(r-\sqrt{r}), \\
\frac{(q-1)}{4 \delta q}(3r+\sqrt{r})=\frac{(q-1)}{\delta q}(r-\sqrt{r}),\\
\frac{3(q-1)}{4 \delta q}(r+\sqrt{r})=\frac{(q-1)}{2 \delta q}(2r-\sqrt{r}).\end{array}\end{aligned}$$
- For $a=1$ we have $(a_1,a_2,a_3)=(7,13,19)$, $(\delta,n)=(1,24)$ and $$h_{a_1}(x)=x^2 + x + 2, \, h_{a_2}(x)= x^2 + 2x + 1, \, h_{a_3}(x)= x^2 + 4x + 2.$$ The parity-check polynomial of $\mathcal{C}$ is then $
h(x)=x^6 + 2x^5 + 4x^4 + x^3+2x^2+3x+4.
$ The code $\mathcal{C}$ is a $[24,6,8]$-cyclic code over ${{\mathbb F}}_5$ with weight enumerator given by $$1+24Y^8+96 Y^{10}+312 Y^{12}+816 Y^{14}+1680 Y^{16}+3456 Y^{18}+5208 z^{20}+3168 Y^{22}+ 864 Y^{24}.$$ This also follows from Table IV. There are 9 distinct non-zero weights because some of the weights in Table IV turn out the same. More precisely, the equations (\[1:eqns\]) still hold true.
Cyclotomy, Gaussian periods and Jacobi sums {#sec-pre}
===========================================
An [*additive character*]{} of $\mathbb{F}_{r}$ is a nonzero function $\phi$ from $\mathbb{F}_{r}$ to the set of complex numbers such that $\phi(x+y)=\phi(x) \phi(y)$ for any pair $(x, y) \in \mathbb{F}_{r}^2$. Let ${{\mathrm{Tr}}}_{r/p}$ denote the trace function from $\mathbb{F}_{r}$ to $\mathbb{F}_{p}$ and $\zeta_p=e^{2\pi \sqrt{-1}/p}$ be the primitive $p$-th complex root of unit. The additive character $\psi$ given by $$\begin{aligned}
\label{dfn-add}
\psi(c)=\zeta_p^{{{\mathrm{Tr}}}_{r/p}(c)} \ \ \mbox{ for any }
c\in\mathbb{F}_{r}\end{aligned}$$ is called the [*canonical additive character*]{} of $\mathbb{F}_{r}$. For any $x\in \mathbb{F}_{r}$, one can easily check the orthogonal property $$\label{add-orth}
\frac{1}{r}\sum\limits_{x\in \mathbb{F}_{r}}\psi(ax)=\left\{
\begin{array}{ll}
1, & \hbox{if $a=0$;} \\
0, & \hbox{if $a\in \mathbb{F}^*_{r}$.}
\end{array}
\right.$$
Let $r-1=l L$ for two positive integers $l, L\geqslant 1$, and let $\gamma$ be a fixed primitive element of $\mathbb{F}_{r}$. Define $C_{i}^{(L,r)}=\gamma^i \langle \gamma^{L} \rangle$ for $i=0,1,...,L-1$, where $\langle \gamma^{L} \rangle$ denotes the subgroup of $\mathbb{F}_{r}^*$ generated by $\gamma^{L}$. The $C_{i}^{(L,r)}$ are called the [*cyclotomic classes*]{} of order $L$ in $\mathbb{F}_{r}$. The [*Gaussian periods*]{} of order $L$ are defined by $$\eta_i^{(L,r)} =\sum_{x \in C_i^{(L,r)}} \psi(x), \quad i=0,1,..., L-1.$$
The values of Gaussian periods are difficult to compute in general. However, they are known in a few cases. We will need the following whose proofs can be found in [@B-E-W] and [@Myer].
\[lem-degree2\] When $L=2$, the Gaussian periods are given by $$\begin{aligned}
\eta_0^{(2,r)}=
\left\{
\begin{array}{ll}
\frac{-1+(-1)^{s\cdot m-1}r^{1/2}}{2}, & \mbox{if $p\equiv 1 \pmod{4}$} \\
\frac{-1+(-1)^{s\cdot m-1}(\sqrt{-1})^{s\cdot m} r^{1/2}}{2}, & \mbox{if $p\equiv 3 \pmod{4}$}
\end{array}
\right.\end{aligned}$$ and $\eta_1^{(2,r)} = -1 - \eta_0^{(2,r)}.$
A [*multiplicative character*]{} of ${{\mathbb F}}_r$ is a nonzero function $\chi$ from ${{\mathbb F}}_r^*$ to the set of complex numbers such that $\chi(xy)=\chi(x)\chi(y)$ for all the pairs $(x, y) \in {{\mathbb F}}_r^*
\times {{\mathbb F}}_r^*$. For $j=1,2,\ldots,r-1$, one can easily check that the functions $\chi^{(j)}$ with $$\begin{aligned}
\label{dfn-mul}
\chi^{(j)}(\gamma^k)=\zeta_{r-1}^{jk} \ \ \mbox{for } k=0,1,\ldots,r-2 \nonumber\end{aligned}$$ give all the multiplicative character of order dividing $r-1$, here $\zeta_{r-1}$ denotes the primitive complex $(r-1)$-th root of unit. When $j=r-1$, $\varepsilon(c):=\chi^{(r-1)}(c)=1 \mbox{ for all }
c\in{{\mathbb F}}_r^*,$ which is called the [*trivial multiplicative character*]{} of ${{\mathbb F}}_r$. One can check the following orthogonal property of multiplicative characters $$\label{equ-orthogonal1}
\frac{1}{r-1}\sum\limits_{x\in \mathbb{F}_{r}^*}\chi(x)=
\left\{
\begin{array}{ll}
1, & \hbox{if\ }\chi=\varepsilon; \\
0, & \hbox{otherwise.}
\end{array}
\right.$$ Furthermore, we may extend the definition of any multiplicative character $\chi$ to $\mathbb{F}_{r}$ as follows, $$\label{equ-chi(0)}
\chi(0)=\left\{
\begin{array}{ll}
0, & \hbox{if }\chi\neq \varepsilon; \\
1, & \hbox{if }\chi=\varepsilon.
\end{array}
\right.\nonumber$$
Let $k\geqslant 2$ and $\chi_1,\cdots ,\chi_k$ be multiplicative characters of $\mathbb{F}_{r}$. The *Jacobi sum* related with $\chi_1,\cdots ,\chi_k$ over $\mathbb{F}_{r}$ is defined by $$J(\chi_1,\cdots ,\chi_k):=\sum\limits_{z_1,\cdots z_k\in \mathbb{F}_{r}\atop z_1+\cdots +z_k=1}\chi_1(z_1) \chi_2(z_2)\cdots \chi_k(z_k).$$
The following ([@B-E-W]) are elementary properties of Jacobi sums.
\[lem\_J\]
(a). $J(\underbrace{\varepsilon,\cdots ,\varepsilon}_k)=q^{k-1}$.
(b). $J(\chi_1,\cdots,\chi_k)=0$ if some but not all of $\chi_1,\cdots ,\chi_k$ are trivial.
(c). When $r$ is odd, let $\rho$ be the quadratic multiplicative character of $\mathbb{F}_{r}$, then $$J(\underbrace{\rho,\cdots,\rho}_k)=
\left\{
\begin{array}{ll}
-\rho(-1)^{\frac{k}{2}}r^{\frac{k-2}{2}}, & \hbox{if $k$ is even;} \\
\rho(-1)^{\frac{k-1}{2}}r^{\frac{k-1}{2}}, & \hbox{if $k$ is odd.}
\end{array}
\right.$$
We now define the *reduced Jacobi sums* below, which is needed in the next section. $$\label{equ-J*}
J^*(\chi_1,\cdots ,\chi_k):=\sum\limits_{z_1,\cdots z_k\in \mathbb{F}^*_{r}\atop z_1+\cdots +z_k=1}\chi_1(z_1) \chi_2(z_2)\cdots \chi_k(z_k).$$ Notice that $J^*(\chi_1,\cdots ,\chi_k)=J(\chi_1,\cdots ,\chi_k)$ if all of $\chi_1,\cdots ,\chi_k$ are non-trivial. The following results give the evaluation of $J^*(\chi_1,\cdots ,\chi_k)$ if some of $\chi_1,\cdots ,\chi_k$ are trivial. The next result is not difficult but may be of independent interest. It is essential in Section \[sec-app\] to establish a complicated combinatorial identity, which is needed in the proofs of Theorems \[thm-e1\] and \[thm-e2\].
\[lem-J\*\] (a). $J^*(\varepsilon,\cdots ,\varepsilon)=\left\{(r-1)^k-(-1)^k\right\}/r$.
(b). Define $J(\chi):=1$ for any multiplicative character $\chi$. Let $u$ be an integer such that $0\leqslant u\leqslant k-1$. If $\chi_1,\cdots ,\chi_{k-u}$ are all nontrivial multiplicative characters, then $$J^*(\chi_1,\cdots ,\chi_{k-u},\underbrace{\varepsilon,\cdots ,\varepsilon}_{u})=(-1)^uJ(\chi_1,\cdots ,\chi_{k-u}).$$
By definition, we have $$J^*(\varepsilon,\cdots ,\varepsilon)=J(\varepsilon,\cdots ,\varepsilon)-\sum\limits_{\mathcal{I}} \sum\limits_{\sum\limits_{i\in \mathcal{I}} z_i=1}\varepsilon(\prod\limits_{i\in \mathcal{I}} z_i),$$ where the subscript $\mathcal{I}$ under the $\sum$ symbol means to sum over all subsets $\mathcal{I}$ such that $\mathcal{I}\subsetneqq\{1,2,\cdots ,k\}$. Using the Inclusion-exclusion principle, Part (a) of Lemma \[lem-J\*\] can be easily proved. Now for Part (b), we have $\mathcal{I}'\subsetneqq\{k-u+1,\cdots ,k\}$, then $$\begin{array}{l}J^*(\chi_1,\cdots ,\chi_{k-u},\underbrace{\varepsilon,\cdots ,\varepsilon}_{u})\\
\ =J(\chi_1,\cdots ,\chi_{k-u},\underbrace{\varepsilon,\cdots ,\varepsilon}_{u})-\sum\limits_{\mathcal{I}'}\sum\limits_{\sum_{j=1}^{k-u} z_j+\sum\limits_{i\in\mathcal{I}'}z_i=1}\chi_1(z_1)\chi_2(z_2)\cdots\chi_k(z_{k-u})\varepsilon(\prod\limits_{i\in \mathcal{I}'}z_i)\\
\ =0-\binom{u}{1}J^*(\chi_1,\cdots ,\chi_{k-u},\underbrace{\varepsilon,\cdots ,\varepsilon}_{u-1})- \binom{u}{2}J^*(\chi_1,\cdots ,\chi_{k-u},\underbrace{\varepsilon,\cdots ,\varepsilon}_{u-2})-\cdots -
\binom{u}{u}J^*(\chi_1,\cdots ,\chi_{k-u}). \end{array}$$ By induction, Part (b) can be also verified.
Proof of Theorem \[thm-e1\] {#sec-main}
===========================
The weight distribution of $\mathcal{C}_{(a_1,\cdots ,a_t)}$ and Summation of Gaussian periods {#sec-Weight of C}
----------------------------------------------------------------------------------------------
We now consider the weight distribution of the cyclic code $\mathcal{C}_{(a_1,\cdots,a_t)}$ given in (\[def\]). Using the orthogonal relation (\[add-orth\]) and some computational techniques, in [@Y-X-D12] we haved expressed the Hamming weight of the codeword $c(x_1,\cdots,x_{t})$ by $$\begin{aligned}
\label{equ-Weight of C}
w_H(c(x_1,\cdots,x_{t})) =\frac{(r-1)(q-1)}{q \delta}-\frac{N(q-1)}{ eq \delta}\sum\limits_{h=0}^{e-1}
\bar\eta^{(N,r)}_{g^{h}\cdot\sum\limits_{\tau=1}^{t} x_\tau \beta_\tau^{h}},\end{aligned}$$ where $g=\gamma^a$, $\beta_\tau=\gamma^{\frac{r-1}{e}\Delta_\tau}$ for $1\leqslant \tau\leqslant t$ and $\bar\eta^{(N,r)}_{v}=\sum\limits_{z\in C_{0}^{(N,r)}}\psi(vz)$ for any $v\in\mathbb{F}_{r}$. These $\bar\eta^{(N,r)}_v$ are called the *modified Gaussian periods*, given by $$\left\{
\begin{array}{l}
\bar\eta^{(N,r)}_0=\frac{r-1}{N}\\
\bar\eta^{(N,r)}_{\gamma^{i}}=\eta_i^{(N,r)}\quad \hbox{ for $0\leqslant i\leqslant N-1$,}
\end{array}
\right.$$ where these $\eta_i^{(N,r)}$ are the ordinary Gaussian periods. Thus, to compute the weight distribution of cyclic code $\mathcal{C}_{(a_1,\cdots,a_t)}$, it suffices to compute the value distribution of the sum $$\label{equ-Tx}
T(x_1,\cdots ,x_{t}):=\sum\limits_{h=0}^{e-1}\bar\eta^{(N,r)}_{g^{h}
\cdot\sum_{\tau=1}^{t} x_\tau \beta_\tau^{h}},\quad (\forall x_1,\cdots ,x_t\in \mathbb{F}_{r}).$$ Now we deal with it under the assumption of $N=2$ and $t=e-1 \ge 2$.
$N=2$ and $t=e-1 \ge 2$
-----------------------
Since $N=2$, it is easy to see that $q$ is odd, $m$ is even and $-1=\gamma^{\frac{q^m-1}{2}}$ is a square. For simplicity, let us write $$\bar\eta_x:=\bar\eta^{(2,r)}_x, \quad \forall x \in {{\mathbb F}}_r.$$ Make a change of variables $$y_h=\sum_{\tau=1}^{t} x_\tau \beta_\tau^{h}, \quad 0 \le h \le t=e-1,$$ which can be written as $$\label{equ-xtoy}
(y_0,\cdots ,y_{t})^T=B(x_1,\cdots ,x_t)^T$$ for some $(t+1) \times t$ matrix $B$. Recall that $\beta=\gamma^{(r-1)/e}$ is an $e$-th root of unity in $\mathbb{F}_{r}$. Since $\beta_{\tau}=\beta^{\triangle_{\tau}}$, the matrix $B$ consists of $t$ columns of the Vandermonde matrix $A$, defined by (\[matrix\]). By [@Y-X-D12 Lemma 18], any $t$ rows of $B$ are linearly independent over ${{\mathbb F}}_q$. This gives a one-to-one correspondence between $(y_1,\ldots,y_t)$ and $(x_1,\ldots,x_t)$ and a relation $$\begin{aligned}
\label{3:xiong} y_0+ \sum_{h=1}^t\lambda_hy_h=0, \quad \mbox{ for some } 0 \ne \lambda_h \in {{\mathbb F}}_{q^{m}} \, \forall h. \end{aligned}$$ We define $\tilde{\lambda}_h\ (1\leqslant h\leqslant t)$ as $$\label{equ-lambda}
\tilde{\lambda}_h=\left\{\begin{array}{ll}1,&\mbox{if $\lambda_hg^h$ is a square in $\mathbb{F}_{r}$;}\\
\gamma,&\mbox{if $\lambda_hg^h$ is a nonsquare in $\mathbb{F}_{r}$,}\end{array}\right.$$ and we change variables again $\lambda_hy_h \to y_h$, then we see that to compute the weight distribution of the cyclic code $\mathcal{C}_{(a_1,\cdots,a_t)}$, it suffices to compute the value distribution of the sum $$\label{equ-Ty}
\widetilde{T}(y_0,\cdots ,y_{t}):=\bar\eta_{y_0}+\sum\limits_{h=1}^{t}\bar\eta_{\tilde{\lambda}_hy_h},\quad \forall \, (y_1,\cdots ,y_t)\in \mathbb{F}_{r}^t,$$ where $y_0:=y_0(y_1,\ldots,y_t)$ satisfies $$\label{eqn-y0}
y_0+\sum\limits_{h=1}^t y_h=0$$
Proof of Theorem \[thm-e1\]. {#sec-sub_e4t3}
----------------------------
When $2|a$, then $g=\gamma^a$ is a square. Moreover, $e|(q^{m/2}-1)$ means that $\beta=\gamma^{(q^{m/2}+1)(q^{m/2}-1)/{e}} \in {{\mathbb F}}_{q^{m/2}}$, hence the matrix $A$ is defined over ${{\mathbb F}}_{q^{m/2}}$, so are all the $\lambda_h$ in (\[3:xiong\]), thus $\lambda_h$ are all squares in ${{\mathbb F}}_{q^m}$, that is, $\tilde{\lambda}_h=1\ (\forall h)$.
To study the value distribution of $\widetilde{T}:=\widetilde{T}(y_0,\ldots,y_t)$, we will divide the space of $(y_1,\ldots,y_t) \in {{\mathbb F}}_r^t$ according to $s$, which counts the number of $i$’s ($0 \le i \le t$) such that $y_i=0$. Obviously $0 \le s \le t+1$.
If $s \ge t$, i.e., at least $t$ terms of $y_0,y_1,\ldots,y_t$ equal to 0, then all of them equal to 0, $\widetilde{T}=(t+1)\bar\eta_0$ and the frequency is 1.
If $s=t-1$, i.e., exactly $(t-1)$ terms of $y_0,y_1,\ldots,y_t$ equal to 0, say for example the two terms which are not 0 are $y_i,y_j$ for some $0 \le i < j \le t$, the number of choices of such $i,j$ is $\binom{t+1}{2}$, and the constraint (\[eqn-y0\]) becomes $y_i+y_j=0$, or $y_i=-y_j$. Hence for this $i,j$ we find that $$\widetilde{T}=(t-1)\bar\eta_0+\bar\eta_{y_j}+\bar\eta_{-y_j}=(t-1)\bar\eta_0+2\bar\eta_{y_j}.$$ So the value distribution of $\widetilde{T}$ for $s=t-1$ is as follows: $$\left.
\begin{array}{ll}
\mbox{Value} \, \, \widetilde{T} & \hbox{Frequency} \\
(t-1)\bar\eta_0+2\eta_0, & \frac{r-1}{2} \cdot \binom{t+1}{2} \\[2mm]
(t-1)\bar\eta_0+2\eta_1, & \frac{r-1}{2} \cdot \binom{t+1}{2} \\ \end{array}
\right.$$
Now suppose in general $s=t-k$ for some $k$ with $1 \le k \le t$. Say the $(k+1)$ terms which are not 0 are $y_{i_0},y_{i_1}, \ldots,y_{i_k}$ for some $0 \le i_0< i_1 < \cdots <i_k \le t$. The number of ways to choose such $i_j$’s is $\binom{t+1}{k+1}$, and for such $i_j$’s, the constraint (\[eqn-y0\]) becomes $$y_{i_0}+y_{i_1}+\cdots+y_{i_k}=0,$$ and we find that $$\widetilde{T}=(t-k)\bar\eta_0+\bar\eta_{y_{i_0}}+\bar\eta_{y_{i_1}}+\cdots+\bar\eta_{y_{i_k}}.$$ In order to compute the value distribution of $\widetilde{T}$ for these cases, it suffices to compute for any positive integer $u$ and any sequence $i_1,\cdots,i_{u},i_{u+1}\in\{0,1\}$ the value $\Omega_{i_1\cdots i_ui_{u+1}}$ given by $$\label{equ-NNN}
\Omega_{i_1\cdots i_ui_{u+1}}:=\#\left\{(x_1,\cdots ,x_u)\in (\mathbb{F}_{r}^*)^u\ \left|\ x_1\in C_{i_1}^{(2,r)},\cdots, x_u\in C_{i_u}^{(2,r)}, \sum\limits_{j=1}^ux_j\in C_{i_{u+1}}^{(2,r)}\right\}\right..$$ We will prove in Section \[sec-app\] that the value $\Omega_{i_1\cdots i_ui_{u+1}}$ depends only on the number of $0$’s and $1$’s in the sequence $i_1,\ldots,i_{u+1}$. More precisely for any $u+v \ge 1$ we have (see Lemma \[thm-NNN3\] in Section \[sec-app\]) $$\Omega_{{\scriptsize \underbrace{0 \cdots 0}_u \underbrace{1\cdots 1}_v}}=\frac{r-1}{r 2^{u+v+1}}\bigg\{2(r-1)^{u+v-1}+(-1)^{u+v} \left\{(1+\sqrt{r})^u(1-\sqrt{r})^v+(1-\sqrt{r})^u(1+\sqrt{r})^v\right\}\bigg\}.$$ Note that the number of ways to choose a fixed $u \ge 0$ is $\binom{k+1}{u}$. So, for the case that $s=t-k$, $1 \le k \le t$, the value distribution of $\widetilde{T}$ is given as follows $$\left.
\begin{array}{ll}
\mbox{Value} \,\, \widetilde{T} & \hbox{Frequency} \, (\forall u,v \ge 0, u+v=k+1)\\
(t-k)\bar\eta_0+u\eta_0+v \eta_1, & \binom{t+1}{k+1}\binom{k+1}{u} \Omega_{{\scriptsize \underbrace{0 \cdots 0}_u \underbrace{1\cdots 1}_v}} \\
\end{array}
\right.$$ As for the values $\bar\eta_0,\eta_0,\eta_1$, we have $\bar\eta_0=\frac{r-1}{2}$ and from Lemma \[lem-degree2\] $$\begin{aligned}
\left\{
\begin{array}{lll}
\eta_0=\frac{-1-\sqrt{r}}{2}, & \eta_1=\frac{-1+\sqrt{r}}{2}, & \mbox{if $q\equiv 1 \pmod{4}$}, \\
\eta_0=\frac{-1-(-1)^{ms/2}\sqrt{r}}{2}, & \eta_1=\frac{-1+(-1)^{ms/2}\sqrt{r}}{2}, & \mbox{if $q\equiv 3 \pmod{4}$}. \\
\end{array}
\right.\end{aligned}$$ Now we have obtained the value distribution of $\widetilde{T}$. Returning to (\[equ-Tx\]) and (\[equ-Weight of C\]) gives us the weight distribution of the cyclic code $\mathcal{C}_{(a_1,\cdots,a_t)}$, which is summarized in Tables \[Table1\] and \[Table2\] in Theorem \[thm-e1\]. This completes the proof of Theorem \[thm-e1\].
Proof of Theorem \[thm-e2\] and Theorem \[thm-e3\] {#sec-mainII}
==================================================
Proof of Theorem \[thm-e3\] {#sec-main3}
---------------------------
Recall from (\[equ-Ty\]) and (\[eqn-y0\]) that to compute the weight distribution of the cyclic code $\mathcal{C}_{(a_1,\cdots,a_t)}$, it suffices to compute the value distribution of the sum $$\label{equ-Ty2}
\widetilde{T}(y_0,y_1,\cdots,y_{l_0},\gamma z_1,\cdots ,\gamma z_{l_1}):=\sum\limits_{h=0}^{l_0}\bar\eta_{y_h}+
\sum\limits_{h=1}^{l_1}\bar\eta_{\gamma z_h},\quad \forall \, (y_1,\cdots,y_{l_0},z_1,\ldots,z_{l_1})\in \mathbb{F}_{r}^{l_0+l_1},$$ where $l_0,l_1$ are defined by (\[equ-l1l2\]) so that $l_0+l_1=t$ and $y_0:=y_0(y_1,\ldots,y_{l_0},z_1,\ldots,z_{l_1})$ satisfies $$\begin{aligned}
\label{eqn-y02}
y_0+y_1+\cdots+y_{l_0}+ z_1+\cdots+ z_{l_1}=0.\end{aligned}$$ To study the value distribution of $\widetilde{T}$ in (\[equ-Ty2\]), we consider the different subcases according to different $(k_0,k_1)$, where $k_0,k_1$ are defined by $$\begin{array}{l}
k_0:=\#\{i \mid 0\leqslant i\leqslant l_0, y_i\neq 0\}; \\
k_1:=\#\{i \mid 1\leqslant i\leqslant l_1, z_i\neq 0\}.
\end{array}$$
If $k_0+k_1 \le 1$, by (\[eqn-y02\]), all of $y_0,y_1,\ldots,y_{l_0},z_1,\ldots,z_{l_1}$ are 0, the frequency is 1 and $\widetilde{T}=(t+1) \bar\eta_0$.
If $k_0+k_1 \ge 2$, the number of ways to choose exactly $k_0$ non-zero terms in $y_0,\ldots,y_{l_0}$ and exactly $k_1$ non-zero terms in $z_1,\ldots,z_{l_1}$ is $\binom{l_0+1}{k_0}\binom{l_1}{k_1}$. Once they are chosen, without loss of generality we may assume that they are $y_1,\ldots,y_{k_0}$ and $z_1,\ldots,z_{k_1}$. Then in this case we have $$\widetilde{T}=(t+1-k_0-k_1) \bar\eta_0+\sum_{i=1}^{k_0}\bar\eta_{y_i}+
\sum_{i=1}^{k_1}\bar\eta_{\gamma z_i},$$ and the constraint (\[eqn-y02\]) becomes $$y_1+\cdots+y_{k_0}+z_1+\cdots+z_{k_1}=0.$$
In order to compute the value distribution of $\widetilde{T}$ for these cases, let us consider for any $i_1,\ldots,i_{k_0},j_1,\ldots,j_{k_1} \in \{0,1\}$ the value $\Omega_{i_1\cdots i_{k_0};j_{1}\cdots j_{k_1}}^{'}$, given by $$\label{equ-NNN2}
\Omega_{i_1\cdots i_{k_0};j_{1}\cdots j_{k_1}}^{'}:=\#\left\{(y_1,\cdots ,y_{k_0};z_1,\cdots ,z_{k_1})\in (\mathbb{F}_{r}^*)^{k_0+k_1}\ \left|\ {y_{u}\in C_{i_{u}}^{(2,r)}, \gamma z_{v} \in C_{j_v}^{(2,r)}, 1 \le u \le k_0,1 \le v \le k_1}\atop {y_1+\cdots+y_{k_0}+z_1+\cdots+z_{k_1}=0}\right\}\right..$$ For any $ i \in \{0,1\}$, define $\bar{i} \in \{0,1\}$ by $\bar{i} \equiv i+1 \pmod{2}$. Clearly $$\Omega_{i_1\cdots i_{k_0};j_{1}\cdots j_{k_1}}^{'}=\Omega_{i_1\cdots i_{k_0}\bar{j}_{1}\cdots \bar{j}_{k_1}},$$ which is defined in (\[equ-NNN\]) and is evaluated in Section \[sec-app\]. In $\{i_1,\ldots,i_{k_0}\}$, let $u_0$ be the number of $0$’s and $u_1$ be the number of $1$’s; similarly, in $\{j_1,\ldots,j_{k_1}\}$, let $v_0$ be the number of $0$’s and $v_1$ be the number of $1$’s. Given such $u_0,u_1,v_0,v_1$, we have $$\widetilde{T}=(t+1-k_0-k_1) \bar\eta_0+(u_0+v_0) \eta_0+(u_1+v_1)\eta_1,$$ and the frequency is $$\binom{l_0+1}{k_0}\binom{l_1}{k_1}\binom{k_0}{u_0}\binom{k_1}{v_0}\Omega_{{\scriptsize \underbrace{0 \cdots 0}_{u_0+v_1} \underbrace{1\cdots 1}_{u_1+v_0}}}.$$ Now let $k$ and $u$ be fixed such that $k_0+k_1=k$ and $u_0+v_0=u$, where $0 \le u \le k_0+k_1=k$ and $2 \le k \le l_0+l_1+1=t+1$, we conclude that $\widetilde{T}$ takes the value $$\begin{aligned}
\label{4:t} \widetilde{T}=(t+1-k) \bar\eta_0+u \eta_0+(k-u)\eta_1, \end{aligned}$$ and the frequency is $$\begin{aligned}
\label{4:fre} \sum_{k_0=0}^k\sum_{u_0=0}^u\binom{l_0+1}{k_0}\binom{l_1}{k-k_0}
\binom{k_0}{u_0}\binom{k-k_0}{u-u_0}\Omega_{{\scriptsize \underbrace{0 \cdots 0}_{2u_0+k-k_0-u,} \underbrace{1\cdots 1}_{k_0+u-2u_0}}}. \end{aligned}$$ This, after returning to (\[equ-Weight of C\]), provides the weight distribution of the cyclic code $\mathcal{C}_{(a_1,\ldots,a_t)}$ for the general case $N=2,t=e-1\geqslant 2$.
Proof of Theorem \[thm-e2\] {#sec-main2}
---------------------------
From $N=2=\gcd\left((q^m-1)/(q-1),4a\right)$ and $t=e-1=3$, it is easy to see that $q \equiv 1 \pmod{4}$, $m \equiv 2 \pmod{4}$ and $e=4\mid(q^{m/2}-1)$. If $2|a$, the weight distribution has been obtained from (i) of Theorem \[thm-e1\] with $t=3$, this is Table \[Table3\] in Theorem \[thm-e2\]. If $2 \nmid a$, we use Theorem \[thm-e3\] to calculate the weight distribution. In this case $l_0=1,l_1=2$, from (\[4:t\]) and (\[4:fre\]), for any $k,u$ with $2 \le k \le 4,\, 0 \le u \le k$, the sum $\widetilde{T}$ takes the value $$\widetilde{T}=(4-k) \bar\eta_0+u \eta_0+(k-u)\eta_1,$$ with frequency $$\sum_{k_0=0 }^k\sum_{u_0=0}^u\binom{2}{k_0}\binom{2}{k-k_0}
\binom{k_0}{u_0}\binom{k-k_0}{u-u_0}\Omega_{{\scriptsize \underbrace{0 \cdots 0}_{2u_0+k-k_0-u,} \underbrace{1\cdots 1}_{k_0+u-2u_0}}}.$$ Using the values $$\left\{\begin{array}{l}
\Omega_{00}=\Omega_{11}=\frac{r-1}{2}; \quad \Omega_{01}=0;\\
\Omega_{000}=\Omega_{111}=\frac{r-1}{8}(r-5);\\
\Omega_{001}=\Omega_{011}=\frac{(r-1)^2}{8};\\
\Omega_{0000}=\Omega_{1111}=\frac{r-1}{16}(r^2-2r+9);\\
\Omega_{0001}=\Omega_{0111} =\frac{r-1}{16}(r^2-4r+3);\\
\Omega_{0011}=\frac{(r-1)^3}{16},
\end{array}\right.$$ which we can obtain from Lemma \[thm-NNN3\] in Section \[sec-app\], we find that for $k=2$, $$\left.
\begin{array}{ll}
\mbox{Value} & \hbox{Frequency} \\
2\bar\eta_0+2\eta_0, & \Omega_{00}+4\Omega_{01}+\Omega_{11}=r-1 \\[2mm]
2\bar\eta_0+2\eta_1, & \Omega_{00}+4\Omega_{01}+\Omega_{11} =r-1 \\[2mm]
2\bar\eta_0+\eta_0+\eta_{1}, & 4\Omega_{00}+4\Omega_{01}+4\Omega_{11} =4(r-1)\\[2mm]
\end{array}
\right.$$ and for $k=3$, $$\left.
\begin{array}{ll}
\mbox{Value} & \hbox{Frequency} \\
\bar\eta_0+3\eta_0, & 2\Omega_{001}+2\Omega_{011}=\frac{(r-1)^2}{2} \\[2mm]
\bar\eta_0+3\eta_1, & 2\Omega_{001}+2\Omega_{011}=\frac{(r-1)^2}{2}\\[2mm]
\bar\eta_0+2\eta_0+\eta_{1},&4\Omega_{011}+2\Omega_{000}+2\Omega_{111}+4\Omega_{001}\\
&\quad =\frac{(r-1)}{2}(3r-7)\\[2mm]
\bar\eta_0+\eta_0+2\eta_{1},&4\Omega_{011}+2\Omega_{000}+2\Omega_{111}+4\Omega_{001}\\
&\quad =\frac{(r-1)}{2}(3r-7)
\end{array}
\right.$$ and for $k=4$, $$\left.
\begin{array}{ll}
\mbox{Value} & \hbox{Frequency} \\
4\eta_0, & \Omega_{0011}=\frac{(r-1)^3}{16} \\[2mm]
4\eta_1, & \Omega_{0011}=\frac{(r-1)^3}{16} \\[2mm]
3\eta_0+\eta_1, & 2\Omega_{0111}+2\Omega_{0001}=\frac{(r-1)}{4}(r^2-4r+3) \\[2mm]
\eta_0+3\eta_1, & 2\Omega_{0111}+2\Omega_{0001}=\frac{(r-1)}{4}(r^2-4r+3) \\[2mm]
2(\eta_0+\eta_{1}),&\Omega_{1111}+4\Omega_{0011}+\Omega_{0000}=\frac{(r-1)}{8}(3r^2-6r+11).
\end{array}
\right.$$
Now we have obtained the value distribution of $\widetilde{T}$. Returning to (\[equ-Tx\]) and (\[equ-Weight of C\]) gives us the weight distribution of the cyclic code $\mathcal{C}_{(a_1,a_2,a_3)}$ with $2 \nmid a$, which is summarized in Table \[Table4\] in Theorem \[thm-e2\]. This completes the proof of Theorem \[thm-e2\].
Appendix: Calculation of $\Omega_{i_1\cdots i_u i_{u+1}}$ {#sec-app}
=========================================================
Recall that for positive integer $u$ and any sequence $i_1,\cdots,i_{u},i_{u+1}\in\{0,1\}$, the value $\Omega_{i_1\cdots i_ui_{u+1}}$ is defined by $$\Omega_{i_1\cdots i_ui_{u+1}}:=\#\left\{(x_1,\cdots ,x_u)\in (\mathbb{F}_{r}^*)^u\ \left|\ x_1\in C_{i_1}^{(2,r)},\cdots, x_u\in C_{i_u}^{(2,r)}, \sum\limits_{j=1}^ux_j\in C_{i_{u+1}}^{(2,r)}\right\}\right..$$
We first prove that the value of $\Omega_{i_1\cdots i_ti_{u+1}}$ is related to reduced quadratic Jacobi sums which were introduced in Section \[sec-pre\] before.
\[thm-NNN\] The number $\Omega_{i_1\cdots i_ui_{u+1}}$ defined above equals to $$\frac{r-1}{2^{u+1}}\sum\limits_{0 \le v_2,\cdots,v_{u+1} \le 1} (-1)^{\sum\limits_{j=2}^{u+1}(i_1+i_j)v_j} \rho\left((-1)^{\sum\limits_{j=2}^{u}v_j}\right) J^*(\rho^{v_2},\cdots,\rho^{v_{u+1}}),$$ where $\rho$ is the quadratic multiplicative character of $\mathbb{F}_{r}$.
For $x\in \mathbb{F}_{r}^*$, let $\chi$ denote a multiplicative character of $\mathbb{F}_{r}$. It is easy to check that $$\label{equ-orthogonal2}
\frac{1}{2}\sum\limits_{\chi^2=\varepsilon}\chi(x \gamma^i)=
\left\{
\begin{array}{ll}
1, & \hbox{if\ }x\in C_i^{(L,r)}; \\
0, & \hbox{otherwise.}
\end{array}
\right.$$ Suppose $\chi_1,\chi_2,\cdots ,\chi_{u+1}$ denote multiplicative characters of $\mathbb{F}_{r}$. By the relation (\[equ-orthogonal2\]), we have $$\begin{array}{l}
\Omega_{i_1\cdots i_ui_{u+1}}\\
=\sum\limits_{x_1,\cdots x_{u}\in\mathbb{F}_{r}^*}\left[\frac{1}{2}\sum\limits_{\chi_1^2=\varepsilon}\chi_1(x_1 \gamma^{i_1})\right]
\cdots
\left[\frac{1}{2}\sum\limits_{\chi_u^2=\varepsilon}\chi_u(x_u \gamma^{i_u})\right]
\left[\frac{1}{2}\sum\limits_{\chi_{u+1}^2=\varepsilon}\chi_{u+1}(\gamma^{i_{u+1}}\sum\limits_{j=1}^{u}x_j)\right].
\end{array}$$ Expanding the right hand side and changing the order of summation we obtain $$\begin{array}{l}
\frac{1}{2^{u+1}}\sum\limits_{\chi_j^2=\varepsilon\atop j=1,\cdots,u+1}\chi_1(\gamma^{i_1})\cdots\chi_{u}(\gamma^{i_{u}})\chi_{u+1}(\gamma^{i_{u+1}})
\\
\qquad \cdot\sum\limits_{x_1, \ldots,x_u\in \mathbb{F}_{r}^*}\chi_1(x_1)\cdots\chi_{u}(x_u)
\chi_{u+1}(x_1+x_2+\cdots+x_u), \end{array}$$ which gives $$\begin{array}{l}
\frac{1}{2^{u+1}}\sum\limits_{\chi_j^2=\varepsilon\atop j=1,\cdots,u+1}\chi_1(\gamma^{i_1})\chi_2(-\gamma^{i_2})\cdots\chi_{u}(-\gamma^{i_{u}})\chi_{u+1}(\gamma^{i_{u+1}})
\\
\qquad \cdot\sum\limits_{x_1, \ldots,x_u\in \mathbb{F}_{r}^*}\chi_1\chi_2 \cdots \chi_{u+1}(x_1) \chi_2(x_2)\cdots\chi_{u}(x_u)
\chi_{u+1}(1-x_2-\cdots-x_u). \end{array}$$ This is $$\begin{array}{l}
\frac{r-1}{2^{u+1}}\sum\limits_{\chi_j^2=\varepsilon\atop j=2,\cdots,u+1}\chi_2(-\gamma^{i_1+i_2})\cdots\chi_u(-\gamma^{i_1+i_u})
\chi_{u+1}(\gamma^{i_1+i_{u+1}})\\
\qquad \cdot\sum\limits_{x_2,\cdots,x_{u}\in \mathbb{F}_{r}^*}\chi_2(x_2)\cdots\chi_u(x_u)\chi_{u+1}(1-x_1-\cdots-x_u).\end{array}$$ So we obtain $$\begin{array}{l}
\Omega_{i_1\cdots i_ui_{u+1}}=\frac{r-1}{2^{u+1}}\sum\limits_{0 \le v_2,\cdots,v_{u+1} \le 1}(-1)^{\sum\limits_{j=2}^{u}(i_1+i_j)v_j} \rho\left((-1)^{\sum\limits_{j=2}^{u}v_j}\right)J^*(\rho^{v_2},\cdots,\rho^{v_{u+1}}).
\end{array}$$ This completes the proof of Lemma \[thm-NNN\].
\[thm-NNN2\] Suppose that $-1$ is a square in ${{\mathbb F}}_r$, then $$\Omega_{i_1\cdots i_ui_{u+1}}=\frac{r-1}{2^{u+1}}\left\{\frac{1}{r}\bigg((r-1)^u-(-1)^u\bigg)-(-1)^u \sum_{1 \le l \le \frac{u+1}{2} }r^{l-1}\sum_{1 \le j_1 <j_2<\cdots<j_{2l} \le u+1} (-1)^{\sum\limits_{k=1}^{2l}i_{j_k}} \right\} .$$
Since $N=\gcd(\frac{q^m-1}{q-1},ea)=2$ implies $2|m$, then $-1=\gamma^{\frac{q^m-1}{2}}$ is always a square in this paper, from Lemma \[thm-NNN\] we have $$\begin{array}{l}
\Omega_{i_1\cdots i_ui_{u+1}}=\frac{r-1}{2^{u+1}}\sum\limits_{0 \le v_2,\cdots,v_{u+1} \le 1}(-1)^{\sum\limits_{j=2}^{u}(i_1+i_j)v_j} J^*(\rho^{v_2},\cdots,\rho^{v_{u+1}}).
\end{array}$$ Note that $J^*(\rho_1,,\cdots,\rho_u)$ does not depend on the order of the characters $\rho_1,\ldots,\rho_u$, so we have $$\begin{array}{l}
\Omega_{i_1\cdots i_ui_{u+1}}=\frac{r-1}{2^{u+1}}\sum\limits_{I \subset \{2,\ldots,u+1\}}(-1)^{\sum\limits_{j \in I}(i_1+i_j)} J^*(\underbrace{\varepsilon,\ldots,\varepsilon}_{u-\#I},\underbrace{\rho,\cdots,\rho}_{\#I}).
\end{array}$$ Separating the cases that $I =\emptyset$, $\#I>0$ is even and $\#I$ is odd and applying Lemmas \[lem\_J\] and \[lem-J\*\], we can obtain $$\begin{array}{l}
\Omega_{i_1\cdots i_ui_{u+1}}=\frac{r-1}{2^{u+1}} \left\{A+B+C\right\},
\end{array}$$ where $$A=J^*(\underbrace{\varepsilon,\cdots,\varepsilon}_{u})=\frac{1}{r}\bigg((r-1)^u-(-1)^u\bigg),$$ $$B=(-1)^{u+1}\sum\limits_{\substack{\emptyset \ne I \subset \{2,\ldots,u+1\}\\
\#I \mbox{\small \, is even}}}(-1)^{\sum\limits_{j \in I}i_j}r^{(\#I-2)/2},$$ and $$C=(-1)^{u+1}\sum\limits_{\substack{I \subset \{2,\ldots,u+1\}\\
\#I \mbox{\small \, is odd}}}(-1)^{i_1+\sum\limits_{j \in I}i_j}r^{(\#I-1)/2}.$$ Setting $\#I=2l$ if $\#$ is even and $\#I=2l-1$ is $\#I$ is odd completes the proof of Lemma \[thm-NNN2\].
It is easy to see from Lemma \[thm-NNN2\] that the value $\Omega_{i_1\cdots i_ui_{u+1}}$ does not depend on the order of the sequence $i_1,\ldots,i_u,i_{u+1}$. Now we can prove
\[thm-NNN3\] Suppose that $-1$ is a square in ${{\mathbb F}}_r$, then $$\Omega_{{\scriptsize \underbrace{0 \cdots 0}_u \underbrace{1\cdots 1}_v}}=\frac{r-1}{r 2^{u+v+1}}\bigg\{2(r-1)^{u+v-1}+(-1)^{u+v} \left\{(1+\sqrt{r})^u(1-\sqrt{r})^v+(1-\sqrt{r})^u(1+\sqrt{r})^v\right\}\bigg\}.$$
From Lemma \[thm-NNN2\], it suffices to compute $$P=\sum_{1 \le l \le \frac{u+v}{2} }r^{l-1}\sum_{1 \le j_1 <j_2<\cdots<j_{2l} \le u+v} (-1)^{\sum\limits_{k=1}^{2l}i_{j_k}}.$$
Since $i_j=0$ for $1 \le j \le u$ and $i_k=1$ for $u+1 \le j \le u+v$, we have $$\sum_{1 \le j_1 <j_2<\cdots<j_{2l} \le u+v} (-1)^{\sum\limits_{k=1}^{2l}i_{j_k}}= \sum_{s=0}^{2l}\binom{u}{2l-s} \binom{v}{s} (-1)^s,$$ and the right hand side is the coefficient of $x^{2l}$ in the expansion of the polynomial $f(x):=(1+x)^u(1-x)^v$. Hence letting $$f(x)=1+\sum_{n=1}^{u+v}a_nx^n,\quad a_n \in \mathbb{R},$$ then $$P=\frac{1}{r}\sum_{1 \le l \le \frac{u+v}{2}} a_{2l} (\sqrt{r})^{2l}.$$ Clearly the right hand side is $$\frac{1}{r} \left\{\frac{f(\sqrt{r})+f(-\sqrt{r})}{2}-1\right\}.$$ This completes the proof of Lemma \[thm-NNN3\].
Conclusions {#sec-conclusion}
===========
In this paper, we determine the weight distributions of a new family of cyclic codes with arbitrary number of zeros, more precisely the cyclic codes $\mathcal{C}_{(a_1,\cdots,a_t)}$ given by (\[def\]) with any $t \ge 2$ zeros under the conditions that $t=e-1$ and $N=2$. Our main results are as follows:
- For $N=2$, $t=e-1 \geqslant 2$, $2|a$ and $e|(q^{m/2}-1)$, we obtain the weight distribution of $\mathcal{C}_{(a_1,\cdots,a_t)}$.
- For $N=2$ and $t=e-1=3$, we obtain the weight distribution of $\mathcal{C}_{(a_1,\cdots,a_t)}$.
- For the general case of $N=2$ and $t=e-1\geqslant 2$, we present a computational formula to determine the weight distribution of $\mathcal{C}_{(a_1,\cdots,a_t)}$.
Except for these cases (in [@Y-X-D12] and this paper), the weight distribution of the code $\mathcal{C}_{(a_1,\cdots,a_t)}$ is open in most cases when $t < e$. It would be good if some of these open cases can be settled.
Acknowledgments {#acknowledgments .unnumbered}
---------------
Maosheng Xiong’s research is supported by the Hong Kong Research Grants Council under Grant Nos. 609513 and 606211. Jing Yang’s research is partly supported by the National Natural Science Foundation of China (No. 11371011). Lingli Xia’s research is partly supported by Beijing Natural Science Foundation(No. 1144012) and Science and Technology on Information Assurance Laboratory (No. KJ-13-005).
[99]{}
Y. Aubry and P. Langevin, “On the weights of binary irreducible cyclic codes,” in: [*Proceedings of the 2005 international conference on Coding and Cryptography*]{}, LNCS 3969, Springer-Verlag, pp. 46–54, 2006.
L. D. Baumert and R. J. McEliece, “Weights of irreducible cyclic codes,” [*Information and Control,*]{} vol. 20, no. 2, pp. 158–175, 1972.
L. D. Baumert and J. Mykkeltveit, “Weight distributions of some irreducible cyclic codes,” [*DSN Progress Report,*]{} No. 16, pp. 128–131, 1973.
B. C. Berndt, R. J. Evans and K. S. Williams, [*Gauss and Jacobi Sums,*]{} J. Wiley and Sons Company, New York, 1997.
R. Calderbankand, W.M. Kantor,“The geometry of two-weight codes,” [*Bull. Lond. Math. Soc.,*]{} vol. 18, pp. 97–122, 1986
P. Delsarte, “On subfield subcodes of modified Reed-Solomon codes," [*IEEE Trans. Inform. Theory,*]{} vol. 21, no. 5, pp. 575–576, 1975.
C. Ding, Y. Liu, C. Ma and L. Zeng, “The weight distributions of the duals of cyclic codes with two zeros," [*IEEE Trans. Inform. Theory,*]{} vol. 57, no. 12, pp. 8000–8006, 2011.
C. Ding and J. Yang, “Hamming weights in irreducible cyclic codes," [*Discrete Mathematics,*]{} vol. 313, pp. 434–446, 2013.
K. Feng and J. Luo, “Weight distribution of some reducible cyclic codes,” [*Finite Fields Appl.*]{}, vol. 14, no. 2, pp. 390–409, 2008.
T. Feng, “On cyclic codes of length $2^{2^r}-1$ with two zeros whose dual codes have three weights," [*Des. Codes Cryptogr.*]{}, vol. 62, pp. 253–258, 2012.
T. Feng and K. Momihara, “Evaluation of the weight distribution of a class of cyclic codes based on index 2 Gauss sums," [*IEEE Trans. Inf. Theory,*]{} vol. 59, no. 9, pp. 5980–5984, 2013.
R. Fitzgerald and J. Yucas, “Sums of Gauss sums and weights of irreducible codes,” [*Finite Fields Appl.,*]{} vol. 11, no. 1, pp. 89–110, 2005.
H. D. L. Hollmann and Q. Xiang, “On binary cyclic codes with few weights,” in [*Proc. Finite Fields Appl. (Augsburg)*]{}, Berline, Germany, pp. 251–275, 1999.
T. Kl[ø]{}ve, [*Codes for Error Detection,*]{} Singapore: World Scientific, 2007.
S. X. Li, S. H. Hu, T. Feng, and G. Ge, “The weight distribution of a class of cyclic codes related to Hermitian forms graphs,” [*IEEE Trans. Inf. Theory,*]{} vol. 59, no. 5, pp. 3064–3067, 2013.
J. Luo and K. Feng, “On the weight distribution of two classes of cyclic codes,” [*IEEE Trans. Inf. Theory,*]{} vol. 54, no. 12, pp. 5332–5344, 2008.
J. Luo and K. Feng, “Cyclic codes and sequences from generalized Coulter-Matthews function,” [*IEEE Trans. Inf. Theory,*]{} vol. 54, no. 12, pp. 5345–5353, 2008.
J. Luo, Y. Tang, and H. Wang, “Cyclic codes and sequences: The generalized Kasami case,” [*IEEE Trans. Inf. Theory,*]{} vol. 56, no. 12, pp. 2130–2142, 2010.
C. Ma, L. Zeng, Y. Liu, D. Feng and C. Ding, “The weight enumerator of a class of cyclic codes," [*IEEE Trans. Inform. Theory,*]{} vol. 57, no. 1, pp. 397–402, 2011.
R. J. McEliece, “Irreducible cyclic codes and Gauss sums," in: [*Combinatorics: Proc. NATO Advanced Study Inst., Breukelen, 1974*]{}, Part 1: Theory of designs, Finite geometry and coding theory, Math. Centre Tracts, No. 55, Math. Centrum, Amsterdam, pp. 179–196, 1974.
R. J. McEliece and J. H. Rumsey, “Euler products, cyclotomy, and coding,” [*J. Number Theory,*]{} vol. 4, pp. 302–311, 1972.
M. Moisio, “Explicit evaluation of some exponential sums," [*Finite Fields Appl.,*]{} vol. 15, no. 6, pp. 644–651, 2009.
M. Moisio, K. Ranto, M. Rintaaho, and K. Väänänen, “On the weight distribution of the duals of irreducible cyclic codes, cyclic codes with two zeros and hyper-Kloosterman codes,” [*Adv. Appl. Discrete Math.*]{}, vol. 3, pp. 155–164, 2009.
G. Myerson, “Period polynomials and Gauss sums for finite fields,” [*Acta Arith.*]{}, vol. 39, pp. 251–264, 1981.
A. Rao and N. Pinnawala, “A family of two-weight irreducible cyclic codes," [*IEEE Trans. Inform. Theory,*]{} vol. 56, no. 6, pp. 2568–2570, June 2010.
B. Schmidt and C. White, ”All two-weight irreducible cyclic codes?,” [*Finite Fields Appl.,*]{} vol. 8, no. 1, pp. 1–17, 2002.
R. Schroof, “Families of curves and weight distribution of codes," [*Bull. Amer. Math. Soc.,*]{} vol. 32, no. 2, 171–183, 1995.
M. van der Vlugt, “Hasse-Davenport curves, Gauss sums, and weight distributions of irreducible cyclic codes,” [*J. Number Theory,*]{} vol. 55, no. 2, pp. 145–159, 1995.
G. Vega, “Determining the number of one-weight cyclic codes when length and dimension are given," in [*Arithmetic of Finite Fields.*]{} Berlin, Germany: Springer, vol. 4547, Lecture Notes Comput. Sci, pp. 284–293, 2007.
G. Vega and J. Wolfmann, “New classes of 2-weight cyclic codes," [*Des. Codes Cryptogr.,*]{} vol. 42, no. 3, pp. 327–334, 2007.
G. Vega, “The weight distribution of an extended class of reducible cyclic codes," [*IEEE Trans. Inform. Theory,*]{} vol. 58, no. 7, pp. 4862–4869, 2012.
B. Wang, C. Tang, Y. Qi, Y. Yang and M. Xu, “The weight distributions of cyclic codes and elliptic curves," [*IEEE Trans. Inform. Theory,*]{} vol. 58, no. 12, pp. 7253–7259, 2012.
J. Wolfmann, “Weight distributions of some binary primitive cyclic codes,” [*IEEE Trans. Inf. Theory,*]{} vol. 40, no. 6, pp. 2068–2071, 1994.
M. Xiong, “The weight distributions of a class of cyclic codes," [*Finite Fields Appl.*]{}, vol. 18, no. 5, pp. 933–945, 2012.
M. Xiong, “The weight distributions of a class of cyclic codes II,” Des. Codes Cryptogr. DOI 10.1007/s10623-012-9785-0, 2012.
M. Xiong, “The weight distributions of a class of cyclic codes III,” [*Finite Fields Appl.,*]{} vol. 21, 84–96, 2012.
J. Yang, M. Xiong, C. Ding, and J. Luo “Weight Distribution of a Class of Cyclic Codes with Arbitrary Number of Zeros," [*IEEE Trans. Inform. Theory,*]{} vol. 59, no. 9, pp. 5985–5993, 2013..
X. Zeng, L. Hu, W. Jiang, Q. Yue, and X. Cao, “Weight distribution of a $p$-ary cyclic code,” [*Finite Fields Appl.,*]{} vol. 16, no. 1, 56-73, 2010.
[^1]: J. Yang is at the Department of Mathematical Sciences, Tsinghua University, Beijing, 100084, China (email: [email protected]).
[^2]: L. Xia is at the Basic Courses Department of Beijing Union University, Beijing, 100101, China & the Department of Mathematical Sciences, Tsinghua University, Beijing, 100084, China (email: [email protected]).
[^3]: M. Xiong is at the Department of Mathematics, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong (email: [email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Event-triggered and self-triggered control have been proposed in recent years as promising control strategies to reduce communication resources in Networked Control Systems (NCSs). Based on the notion of set-invariance theory, this note presents new self-triggered control strategies for linear discrete-time systems subject to input and state constraints. The proposed schemes not only achieve communication reduction for NCSs, but also ensure both asymptotic stability of the origin and constraint satisfactions. A numerical simulation example validates the effectiveness of the proposed approaches.'
author:
- 'Kazumune Hashimoto, , Shuichi Adachi, and Dimos V. Dimarogonas, [^1] [^2]'
title: 'Aperiodic Sampled-Data Control via Explicit Transmission Mapping: A Set Invariance Approach'
---
Event-triggered and self-triggered control, Constrained control, Set-invariance theory.
Introduction
============
Efficient network utilization and energy-aware communication protocols between sensors, actuators and controllers have been recent challenges in the community of Networked Control Systems (NCSs). To tackle such challenges, event and self-triggered control schemes have been proposed as alternative approaches to the typical time-triggered controllers, see e.g., [@dimos2010a; @heemels2011a; @tabuada2010a]. In contrast to the time-triggered case where the control signals are executed periodically, event and self-triggered strategies trigger the executions based on the violation of prescribed control performances, such as Input-to-State Stability (ISS) [@dimos2010a] and ${\cal L}_\infty$ gain stability [@heemels2011a]. In particular, we are interested in designing self-triggered strategies for *constrained* control systems, where certain constraints such as physical limitations and actuator saturations need to be explicitly taken into account. One of the most popular control schemes to deal with such constraints is Model Predictive Control (MPC) [@Mayne2000a]. In the MPC strategy, the current control action is determined by solving a constrained optimal control problem online, based on the knowledge of current state information and dynamics of the plant. Moreover, applications of the event and self-triggered control to MPC have been recently proposed to reduce the frequency of solving optimal control problems, see e.g., [@evmpc_linear6; @evmpc_linear9; @evmpc_linear10; @hashimoto2015c; @hashimoto2017a; @hashimoto2017c].
The main contribution of this note is to provide novel self-triggered strategies for constrained systems from an alternative perspective to the afore-cited papers, namely, a perspective from *set-invariance theory* [@blanchini1999a]. Set invariance theory has been extensively studied for the past two decades [@blanchini1994a; @bitsoris1988a; @gilbert1991a], and it provides a fundamental tool to design controllers for constrained control systems. Two established concepts are those of a *controlled invariant set* and $\lambda$-*contractive set*. While a controlled invariant set implies that the state stays inside the set for all time, a $\lambda$-contractive set guarantees the more restrictive condition that the state is asymptotically stabilized to the origin. In this note, two different types of set-invariance based self-triggered strategies are presented. In the first approach, we formulate an optimal control problem such that the controller obtains stabilizing control inputs under *multiple candidates* of transmission time intervals. Among the multiple solutions, the controller selects a suitable one such that both control performance and communication load are taken into account. Asymptotic stability of the origin is ensured by using Lyapunov techniques, where the Lyapunov function is induced by a $\lambda$-contractive set obtained offline. Although the first approach guarantees asymptotic stability, it may lead to a high computation load as it requires to solve multiple optimization problems online. Therefore, we secondly propose an alternative strategy that aims to overcome the computational drawback of the first proposal. Similarly to the concept of explicit MPC [@bemporad2002b], we provide an *offline*, explicit mapping that sends the state information to the desired transmission time interval. As we will see in later sections, the state-space is decomposed into a finite number of subsets, to which appropriate transmission time intervals are assigned. The rest of the paper is organized as follows. In Section II, the system description and some preliminaries of invariant set theory are given. In Section III, we propose the first approach of the self-triggered strategy. In Section IV, the second approach of the self-triggerd strategy is presented. In Section V, a illustrative simulation example is given. We finally conclude in Section VI.\
*(Nomenclature)*: Let $\mathbb{R}$, $\mathbb{R}_+$, $\mathbb{N}$, $\mathbb{N}_+$ be the *non-negative reals, positive reals, non-negative* and *positive integers*, respectively. The *interior* of the set ${\cal S} \subset \mathbb{R}^n$ is denoted as ${\rm int} \{ {\cal S} \}$. A set ${\cal S} \subset \mathbb{R}^n$ is called *${\cal C}$-set* if it is compact, convex, and $0 \in {\rm int} \{{\cal S}\}$. For vectors $v_1, \ldots, v_N$, ${\rm co} \{ v_1, \ldots, v_N \}$ denotes their *convex hull*. A set of vectors $\{v_1, \ldots, v_N \}$ whose convex hull gives a set ${\cal P}$ (i.e., ${\cal P} = {\rm co} \{ v_1, \ldots, v_N \}$), and each $v_n$, $n\in \{1, 2, \ldots, N\}$ is not contained in the convex hull of $v_1, \ldots, v_{n-1}, v_{n+1}, \ldots, v_{N}$ is called a set of *vertices* of ${\cal P}$. Given a ${\cal C}$-set ${\cal S} \subset \mathbb{R}^n $, denote by $\partial {\cal S} \subset \mathbb{R}^n$ the boundary of ${\cal S}$. For a given $\lambda \in \mathbb{R}$ and a ${\cal C}$-set ${\cal S}\subset \mathbb{R}^n$, denote $\lambda {\cal S}$ as $\lambda {\cal S} = \{ \lambda x \in \mathbb{R}^n : x \in {\cal S}\}$. Given a set ${\cal S} \subset \mathbb{R}^n$, the function $\Psi_{\cal S} : \mathbb{R}^n \rightarrow \mathbb{R}_+$ with $\Psi_{{\cal S}}(x) = {\rm inf} \{\mu : x \in \mu {\cal S}, \mu \geq 0 \}$ is called a *gauge function*. For given two sets ${\cal S}_1, {\cal S}_2 \subset \mathbb{R}^n$, define ${\cal S}_1 \backslash {\cal S}_2$ as ${\cal S}_1 \backslash {\cal S}_2 = \{ x\in \mathbb{R}^n : x \in {\cal S}_1, x \notin {\cal S}_2 \}$.
Problem formulation and some preliminaries {#strategy_sec}
==========================================
In this section, the system description and some established results of set-invariance theory are provided.
System description and control strategy {#sys_desc_sec}
---------------------------------------
Consider a networked control system illustrated in [Fig.\[network\]]{}. We assume that the dynamics of the plant are given by $$\label{sys}
x ({k+1}) = A x (k) + Bu (k)$$ for $k\in \mathbb{N}$, where $x (k) \in \mathbb{R}^{n}$ is the state and $u (k) \in \mathbb{R}^m$ is the control variable. The state and control input are assumed to be constrained as $x(k) \in {\cal X},\ u(k) \in {\cal U}$, $\forall k \in \mathbb{N}$, where ${\cal X} \subset \mathbb{R}^n, \ {\cal U} \subset \mathbb{R}^m$ are both polyhedral ${\cal C}$-sets described as $$\label{constraint}
\begin{array}{lll}
{\cal X} = \{x\in \mathbb{R}^n: H_x x\leq h_x \}, \\
{\cal U} = \{u \in \mathbb{R}^m :H_u u\leq h_u \},
\end{array}$$ where $H_x \in \mathbb{R}^{n_x \times n}$, $H_u \in \mathbb{R}^{n_u \times m}$ and $h_x$, $h_u$ are appropriately sized vectors having positive components. The control objective is to steer the state to the origin, i.e., $x(k) \rightarrow 0$ as $k\rightarrow \infty$. Let $k_m$, $m\in\mathbb{N}$ with $k_0 = 0$ be the transmission time instants when the plant transmits the state information $x(k_m)$ to the controller and updates the control input. In the self-triggered strategy, the transmission times are determined as $$\label{transmission_times}
k_{m +1} = k_m + \Gamma (x(k_m)), \ \ m \in \mathbb{N},$$ where $\Gamma : {\cal X} \rightarrow \{ 1, 2, \ldots, j_{\max} \}$ denotes a mapping that sends the state information to the corresponding transmission time interval. Here, a maximal transmission time interval $j _{\max} \in \mathbb{N}_+$ is set apriori in order to formulate the self-triggered strategy. Due to the limited nature of communication bandwidth, we assume that only one control sample (not a sequence of control samples) is allowed to be transmitted at each transmission time. Namely, the control input is constant between two consecutive inter-transmission times, i.e., $$\label{controller}
u (k) = \kappa (x(k_m)) \in {\cal U}, \ \ k \in [k_m , k_{m+1} ),$$ where $\kappa : {\cal X} \rightarrow {\cal U}$ denotes the state-feedback control law. The following assumptions are made throughout the paper (see e.g., [@blanchini1994a]):
The pair $(A, B)$ is controllable.
[The matrix $B$ has full column rank.]{}
![Networked Control System[]{data-label="network"}](networksystem-eps-converted-to.pdf){width="7cm"}
Set-invariance theory {#set_invariance_sec}
---------------------
In the following, we define the standard notions of *controlled invariant set* and $\lambda$*-contractive set* [@blanchini1994a], which are important concepts to characterize invariance and convergence properties for constrained control systems.
\[lambda\_contractive\] For a given ${\cal C}$-set ${\cal S} \subseteq {\cal X}$, ${\cal S}$ is said to be a controlled invariant set in ${\cal X}$, if and only if there exists a control law $g(x)\in {\cal U}$ such that $Ax + Bg(x) \in {\cal S}$ for all $x\in {\cal S}$.
\[lambda\_contractive\] For a given ${\cal C}$-set ${\cal S} \subseteq {\cal X}$, ${\cal S}$ is said to be a $\lambda$-contractive set in ${\cal X}$ for $\lambda\in [0, 1] $, if and only if there exists a control law $g(x)\in {\cal U}$ such that $A x + B g(x) \in \lambda {\cal S}$ for all $x\in {\cal S}$.
Roughly speaking, a set ${\cal S}$ is called $\lambda$-contractive set if all states in ${\cal S}$ can be driven into a tighter (or equivalent) region $\lambda {\cal S}$ by applying a one-step control input. From the definition, a controlled invariant set implies a $\lambda$-contractive set with $\lambda = 1$. We review several established results for obtaining a contractive set and the corresponding properties. For given $\lambda \in [0, 1)$ and ${\cal C}$-set ${\cal X}\subset \mathbb{R}^n$, there are several ways to efficiently construct a $\lambda$-contractive set in ${\cal X}$. For a given ${\cal C}$-set ${\cal D} \subset \mathbb{R}^n$, let ${\cal Q}_{\lambda}: \mathbb{R}^n \rightarrow \mathbb{R}^n$ be the mapping $$\label{kstep_contrl}
{\cal Q}_{\lambda} ({\cal D}) = \{ x\in {\cal X} : \exists u \in {\cal U},\ A x + B u \in \lambda {\cal D} \}.$$ A simple algorithm to obtain a $\lambda$-contractive set in ${\cal X}$ is to compute $\Omega_{j} \subset \mathbb{R}^n$, $j \in \mathbb{N}$ as $$\label{iterative_procedure}
\Omega_{0} = {\cal X}, \quad \Omega_{j+1} = {\cal Q} _{\lambda } (\Omega_{j} ) \cap {\cal X},$$ and then it holds that the set ${\cal S} = \lim_{j\rightarrow \infty} \Omega_j $ is $\lambda$-contractive, see e.g., [@blanchini1994a]. If $\Omega_{j+1} =\Omega_j$ for some $j$, the $\lambda$-contractive set is obtained as ${\cal S} = \Omega_j $, which requires only a finite number of iterations. Although such condition does not hold in general, it is still shown, under Assumptions 1 and 2, that the algorithm converges in the sense that for every $\lambda < \bar{\lambda} <1$, there exists a finite $j \in \mathbb{N}_+$ such that the set $\Omega_j $ is $\bar{\lambda}$-contractive (see *Theorem 3.2* in [@blanchini1994a]). Several other algorithms have been recently proposed, see e.g., [@hovd2014a; @hovd2016a] and see also [@darup2017a] for a detailed convergence analysis. The following lemma illustrates the existence of a (non-quadratic) Lyapunov function in a given $\lambda$-contractive set:
\[stability\_lem\][[@blanchini1994a]:]{} Let ${\cal S} \subset {\cal X}$ be a $\lambda$-contractive ${\cal C}$-set with $\lambda \in [0, 1]$ and the associated gauge function $\Psi_{\cal S} : {\cal S} \rightarrow \mathbb{R}_+$. Then, there exists a control law $g : {\cal X} \rightarrow {\cal U}$ such that $$\label{set_induced_lyapunov}
\Psi_{\cal S} ( Ax + B g (x)) \leq \lambda \Psi_{\cal S} ( x),$$ for all $x \in {\cal S}$.
[Lemma\[stability\_lem\]]{} follows immediately from [Definition\[lambda\_contractive\]]{}. If $\lambda < 1$, implies the existence of a stabilizing controller in ${\cal S}$ in the sense that the output of the gauge function $\Psi_{\cal S} (\cdot)$ is guaranteed to decrease. The gauge function $\Psi_{\cal S} (\cdot)$ defined in ${\cal S}$ is known as *set-induced Lyapunov function* in the literature; for a detailed discussion, see e.g., [@blanchini1994a].
Self-triggered strategy
=======================
As described in the introduction, we propose two different types of self-triggered controllers; in this section, the first approach is presented.
Designing a stabilizing controller {#stabilize_control_sec}
----------------------------------
For a given $\lambda \in [0, 1)$, let us first construct a $\lambda$-contractive set ${\cal S}$ in ${\cal X}$. Note that since ${\cal X}$ is a polyhedral ${\cal C}$-set, one can efficiently compute the $\lambda$-contractive set through polyhedral operations according to [^3]. The obtained $\lambda$-contractive set ${\cal S}$ can be denoted as $$\label{setS}
{\cal S} = {\rm co} \{v_1, v_2, \ldots, v_N \} \subseteq {\cal X},$$ where $v_n, n\in \{1, 2, \ldots, N \}$ represent the vertices of ${\cal S}$, and $N$ represents the number of them.
\[initial\_cond\] The initial state is inside ${\cal S}$, i.e., $x(k_0) \in {\cal S}$.
Based on [Assumption\[initial\_cond\]]{}, we will design the self-triggered strategy such that the state remains in ${\cal S}$ for each transmission time instant. Suppose that at a certain transmission time $k_m$, $m \in \mathbb{N}$, the plant transmits the state information $x(k_m)$ to the controller. Based on $x(k_m)$, the controller needs to compute both a suitable controller to be applied and a transmission time interval, such that the state is stabilized to the origin. To this end, we first propose an approach to obtain stabilizing controllers under *multiple candidates* of transmission time intervals. More specifically, we obtain different control actions under different transmission time intervals, and the controller selects a suitable one among them. To obtain the stabilizing controllers, we formulate the following optimal control problem for each $j\in \{1, \ldots, j_{\max} \}$:
\[control\_problem\] For given $x(k_m)$, $j \in \{1, \ldots, j_{\max} \}$ and the $\lambda$-contractive set ${\cal S}$, find $u \in {\cal U}$ and $\varepsilon \in \mathbb{R}$ by solving the following problem:\
$$\label{cost_epsilon}
\underset{u \in {\cal U}} {\min}\ \ \varepsilon,$$
subject to $\varepsilon \in [0, \lambda]$, and
1. $A^{j'} x(k_m) + \sum^{j'} _{i=1} A^{i-1} B u \in {\cal X}$, $\forall j' \in \{ 1, \ldots, j \}$;
2. $A^j x(k_m) + \sum^{j} _{i=1} A^{i-1} B u \in \varepsilon \varepsilon_{x} {\cal S}$;
where $\varepsilon_x = \Psi_{\cal S}(x(k_m))$. [$\Box$]{}
In (C.2), $A^j x(k_m) + \sum^{j} _{i=1} A^{i-1} B u$ represents a state by applying the control input $u \in {\cal U}$ *constantly* for $j$ time steps. Moreover, from the definition of the gauge function $\Psi_{\cal S}(\cdot)$ we have $x(k_m) \in \varepsilon_{x} {\cal S}$. Thus, [Problem\[control\_problem\]]{} aims to find the smallest possible scaled set $\varepsilon \varepsilon_{x} {\cal S}$, such that the state enters $\varepsilon \varepsilon_{x} {\cal S}$ (from $\varepsilon_x {\cal S}$) by applying a $j$-step constant control input. This means that a stabilizing controller is found under the transmission time interval $j$. The constraint in (C.1) implies that the state must remain inside ${\cal X}$ while applying a $j$-step constant controller, which is imposed to guarantee the constraint satisfaction. Note that [Problem\[control\_problem\]]{} is a linear program, since all constraints imposed in (C.1), (C.2), as well as the cost in are all linear.
For given $x(k_m)$ and $j$, let $(u^* _j, \varepsilon^* _j )$ be a pair of optimal solutions obtained by solving [Problem\[control\_problem\]]{}. From (C.2), the state enters $\varepsilon^* _j \varepsilon_x {\cal S}$ if $u^* _j$ is applied constantly for $j$ steps, i.e., $A^j x(k_m) + \sum^{j} _{i=1} A^{i-1} B u^* _j \in \varepsilon^* _j \varepsilon _{x} {\cal S}$, which means that we have $\Psi_{\cal S} (A^j x(k_m) + \sum^{j} _{i=1} A^{i-1} B u^* _j ) \leq \varepsilon^* _j \Psi_{\cal S} (x(k_m))$, or $$\label{lyapunov2}
\begin{aligned}
\Psi_{\cal S} (A^j x(k_m) + & \sum^{j} _{i=1} A^{i-1} B u^* _j ) -\Psi_{\cal S} (x(k_m)) \\
& \leq - (1-\varepsilon^* _j) \Psi_{\cal S} (x(k_m))
\end{aligned}$$ with $0 \leq \varepsilon^* _j \leq \lambda < 1$. Thus, $1-\varepsilon^* _j$ represents how much the output of the gauge function (as a Lyapunov function candidate) decreases by applying the optimal controller $u^* _j$ constantly for $j$ steps. That is, if $1-\varepsilon^* _j$ becomes larger (i.e., $\varepsilon^* _j$ becomes smaller), then the state will be closer to the origin and a better control performance is achieved. Now, consider solving [Problem\[control\_problem\]]{} for all $j \in \{1, \ldots, j_{\max} \}$, which provides different solutions under different transmission time intervals. In the following, let ${\cal J} (x(k_m))$ be the set of indices (transmission time intervals) where [Problem\[control\_problem\]]{} provides a feasible solution. That is, $$\label{calJ}
\begin{aligned}
{\cal J} (x(k_m)) = \{j \in & \{1, \ldots, \ j_{\max}\} : \\
& {\rm {Problem\,\ref{control_problem}}\ is\ feasible\ for\ } j \}.
\end{aligned}$$
\[feasibility\_problem\] If $x(k_m) \in {\cal S}$, there always exists $u \in {\cal U}$ such that $A x(k_m) + B u \in \lambda \varepsilon_x {\cal S} \subseteq {\cal X}$ holds from the properties of the $\lambda$-contractive set (see [@blanchini1994a]). Thus, [Problem\[control\_problem\]]{} has a solution with $j=1$ for any $x(k_m) \in {\cal S}$, and hence, ${\cal J} (x(k_m))$ is non-empty for any $x(k_m) \in {\cal S}$. [$\Box$]{}
An overall algorithm {#transmission_sec}
--------------------
In this subsection an overall self-triggered algorithm is presented. After solving [Problem\[control\_problem\]]{} for all $j \in \{1,\ldots, j_{\max} \}$, which provides the optimal (feasible) sets of solutions $(u^* _j, \varepsilon^* _j )$ for all $j \in {\cal J} (x(k_m))$, the controller selects a suitable transmission time interval among them. The transmission time interval is selected such that both control performance and the communication load are taken into account. A more specific way to achieve this is given in the following overall strategy:\
*[Algorithm 1]{}* (Self-triggered strategy): For any transmission time $k_m$, $m \in \mathbb{N}$, do the following:
1. The plant transmits the current state information $x(k_m)$ to the controller.
2. Based on $x(k_m)$, the controller solves Problem 1 for all $j \in \{1, \ldots, j_{\max} \}$, which provides the optimal (feasible) solutions $(\varepsilon^* _j, u^* _j)$ for all $j \in {\cal J} (x(k_m))$.
3. The controller picks up an optimal index $j_m \in {\cal J} (x(k_m))$ by solving the following problem: $$\label{cost_func}
j_m = \underset{j \in {\cal J}(x(k_m))}{\rm argmax}\ w_1 (1-\varepsilon^* _j)/j + w_2 j ,$$ where $w_1, w_2 \geq 0$ represent given tuning weight parameters. Then, set $k_{m+1} = k_m + j_m$ and $u^* (k_m) = u^* _{j_m}$, and the controller transmits $u^* (k_m)$ and $k_{m+1}$ to the plant.
4. The plant applies $u^* (k_m)$ for all $k \in [k_m, k_{m+1})$. Set $m \leftarrow m+1$, and then go back to step (1). [$\Box$]{}
As shown in Algorithm 1, for each $k_m$ we select the transmission time interval $j_m$ according to . As described in the previous subsection, the term $(1-\varepsilon^* _j)$ represents how much the output of the gauge function decreases by applying the optimal controller $u^* _j$ constantly for $j$ steps. Thus, the first term $(1-\varepsilon^* _j)/j$ represents a *reward* due to the rate of decrease of the gauge function per [one time step]{}, and a better control performance can be achieved when this term becomes larger. On the other hand, from a self-triggered control viewpoint, less control updates will be obtained when control inputs can be applied constantly longer (i.e., when $j$ becomes larger). Thus, the second part in involves $j$ to represent some [reward]{} for alleviating the communication load; as $j$ gets larger, then we obtain less communication load and a larger reward is obtained.
Some remarks are in order regarding Algorithm 1:
The proposed algorithm is related to *move-blocking MPC*[@move_blocking], in the sense that the optimal control inputs are restricted to be constant for some time period. Note that move-blocking MPC aims at reducing the computational complexity by decreasing the degrees of freedom of the optimal control problem[@move_blocking]; the proposed approach, on the other hand, aims at reducing the *communication load* through the move-blocking technique, and the reduction of computation load is not a primary objective here. [$\Box$]{}
In Algorithm 1, the controller solves [Problem\[control\_problem\]]{} for all $j \in \{ 1, \ldots, j_{\max} \}$ for each transmission time instant. While we can potentially achieve longer transmission intervals if $j_{\max}$ is selected larger, the computation load of solving [Problem\[control\_problem\]]{} becomes heavier. Thus, in practical implementation, the user may carefully select a suitable $j_{\max}$ by considering the trade-off between the communication load and the calculation time of solving the optimal control problem. [$\Box$]{}
\[stability1\] Suppose that Assumption 3 holds, and Algorithm 1 is implemented. Then, it holds that $x (k) \rightarrow 0$ as $k \rightarrow \infty$. [$\Box$]{}
We first show that is always feasible (i.e., we can always pick up a transmission time interval according to ), by proving that ${\cal J} (x(k_m))$ is non-empty for all $m\in \mathbb{N}$. By [Assumption\[initial\_cond\]]{} we obtain $x(k_0) \in {\cal S}$ and thus ${\cal J} (x(k_0))$ is non-empty (see [Remark\[feasibility\_problem\]]{}). Since $j_0$ is obtained from , we have $j_0 \in {\cal J} (x(k_0))$ which means that [Problem\[control\_problem\]]{} has a feasible solution for $j = j_0$. Thus, from the constraint (C.2) in [Problem\[control\_problem\]]{}, we obtain $x (k_1) = A^{j_0} x(k_0) + \sum^{j_0} _{i=1} A^{i-1} B u^* (k_0) \in {\cal S}$, which means that ${\cal J} (x(k_1))$ is non-empty. By recursively following this argument, it is shown that $x (k_m) \in {\cal S}$ for all $m\in \mathbb{N}$, which follows that ${\cal J} (x(k_m))$ is non-empty for all $m\in \mathbb{N}$.
Now, it is shown that $x (k) \rightarrow 0$ as $k \rightarrow \infty$. Since $j_m \in {\cal J} (x_m)$, $\forall m\in \mathbb{N}$, it holds from (C.2) in [Problem\[control\_problem\]]{} that: $$\label{gauge_decrease}
\begin{aligned}
\Psi_{\cal S} (x(k_{m+1}) ) & \leq \varepsilon^* _{j_m} \Psi_{\cal S} (x(k_m)) \\
& \leq \lambda \ \Psi_{\cal S} (x(k_m)).
\end{aligned}$$ with $\lambda < 1$. Therefore, by regarding $\Psi_{\cal S} (\cdot)$ as a set-induced Lyapunov function candidate (see [Lemma\[stability\_lem\]]{}), the Lyapunov function is strictly decreasing and the state trajectory is asymptotically stabilized to the origin. This completes the proof.
Although Theorem 1 states only asymptotic stability of the origin, exponential stability can be achieved by imposing an additional constraint when evaluating the reward function in . Specifically, let $j_\ell$ be chosen according to , subject to the constraint $\varepsilon^* _{j} \leq \lambda^{j}$. Indeed, imposing this constraint yields that $\Psi_{\cal S} (x(k_{m+1}) ) \leq \lambda^{j_{m}} \Psi_{\cal S} (x(k_{m}))$, $\forall m \in \mathbb{N}$ (instead of $\Psi_{\cal S} (x(k_{m+1}) ) \leq \lambda \Psi_{\cal S} (x(k_{m}))$ as in ). Thus, we obtain $$\label{exponential_stability}
\Psi_{\cal S} (x(k_{m}) ) \leq \lambda^{j_{m-1}} \Psi_{\cal S} (x(k_{m-1})) \leq \cdots \leq \lambda^{k_m} \Psi_{\cal S} (x(0)),$$ which implies that exponential stability is guaranteed (see e.g., [@blanchini1994a]).
[$\Box$]{}
Self-triggered control via explicit mapping $\Gamma$
====================================================
In the previous section, the self-triggered strategy has been presented by solving [Problem\[control\_problem\]]{} for all $j \in \{ 1, \ldots, j_{\max} \}$. However, solving [Problem\[control\_problem\]]{} for all candidates of transmission time intervals may lead to a high computation load, which may induce computational delays to transmit control samples to the plant. A more preferred approach may be that the transmission mapping $\Gamma: {\cal X} \rightarrow \{1, \ldots, j_{\max}\}$ given in , which sends the state to the desired transmission time interval, is obtained *offline*. That is, with the mapping $\Gamma$ provided explicitly offline, the next transmission time can be directly determined from the (current) state information, without having to solve Problem 1 for all $j \in \{ 1, \ldots, j_{\max} \}$. The approach presented in this section is related to explicit MPC framework [@bemporad2002b], in which an offline characterization of the control strategy (but here, the transmission time intervals) is given via state-space decomposition. A more specific formulation is given below.
Construction of $\Gamma$ via state-space decomposition
------------------------------------------------------
In order to create the explicit mapping of $\Gamma$, we first *decompose* the contractive set ${\cal S}$ into a finite number of disjoint subsets ${\cal S}_1, \ldots, {\cal S}_L \subset {\cal S}$, i.e., $$\label{decomposition_subsets}
{\cal S} = \bigcup^{L} _{\ell =1} {\cal S}_\ell,$$ where it holds that ${\cal S}_\ell \cap {\cal S}_{\ell'} = \emptyset$ for all $(\ell, \ell') \in \{1, \ldots, L\} \times \{1, \ldots, L\}$ $(\ell \neq \ell')$. Based on the decomposition, we will then assign a specific transmission time interval to each ${\cal S}_\ell$, $\ell \in \{1, \ldots, L \}$, so that the controller directly determines the next transmission time. Intuitively, if the state is located far from the origin we would like to assign a short transmission time interval to achieve stability of the origin (or achieve good control performance). In particular, in the case of un-stable systems, applying a constant control signal may lead to a divergence of states, especially if the state is far from the origin. On the other hand, if the state is close to the origin, a small control effort may be sufficient to stabilize the system. That is, assigning a long transmission time interval may be allowable to achieve both stability and communication reduction.
Motivated by the above intuition, we decompose the contractive set as follows. First, for a given $L \in \mathbb{N}_+$, define a set of scalars $\rho_1, \ldots \rho_L \in (0, 1]$, with $$0 < \rho_1 < \rho_2 < \cdots < \rho_{L-1} < \rho_{L} = 1.$$ Then, consider the following sequence of $L$ sets ${\cal S}_1, {\cal S}_2, \ldots , {\cal S}_{L} \subset {\cal S}$: $$\label{set_sequence}
\begin{aligned}
{\cal S}_1 &= \rho_1 {\cal S}, \\
{\cal S}_\ell & = \rho_\ell {\cal S} \backslash {\cal S}_{\ell-1},\ \ \forall \ell \in \{2, \ldots, L\}.
\end{aligned}$$
![Illustration of the decomposed subsets ${\cal S}_\ell$, $\ell \in \{1, \ldots, L\}$ (for the case $L=4$) according to . In the figure, the pentagon represents the contractive set ${\cal S}$ and the subsets ${\cal S}_1, \ldots, {\cal S}_4$ are illustrated with different shades of blue. []{data-label="decomposition"}](decomposition-eps-converted-to.pdf){width="5cm"}
The illustration of the sequence of sets is depicted in [Fig.\[decomposition\]]{}. It can be easily shown that the set sequence defined in yields a [decomposition]{} as in , which satisfies the disjoint property as described above. Now, ${\cal S}$ has been decomposed into a finite number of $L$ sets ${\cal S}_1, \ldots, {\cal S}_L$, to which we next assign suitable transmission time intervals. To this end, we formulate the following optimal control problem for each pair $(\ell, j) \in \{1, \ldots, L \} \times \{1, \ldots, j_{\max}\}$:
\[control\_problem2\] For a given pair $(\ell, j) \in \{1, \ldots, L\}\times \{1, \ldots, j_{\max} \}$, find $u_{1}, \ldots, u_{N} \in {\cal U}$ and $\varepsilon \in \mathbb{R}$ by solving the following problem:\
$$\underset{u_1, \ldots, u_{N} \in {\cal U}} {\min}\ \ \varepsilon,$$
subject to $\varepsilon \in [0, \lambda]$, and
1. For all $n \in \{ 1, \ldots, N \}$, $j' \in \{ 1, \ldots, j \}$, $$\begin{aligned}
\label{state_constraint}
A^{j'} v_{\ell, n} + \sum^{j'} _{i=1} A^{i-1} B u_n \in {\cal X}, \end{aligned}$$ where $v_{\ell, n} = \rho_\ell v_{n}$, $n \in \{ 1, \ldots, N \}$, $j' \in \{ 1, \ldots, j \}$.
2. For all $n \in \{ 1, \ldots, N \}$, $$\label{stability_constraint}
\begin{aligned}
& A^j v_{\ell, n} + \sum^{j} _{i=1} A^{i-1} B u_n \in \varepsilon \rho_\ell {\cal S}.
\end{aligned}$$
[$\Box$]{}
Recall that $v_n$, $n \in \{1, \ldots, N\}$ represent the vertices of ${\cal S}$ (see ). Thus, $v_{\ell, n} = \rho_\ell v_n$, $n \in \{1, \ldots, N\}$ represent the extreme points on the outer boundary of ${\cal S}_\ell$ (see the illustration in [Fig.\[decomposition\]]{}). [Problem\[control\_problem2\]]{} for $(\ell, j)$ aims at finding a set of controllers $u_1, \ldots, u_{N}$ and a scalar $\varepsilon$, such that all the extreme points $v_{\ell, n}$, $n \in \{1, \ldots, N\}$ can be driven into $\varepsilon \rho_\ell {\cal S}$ under the $j$-step constant control inputs. Note that [Problem\[control\_problem2\]]{} is solved [offline]{} for all $\ell \in \{1,\ldots, L\}$, $j\in \{1, \ldots, j_{\max} \}$, since it can be solved by evaluating the extreme points $v_{\ell, n}$, $n \in \{1, \ldots, N\}$ that are given offline. Now, suppose that [Problem\[control\_problem2\]]{} has a solution for $(\ell, j)$, which provides optimal control inputs and a scalar denoted as $u^* _{\ell, j}(1), u^* _{\ell, j}(2), \ldots, u^* _{\ell, j} (N) \in {\cal U}$, $\varepsilon^* _{\ell, j} \in [0, \lambda]$, respectively. The following lemma describes that the feasibility of [Problem\[control\_problem2\]]{} for $(\ell, j)$ implies the existence of a stabilizing controller for all $x \in {\cal S}_{\ell}$:
\[stabilize\_controller\] Suppose that [Problem\[control\_problem2\]]{} finds a solution for $(\ell, j)$. Then, for every $x \in {\cal S}_\ell$, there exists $u \in {\cal U}$ such that: (i) $A^{j'} x + \sum^{j'} _{i=1} A^{i-1} B u \in {\cal X}$ for all $j' \in \{ 1, \ldots, j \}$; (ii) $A^j x + \sum^{j} _{i=1} A^{i-1} B u \in \varepsilon^* _{\ell, j} \varepsilon_x \ {\cal S}$ with $\varepsilon_x = \Psi_{\cal S} (x)$. [$\Box$]{}
Since [Problem\[control\_problem2\]]{} has a solution for $(\ell, j)$, from and we obtain $A^{j'} v_{\ell, n} + \sum^{j'} _{i=1} A^{i-1} B u^* _{\ell, j} (n) \in {\cal X}$ for all $n \in \{ 1, \ldots, N \}$, $j' \in \{ 1, \ldots, j \}$, and $A^j v_{\ell, n} + \sum^{j} _{i=1} A^{i-1} B u^* _{\ell, j} (n) \in \varepsilon^* _{\ell, j} \rho_\ell {\cal S}$ for all $n \in \{1, \ldots, N \}$.
Suppose $x \in {\cal S}_\ell$ and let $\varepsilon_x = \Psi_{{\cal S}} (x) \in [0, 1]$. Since $x \in {\cal S}_\ell \subseteq \rho_\ell {\cal S}$, we have $\varepsilon_x \leq \rho_\ell$. Moreover, since $x \in \varepsilon_x {\cal S}$, there exist $\lambda_n \in [0, 1]$, $n\in \{1, \ldots, N \}$ such that $x = \varepsilon_x \sum^{N} _{n=1} \lambda_n v_{n} = (\varepsilon_x/\rho_\ell) \sum^{N} _{n=1} \lambda_n v_{\ell, n}$, $\sum^{N} _{n=1} \lambda_n = 1$, where we have used $v_{\ell, n} = \rho_\ell v_n$. Let $u \in \mathbb{R}^m$ be given by $$\label{feasible_input}
u = \frac{\varepsilon_x}{\rho_\ell} \sum^{N} _{n=1} \lambda_n u^* _{\ell, j} (n) \in {\cal U}.$$ Then, for all $j' \in \{1, \ldots, j \}$, we obtain $$\begin{aligned}
& A^{j'} x + \sum^{j'} _{i=1} A^{i-1} B u \\
&= \frac{\varepsilon_x}{\rho_\ell} \sum^{N} _{n=1} \lambda_n ( A^{j'} v_{\ell, n} + \sum^{j'} _{i=1} A^{i-1} B u^* _{\ell, j} (n)) \in \frac{\varepsilon_x}{\rho_\ell} {\cal X} \subseteq {\cal X},
\end{aligned}$$ where the first inclusion holds since $A^{j'} v_{\ell, n} + \sum^{j'} _{i=1} A^{i-1} B u^* _{\ell, j} (n) \in {\cal X}$ for all $n \in \{ 1, \ldots, N \}$, $j' \in \{ 1, \ldots, j \}$, and the last inclusion holds since $\varepsilon_x \leq \rho_\ell$. Moreover, we have $$\begin{aligned}
& A^j x + \sum^j _{i=1} A^{i-1} B u \\
&= \frac{\varepsilon_x}{\rho_\ell} \sum^{N} _{n=1} \lambda_n ( A^j v_{\ell, n} + \sum^j _{i=1} A^{i-1} B u^* _{\ell, j} (n)) \\
& \in \frac{\varepsilon_x}{\rho_\ell}\ \rho_\ell \ \varepsilon^* _{\ell, j} {\cal S} =\varepsilon^* _{\ell, j} \varepsilon_x {\cal S},
\end{aligned}$$ where the inclusion holds since $A^j v_{\ell, n} + \sum^{j} _{i=1} A^{i-1} B u^* _{\ell, j} (n) \in \rho_\ell \varepsilon^* _{\ell, j} {\cal S}$ for all $n \in \{1, \ldots, N \}$. Hence, we obtain $A^j x + \sum^{j} _{i=1} A^{i-1} B u \in \varepsilon^* _{\ell, j} \varepsilon_x \ {\cal S}$. This completes the proof.
[Lemma\[stabilize\_controller\]]{} implies that for every $x \in {\cal S}_\ell$ there exists $u\in{\cal U}$ such that $\Psi_{\cal S} (A^j x + \sum^j _{i=1} A^{i-1} B u) \leq \varepsilon^* _{\ell, j} \Psi_{\cal S} (x)$ holds. Thus, this means that if $x \in {\cal S}_\ell$, then there exists a $j$-step stabilizing controller such that the output of the gauge function decreases. As mentioned previously in [Section\[stabilize\_control\_sec\]]{}, $(1-\varepsilon^* _{\ell, j})$ represents the decreasing rate of the gauge function. Thus, we can evaluate the control performance by $\varepsilon^* _{\ell, j}$ similarly to the self-triggered strategy presented in the previous section.
Now, suppose that for each $\ell$ we solve [Problem\[control\_problem2\]]{} for all $j \in \{1, \ldots, j_{\max} \}$. Let ${\cal J}_{\ell}$ be a set of indices (transmission time intervals) where [Problem\[control\_problem2\]]{} has a feasible solution for $\ell$, i.e., $$\label{calJ2}
\begin{aligned}
{\cal J}_{\ell} = \{j \in & \{1, \ldots, \ j_{\max}\} : \\
& {\rm {Problem\,\ref{control_problem2}}\ is\ feasible\ for\ } (\ell, \ j) \}.
\end{aligned}$$ Regarding the feasible set ${\cal J}_\ell$, we obtain the following:
\[feasible\_lem2\] ${\cal J}_{\ell}$ is non-empty for all $\ell \in \{ 1, \ldots, L \}$.
The proof immediately follows from [Definition\[lambda\_contractive\]]{} and is given in the Appendix. By evaluating the feasible solutions obtained above, we now assign to each ${\cal S}_\ell$ a suitable transmission time interval. In Algorithm 1, we presented the self-triggered strategy by determining the transmission interval according to the reward function in . Motivated by this, we similarly now consider the following assignment of the transmission time interval to ${\cal S}_\ell$, by taking both control performance and communication load into account: $$\label{cost_func2}
j^* _{\ell} = \underset{j \in {\cal J}_{\ell} }{\rm argmax}\ w_1 (1-\varepsilon^* _{\ell, j})/j + w_2 j,$$ where $w_1, w_2 \geq 0$ denote the given tuning weights associated to each part of the reward similarly to . Note that in contrast to the previous self-triggered strategy where a suitable transmission time interval is obtained online, is now given in an offline fashion. Suppose that we compute $j^* _\ell$ according to for all $\ell \in \{1, \ldots, L \}$. Then, each $j^* _{\ell}$ is assigned to ${\cal S}_{\ell}$ as the suitable transmission time interval. That is, if $x(k_m) \in {\cal S}_{\ell}$ for a certain transmission time $k_m$, the controller directly sets the next transmission time as $k_{m+1} = k_m + j^* _{\ell}$. Let ${\cal T} : {\cal S} \rightarrow \{1, \ldots, j_{\max} \}$ be a mapping from ${\cal S}_\ell$ to the assigned transmission time interval, i.e., $ j^* _\ell = {\cal T} ({\cal S}_\ell )$. Moreover, let ${\cal R}: {\cal X} \rightarrow {\cal S}$ be a mapping from $x$ to the corresponding subset that $x$ belongs to, i.e., ${\cal R} (x) = {\cal S}_\ell, \ \ {\rm iff} \ x\in {\cal S}_\ell,\ \ell \in \{1, \ldots, L\}. $ Then, the overall transmission mapping $\Gamma: {\cal X} \rightarrow \{1, \ldots, j_{\max} \}$ is given by $$\label{gamma}
\Gamma (x) = ({\cal T} \circ {\cal R}) (x).$$
An overall algorithm {#an-overall-algorithm}
--------------------
Given the explicit transmission mapping obtained in , the overall self-triggered algorithm is now provided below:\
*Algorithm 2 (Self-triggered strategy via explicit mapping $\Gamma$)*: Given the explicit mapping $\Gamma$ obtained by and for any transmission time $k_m$, $m \in \mathbb{N}$, do the following:
1. The plant transmits the current state information $x(k_m)$ to the controller.
2. Based on $x(k_m)$, the controller sets the transmission time interval as $j_m =\Gamma (x(k_m))$. Then, set the next trasmission time as $k_{m+1} = k_m + j_m$.
3. Suppose that $x(k_m) \in {\cal S}_{\ell_m}$ for some $\ell_m \in \{1, \ldots, L \}$. For a given $j_m$ obtained in step 2), the controller sets $u^* (k_m) = (\varepsilon_x / \rho_{\ell_m}) \sum^{N} _{n=1} \lambda_n u^* _{\ell_m, j_m} (n) \in {\cal U}$, where $\varepsilon_x = \Psi_{\cal S}(x(k_m))$ and $u^* _{\ell_m, j_m} (1), \ldots, u^* _{\ell_m, j_m} (N) \in {\cal U}$ are the solution to [Problem\[control\_problem2\]]{} for $(\ell_m, j_m)$. Then, the controller transmits $u^* (k_m)$ and $k_{m+1}$ to the plant.
4. The plant applies $u^* (k_m)$ for all $k \in [k_m, k_{m+1})$. Set $m \leftarrow m+1$ and then go back to step 1). [$\Box$]{}
As shown in Algorithm 2, in contrast to the first approach the controller only needs to compute the control input for a given transmission time interval from the explicit mapping $\Gamma$.
For each transmission time $k_m$, the controller needs to find a suitable subset ${\cal S}_\ell$ such that $x(k_m)\in {\cal S}_\ell$ holds to determine the assigned transmission time interval according to . This problem, which we call *the point location problem*, can be easily solved by using the following property: we have $x(k_m) \in {\cal S}_1 \Leftrightarrow x(k_m) \in \rho_1 {\cal S}$, and for all $\ell \in \{2, \ldots, L\}$, $$x(k_m) \in {\cal S}_\ell\ \Leftrightarrow\ x(k_m) \notin \rho_{\ell-1} {\cal S},\ x(k_m) \in \rho_{\ell} {\cal S}.$$ Hence, the point location problem can be solved by checking if $x(k_m) \in \rho_\ell {\cal S}$, $\ell \in \{1, \ldots, L\}$ sequentially in that order, and takes the first index $\ell$ such that $x (k_m) \in \rho_\ell {\cal S}$ holds.
[$\Box$]{}
Suppose that Assumptions 3 holds, and Algorithm 2 is implemented. Then, it holds that $x (k) \rightarrow 0$ as $k \rightarrow \infty$. [$\Box$]{}
We first show that selecting $j_m$ as $j_m = \Gamma (x (k_m) )$ is always feasible by proving that $x(k_m) \in {\cal S}$ for all $m \in \mathbb{N}$ (if $x(k_m) \notin {\cal S}$, the controller cannot determine $j_m$ since the mapping is not defined). By [Assumption\[initial\_cond\]]{}, we obtain $x(k_0) \in {\cal S}$. To prove by induction, assume $x(k_m) \in {\cal S}$ for some $m\in \mathbb{N}_+$, and we will show $x(k_{m+1}) \in {\cal S}$. Suppose that $x(k_m) \in {\cal S}_{\ell_m} \subseteq {\cal S}$ for some $\ell_m \in \{ 1,\ldots, L\}$, which means from that $j_m = j^* _\ell = \Gamma (x (k_m))$ with $\ell = \ell_m$. Since $j_m = j^* _\ell \in {\cal J}_\ell$ with $\ell = \ell_m$, [Problem\[control\_problem2\]]{} has a solution for the pair $(\ell_m, j_m)$. Let $\varepsilon^* _{\ell_m, j_m} \in [0, \lambda]$ be the optimal $\varepsilon$ as a solution to [Problem\[control\_problem2\]]{} for $(\ell_m, j_m)$. Then, from the proof of [Lemma\[stabilize\_controller\]]{}, setting $u^*(k_m) = (\varepsilon_x /\rho_{\ell_m}) \sum^{N} _{n=1} \lambda_n u^* _{\ell_m, j_m} (n)$ yields that $x(k_{m+1}) = A^{j_m} x(k_m) + \sum^{j_m} _{i=1} A^{i-1} B u^* (k_m) \in \varepsilon_x \varepsilon^* _{\ell_m, j_m} {\cal S} \subseteq {\cal S}$. Therefore, we have $x(k_m) \in {\cal S}$ for all $m\in \mathbb{N}$.
Now, it is shown that $x (k) \rightarrow 0$ as $k \rightarrow \infty$. Since $x(k_m) \in {\cal S}$ for all $m\in \mathbb{N}$, we obtain from [Lemma\[stabilize\_controller\]]{} that $$\begin{aligned}
\Psi_{\cal S} (x(k_{m+1}) ) & \leq \varepsilon^* _{\ell_m, j_m} \Psi_{\cal S} (x(k_m)) \\
& \leq \lambda \ \Psi_{\cal S} (x(k_m)),
\end{aligned}$$ with $\lambda < 1$. Therefore, by considering $\Psi_{\cal S}(\cdot)$ as a set-induced Lyapunov function candidate, the state trajectory is asymptotically stabilized to the origin. This completes the proof.
Comparisons between first and second approach {#discuss_alg2_sec}
---------------------------------------------
In this subsection we discuss both advantages and drawbacks of the second approach (Algorithm 2), by making some comparisons with the first one (Algorithm 1). As stated previously, the second approach is advantageous over the first one in terms of the computation load, since the transmission mapping $\Gamma$ is given offline according to the procedure presented in the previous subsection. Note, however, that in the second approach, each $j^* _\ell$ is computed by solving [Problem\[control\_problem2\]]{} that evaluates *the extreme points* of ${\cal S}_\ell$ (i.e., $v_{\ell, n}$, $n \in \{1, \ldots, N\}$). This means that, while $x(k_m)$ is in the *interior* of ${\cal S}_\ell$, which is not on some extreme point of ${\cal S}_\ell$, there may exist some $j \in {\cal J}_\ell \ (j\neq j^* _\ell)$, such that applying $u^* (k_m) = (\varepsilon_x/\rho_\ell) \sum^{N} _{n=1} \lambda_n u^* _{\ell, j} (n) \in {\cal U}$ [could]{} yield a *larger* reward in than the one obtained with $\varepsilon^* _{\ell, j^* _\ell}$. In this sense, the second approach yields a suboptimal (or conservative) solution compared with Algorithm 1 on the selection of transmission time intervals. This observation is also illustrated in the simulation example, where it is shown that Algorithm 1 achieves less communication load than Algorithm 2 for the case $w_1 = 0$ (for details, see Section V).
Simulation results
==================
In this section we provide an illustrative example to validate our control schemes. The simulation was conducted on Matlab 2016a under Windows 10, Intel(R) Core(TM) 2.40 GHz, 8 GB RAM, using Multi-Parametric Toolbox (MPT3) to compute the $\lambda$-contractive set. We consider a control problem of a *batch reactor system*, which is often utilized as a benchmark in the NCSs community (see, e.g., [@heemels2012a]). The linearized model is given in the continuous-time domain as $\dot{x} (t) = A_c x(t) + B_c u(t)$, where $A_c, B_c$ are given by $${\small
\begin{aligned}
A_c & = \left [
\begin{array}{cccc}
1.380 & -0.208 & 6.715 & -5.676 \\
-0.581 & -4.290 & 0 & 0.675 \\
1.067 & 4.273 & -6.654 & 5.893 \\
0.048 & 4.273 & 1.343 & -2.104 \\
\end{array}
\right ] \\
B_c & = \left [
\begin{array}{cc}
0 & 0 \\
5.679 & 0 \\
1.136 & -3.146 \\
1.136 & 0
\end{array}
\right ].
\end{aligned}
}$$ The system is unstable having unstable poles $1.9911, 0.0633$. We assume ${\cal X} = \{ x\in \mathbb{R}^4 : ||x||_\infty \leq 2 \}$, ${\cal U} = \{ u \in \mathbb{R}^2 : ||u||_{\infty} \leq 5 \}$ and $j_{\max} = 30$. We obtain the corresponding discrete-time system under a zero-order-hold controller with a sampling time interval $0.1$, and the $\lambda$-contractive set ${\cal S}$ is obtained with $\lambda=0.99$ according to the procedure presented in Section II. [Fig.\[state\_trajectory\]]{} illustrates the resulting state trajectories and the corresponding control inputs by implementing Algorithm 1, starting from the initial state $x(k_0) = [1;\ 2;\ 2;\ 0.5]$ and the weights $(w_1, w_2) = (50, 1)$. The figure shows that the resulting state trajectories are asymptotically stabilized to the origin, and control inputs are updated only when necessary.
To analyse the effect of weights, we again simulate Algorithm 1 with $x(k_0) = [1;\ 2;\ 2;\ 0.5]$ under different selection of weights $(w_1, w_2) =(0, 1), (50,1), (100, 1)$. We then compute the convergence time steps when the state enters the small region around the origin (the region satisfying $||x|| \leq 0.001$), and the total number of transmission instances during the time period $k \in [0, 100]$. The results are shown in [Table\[result\_alg1\]]{}. From the table, $(w_1, w_2) = (100, 1)$ achieves the fastest speed of convergence. This is due to the fact that by selecting $w_1$ larger, the reward for the control performance (i.e., the first term in ) is emphasized to be obtained. On the other hand, the number of transmission instances is the smallest for the case $(w_1, w_2) =(0, 1)$, which means that the smallest communication load is obtained. Therefore, it is shown that there exists a trade-off between control performance and communication load, and such trade-off can be regulated by tuning the weights $(w_1, w_2)$.
($w_1, w_2$) (0,1) $(50,1)$ $(100, 1)$
------------------------ ------- ---------- ------------
Convergence (steps) 141 93 69
Transmission instances 5 6 10
: Convergence time and number of transmission instances[]{data-label="result_alg1"}
![Simulation results of state trajectories (upper) and the control inputs (lower) by applying Algorithm 1.[]{data-label="state_trajectory"}](state_input_ver2-eps-converted-to.pdf){width="7.8cm"}
![$j^* _\ell$ as a function of $\ell$ with $(w_1, w_2) = (0, 1), (400, 1), (600,1)$. []{data-label="jmax"}](jmaxmapping4-eps-converted-to.pdf){width="7.8cm"}
To implement the second proposal, we decompose ${\cal S}$ into $L=10$ subsets with $\rho_\ell = 0.1 \ell$, $ \ell \in \{1, \ldots, 10 \}$ and the mapping $\Gamma$ is constructed according to Section IV. The selected transmission time intervals $j^* _\ell$ as a function of $\ell$ are illustrated in [Fig.\[jmax\]]{} under different selections of the weights $(w_1, w_2) = (0, 1), (400, 1), (600,1)$. The figure shows that the transmission time interval tends to be smaller as the weight $w_1$ increases. This means that attaining control performance is emphasized more than attaining communication reduction, if the weight for achieving the control performance $w_1$ is selected larger. In [Fig.\[jmax\]]{}, we also illustrate the selected transmission time intervals for $(w_1, w_2) = ( 0 , 1 )$ (i.e., the blue dashed line). Since $w_1 = 0$, each $j^* _\ell$ corresponds to the *largest* transmission time interval such that [Problem\[control\_problem2\]]{} becomes feasible for $\ell$ (i.e., the maximal index in the feasible set ${\cal J}_\ell$). As shown in the figure, the feasible transmission time interval gets smaller as $\ell$ increases. Intuitively, this is due to that for unstable systems, applying a constant control input leads to a divergence of states, especially if the extreme points to solve [Problem\[control\_problem2\]]{} are located far from the origin. To illustrate the calculation time and the optimality of Algorithm 1 and 2 as discussed in [Section\[discuss\_alg2\_sec\]]{}, we again simulate the two algorithms with $(w_1, w_2) = (0, 1)$ and the initial state $x(k_0) = [1;\ 2;\ 2;\ 0.5]$. As previously described, setting $w_1 = 0$ corresponds to selecting the largest index in ${\cal J}_\ell$. [Table\[result\_tab\]]{} illustrates the total number of transmission instances and the average calculation time to compute the control input for each transmission instance (i.e., the calculation time from step 2) to step 3) in Algorithms 1, 2). Here, both the total number of transmission instances and the average calculation time are computed over the time period $k \in [0, 100 ]$. From the table, Algorithm 1 achieves less communication load than Algorithm 2. As already discussed in [Section\[discuss\_alg2\_sec\]]{}, this is because of the sub-optimality of Algorithm 2; while Algorithm 1 solves [Problem\[control\_problem\]]{} based on the current state information online, [Problem\[control\_problem2\]]{} is solved offline by evaluating the extreme points of the subsets. On the other hand, Algorithm 2 achieves less calculation time than Algorithm 1, as the transmission mapping $\Gamma$ is explicitly given offline.
Algorithm 1 Algorithm 2
------------------------ ------------- -------------
Transmission instances 5 7
Calculation time (sec) 2.98 0.83
: Number of transmission instances and average calculation time with $(w_1, w_2) = (0, 1)$. []{data-label="result_tab"}
Conclusion
==========
In this note, we present two different types of self-triggered strategies based on the notion of set-invariance theory. In the first approach, we formulate an optimal control problem such that suitable transmission time intervals are selected by evaluating both the control performance and the communication load. The second approach aims to overcome the computation drawback of the first one by providing an offline characterization of the mapping $\Gamma$. In this approach, the state space is decomposed into a finite number of subsets, to which suitable transmission time intervals are assigned. Finally, the proposed self-triggered strategies are illustrated through a numerical example of controlling a batch reactor system.
[10]{}
A. Eqtami, D. V. Dimarogonas, and K. J. Kyriakopoulos, “Event-triggered control for discrete time systems,” in *Proceedings of American Control Conference (ACC)*, 2010, pp. 4719–4724.
M. C. F. Donkers and W. P. M. H. Heemels, “Output-based event-triggered control with guaranteed [${\mathcal L}_{\infty}$]{} gain and decentralized event-triggering,” *IEEE Transactions on Automatic Control*, vol. 57, no. 6, pp. 1362–1376, 2011.
A. Anta and P. Tabuada, “To sample or not to sample: Self-triggered control for nonlinear systems,” *IEEE Transactions on Automatic Control*, vol. 55, no. 9, pp. 2030–2042, 2010.
D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, “Constrained model predictive control: Stability and optimality,” *Automatica*, vol. 36, pp. 789–814, 2000.
E. Henriksson, D. E. Quevedo, E. G. W. Peters, H. Sandberg, and K. H. Johansson, “Multiple-loop self-triggered model predictive control for network scheduling and control,” *IEEE Transactions on Control Systems Technology*, vol. 23, no. 6, pp. 2167–2181, 2015.
K. Kobayashi and K. Hiraishi, “Self-triggered model predictive control with delay compensation for networked control systems,” in *Proceedings of the 38th Annual Conference of the IEEE Industrial Electronics Society*, 2012, pp. 3182–3187.
F. D. Brunner, W. P. M. H. Heemels, and F. Allg[ö]{}wer, “Robust self-triggered mpc for constrained linear systems: A tube-based approach,” *Automatica*, vol. 72, pp. 73–83, 2016.
K. Hashimoto, S. Adachi, and D. V. Dimarogonas, “Distributed aperiodic model predictive control for multi-agent systems,” *IET Control Theory and Applications*, vol. 9, no. 1, pp. 11–20, 2015.
——, “Self-triggered model predictive control for nonlinear input-affine dynamical systems via adaptive control samples selection,” *IEEE Transactions on Automatic Control*, vol. 62, no. 1, pp. 177–189, 2017.
——, “Event-triggered intermittent sampling for nonlinear model predictive control,” *Automatica*, vol. 81, pp. 148–155, 2017.
F. Blanchini, “Set invariance in control,” *Automatica*, vol. 35, no. 11, pp. 1747–1767, 1999.
——, “Ultimate boundedness control for uncertain discrete-time systems via set-induced lyapunov functions,” *IEEE Transactions on Automatic Control*, vol. 39, no. 2, pp. 428–433, 1994.
G. Bitsoris, “Positively invariant polyhedral sets of discrete-time linear systems,” *International Journal of Control*, vol. 47, no. 6, pp. 1713–1726, 1988.
E. G. Gilbert and K. T. Tan, “Linear systems with state and control constraints: The theory and application of maximal output admissible sets,” *IEEE Transactions on Automatic Control*, vol. 36, no. 9, pp. 1008–1020, 1991.
, “The explicit linear quadratic regulator for constrained systems,” *Automatica*, vol. 38, no. 1, pp. 3–20, 2002.
M. Hovd, S. Olaru, and G. Bitsoris, “Low complexity constraint control using contractive sets,” in *The IFAC World Congress*, 2014, pp. 2933–2938.
S. Munir, M. Hovd, G. Sandou, and S. Olaru, “Controlled contractive sets for low-complexity constrained control,” in *Proceedings of 2016 IEEE Conference on Computer Aided Control System Design (Part of 2016 IEEE Multi-Conference on Systems and Control)*, 2016, pp. 856–861.
M. S. Darup and M. Cannon, “On the computation of lambda-contractive sets for linear constrained systems,” *IEEE Transactions on Automatic Control*, vol. 62, no. 3, pp. 1498–1504, 2017.
R. Cagienard, P. Grieder, E. C. Kerrigan, and M. Morari, “[Move blocking strategies in receding horizon control]{},” *[Journal of Process Control]{}*, vol. 17, no. 6, pp. 563–570, 2007.
W. P. M. H. Heemels, K. H. Johansson, and P. Tabuada, “An introduction to event-triggered and self-triggered control,” in *Proceedings of the 51st IEEE Conference on Decision and Control (IEEE CDC)*, 2012, pp. 3270–3285.
*(Proof of [Lemma\[feasible\_lem2\]]{}) :* We show that ${\cal J}_\ell$ is non-empty for all $\ell \in \{1, \ldots, L \}$, by proving that [Problem\[control\_problem2\]]{} is feasible for $(\ell, 1)$. Since $v_{\ell, 1}, \cdots, v_{\ell, N} \in \rho_\ell {\cal S}$, it holds that there exist a set of controllers $\tilde{u}_{\ell, 1}, \cdots, \tilde{u}_{\ell, N}\in {\cal U}$, such that $A v_{\ell, n} + B \tilde{u}_{\ell, n} \in \lambda \rho_\ell {\cal S} \subseteq {\cal X}$, $\forall n \in \{1, \ldots, N \}$ from the properties of the $\lambda$-contractive set. Thus, this directly means from , that [Problem\[control\_problem2\]]{} has a feasible solution for $(\ell, 1)$, with $\varepsilon = \lambda$ and $u_{\ell, n} = \tilde{u}_{\ell, n}$, $\forall n \in \{1, \ldots, N \}$. This completes the proof. [$\Box$]{}
[^1]: Kazumune Hashimoto and Shuichi Adachi are with Department of Applied Physics and Physico-Informatics, Keio University, Japan. Hashimoto’s work is supported by Grant-in-Aid for JSPS Research Fellow (DC2).
[^2]: Dimos V. Dimarogonas is with the department of Electrical Engineering, KTH Royal Institute of Technology, Sweden. His work was supported by the Swedish Research Council (VR), the Swedish Foundation for Strategic Research (SSF), ERC Starting Grant BUCOPHSYS, and Knut och Alice Wallenberg foundation (KAW).
[^3]: If the iterative procedure in does not converge in finite time, one can stop the procedure to obtain a $\bar{\lambda}$-contractive set ($\lambda < \bar{\lambda} <1$) in a finite number of iterations. In such case, we use $\bar{\lambda}$ (instead of $\lambda$) as the parameter to design the self-triggered strategies provided throughout the paper.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the problem of designing a data structure that reports the positions of the distinct $\tau$-majorities within any range of an array $A[1,n]$, without storing $A$. A $\tau$-majority in a range $A[i,j]$, for $0<\tau< 1$, is an element that occurs more than $\tau(j-i+1)$ times in $A[i,j]$. We show that $\Omega(n\log(1/\tau))$ bits are necessary for any data structure able just to count the number of distinct $\tau$-majorities in any range. Then, we design a structure using $O(n\log(1/\tau))$ bits that returns one position of each $\tau$-majority of $A[i,j]$ in $O((1/\tau)\log\log_w(1/\tau)\log n)$ time, on a RAM machine with word size $w$ (it can output any further position where each $\tau$-majority occurs in $O(1)$ additional time). Finally, we show how to remove a $\log n$ factor from the time by adding $O(n\log\log n)$ bits of space to the structure.'
address:
- 'Department of Computer Science, University of Chile,Chile.'
- 'Georgia Institute of Technology, USA.'
author:
- Gonzalo Navarro
- 'Sharma V. Thankachan'
bibliography:
- 'paper.bib'
title: Optimal Encodings for Range Majority Queries
---
Introduction
============
Given an array $A[1, n]$ of $n$ arbitrary elements, an *array range query* problem asks us to build a data structure over $A$, such that whenever a range $[i, j]$ with $1 \leq i \leq j \leq n$ arrives as an input, we can efficiently answer queries on the elements in $A[i, j]$ [@skala]. Many array range queries arise naturally as subproblems of combinatorial problems, and are also of direct interest in data mining applications. Well-known examples are range minimum queries (RMQs, which seek the smallest element in $A[i, j]$) [@BV93], top-$k$ queries (which report the $k$ largest elements in $A[i, j]$) [@bro], range selection queries (which report the $k$th largest element in $A[i,j]$) [@CW13], and colored top-$k$ queries (which report the $k$ largest distinct elements in $A[i,j]$) [@NNN].
An [*encoding*]{} for array range queries is a data structure that answers the queries without accessing $A$. This is useful when the values of $A$ are not of interest themselves, and thus $A$ may be deleted, potentially saving a lot of space. It is also useful when array $A$ does not fit in main memory, so it can be kept in secondary storage while a much smaller encoding can be maintained in main memory, speeding up queries. In this setting, instead of reporting an element in $A$, we only report a position in $A$ containing the element. Otherwise, in many cases we would be able to reconstruct $A$ via queries on the encodings, and thus these could not be small (e.g., $A[i]$ would be the only answer to the range query $A[i,i]$ for all the example queries given above). As examples of encodings, RMQs can be solved in constant time using just $2n+o(n)$ bits [@RMQ1] and, using $O(n\log k)$ bits, top-$k$ queries can be solved in $O(k)$ time [@grossi] and range selection queries in $O(\log k / \log\log n)$ time [@NRR14].
Frequency-based array range queries, in particular variants of heavy-hitter-like problems, are very popular in data mining. Queries such as finding the most frequent element in a range (known as the range mode query) are known to be harder than problems like RMQs. For range mode queries, known data structures with constant query time require nearly quadratic space [@R2]. The best known linear-space solution requires $O(\sqrt{n/\log n})$ query time [@stacs], and conditional lower bounds given in that paper show that a significant improvement is highly unlikely.
Still, efficient solutions exist for some useful variants of the range mode problem. An example are approximate range mode queries, where we are required to output an element whose number of occurrences in $A[i, j]$ is at least $1/(1+\epsilon)$ times the number of occurrences of the mode in $A[i,
j]$ [@lbo; @lboo].
In this paper we focus on a popular variant of range mode queries called *range $\tau$-majority queries*, which ask to report any element that occurs more than $\tau(j-i+1)$ times in $A[i,j]$. A version of the problem useful for encodings can be stated as follows (other variants are possible).
Given an array $A[1,n]$, a [*range $\tau$-majority query*]{} receives a range $[i,j]$ and returns one position in the range where each $\tau$-majority in $A[i,j]$ occurs. A [*$\tau$-majority*]{} is any element that occurs more than $\tau(j-i+1)$ times in $A[i,j]$. When $\tau=1/2$ we simply call it a [*majority*]{}.
Range majority queries can be answered in constant time by maintaining a linear space (i.e., $O(n)$-word or $O(n\log n)$-bit) data structure [@Steph]. Similarly, range $\tau$-majority queries can be solved in time $O(1/\tau)$ and linear space if $\tau$ is fixed at construction time, or $O(n\log\log n)$ space (i.e., $O(n\log n\log\log n)$ bits) if $\tau$ is given at query time [@wads].
In this paper, we focus for the first time on [*encodings for range $\tau$-majority queries*]{}. In this scenario, a valid question is how much space is necessary for an encoding that correctly answers such queries (we recall that $A$ itself is not available at query time). We answer that question in Section \[sec:lb\], proving a lower bound for any encoding that solves even a weaker query.
\[thm:lb\] Given a real number $0<\tau<1$, any encoding able to count the number of range $\tau$-majorities in any range $A[i,j]$ must use $\Omega(n\log(1/\tau))$ bits.
Since when using $O(n\log n)$ bits we have sufficient space to store $A[1,n]$[^1] (and achieve the optimal $O(1/\tau)$ time [@wads]), encodings for range $\tau$-majorities are asymptotically interesting only for $\log(1/\tau) = o(\log n)$.
In Section \[sec:nlglgn\] we show how range $\tau$-majority queries can be solved using $O((n/\tau)\log\log n)$ bits of space and $O((1/\tau)\log n)$ query time. In Section \[sec:nbits\] we reduce the space to the optimal $O(n\log(1/\tau))$ bits and slightly increase the time. After spending this time, the structure can report [*any*]{} of the positions of any majority in optimal time (e.g., the leftmost position of each $\tau$-majority in a negligible $O(1/\tau)$ time). In Section \[sec:constr\] we show how to build our structure in $O(n\log n)$ time. All the results hold on the RAM model with word size $w=\Omega(\log n)$ bits.
\[thm:main\] Given a real number $0<\tau<1$, there exists an encoding using the optimal $O(n\log(1/\tau))$ bits that answers range $\tau'$-majority queries, for any $\tau \le \tau' < 1$, in time $O((1/\tau)\log\log_w(1/\tau)\log n)$, where $w=\Omega(\log n)$ is the RAM word size in bits. It can report any $occ$ further occurrence positions of the majorities in $O(occ)$ time. The encoding can be built in $O(n\log n)$ time.
We note that the query time is simply $O((1/\tau)\log n)$ for polylogarithmic values of $1/\tau$. We also note that the time depends on $\tau$, not $\tau'$. In Section \[sec:constr\] we also show how to obtain a query time that is a function of $\tau'$, yet using $O(n\log^2(1/\tau))$ bits of space.
Finally, in Section \[sec:optimal\] we derive a new variant that may use more space but slashes the $\log n$ term from the time complexity.
\[thm:optimal\] Given a real number $0<\tau<1$, there exists an encoding using $O(n\log(1/\tau)+n\log\log n)$ bits that answers range $\tau'$-majority queries, for any $\tau \le \tau' < 1$, in time $O((1/\tau)\log\log_w(1/\tau))$, where $w=\Omega(\log n)$ is the RAM word size in bits. It can report any $occ$ further occurrence positions of the majorities in $O(occ)$ time. The encoding can be built in $O(n\log n)$ time.
By combining the results of Theorems \[thm:main\] and \[thm:optimal\], we obtain the combinations given in Table \[tab:final\].
------------------------------------------------------- ------------------------- ---------------------------------
Condition Space (bits) Query time
\[-1.5ex\]
\[-1.5ex\] $1/\tau = \omega(\operatorname{polylog}n)$ $O(n\log(1/\tau))$ \* $O((1/\tau)\log\log_w(1/\tau))$
\[-1.5ex\] $1/\tau = \Theta(\operatorname{polylog}n)$ $O(n\log(1/\tau))$ \* $O(1/\tau)~*$
\[-1.5ex\] $1/\tau = o(\operatorname{polylog}n)$ $O(n\log(1/\tau))$ \* $O((1/\tau)\log n)$
\[-1.5ex\] $1/\tau = o(\operatorname{polylog}n)$ $O(n\log\log n)$ $O(1/\tau)~*$
------------------------------------------------------- ------------------------- ---------------------------------
: Space-time tradeoffs achieved. We mark the optimal spaces and times with a \*.[]{data-label="tab:final"}
Related Work
============
In this section we first cover the state of the art for answering range $\tau$-majority queries. Then, we survey a few results on bitmap representation, and give a new result that will be useful for this paper. Again, all these results hold on the RAM model with word size $w=\Omega(\log n)$ bits.
Range Majorities
----------------
Range $\tau$-majority queries were introduced by Karpinski and Nekrich [@KarpinskiN08], who presented an $O(n/\tau)$-words structure with $O((1/\tau)(\log\log n)^2)$ query time. Durocher et al. [@Steph] improved their word-space and query time to $O(n\log(1/\tau))$ and $O(1/\tau)$, respectively. Gagie et al. [@spire] presented another trade-off, where the space is $O(n(H+1))$ [*bits*]{} and the query time is $O((1/\tau)\log\log n)$. Here $H\le\lg n$ denotes the empirical entropy of the distribution of elements in $A$ (we use $\lg$ to denote the logarithm in base 2). The best current result in general is by Belazzougui et al. [@wads], where the space is $O(n)$ words and the query time is $O(1/\tau)$. All these results assume that $\tau$ is fixed at construction time.
For the case where $\tau$ is also a part of the query input, data structures of space (in words) $O(n(H+1))$ and $O(n\log n)$ were proposed by Gagie et al. [@spire] and Chan et al. [@swat], respectively. Very recently, Belazzougui et al. [@wads] brought down the space occupancy to $O(n\log\log\sigma)$ words, where $\sigma$ is the number of distinct elements in $A$. The query time is $O(1/\tau)$ in all cases. Belazzougui et al. [@wads] also presented a compressed solution using $nH+o(n\log \sigma)$ bits, with slightly higher query time. All these solutions include a (sometimes compressed) representation of $A$, thus they are not encodings. As far as we know, ours is the first encoding for this problem.
For further reading, we recommend the recent survey by Skala [@skala].
Bitmap Representations {#sec:bitmaps}
----------------------
Given a bitmap $B[1,m]$ with $n$ 1s, the operation $rank(B,i)$ returns the number of 1s in $B[1,i]$, whereas operation $select(B,j)$ gives the position of the $j$th 1 in $B$. Both operations can be solved in constant time by storing $o(m)$ bits in addition to $B$ [@Mun96; @Cla96]. When $n$ is significantly smaller than $m$, another useful representation [@RRR07] compresses $B$ to $n\lg\frac{m}{n}+O(n)+o(m)$ bits and retains constant time for both operations.
When $n$ is much smaller than $m$, even the $o(m)$ extra bits of that compressed representation [@RRR07] are troublesome, and an Elias-Fano-based [@Fan71; @Eli74] compressed representation [@OS07] is useful. It requires $n\lg\frac{m}{n}
+ O(n)$ bits, solves $select$ in $O(1)$ time and $rank$ in $O(\log\frac{m}{n})$ time. The representation considers the positions of all the 1s in $B$, $x_i = select(B,i)$, and encodes the lowest $b =
\lceil \lg\frac{m}{n} \rceil$ bits of each $x_i$ in an array $L[1,n]$, $L[i] = x_i~\mathrm{mod}~2^b$. Then it defines a bitmap $H[1,2n]$ that encodes the highest bits of the $x_i$ values: all the bits at positions $i+(x_i ~\mathrm{div}~ 2^b)$ are set in $H$. Bitmap $H$ is indexed for constant-time $rank$ and $select$ queries [@Mun96; @Cla96]. The space for $L[1,n]$ is $n\lceil \lg\frac{m}{n} \rceil$ and $H$ uses $2n+o(n)$ bits.
Now, $select(B,j) = 2^b (select(H,j)-j) + L[i]$ can be computed in constant time. For $rank(B,i)$, we observe that the $h$th 0 in $H$ represents the point where the position $B[2^b h]$ is reached in the process of setting the 1s at positions $i+(x_i ~\mathrm{div}~ 2^b)$, that is, $x_{i-1} < 2^b h \le x_i$. The number of 1s in $H$ up to that position is $rank(B,2^b h)$. Therefore, if we write $i = 2^b h + l$, then $rank(B,i)$ is between $j_1 = rank(H,select_0(H,h))+1$ and $j_2 =
rank(H,select_0(H,h+1))$, where $select_0(H,h)$ gives the position of the $h$th 0 in $H$ and is also computed in constant time and $o(n)$ bits [@Mun96; @Cla96]. Now we binary search for $l$ in $L[j_1,j_2]$, which is increasing in that range. The range is of length at most $2^b$, so the search takes $O(b)=O(\log\frac{m}{n})$ time. The final position $j$ returned by the search is $rank(B,i)$.
The time can be improved to $O(\log\log_w\frac{m}{n}+\log s)$ on a RAM machine of $w$ bits by sampling, for each increasing interval of $L$ of length more than $s$, one value out of $s$. Predecessor data structures are built on the samples of each interval, taking at most $O((n/s)\log\frac{m}{n})$ bits. Then we first run a predecessor query on $L[j_1,j_2]$, which takes time $O(\log\log_w\frac{m}{n})$ [@PT08], and finish with an $O(\log s)$-time binary search between the resulting samples.
\[lem:bitmap\] A bitmap $B[1,m]$ with $n$ 1s can be stored in $n\log\frac{m}{n} +
O((n/s)\log\frac{m}{n} + n)$ bits, so that $select$ queries take $O(1)$ time and $rank$ queries take $O(\log\log_w\frac{m}{n}+\log s)$, for any $s$, on a RAM machine of $w$ bits.
Lower Bounds {#sec:lb}
============
We derive a lower bound on the minimum size range $\tau$-majority encodings may have, even if we just ask them to count the number of distinct $\tau$-majorities present in any range. The idea is to show that we can encode a certain combinatorial object in the array $A$, so that the object can be recovered via range $\tau$-majority queries. Therefore, in the worst case, the number of bits needed to solve such queries must be at least the logarithm of the number of distinct combinatorial objects that can be encoded.
Consider a sequence of $m$ permutations on $[3k]$. There are $(3k)!^m$ such sequences, thus any encoding for them must use at least $m\lg ((3k)!)$ bits in the worst case. Now consider the following encoding. Array $A$ will have length $n=36\cdot k\cdot m$. To encode the $i$th permutation, $\pi_i =
(x_1~x_2~\ldots~x_{3k})$, we will write 9 chunks on $A[36k(i-1)+1,36ki]$:
$$\begin{aligned}
1,2,3,\ldots,k, & {-}1,{-}2,{-}3,\ldots,{-}2k, & x_1,x_2,x_3,\ldots,x_k \\
k{+}1,k{+}2,k{+}3,\ldots,2k,& {-}1,{-}2,{-}3,\ldots,{-}2k,& x_1,x_2,x_3,\ldots,x_k \\
2k{+}1,2k{+}2,2k{+}3,\ldots,3k,& {-}1,{-}2,{-}3,\ldots,{-}2k,& x_1,x_2,x_3,\ldots,x_k \\
1,2,3,\ldots,k,& {-}1,{-}2,{-}3,\ldots,{-}2k,& x_{k{+}1},x_{k{+}2},x_{k{+}3},\ldots,x_{2k} \\
k{+}1,k{+}2,k{+}3,\ldots,2k,& {-}1,{-}2,{-}3,\ldots,{-}2k,& x_{k{+}1},x_{k{+}2},x_{k{+}3},\ldots,x_{2k} \\
2k{+}1,2k{+}2,2k{+}3, \ldots,3k,& {-}1,{-}2,{-}3,\ldots,{-}2k,& x_{k{+}1},x_{k{+}2},x_{k{+}3},\ldots,x_{2k} \\
1,2,3,\ldots,k,& {-}1,{-}2,{-}3,\ldots,{-}2k,& x_{2k{+}1},x_{2k{+}2},x_{2k{+}3},\ldots,x_{3k} \\
k{+}1,k{+}2,k{+}3,\ldots,2k,& {-}1,{-}2,{-}3,\ldots,{-}2k,& x_{2k{+}1},x_{2k{+}2},x_{2k{+}3},\ldots,x_{3k} \\
2k{+}1,2k{+}2,2k{+}3,\ldots,3k,& {-}1,{-}2,{-}3,\ldots,{-}2k,& x_{2k{+}1},x_{2k{+}2},x_{2k{+}3},\ldots,x_{3k} \\\end{aligned}$$
We will set $\tau = 1/(2k+2)$ and perform $\tau$-majority queries on parts of $A$ to recover any permutation.
Let us start obtaining $\pi_i(1)=x_1$. Let $C[1,36] = A[36k(i-1)+1,36ki]$. Consider an interval of the form $$C[\ell,3k+g] = \ell,\ell+1,\ldots,k,
{-}1,{-2},\ldots,{-}2k,x_1,x_2,\ldots,x_g,$$ for $1 \le \ell,g \le k$. Note that $x_1,\ldots,x_g$ are the only values that may appear twice in $C[\ell,3k+g]$, precisely, if they belong to $\{\ell,\ldots,k\}$. Note that elements appearing once in $C[\ell,3k+g]$ are not $\tau$-majorities, since $1 \le \tau(3k+g-\ell+1)$ for any values $k,\ell,g$. On the other hand, if an element appears twice in $C[\ell,3k+g]$, then it is a $\tau$-majority, since $2 > \tau(3k+g-\ell+1)$ for any values $k,\ell,g$.
With this tool, we can discover $x_1$ as follows. First, we ask whether there is a $\tau$-majority in $C[1,3k+1]$. If there is none, then $x_1 \not\in
\{1,\ldots,k\}$, and we have to look for it elsewhere (in $C[4k+1,8k]$ or $C[8k+1,12k]$). Assume there is a $\tau$-majority in $C[1,3k+1]$; then $x_1 \in \{1,\ldots,k\}$. Now we query the range $C[2,3k+1]$. If there is no $\tau$-majority, then $x \not\in \{2,\ldots,k\}$, and we conclude that $x_1=1$. If there is, then $x \in \{2,\ldots,k\}$ and we query the range $C[3,3k+1]$. If there is no $\tau$-majority, then $x \not\in \{3,\ldots,k\}$ and we conclude that $x_1=2$, and so on. The process is continued, if necessary, until querying the range $C[k,3k+1]$. If, instead, we had originally found out that $x \not\in \{1,\ldots,k\}$, then we look for it analogously in $C[4k+1,8k]$ or $C[8k+1,12k]$.
To look for $x_2$, we consider similarly ranges of the form $C[\ell,3k+2]$, with identical reasoning. This time, it is possible that element $x_1$ is also counted as an answer, but since we already know the value of $x_1$, we simply subtract 1 from the count in any range $C[\ell,3k+2]$ with $\ell \le x_1$. This process continues analogously until we identify $x_k$. The other two thirds of $\pi_i$ are extracted analogously from $C[12k+1,24k]$ and $C[24k+1,36k]$.
Consider encoding $m=1$ permutation $\pi = (1~5~3~9~2~4~6~8~7)$, of size $3k=9$. Then we set $\tau=1/8$ and the array $A[1,108]$ is as follows: $$\begin{aligned}
1,2,3, & {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 1,5,3 \\
4,5,6,& {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 1,5,3 \\
7,8,9,& {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 1,5,3 \\
1,2,3,& {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 9,2,4 \\
4,5,6,& {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 9,2,4 \\
7,8,9,& {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 9,2,4 \\
1,2,3,& {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 6,8,7 \\
4,5,6,& {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 6,8,7 \\
7,8,9,& {-}1,{-}2,{-}3,{-4},{-}5,{-}6,& 6,8,7 \\\end{aligned}$$
Now we will find $x_1$ (which is 1, but we do not know it yet). Since $A[1,10]$ has a $\tau$-majority, we know that $x_1 \in \{ 1,2,3 \}$. Since $A[2,10]$ has no $\tau$-majority, we know that $x_2 \not\in \{ 2,3\}$, thus we learn $x_1=1$.
Now let us find $x_2$. Since $A[1,11]$ has one $\tau$-majority, which we know corresponds to $x_1=1$, we conclude that $x_2 \not\in \{ 1,2,3 \}$. Thus we will have to find it analogously in $A[13,24]$ or in $A[25,36]$.
Now let us find $x_3$. Since $A[1,12]$ has two $\tau$-majorities, one of which we know corresponds to $x_1=1$, and the other we know does not correspond to $x_2$, we conclude that $x_3 \in \{ 1,2,3\}$. Now $A[2,12]$ has one $\tau$-majority. We know it does not correspond to $x_1=1$ (as it falls outside the range) nor to $x_2$ (as it is not in this subset). Then it follows that $x_3 \in \{ 2,3\}$. Finally, since $A[3,12]$ still has one $\tau$-majority, we conclude $x_3 \in \{ 3 \}$, thus $x_3=3$.
Element $x_2$ will be found in $A[13,24]$. Elements $x_4,x_5,x_6$ will be obtained from $A[37,72]$ and elements $x_7,x_8,x_9$ from $A[73,108]$.
Now, since $n=36km$ and $\tau=1/(2k+2)$, we have that any encoding able to answer the above queries requires at least $$m\lg \left((3k)!\right) > m\left(3k \lg (3k) - 3k \lg e + 1\right) >
\frac{n}{12}\left(\lg\left(\frac{3}{2}\frac{1}{\tau}-3\right) - \lg e\right)$$ bits[^2]. This is $\Omega(n\log(1/\tau))$ unless $1/\tau \le 2+\frac{2}{3}e = O(1)$, thus it is sufficient that we show that $\Omega(n)$ is a lower bound for any constant $\tau \ge 1/(2+\frac{2}{3}e) > 1/4$.
To show that $\Omega(n)$ bits are necessary for any $\tau \ge 1/4$, consider encoding a bitmap $B[1,m]$ in an array $A[1,4m]$ so that, if $B[i]=0$, then $A[4(i-1)+1]=1$, $A[4(i-1)+2]=2$, $A[4(i-1)i+3]=3$, and $A[4i]=4$. Instead, if $B[i]=1$, then $A[4(i-1)+1,4i]=1$. Then, for any $\tau \ge 1/4$, there is a $\tau$-majority in $A[4(i-1)+1,4i]$ iff $B[i]=1$. As there are $2^m$ possible bitmaps $B$ and our array is of length $n=4m$, we need at least $m = n/4 = \Omega(n)$ bits for any encoding. Then the proof of Theorem \[thm:lb\] is complete.
An $O((n/\tau)\log\log n)$ Bits Encoding for Range $\tau$-Majorities {#sec:nlglgn}
====================================================================
In this section we obtain an encoding using $O((n/\tau)\log\log n)$ bits and solving $\tau$-majority queries in $O((1/\tau)\log n)$ time. In the next section we improve the space usage. We assume that $\tau$ is fixed at construction time. At query time, we will be able to solve any $\tau'$-majority query for any $\tau \le \tau' < 1$.
The Basic Idea
--------------
Consider each distinct symbol $x$ appearing in $A[1,n]$. Now consider the set of all the segments $S_x$ within $[1,n]$ where $x$ is a $\tau$-majority (this includes, in particular, all the segments $[k,k]$ where $A[k]=x$). Segments in $S_x$ may overlap each other. Now let $A_x[1,n]$ be a bitmap such that $A_x[k]=1$ iff position $k$ belongs to some segment in $S_x$. We define a second bitmap related to $x$, $M_x$, so that if $A_x[k]=1$, then $M_x[rank(A_x,k)]=1$ iff $A[k]=x$, where operation $rank$ was defined in Section \[sec:bitmaps\].
Let our running example array be $A[1,7] = \langle 1~3~2~3~3~1~1\rangle$, and $\tau=1/2$. Then we have the segments $S_x$: $$\begin{aligned}
S_1 &=& \{ [1,1], [6,6], [7,7], [6,7], [5,7] \},\\
S_2 &=& \{ [3,3] \},\\
S_3 &=& \{ [2,2], [4,4], [5,5], [4,5], [2,4], [3,5], [4,6], [2,5], [1,5],
[2,6] \},\end{aligned}$$ and the corresponding bitmaps $A_x$: $$A_1 ~=~ \langle 1~0~0~0~1~1~1 \rangle, ~~~
A_2 ~=~ \langle 0~0~1~0~0~0~0 \rangle, ~~~
A_3 ~=~ \langle 1~1~1~1~1~1~0 \rangle.$$ Finally, the corresponding bitmaps $M_x$ are: $$M_1 ~=~ \langle 1~0~1~1 \rangle, ~~~
M_2 ~=~ \langle 1 \rangle, ~~~
M_3 ~=~ \langle 0~1~0~1~1~0 \rangle.$$
Then, the following result is not difficult to prove.
\[lem:Ax\] An element $x$ is a $\tau'$-majority in $A[i,j]$ iff $A_x[k]=1$ for all $i\le k\le j$, and 1 is a $\tau'$-majority in $M_x[rank(A_x,i),rank(A_x,j)]$.
If $x$ is a $\tau'$-majority in $A[i,j]$, then it is also a $\tau$-majority. Thus, by definition, $[i,j] \in S_x$, and therefore all the positions $k \in [i,j]$ are set to 1 in $A_x$. Therefore, the whole segment $A_x[i,j]$ is mapped bijectively to $M_x[rank(A_x,i),rank(A_x,j)]$, which is of the same length. Finally, the number of occurrences of $x$ in $A[i,j]$ is the number of occurrences of 1 in $M_x[rank(A_x,i),$ $rank(A_x,j)]$, which establishes the result.
Conversely, if $A_x[k]=1$ for all $i\le k\le j$, then $A[i,j]$ is bijectively mapped to $M_x[rank(A_x,i),rank(A_x,j)]$, and the 1s in this range correspond one to one with occurrences of $x$ in $A[i,j]$. Therefore, if 1 is a $\tau'$-majority in $M_x[rank(A_x,i),$ $rank(A_x,j)]$, then $x$ is a $\tau'$-majority in $A[i,j]$.
Value 1 is a majority in $A[5,7]$, and it holds that $A_1[5,7] = \langle
1~1~1 \rangle$ and $M_1[rank(A_1,5),rank(A_1,7)] = M_1[2,4] = \langle 0~1~1
\rangle$, where 1 is a majority.
Thus, with $A_x$ and $M_x$ we can determine whether $x$ is a majority in a range.
\[lem:algox\] It is sufficient to have $rank$-enabled bitmaps $A_x$ and $M_x$ to determine, in constant time, whether $x$ is a $\tau'$-majority in any $A[i,j]$.
We use Lemma \[lem:Ax\]. We compute $i'=rank(A_x,i)$ and $j'=rank(A_x,j)$. If $j'-i' \not= j-i$, then $A_x[k]=0$ for some $i \le k \le j$ and thus $x$ is not a $\tau$-majority in $A[i,j]$, hence it is also not a $\tau'$-majority. Otherwise, we find out whether 1 is a $\tau'$-majority in $M_x[i',j']$, by checking whether $rank(M_x,j')-rank(M_x,i'-1)>\tau'(j'-i'+1)$.
To find any position $i \le k \le j$ where $A[k]=x$, we need the operation $select(B,j)$, defined in Section \[sec:bitmaps\]. Then, for example, if $x$ is a $\tau'$-majority in $A[i,j]$, its leftmost occurrence in $A[i,j]$ is $i-i'+select(M_x,rank(M_x,i'-1)+1)$. In general, for any $1 \le t \le rank(M_x,j')-rank(M_x,i'-1)$, we can retrieve the $t$th occurrence with $i-i'+select(M_x,rank(M_x,i'-1)+t)$.
Coalescing the Bitmaps
----------------------
We cannot afford to store (and probe!) all the bitmaps $A_x$ and $M_x$ for all $x$, however. The next lemma is the first step to reduce the total space to slightly superlinear.
\[lem:five1s\] For any position $A[k]=x$ there are at most $2\lceil 1/\tau\rceil$ 1s in $A_x$.
Consider a process where we start with $A[k]=\,\perp$ for all $k$, and set the values $A[k]=x$ progressively. We will distinguish three kinds of changes.
#### [(1) New segments around $A[k]$ are created in $S_x$]{}
Setting $A[k]=x$ creates in $S_x$ all the segments of the form $[k-k_l,k+k_r]$ for $1>\tau(k_r+k_l+1)$, or $k_l+k_r < 1/\tau-1$. Their union is the area $A_x[k-\lceil 1/\tau\rceil+2,\ldots, k+\lceil 1/\tau\rceil-2]=1$, which may increase the number of 1s in $A_x$ by up to $2\lceil 1/\tau\rceil-3$.
#### [(2) Segments already covering $A[k]$ are extended]{}
Any maximal segment $[l,r] \in S_x$ covering $A_x[k]$ contains $c > \tau(r-l+1)$ occurrences of $x$, but it holds that $c \le \tau(r-l+2)$, otherwise there would also exist segments $[l-1,r]$ and $[l,r+1]$ in $S_x$, and $[l,r]$ would not be maximal. Therefore, adding one more occurrence, $A[k]=1$, we get $c+1 \le \tau(r-l+2+1/\tau)$ occurrences in $[l,r]$. Now it holds that $x$ may be a $\tau$-majority in segments $[l-k_l,r+k_r]$ for all $0 \le k_l+k_r <
1+1/\tau$ (i.e., where $c+1 > \tau(r-l+1+k_l+k_r)$, using only that $c+1 \le \tau(r-l+2+1/\tau)$), and therefore we can extend $[l,r]$ to the left by up to $\lceil 1/\tau\rceil$, or to the right by up to $\lceil 1/\tau\rceil$.
#### [(3) Segments reaching close to $A[k]$ are extended]{}
The same reasoning as for the previous case applies, even if $[l,r]$ does not originally contain position $k$. There are more restrictions, since now $[l-k_l,r+k_r]$ must be so that it contains $k$, and the same limit $0 \le k_l+k_r < 1+1/\tau$ applies. Thus, in addition to being possible to extend them by at most $\lceil 1/\tau\rceil$ cells in either direction, position $k$ must lie within the extended area.
#### Total extension
The three cases above are superimposed. Let $\ell_l$ and $\ell_r$ the closest positions $\ell_l \le k \le \ell_r$ where $A_x[\ell_l] = A_x[\ell_r] = 1$. Then, if $\ell_l=k$, we can set at most $\lceil 1/\tau\rceil$ new 1s in $A_x$ to the left of $k$ by extending segments using case (2). Otherwise, if $k-\ell_l \le \lceil 1/\tau\rceil$, we can cover the area $A_x[\ell_l+1,\ldots,k]$ and add up to $\lceil 1/\tau\rceil-(k-\ell_l)$ further cells to the left, using case (3). Otherwise, if $k-\ell_l > \lceil
1/\tau\rceil$, we set $\lceil 1/\tau\rceil-2$ cells to the left, apart from $k$, using case (1). The same reasoning applies to the right, and therefore $2 \lceil
1/\tau\rceil$ is an upper bound to the number of 1s in $A_x$ produced by each new occurrence of $x$ in $A$.
The lemma shows that all the $A_x$ bitmaps add up to $O(n/\tau)$ 1s, and thus the lengths of all the $M_x$ bitmaps add up to $O(n/\tau)$ as well (recall that $M_x$ has one position per 1 in $A_x$). Therefore, we can store all the $M_x$ bitmaps within $O(n/\tau)$ bits of space. We cannot, however, store all the $A_x$ bitmaps, as they may add up to $O(n^2)$ 0s (note there can be $O(n)$ distinct symbols $x$), and we still cannot probe all the $A_x$ bitmaps for all $x$ in $o(n)$ time.
Instead, we will [*coalesce*]{} all the bitmaps $A_x$ into a smaller number of bitmaps $A'_r$ (which will be called coalesced bitmaps). Coalescing works as follows. Let us write $A[i,j]=b$ to mean $A[\ell]=b$ for all $i \le \ell \le j$. We start with all $A'_r[1,n]=0$ for all $r$. Then we take each maximal area of all 1s of each bitmap, $A_x[i,j]=1$, choose some $r$ such that $A'_r[i-1,j+1]=0$, and set $A'_r[i,j]=1$. That is, we copy the run of 1s from $A_x$ to some coalesced bitmap $A'_r$ such that the run does not overlap nor touch other previous runs already copied (i.e., there must be at least one 0 between any two copied runs of 1s). We associate to each such $A'_r$ a bitmap $M'_r$ where the areas of each $M_x$ corresponding to each coalesced area of $A_x$ are concatenated, in the same order of the coalesced areas. That is, if $A'_r[i_t,j_t]=1$, the $t$th left-to-right run of 1s in $A'_r$, was copied from $A_x$, then $M_x[rank(A_x,i_t),rank(A_x,j_t)]$ will be the $t$th segment appended to $M'_r$.
We can coalesce the whole bitmaps $A_1$ and $A_2$ into $A' = \langle
1~0~1~0~1~1~1 \rangle$, with the corresponding bitmap $M' = \langle 1~1~0~1~1
\rangle$.
The coalesced bitmaps $A'_r$ and $M'_r$ will replace the original bitmaps $A_x$ and $M_x$. At query time, we check for the area $[i,j]$ of each coalesced bitmap using Lemma \[lem:algox\]. We cannot confuse the areas of different symbols $x$ because we force that there is at least one 0 between any two areas. We cannot report the same $\tau'$-majority $x$ in more than one coalesced bitmap, as both areas should overlap on $[i,j]$ and then they would have been merged as a single area in $A_x$. If we find one $\tau'$-majority in one coalesced bitmap, we know that there is a $\tau'$-majority $x$ and can spot all of its occurrences (or the leftmost, if desired) in optimal time, even if we cannot know the identity of $x$. Moreover, we will find all the distinct $\tau'$-majorities in this way.
Bounding the Number of Coalesced Bitmaps
----------------------------------------
This scheme will work well if we obtain just a few coalesced bitmaps overall. Next we show how to obtain only $O((1/\tau)\log n)$ coalesced bitmaps.
\[lem:2lgn\] At most $2\log_{1+\tau} n$ distinct values of $x$ can have $A_x[k]=1$ for a given $k$.
First, $A[k]=x$ is a $\tau$-majority in $A[k,k]$, thus $A_x[k]=1$. Now consider any other element $x' \not=x$ such that $A_{x'}[k]=1$. This means that $x'$ is a $\tau$-majority in some $[i,j]$ that contains $k$. Since $A[k]\not=x'$, it must be that $x'$ is a $\tau$-majority in $[i,k-1]$ or in $[k+1,j]$ (or in both). We say $x'$ is a left-majority in the first case and a right-majority in the second. Let us call $y_1, y_2, \ldots$ the $x'$ values that are left-majorities, and $i_1, i_2, \ldots$ the starting points of their segments (if they are $\tau$-majorities in several segments covering $k$, we choose one arbitrarily). Similarly, let $z_1, z_2, \ldots$ be the $x'$ values that are right-majorities, and $j_1, j_2, \ldots$ the ending points of their segments. Assume the left-majorities are sorted by decreasing values of $i_r$ and the right-majorities are sorted by increasing values of $j_r$. If a same value $x'$ appears in both lists, we arbitrarily remove one of them. As an exception, we will start both lists with $y_0 = z_0 = x$, with $i_0=j_0=k$.
It is easy to see by induction that $y_r$ must appear at least $(1+\tau)^r$ times in the interval $[i_r,k]$ (or in $[i_r,k-1]$, which is the same). This clearly holds for $y_0 = x$. Now, by the inductive hypothesis, values $y_0, y_1, \ldots, y_{r-1}$ appear at least $(1+\tau)^0, (1+\tau)^1,\ldots,(1+\tau)^{r-1}$ times within $[i_{r-1},k-1]$ (which contains all the intervals), adding up to $\frac{(1+\tau)^r-1}{\tau}$ occurrences. Thus $k-1-i_{r-1}+1 \ge \frac{(1+\tau)^r-1}{\tau}$. In order to be a left-majority, element $y_r$ must appear strictly more than $\tau(k-i_{r-1}) \ge (1+\tau)^r-1$ times in $[i_r,k-1]$, to outweight all the occurrences of the previous symbols. The case of right-majorities is analogous. This shows that there cannot be more than $\log_{1+\tau} n$ left-majorities and $\log_{1+\tau} n$ right-majorities.
In the following it will be useful to define $C_x$ as the set of maximal contiguous areas of 1s in $A_x$. That is, $C_x$ is obtained by merging all the segments of $S_x$ that touch or overlap. Note that segments of $C_x$ do not overlap, unlike those of $S_x$. Since a segment of $C_x$ covers a position $k$ iff some segment of $S_x$ covers position $k$ (and iff $A_x[k]=1$), it follows by Lemma \[lem:2lgn\] that any position is covered by at most $2\log_{1+\tau} n$ segments of $C_x$ of distinct symbols $x$.
Note that a pair of consecutive positions $A[k]=x$ and $A[k+1]=y$ is also covered by at most $2\log_{1+\tau} n$ such segments: the right-majorities for $A[k]$ either are $y$ or are also right-majorities for $A[k+1]$, and those are already among the $\log_{1+\tau} n$ right-majorities of $A[k+1]$. And vice versa.
We obtain $O(\log_{1+\tau} n)$ coalesced bitmaps as follows. We take the union of all the sets $C_x$ of all the symbols $x$ and sort the segments by their starting points. Then we start filling coalesced bitmaps. We check if the current segment can be added to an existing bitmap without producing overlaps (and leaving a 0 in between). If we can, we choose any appropriate bitmap, otherwise we start a new bitmap. If at some point we need more than $2\log_{1+\tau} n$ bitmaps, it is because all the last segments of the current $2\log_{1+\tau} n$ bitmaps overlap either the starting point of the current segment or the previous position, a contradiction.
We have $C_1 = \{ [1,1], [5,7]\}$, $C_2 = \{ [3,3] \}$, and $C_3 = \{ [1,6] \}$. Now, we take $C_1 \cup C_2 \cup C_3 = \{ [1,1], [1,6], [3,3],
[5,7] \}$, and the process produces precisely the coalesced bitmaps $A'$, corresponding to the set $\{ [1,1], [3,3], [5,7] \}$, and $A_3$, corresponding to $\{ [1,6] \}$.
Note that in general the coalesced bitmaps may not correspond to the union of complete original bitmaps $A_x$, but areas of a bitmap $A_x$ may end up in different coalesced bitmaps.
Therefore, the coalescing process produces $O(\log_{1+\tau} n) =
O((1/\tau)\log n)$ bitmaps. Consequently, we obtain $O((1/\tau)\log n)$ query time by simply checking the coalesced bitmaps one by one using Lemma \[lem:algox\].
Finally, representing the $O((1/\tau)\log n)$ coalesced bitmaps $A'$, which have total length $O((n/\tau)\log n)$ and contain $O(n/\tau)$ 1s, requires $O((n/\tau)\log\log n)$ bits if we use a compressed bitmap representation [@RRR07] that still offers constant-time $rank$ and $select$ queries (recall Section \[sec:bitmaps\]). The coalesced bitmaps $M'$ still have total length $O(n/\tau)$.
This completes the first part of our result. Next, we will reduce the space usage of our encoding.
Reducing the Space to $O(n\log(1/\tau))$ Bits {#sec:nbits}
=============================================
We introduce a different representation of the coalesced bitmaps that allows us to store them in $O(n\log(1/\tau))$ bits, while retaining the same mechanism described above. We note that, although there can be $O(n/\tau)$ bits set in the bitmaps $A_x$, each new element $x$ produces at most one new [*run*]{} of contiguous 1s (case (1) in the proof of Lemma \[lem:five1s\]). Therefore there are at most $n$ runs in total. We will use a representation of coalesced bitmaps that takes advantage of these runs.
We will distinguish segments of $C_x$ by their lengths, separating lengths by ranges between $\lceil 2^\ell/\tau \rceil$ and $\lceil
2^{\ell+1}/\tau\rceil -1$, for any [*level*]{} $0 \le \ell \le \lg(\tau n)$ (level $0$ is special in that it contains lengths starting from 1). In the process of creating the coalesced bitmaps described in the previous section, we will have separate coalesced bitmaps for inserting segments within each range of lengths; these will be called bitmaps of level $\ell$. There may be several bitmaps of the same level. It is important that, even with this restriction, our coalescing process will still generate $O((1/\tau)\log n)$ bitmaps, because only $O(1/\tau)$ coalesced bitmaps of each level $\ell$ will be generated.
\[lem:8lev\] There can be at most $4/\tau$ segments of any $C_x$, of length between $\lceil 2^\ell/\tau \rceil$ and $\lceil 2^{\ell+1}/\tau \rceil -1$, covering a given position $k$, for any $\ell$.
Any such segment must be contained in the area $A[k-\lceil
2^{\ell+1}/\tau\rceil+1, k+\lceil 2^{\ell+1}/\tau\rceil-1]$, and if $x$ is a $\tau$-majority in it, it must appear more than $\tau \lceil 2^\ell / \tau \rceil \ge 2^\ell $ times. There can be at most $4/\tau$ different values of $x$ appearing more than $2^\ell$ times in an area of length less than $2^{\ell+2}/\tau$.
Consider a coalesced bitmap $A'[1,n]$ of level $\ell$. All of its 1s come in runs of lengths at least $b = \lceil 2^\ell/\tau\rceil$. We cut $A'$ into [*chunks*]{} of length $b$ and define two bitmaps: $A'_1[1,n/b]$ will have $A'_1[i]=1$ iff the $i$th chunk of $A'$ is all 1s, and $A'_2[1,n/b]$ will have $A'_2[i]=1$ iff the $i$th chunk of $A'$ has 0s and 1s. Note that, since the runs of 1s are of length at least $b$, inside a chunk with 0s and 1s there can be at most one 01 and at most one 10, and the 10 can only come before the 01. Let $p_{10}[j]$ be the position, in the $j$th chunk with 0s and 1s, of the 1 preceding a 0, where $p_{10}[j]=0$ if the chunk starts with a 0. Similarly, let $p_{01}[j]$ be the position of the 0 preceding a 1, with $p_{01}[j]=b$ if the chunk ends with a 0. It always holds that $p_{10}[j] <
p_{01}[j]$, and the number of 1s in the chunk is $r(j)=p_{10}[j]+(b-p_{01}[j])$. Also, the rank up to position $k$ in the chunk, $r(j,k)$, is $k$ if $k \le p_{10}[j]$, $p_{10}[j]$ if $p_{10}[j] < k \le p_{01}[j]$, and $p_{10}[j]+(k-p_{01}[j])$ if $k > p_{01}[j]$. Then it holds that $$\begin{aligned}
&& rank(A',i) ~~=~~ b\cdot r_1 ~+~
\sum_{j=1}^{r_2} r(j) ~+~ \\
&& ~~~~~~~~~~~~~~~
[\mathbf{if}~A'_2[1+\lfloor i/b \rfloor]=1~\mathbf{then}~
r(r_2+1,k)~\mathbf{else}~
A'_1[1+\lfloor i/b \rfloor]\cdot k],\end{aligned}$$ where $r_1 = rank(A'_1,\lfloor i/b\rfloor)$, $r_2 = rank(A'_2,\lfloor i/b\rfloor)$, and $k = i~\mathrm{mod}~b$. Note this can be computed in constant time as long as we have constant-time $rank$ data structures on $A'_1$ and $A'_2$, and constant-time access and sums on $p_{10}$ and $p_{01}$.
Using $b=2^\ell$ to make it more interesting, we would have three coalesced bitmaps: $A' = \langle
1~0~1~0~0~0~0 \rangle$, of level $\ell=0$, for the segments $[1,1]$ and $[3,3]$; $A'' = \langle 0~0~0~0~1~1~1 \rangle$, of level $\ell=1$, for the segment $[5,7]$; and $A''' = \langle 1~1~1~1~1~1~0 \rangle$, of level $\ell=2$, for the segment $[1,6]$. Consider level $\ell=0$ and $b=2$, and let us focus on $A'$. Then, we would have $A'_1 = \langle 0~0~0~0 \rangle$, $A'_2 = \langle 1~1~0~0 \rangle$, $p_{10} = \langle 1~1 \rangle$, and $p_{01} = \langle 2~2 \rangle$.
To have constant-time sums on $p_{10}$ ($p_{01}$ is analogous), we store its values in a bitmap $A'_{10}$, where we set all the bits at positions $r + \sum_{j=1}^r p_{10}[j]$ to 1, for all $r$. Then we can recover $\sum_{j=1}^r p_{10}[j] = select(A'_{10},r)-r$. We use a bitmap representation [@OS07] that solves $select$ in constant time (recall Section \[sec:bitmaps\]). Let $n'$ be the number of segments $C_x$ represented in bitmap $A'$. Then there are at most $2n'$ chunks with 0s and 1s, and $A'_{10}$ contains at most $2n'$ 1s and $2n'b$ 0s (as $0 \le p_{10}[j] \le b$). The size of the bitmap representation [@OS07] is in this case $O(n'\log b) = O(n'(\ell + \log(1/\tau)))$ bits. On the other hand, bitmaps $A'_1$ and $A'_2$ are represented in plain form [@Mun96; @Cla96], requiring $O(n/b) = O(n\tau/2^\ell)$ bits.
Considering that there are $O(n/\tau)$ 1s overall, and that the runs of level $\ell$ are of length at least $2^\ell/\tau$, we have that there can be at most $n/2^\ell$ runs across the $O(1/\tau)$ bitmaps of level $\ell$. Therefore, adding up the space over the bitmaps of level $\ell$, we have $O(n(\ell+\log(1/\tau))/2^\ell)$ bits. Added over all the levels $\ell$, this gives $O(n\log(1/\tau))$ bits.
Let us now consider the representation of the coalesced bitmaps $M'$. They have total length $O(n/\tau)$ and contain $n$ 1s overall, therefore using the representation of Lemma \[lem:bitmap\] with $s=1$, we have $O(n\log(1/\tau))$ bits of space. They solve $rank$ queries in time $O(\log\log_w(1/\tau))$, and $select$ in constant time.
As we have to probe $O((1/\tau)\log n)$ coalesced bitmaps $M'$ in the worst case, this raises our query time to $O((1/\tau)\log\log_w(1/\tau)\log n)$. This concludes the proof of Theorem \[thm:main\], except for the construction time (see the next section).
In our previous work [@NT14], we had obtained $O((1/\tau)\log n)$ time, but using $O((n/\tau)\log^*n)$ bits of space. It is not hard to obtain that time, using $O(n/\tau)$ bits, by simply representing the coalesced bitmaps $M'$ using plain $rank$/$select$ structures [@Cla96; @Mun96], or even using $O(n\log(1/\tau) +
(n/\tau)/\operatorname{polylog}n)$ bits, for any $\operatorname{polylog}n$, using compressed representations [@Pat08]. The extra $O(\log\log_w(1/\tau))$ time factor arises when we insist in obtaining the optimal $O(n\log(1/\tau))$ bit space. We note that this time penalty factor vanishes when $1/\tau = w^{O(1)}$, which includes the case where $1/\tau$ grows polylogarithmically with $n$.
Construction {#sec:constr}
============
The most complex part of the construction of our encoding is to build the sets $C_x$. Once these are built, the structures described in Section \[sec:nbits\] can be easily constructed in $o(n\log n)$ time:
1. The $O(n)$ segments $C_x$ belong to $[1,n]$, so they are sorted by starting point in $O(n)$ time.
2. We maintain a priority queue for each level $\ell$, containing the last segment of each coalesced bitmap. We use the queue to find the segment that finishes earliest in order to try to add the new segment of $C_x$ after it. We carry out, in total, $O(n)$ operations on those queues, and each contains $O(1/\tau)$ elements, thus they take total time $O(n\log(1/\tau))=o(n\log n)$.
3. The bitmaps $A'$ of each level $\ell$, represented with $A'_1$, $A'_2$, $A'_{01}$ and $A'_{10}$, are easily built in $O(n/b)=O(n\tau/2^\ell)$ time. Added over the $O(1/\tau)$ coalesced bitmaps of level $\ell$ this is $O(n/2^\ell)$, and added over all the levels $\ell$ this gives $O(n)$ total time.
4. The coalesced bitmaps $M'$ have $O(n)$ 1s overall, so their representation (Lemma \[lem:bitmap\]) is also built in $O(n)$ time, except for the predecessor structures, which need construction of deterministic dictionaries. This can be done in $o(n\log n)$ total time [@Ruz08].
Now we show that the sets $C_x$ can be built in $O(n\log n)$ time, thus finishing the proof of Theorem \[thm:main\].
We build the set of increasing positions $P_x$ where $x$ appears in $A$, for each $x$, in $O(n\log n)$ total time (the elements of $A$ can be of any atomic type, so we only rely on a comparison-based dictionary to maintain the set of different $x$ values and their $P_x$ lists). Now we build $C_x$ from each $P_x$ using a divide-and-conquer approach, in $O(|P_x|\log|P_x|)$ time, for a total construction time of $O(n\log n)$.
We pick the middle element $k \in P_x$ and compute in linear time the segment $[l,r] \in C_x$ that contains $k$. To compute $l$, we find the leftmost element $p_l \in P_x$ such that $x$ is a $\tau$-majority in $[p_l,k_r]$, for some $k_r \in P_x$ with $k_r \ge k$.
To find $p_l$, we note that it must hold that $(w(p_l,k-1)+w(k,k_r))/(k_r-p_l+1) >
\tau$, where $w(i,j)$ is the number of occurrences of $x$ in $A[i,j]$. The condition is equivalent to $w(p_l,k-1)/\tau + p_l - 1 > k_r - w(k,k_r)/\tau$. Thus we compute in linear time the minimum value $v$ of $k_r - w(k,k_r)/\tau$ over all those $k_r \in P_x$ to the right of $k$, and then traverse all those $p_l \in P_x$ to the left of $k$, left to right, to find the first one that satisfies $w(p_l,k-1)/\tau + p_l + 1 > v$, also in linear time. Once we find the proper $p_l$ and its corresponding $k_r$, the starting position of the segment is slightly adjusted to the left of $p_l$, to be the smallest value that satisfies $w(p_l,k_r)/(k_r-l+1) > \tau$, that is, $l$ satisfies $l > -w(p_l,k_r)/\tau +k_r+1$, or $l = k_r - \lceil w(p_l,k_r)/\tau
\rceil + 2$.
Once $p_r$ and then $r$ are computed analogously, we insert $[l,r]$ into $C_x$ and continue recursively with the elements of $P_x$ to the left of $p_l$ and to the right of $p_r$. Upon return, it might be necessary to join $[l,r]$ with the rightmost segment of the left part and/or with the leftmost segment of the right part, in constant time. The total construction time is $T(n) = O(n) +
2T(n/2) = O(n\log n)$.
#### Building multiple structures
In order to answer $\tau'$-majority queries for any $\tau \le \tau' < 1$ in time related to $1/\tau'$ and not to $1/\tau$, we build the encoding of Theorem \[thm:main\] for values $\tau'' = 1/2, 1/4, 1/8, \ldots,
1/2^{\lceil \lg 1/\tau \rceil}$. Then, a $\tau'$-majority query is run on the structure built for $\tau'' = 1/2^{\lceil \lg 1/\tau' \rceil}$. Since $\tau'/2 < \tau'' \le \tau'$, the query time is $O((1/\tau'')\log\log_w(1/\tau'')\log
n)=O((1/\tau')\log\log_w(1/\tau')\log n)$.
As for the space, we build $O(\log(1/\tau))$ structures, so we use $O(n\log^2(1/\tau))$ bits, and the construction time is $O(n\log(1/\tau)\log n)$.
\[cor:vartau\] Given a real number $0<\tau<1$, there exists an encoding using $O(n\log^2(1/\tau))$ bits that answers range $\tau'$-majority queries, for any $\tau \le \tau' < 1$, in time $O((1/\tau')\log\log_w(1/\tau')\log
n)$, where $w=\Omega(\log n)$ is the RAM word size in bits. The structure can be built in time $O(n\log(1/\tau)\log n)$.
A Faster Data Structure {#sec:optimal}
=======================
In this section we show how, by adding $O(n\log\log n)$ bits to our data structure, we can slash a $\log n$ factor from the query time, that is, we prove Theorem \[thm:optimal\]. The result, as discussed in the Introduction, yields the optimal query time $O(1/\tau)$ when $1/\tau = O(\operatorname{polylog}n)$, although the resulting space may not be optimal anymore.
The idea is inspired in a previous non-encoding data structure for majority queries [@wads]. Consider a value $\ell$. Then we will cut $A$ into consecutive pieces of length $2^\ell$ (said to be of [*level*]{} $\ell$) in two overlapped ways: $A[2^\ell k+1,2^\ell(k+1)]$ and $A[2^\ell
k+2^{\ell-1}+1,2^\ell(k+1)+2^{\ell-1}]$, for all $k \ge 0$. We carry out this partitioning for every $\lceil \lg(1/\tau) \rceil \le \ell
\le \lceil \lg n \rceil$.
Note that there are $O(n/2^\ell)$ pieces of level $\ell$, and any interval $A[i,j]$ of length up to $2^\ell/2$ is contained in some piece $P$ of level $\ell$. Now, given a query interval $A[i,j]$, let $\ell = \lceil \lg(j-i+1)\rceil+1$. Then, not only $A[i,j]$ is contained in a piece $P$ of level $\ell$, but also any $\tau$-majority $x$ in $A[i,j]$ must be a $\tau/4$-majority in $P$: Since $j-i+1 > 2^\ell/4$, $x$ occurs more than $\tau(j-i+1) > (\tau/4)2^\ell$ times in $A[i,j]$, and thus in $P$.
Consider a $\tau/4$-majority $x$ in a given piece $P$ of level $\ell$ that is also a $\tau$-majority for some range $A[i,j]$ within $P$, where $2^\ell/4 <
j-i+1 \le 2^\ell/2$. By construction of our previous structures, there exists a maximal segment $C_x$ that contains the range $[i,j]$. If there is another range $A[i',j']$ within $P$ where $x$ is a $\tau$-majority, then there exists another maximal segment $C'_x$ for the same $x$ within $P$. By our construction, if $C'_x \neq C_x$, then $C'_x$ is disjoint with $C_x$, and thus each of them contains at least $(\tau/4)2^\ell$ distinct occurrences of $x$. Obviously, segments $C_y$ for $\tau$-majorities $y \neq x$ contain other $(\tau/4)2^\ell$ occurrences disjoint from those of $x$. Therefore, the number of distinct maximal segments $C$ that contain $\tau$-majorities at any range $A[i,j]$ (with $j-i+1 > 2^\ell/4$) within $P$ is upper bounded by $4/\tau$. We will say those segments $C$ are [*relevant*]{} to $P$.
Therefore, for each piece $P$ of level $\ell$, we will store the index $r$ of the coalesced bitmap $A'_r$ (and its companion $M'_r$) to which each maximal segment $C$ that is relevant to $P$ belongs. Since there are at most $4/\tau$ such coalesced bitmaps to record, out of a total of $O((1/\tau)\log n)$ coalesced bitmaps, $\gamma$-codes on a differential encoding of the subset values requires $O((1/\tau)\log\log n)$ bits.[^3] Added up over the $O(n/2^\ell)$ pieces of level $\ell \ge \lceil \lg(1/\tau)\rceil$, this yields $\sum_{\ell \ge \lceil\lg(1/\tau)\rceil} O((n/2^\ell)
(1/\tau) \log\log n) = O(n\log\log n)$ bits.
This information reduces the search effort to that of verifying $O(1/\tau)$ coalesced bitmaps $A'_r$ and $M'_r$ for the range $[i,j]$, and thus to $O((1/\tau)\log\log_w (1/\tau))$ query time. However, for ranges shorter than $1/\tau$, where no piece structure has been built, we still have the original query time. To speed up this case, we build a second structure where, for each element $A[k]$, we identify the coalesced bitmap where the maximal segment $C_{A[k]}$ containing the segment $A[k,k]$ belongs, and store the identifier $r$ of the corresponding coalesced bitmap $A'_r$ (and $M'_r$) associated to $k$. This requires $O(n\log((1/\tau)\log n)) = O(n\log(1/\tau) + n\log\log n)$ further bits, and allows checking only one coalesced bitmap $A'_r$ (and $M'_r$) for each of the $O(1/\tau)$ positions that need to be checked.
To finish the proof we must consider the construction time. The second structure (for short ranges) is easily built with the general structure, taking no additional time, by keeping track of which maximal segment $C_{A[k]}$ contains each segment $A[k,k]$ and which coalesced bitmap it is assigned. With this, the structure for long ranges can be built as follows: for each position $A[k]$ contained in a piece $P$ of level $\ell$, consider the maximal segment $C_{A[k]}$ that contains it and determine whether it is relevant to $P$. A weak test for this is to consider the coalesced bitmap $M'$ where $C_{A[k]}$ is represented (which is precisely what the first structure stores associated to $k$) and ask whether $M'$ contains more than $(\tau/4)2^\ell$ 1s in the range of $P$. This must be the case if $C_{A[k]}$ is relevant to $P$. Although including the identifier of each $M'$ that passes the test may add some nonrelevant ones, we still cannot include more than $4/\tau$ coalesced bitmaps in the set, as the 1s in the $M'$ bitmaps are disjoint.
The $rank$ operations on bitmaps $M'$ take $O(\log\log_w(1/\tau))$ time, so we avoid them to count how many 1s does $M'$ contain in the range of $P$. Instead, we perform a preprocessing pass over $P$ as follows: We initialize to zero a set of $O((1/\tau)\log n)$ counters, one per coalesced bitmap $M'$, and process $P$ left to right. We increase the counter associated to the bitmap $M'$ of each element $A[k]$ in $P$. At the end, we know all the desired values. This takes $O(2^\ell)$ time, and a similar postprocessing pass clears the counter for the next piece.
Therefore, we process all the pieces $P$ of level $\ell$ in time $O(2^\ell)$, which amounts to $O(n)$ time per level. Added over all the levels, this gives $O(n\log n)$ total time. This concludes the proof of Theorem \[thm:optimal\].
Conclusions {#sec:concl}
===========
A $\tau$-majority query on array $A[1,n]$ receives a range $[i,j]$ and returns all the elements appearing more than $\tau(j-i+1)$ times in $A[i,j]$. We have obtained the first results about [*encodings*]{} for answering range $\tau$-majority queries. Encodings are data structures that use less space than what is required to store $A$ and answer queries without accessing $A$ at all. In the encoding scenario we do not report the $\tau$-majorities themselves, but one of their positions in $A[i,j]$.
We have proved that $\Omega(n\log(1/\tau))$ bits are necessary for any such encoding, even if it can only count the number of $\tau$-majorities in any range. Then we presented an encoding that uses the optimal $O(n\log(1/\tau))$ bits, and answers queries in $O((1/\tau)\log\log_w(1/\tau)\log n)$ time in the RAM model with word size $w=\Omega(\log n)$ bits. We also showed that this time can be divided by $\log n$ if we add $O(n\log\log n)$ bits to the space. This yields various space/time tradeoffs, shown in Table \[tab:final\]. Our encoding can actually report any occurrence of each $\tau$-majority, in optimal extra time. The structure is built in $O(n\log n)$ time.
An open question is whether it is possible to achieve optimal query time within optimal space for all values of $1/\tau$. As seen in Table \[tab:final\], we reach this only for $\log(1/\tau)=\Theta(\log\log n)$. This is also possible when $\log (1/\tau) = \Omega(\log n)$, where we leave the non-encoding scenario [@wads]. Instead, our results for $\log(1/\tau)$ between $\log\log n$ and $\log n$ have a small factor $O(\log\log_w (1/\tau))$ over the optimal time, and those for $\log(1/\tau)$ below $\log\log n$ either require nonoptimal $O(n\log\log n)$ bits of space, or an $O(\log n)$ factor over the optimal time. It is not clear whether combined optimality can be reached.
Another open question is whether we can do better for weaker versions of the problem we have not studied. For example, if we are only required to report [*any*]{} occurrence of [*any*]{} $\tau$-majority (or, even less, telling whether or not there exists a $\tau$-majority), our lower bound based on representing a bitmap $B$ shows that $\Omega(n)$ bits are necessary, but we do not know if this bound is tight.
[^1]: Or an equivalent array where each element is replaced by an identifier in $[1,n]$.
[^2]: Bounding $\lg (3k)!$ with integrals one obtains $3k \lg(3k/e) + 1 \le \lg (3k)! \le (3k+1)\lg((3k+1)/e)+1$.
[^3]: We could also afford to store them in plain form using $O((1/\tau)(\log(1/\tau)+\log\log n))$ bits.
| {
"pile_set_name": "ArXiv"
} |
\[section\] \[Theorem\][Proposition]{} \[Theorem\][Remark]{} \[Theorem\][Lemma]{} \[Theorem\][Corollary]{} \[Theorem\][Definition]{} \[Theorem\][Example]{}
[^1]\
BORIS S. MORDUKHOVICH [^2] and T. T. A. NGHIA[^3]\
[**Dedicated to Jon Borwein in honor of his 60th birthday**]{}\
The paper concerns the study of new classes of nonlinear and nonconvex optimization problems of the so-called infinite programming that are generally defined on infinite-dimensional spaces of decision variables and contain infinitely many of equality and inequality constraints with arbitrary (may not be compact) index sets. These problems reduce to semi-infinite programs in the case of finite-dimensional spaces of decision variables. We extend the classical Mangasarian-Fromovitz and Farkas-Minkowski constraint qualifications to such infinite and semi-infinite programs. The new qualification conditions are used for efficient computing the appropriate normal cones to sets of feasible solutions for these programs by employing advanced tools of variational analysis and generalized differentiation. In the further development we derive first-order necessary optimality conditions for infinite and semi-infinite programs, which are new in both finite-dimensional and infinite-dimensional settings.
Introduction
============
The paper mainly deals with constrained optimization problems formulated as follows: $$\begin{aligned}
\label{1.1}
\left\{\begin{array}{ll}
\mbox{minimize }\;f(x)\;\mbox{ subject to}\\
g_t(x)\leq 0\;\mbox{ with }\;\;t\in T\;\mbox{ and }\;h(x)=0,
\end{array}\right.\end{aligned}$$ where $f:X\to \Bar\R:=(-\infty,\infty]$ and $g_t:X\to \Bar \R$ as $t\in T$ are extended-real-valued functions defined on Banach space $X$, and where $h:X\to Y$ is a mapping between Banach spaces. An important feature of problem (\[1.1\]) is that the index set $T$ is [*arbitrary*]{}, i.e., may be infinite and also noncompact. When the spaces $X$ and $Y$ are finite-dimensional, the constraint system in (\[1.1\]) can be formed by finitely many equalities and infinite inequalities. These optimization problems belong to the well-recognized area of [*semi-infinite programming*]{} (SIP); see, e.g., the books [@GL; @GL1] and the references therein. When the dimension of the decision space $X$ as well as the cardinality of $T$ are infinite, problem belongs to the so-called [*infinite programming*]{}; cf. the terminology in [@an; @DGLS] for linear and convex problems of this type. We also refer the reader to more recent developments [@CLMP1; @CLMP2; @DMN1; @DMN2; @FLN; @LNP] concerning linear and convex problems of infinite programming with inequality constraints.
To the best of our knowledge, this paper is the first one in the literature to address [*nonlinear*]{} and [*nonconvex*]{} problems of infinite programming. Our primary goal in what follows is to find verifiable [*constraint qualifications*]{} that allow us to establish efficient [*necessary optimality conditions*]{} for local optimal solutions to nonconvex infinite programs of type under certain differentiability assumptions on the constraint (while not on the cost) functions. In this way we obtain a number of results, which are new not only for infinite programs, but also for SIP problems with noncompact (e.g., countable) index sets.
It has been well recognized in semi-infinite programming that the [*Extended Mangasarian-Fromovitz Constraint Qualification*]{} (EMFCQ), first introduced in [@JTW], is particularly useful when the index set $T$ is a compact subset of a finite-dimensional space and when $g(x,t):=g_t(x)\in\mathcal{C}(T)$ for each $x\in
X$; see, e.g., [@BS; @CHY; @jrs; @HK; @KH; @LS; @S; @st; @zy] for various applications of the EMFCQ in semi-infinite programming. Without the compactness of the index set $T$ and the continuity of the inequality constraint function $g(x,t)$ with respect to the index variable $t$, problem (\[1.1\]) changes dramatically and–as shown below–does not allow us to employ the EMFCQ condition anymore. That motivates us to seek for new qualification conditions, which are more appropriate in applications to infinite programs as well as to SIP problems with noncompact index sets and infinite collections of inequality constraints defined by discontinuous functions.
In this paper we introduce two new qualification conditions, which allow us to deal with infinite and semi-infinite programs of type without the convexity/linearity and compactness assumptions discussed above. The first condition, called the [*Perturbed Mangasarian-Fromovitz Constraint Qualification*]{} (PMFCQ), turns out to be an appropriate counterpart of the EMFCQ condition for infinite and semi-infinite programs with noncompact index sets $T$ and discontinuous functions $g(x,\cdot)$. The second condition, called the [*Nonlinear Farkas-Minkowski Constraint Qualification*]{} (NFMCQ), is a new qualification condition of the [*closedness*]{} type, which is generally independent of both EMFCQ and PMFCQ conditions even for countable inequality constraints in finite dimensions.
Our approach is based on advanced tools of variational analysis and generalized differentiation that can be found in [@M1; @M2]. Considerably new ingredients of this approach relate to computing appropriate [*normal cones*]{} to the set of [*feasible solutions*]{} for the infinite/semi-infinite program given by $$\label{1.2}
\O:=\big\{x\in X\big|\;h(x)=0,\;g_t(x)\le 0\;\mbox{ as }\;t\in
T\big\}.$$
Since the feasible solution set $\O$ is generally nonconvex, we need to use some normal cone constructions for nonconvex sets. In this paper we focus on the so-called [*Fréchet/regular normal cone*]{} and the [*basic/limiting normal cone*]{} introduced by Mordukhovich; see [@M1] with the references and commentaries therein. Developing general principles of variational analysis, we employ this approach to derive several necessary optimality conditions for the class of nonlinear infinite programs under consideration.
The rest of the paper is organized as follows. In Section 2 we present basic definitions as well as some preliminaries from variational analysis and generalized differentiation widely used in this paper. Section 3 is mainly devoted to the study of the new PMFCQ and NFMCQ conditions for infinite programs in Banach spaces. Relationships between the new qualification conditions and other well-recognized constraint qualifications for SIP and infinite programs are discussed here.
In Section 4, we provide exact computations for the Fréchet and limiting normal cones to the feasible set of under the PMFCQ and NFMCQ conditions. This part plays a crucial role for the subsequent results of the paper. Following this way, Section 5 concerns the derivation of necessary optimality conditions for local minimizers of the infinite and semi-infinite programs under consideration.
Our notation and terminology are basically standard and conventional in the area of variational analysis and generalized differentials.; see, e.g., [@M1; @rw]. As usual, $\|\cdot\|$ stands for the norm of Banach space $X$ and $\la\cdot,\cdot\ra$ signifies for the canonical pairing between $X$ and its topological dual $X^*$ with the symbol $\st{w^*}\to$ indicating the convergence in the weak$^*$ topology of $X^*$ and the symbol $\cl$ standing for the weak$^*$ topological closure of a set. For any $x\in X$ and $r>0$, denote by $\B_r(x)$ the closed ball centered at $x$ with radius $r$ while $\B_X$ stands for the closed unit ball in $X$.
Given a set $\O\subset X$, the notation $\co\O$ signifies the convex hull of $\O$ while that of $\cone\O$ stands for the [*convex conic hull*]{} of $\O$, i.e., for the convex cone generated by $\O\cup\{0\}$. Depending on the context, the symbols $x\st{\O}{\to}\ox$ and $x\st{\ph}{\to}\ox$ mean that $x\to\ox$ with $x\in\O$ and $x\to\ox$ with $\ph(x)\to\ph(\ox)$ respectively. Given finally a set-valued mapping $F\colon X\tto X^*$ between $X$ and $X^*$, recall that the symbol $$\begin{aligned}
\label{1.4}
{\mathop{{\rm Lim}\,{\rm sup}}}_{x\rightarrow \ox} F(x):=\Big\{x^*\in
X^*\Big|\;\exists\,x_n\to\ox,\;\exists\,
x^*_n\xrightarrow{w^*}x^*\;\mbox{ with }\;x^*_n\in F(x_n),\quad
n\in\IN\Big\}\end{aligned}$$ stands for the [*sequential Painlevé-Kuratowski outer/upper limit*]{} of $F$ as $x\to\ox$ with respect to the norm topology of $X$ and the weak$^*$ topology of $X^*$, where $\IN:=\{1,2,\ldots\}$.
Preliminaries from Generalized Differentiation
==============================================
In this preliminary section we briefly review some constructions of generalized differentiation used in what follows; see [@BZ; @M1; @rw; @sc] for more details and related material. Throughout this paper, unless otherwise stated, all the spaces under consideration are Banach.
Given an extended-real-valued function $\ph\colon
X\to\oR:=(-\infty,\infty]$, we always assume that it is proper, i.e., $\ph\not\equiv\infty$. The notation $$\begin{aligned}
\dom\ph:=\big\{x\in X\big|\;\ph(x)<\infty\big\}\;\mbox{ and }\;\epi\ph:=
\big\{(x,r)\in X\times\R\big|\;r\ge\ph(x)\big\}\end{aligned}$$ are used for the domain and the epigraph of $\ph$, respectively,
Define the [*analytic $\ve$-subdifferential*]{} of $\ph$ at $\ox\in\dom\ph$ by $$\begin{aligned}
\label{2.1}
\Hat\partial_\ve\ph(\ox):=\Big\{x^*\in
X^*\Big|\;\liminf_{x\to\ox}\frac {\ph(x)-\ph(\ox)-\la
x^*,x-\ox\ra}{\|x-\ox\|}\ge-\ve\Big\},\quad\ve\ge 0\end{aligned}$$ and let $\Hat\partial_\ve\ph(\ox):=\emp$ for $\ox\notin\dom\ph$. If $\ve=0$, the construction $\Hat\partial
\ph(\ox):=\Hat\partial_0\ph(\ox)$ in (\[2.1\]) is known as the [*Fréchet*]{} or [*regular subdifferential*]{} of $\ph$ at $\ox$; it reduces in the convex case to the classical subdifferential of convex analysis. The sequential regularization of defined via the outer limit by $$\begin{aligned}
\label{2.2}
\partial\ph(\ox):={\mathop{{\rm Lim}\,{\rm sup}}}_\substack{x\st{\ph}{\to}\ox}{\ve\dn 0}\Hat\partial_\ve\ph(x),\end{aligned}$$ is known as the [*limiting*]{}, or [*basic*]{}, or [*Mordukhovich subdifferential*]{} of $\ph$ at $\ox\in\dom\ph$. It can be equivalently described with $\ve=0$ in (\[2.2\]) if $\ph$ is lower semicontinuous (l.s.c.) around $\ox$ and if $X$ is an Asplund space, i.e., each of its separable subspace has a separable dual (in particular, any reflexive space is Asplund; see, e.g., [@BZ; @M1] for more details and references). We have $\partial\ph(\ox)\ne\emp$ for every locally Lipschitzian function on an Asplund space.
A complementary construction to , known as the [*singular*]{} or [*horizontal subdifferential*]{} of $\ph$ at $\ox$, is defined by $$\begin{aligned}
\label{2.3}
\partial^\infty\ph(\ox):={\mathop{{\rm Lim}\,{\rm sup}}}_\substack{x\st{\ph}{\to}\ox}{\lm,\ve\dn
0}\lm\Hat\partial_\ve\ph(x),\end{aligned}$$ where we can equivalently put $\ve=0$ if $\ph$ is l.s.c. around $\ox$ and $X$ is Asplund. Note that $\partial^\infty\ph(\ox)=\{0\}$ if $\ph$ is locally Lipschitzian around $\ox$. The converse implication also holds provided that $\ph$ is l.s.c. around $\ox$, that $X$ is Asplund, and that $\ph$ satisfies the so-called “sequential normal epi-compactness” property at $\ox$ (see below), which is always the case when $X$ is finite-dimensional.
Given a set $\O\subset X$ with its indicator function $\dd(\cdot;\O)$ defined by $\delta(x;\O):=0$ for $x\in\O$ and by $\delta(x;\O):=\infty$ otherwise, we construct the [*Fréchet/regular normal cone*]{} and [*limiting/basic/Mordukhovich normal cone*]{} to $\O$ at $\ox\in\O$ by, respectively, $$\label{2.4}
\Hat N(\ox;\O):=\Hat\partial\delta(\ox;\O)\quad\textrm{and}\quad
N(\ox;\O):=\partial\delta(\ox;\O)$$ via the corresponding subdifferential of the indicator function. If follows from (\[2.4\]) that $\Hat N(\ox;\O)\subset N(\ox;
\O)$. A set $\O$ is [*normally regular*]{} at $\ox$ if $\Hat
N(\ox;\O)= N(\ox;\O)$; the latter is the case of convex and some other “nice" sets.
Recall further that $\O$ is [*sequentially normally compact*]{} (SNC) at $\ox\in\O$ if for any sequences $\ve_n\dn 0$, $x_n\st{\O}\to\ox$, and $x^*_n\in\Hat N_{\ve_n}(x_n;\O):=\Hat \partial_{\ve_n} \delta(\ox;\O)$ we have $$\begin{aligned}
\big[x^*_n\st{w^*}\to 0\big]\Longrightarrow\big[\|x^*_n\|\to
0\big]\;\mbox{ as }\;n\to\infty,\end{aligned}$$ where $\ve_n$ can be omitted if $\O$ is locally closed around $\ox$ and the space $X$ is Asplund. A function $\varphi:X\to
\Bar\R$ is [*sequentially normally epi-compact*]{} (SNEC) at a point $\ox\in \dom \varphi$ if its epigraph is SNC at $(\ox,\varphi(\ox))$. Besides the finite dimensionality, the latter properties hold under certain Lipschitzian behavior; see, e.g., [@M1 Subsections 1.1.4 and 1.2.5].
Having an arbitrary (possibly infinite and noncompact) index set $T$ as in , we consider the product space of multipliers $\R^T:=\{\lm=(\lm_t)|\;t\in T\}$ with $\lm_t\in\R$ for $t\in T$ and denote by $\Tilde\R^T$ the collection of $\lm\in\R^T$ such that $\lm_t\ne 0$ for finitely many $t\in T$. The [*positive cone*]{} in $\Tilde\R^T$ is defined by $$\begin{aligned}
\label{2.6}
\tR:=\big\{\lm\in\Tilde\R^T\big|\;\lm_t\ge 0\;\mbox{ for all }\;t\in
T\big\}.\end{aligned}$$
Qualification Conditions for Infinite Constraint Systems
========================================================
This section is devoted to studying the set of [*feasible solutions*]{} to the original optimization problem defined by the infinite constraint systems of inequalities and equalities $$\begin{aligned}
\label{3.1}
\left\{\begin{array}{ll}
g_t(x)\le 0,\;\;t\in T,\\
h(x)=0,
\end{array}\right.\end{aligned}$$ where $T$ is an arbitrary index set, and where the functions $g_t:X\to\Bar\R$, $t\in T$, and the mapping $h:X\to Y$ are differentiable but may not be linear and/or convex. As in (\[1.2\]), the set of feasible solutions to , i.e., those $x\in X$ satisfying , is denoted by $\O$.
Our [*standing assumptions*]{} throughout the paper (unless otherwise stated) are as follows:\
.
Recall that a mapping $h\colon X\to Y$ is [*strictly differentiable*]{} at $\ox$ with the (strict) derivative $\nabla
h(\ox)\colon X\to Y$ if $$\lim_{x,x'\to\ox}\frac{h(x)-h(x')-\nabla
h(\ox)(x-x')}{\|x-x'\|}=0.$$ The latter holds automatically when $h$ is continuously differentiable around $\ox$.
In addition to the standing assumptions (SA), we often impose some stronger requirements on the inequality constraint functions $g_t$ that postulate a certain uniformity of their behavior with respect to the index parameter $t\in T$. We say that the functions $\{g_t\}_{t\in T}$ are [*uniformly Fréchet differentiable*]{} at $\ox$ if $$\begin{aligned}
\label{3.3}
s(\eta):=\sup_{t\in T}\sup_\substack{x\in\B_{\eta}(\ox)}{x\ne
\ox}\frac{|g_t(x)-g_t(\ox)-\la\nabla
g_t(\ox),x-\ox\ra|}{\|x-\ox\|}\to 0\;\mbox{ as }\;\eta\dn 0.\end{aligned}$$ Similarly, the functions $\{g_t\}_{t\in T}$ are [*uniformly strictly differentiable*]{} at $\ox$ if condition (\[3.3\]) above is replaced by a stronger one: $$\begin{aligned}
\label{3.4}
r(\eta):=\sup_{t\in T}\sup_\substack{x,x'\in\B_{\eta}(\ox)}{x\ne
x'}\frac{|g_t(x)-g_t(x')-\la\nabla
g_t(\ox),x-x'\ra|}{\|x-x'\|}\to 0\;\mbox{ as }\;\eta\dn 0,\end{aligned}$$ which clearly implies the strict differentiability of each function $g_t$, $t\in T$, at $\ox$.
Let us present some sufficient conditions ensuring the fulfillment of all the assumptions formulated above for infinite families of inequality constraint functions.
\[p3\] [**(compact index sets).**]{} Let $T$ be a compact metric space, let the functions $g_t$ in be Fréchet differentiable around $\ox$ for each $t\in T$, and let the mapping $(x,t)\in X\times T\mapsto\nabla g_t(x)\in X^*$ be continuous on $\B_\eta(\ox)\times T$ for some $\eta>0$. Then the standing assumptions [(SA)]{} as well as [(\[3.3\])]{} and [(\[3.4\])]{} are satisfied.
[**Proof.**]{} It is easy to see that our standing assumptions (SA) hold, since $\|\nabla g_t(\ox)\|$ is assumed to be continuous on the compact space $T$ being hence bounded. It suffices to prove that (\[3.4\]) holds, which surely implies .
Arguing by contradiction, suppose that (\[3.4\]) fails. Then there are $\ve>0$, sequences $\{t_n\}\subset T$, $\{\eta_n\}\dn
0$, and $\{x_n\}$, $\{x^\prime_n\}\subset\B_{\eta_n}(\ox)$ such that $$\label{3.5}
\frac{|g_{t_n}(x_n)-g_{t_n}(x'_n)-\la \nabla
g_{t_n}(\ox),x_n-x'_n\ra|}{\|x_n-x'_n\|}\ge
\ve-\frac{1}{n}\;\mbox{ for all large}\;n\in\IN.$$ Since $T$ is a compact metric space, there is a subsequence of $\{t_n\}$ converging (without relabeling) to some $\bar t\in T$. Applying the classical Mean Value Theorem to (\[3.5\]), we find $\theta_n\in[x_n,x^\prime_n]:=\co\{x_n,x^\prime_n\}$ such that $$\begin{aligned}
\frac{\ve}{2}&<&\frac{|\la \nabla
g_{t_n}(\theta_n),x_n-x'_n\ra-\la
\nabla g_{t_n}(\ox),x_n-x'_n\ra|}{\|x_n-x'_n\|}\le\|\nabla g_{t_n}(\theta_n)-\nabla g_{t_n}(\ox)\|\\
&<&\|\nabla g_{t_n}(\theta_n)-\nabla g_{\bar t}(\ox)\|+\|\nabla g_{\bar t}(\ox)-\nabla g_{t_n}(\ox)\|\end{aligned}$$ for all large $n\in\IN$. This contradicts the continuity of the mapping $(x,t)\in X\times T\mapsto \nabla g_t(x)$ on $\B_\eta(\ox)\times T$ and thus completes the proof of the proposition. $\h$
Next we recall a well-recognized constraint qualification condition, which is often used in problems of nonlinear and nonconvex semi-infinite programming.
\[cd1\] [**(Extended Mangasarian-Fromovitz Constraint Qualification).**]{} The infinite system [(\[3.1\])]{} satisfies the [Extended Mangasarian-Fromovitz Constraint Qualification (EMFCQ)]{} at $\ox\in
\O$ if the derivative operator $\nabla h(\ox)\colon X\to Y$ is surjective and if there is $\Tilde x\in X$ such that $\nabla
h(\ox)\tilde x=0$ and that $$\begin{aligned}
\label{3.6}
\la\nabla g_t(\ox),\tilde x\ra<0\;\mbox{ for all }\;t\in
T(\ox):=\big\{t\in T\big|\;g_t(\ox)=0\big\}.\end{aligned}$$
It is clear that in the case of a finite index set $T$ and a finite-dimensional space $Y$ in the EMFCQ condition reduced to the classical Mangasarian-Fromovitz Constraint Qualification (MFCQ) in nonlinear programming. In the case of SIP problems the EMFCQ was first introduced in [@JTW] and then extensively studied and applied in semi-infinite frameworks with $X=\R^m$ and $Y=\R^n$; see, e.g., [@HK; @KH; @LS; @Sh], where the reader can find its relationships with other constraint qualifications for SIP problems.
To the best of our knowledge, the vast majority of nonconvex semi-infinite programs are usually considered with the general assumptions that the index set $T$ is compact, the functions $g_t$ are continuously differentiable, and the mapping $(x,t)\mapsto
\nabla g_t(x)$ is continuous on $X\times T$. Under these assumptions and the EMFCQ formulated above, several authors derive the Karush-Kuhn-Tucker (KKT) necessary optimality conditions of the following type: If $\ox$ is an optimal solution to (\[1.1\]) with $f\in{\cal C}^1$ and $h=(h_1, h_2,\ldots,h_n)$, then there are $\lm\in\tR$ from (\[2.6\]) and $\mu\in \R^n$ such that $$\begin{aligned}
\label{kkt}
0=\nabla f(\ox)+\sum_{t\in T(\ox)}\lm_t\nabla
g_t(\ox)+\sum_{j=1}^n \mu_j\nabla h_j(\ox).\end{aligned}$$
We are not familiar with any results in the literature on nonconvex infinite programming that apply to problems with noncompact index sets $T$. The following example shows that the KKT optimality conditions in form may fail for nonconvex SIP with countable constraints even under the fulfillment of the EMFCQ.
\[ex1\] [**(violation of KKT for nonconvex SIP with countable sets under EMFCQ).**]{} [Consider problem (\[1.1\]) with countable inequality constraints given by $$\begin{aligned}
\label{3.7}
\left\{\begin{array}{ll}
\mbox{minimize }\;(x_1+1)^2+x_2\;\mbox{ subject to}\\
x_1+1\le 0\;\mbox{ and }\;\disp\frac{1}{3n}x_1^3-x_2\le
0\;\mbox{ for all }\;n\in\IN\setminus\{1\}\;\mbox{ with }\;
(x_1,x_2)\in\R^2.
\end{array}\right.\end{aligned}$$ Let $X:=\R^2$, $Y:=\{0\}$, $f(x_1,x_2):=(x_1+1)^2+x_2$, $T:=\IN$, $g_1(x_1,x_2):=x_1+1$, and $g_n(x_1,x_2):=\disp\frac{1}{3n}x^3_1-x_2$ for all $n\in
\IN\setminus\{1\}$. Observe that $\ox:=(-1,0)$ is a global minimizer for problem (\[3.7\]) and that $T(\ox)=\{1\}$ for the active index set in . It is easy to check that the EMFCQ holds at $\ox$ while there is no Lagrange multiplier $\lm\in\R_+$ satisfying the KKT optimality condition at $\ox$. Indeed, we have $\la \nabla g_1(\ox),\;(-1,0)\ra=-1<0$, and the following equation does not admit any solution for $\lm\ge
0$: $$(0,0)=\nabla f(\ox)+\lm\nabla g_1(\ox)=(0,1)+(\lm,0).$$]{}
Now we introduce a new extension of the MFCQ condition to the infinite programs under consideration, which plays a crucial role throughout the paper.
\[cd2\] [**(Perturbed Mangasarian-Fromovitz Constraint Qualification).**]{} We say that the infinite system [(\[3.1\])]{} satisfies the [Perturbed Mangasarian-Fromovitz Constraint Qualification (PMFCQ)]{} at $\ox\in\O$ if the derivative operator $\nabla h(\ox)\colon X\to Y$ is surjective and if there is $\tilde x\in X$ such that $\nabla h(\ox)\tilde x=0$ and that $$\begin{aligned}
\label{3.8}
\inf_{\ve>0}\sup_{t\in T_\ve(\ox)}\la \nabla g_t(\ox),\tilde
x\ra<0\;\mbox{ with }\;T_\ve(\ox):=\big\{t\in T\big|\;g_t(\ox)\ge
-\ve\big\}.\end{aligned}$$
In contrast to the EMFCQ, the active index set in (\[3.8\]) is [*perturbed*]{} by a small $\ve>0$. Since $T(\ox)\subset
T_\ve(\ox)$ for all $\ve>0$, the PMFCQ is stronger than the EMFCQ. However, as shown in Section 4 and Section 5, the new condition is much more appropriate for applications to semi-infinite and infinite programs with general (including compact) index sets than the EMFCQ.
The following proposition reveals some assumptions on the initial data of ensuring the equivalence between the PMFCQ and EMFCQ.
\[p4\] [**(PMFCQ from EMFCQ).**]{} Let $T$ be a compact metric space, and let $\ox\in\O$ in . Assume that the function $t\in T\mapsto g_t(\ox)$ is upper semicontinuous $($u.s.c.$)$ on $T$, that the derivative mapping $\nabla
h(\ox)\colon X\to Y$ is surjective, and that there is $\tilde x\in
X$ with the following properties: $\nabla h(\ox)\tilde x=0$, the function $t\in T\mapsto\la\nabla g_t(\ox),\tilde x\ra$ is u.s.c., and $\la\nabla g_t(\ox),\tilde x\ra<0$ for all $t\in T(\ox)$. Then the PMFCQ condition holds at $\ox$, being thus equivalent to the EMFCQ condition at this point.
[**Proof.**]{} Arguing by contradiction, suppose that the PMFCQ fails at $\ox$. Then it follows from (\[3.8\]) that there exist sequences $\{\ve_n\}\dn 0$ and $\{t_n\}\subset T$ such that $t_n\in T_{\ve_n}(\ox)$ and $$\begin{aligned}
\la\nabla g_{t_n}(\ox),\tilde x\ra\ge-\frac{1}{n}\;\mbox{ for all
}\;n\in\IN.\end{aligned}$$ Since $T$ is a compact metric space, we find a subsequence of $\{t_n\}$ (no relabeling), which converges to some $\bar t\in T$. Observe from the continuity assumptions made imply that $$\begin{aligned}
g_{\bar
t}(\ox)\ge\limsup_{n\to\infty}g_{t_n}(\ox)\ge\limsup_{n\to\infty}-\ve_n=0\;\mbox{
and}\\ \la \nabla g_{\bar t}(\ox),\tilde x\ra\ge\limsup_{n\to\infty}
\la\nabla g_{t_n}(\ox),\tilde x\ra\ge
\limsup_{n\to\infty}-\frac{1}{n}=0.\end{aligned}$$ Thus we have that $\bar t\in T(\ox)$ and $\la\nabla g_{\bar
t}(\ox),\tilde x\ra\ge 0$, which is a contradiction that completes the proof of the proposition. $\h$
The following example shows that the EMFCQ does not imply the PMFCQ (while not ensuring in this case the validity of the required necessary optimality conditions as will be seen in Sections 4 and 5) even for simple frameworks of nonconvex semi-infinite programs with [*compact*]{} index sets.
\[exa1\] [**(EMFCQ does not imply PMFCQ for semi-infinite programs with compact index sets).**]{} [Let $X=\R^2$ and $T=[0,1]$ in with $h=0$ and $$\begin{aligned}
g_0(x):=x_1+1\le 0,\quad g_t(x):=tx_1-x^3_2\le 0\;\mbox{ for
}\;t\in T\setminus\{0\}.\end{aligned}$$ It is easy to check that the functions $g_t$, $t\in T$, satisfy our standing assumptions and that they are strictly uniformly differentiable at the feasible point $\ox=(-1,0)$. Observe furthermore that $T(\ox)=\{0\}$, that $T_\ve(\ox)=[0,\ve]$ for all $\ve\in(0,1)$, and that the EMFCQ holds at $\ox$. However, for any $d=(d_1,d_2)\in \R^2$ we have $$\begin{aligned}
\disp\inf_{\ve>0}\sup_{t\in T_{\ve(\ox)}}\la\nabla
g_t(\ox),d\ra&=&\inf_{\ve>0}
\sup\Big\{\la \nabla g_0(\ox),d\ra,\sup\big\{\la\nabla g_t(\ox),d\ra\big|\;t\in(0,\ve]\big\}\Big\}\\
&=&\inf_{\ve>0}\sup\Big\{d_1,\sup\{td_1\big|\;t\in
(0,\ve]\}\Big\}\ge 0,\end{aligned}$$ which shows that the PMFCQ does not satisfy at $\ox$. Note that the u.s.c. assumption with respect of $t$ in Propositions \[p4\] does not hold in this example.]{}
It is well known in the classical nonlinear programming (when the index set $T$ in is finite), that the MFCQ condition is equivalent to the Slater condition provided that all the functions $g_t$ are convex and differentiable and that $h$ is a linear operator. The next proposition shows that a similar equivalence holds in the semi-infinite and infinite programming frameworks with replacing the MFCQ by our new PMFCQ condition and replacing the Slater by its strong counterpart well recognized in the SIP community; see, e.g., [@GL] and [@CLMP1] for more references and discussions.
\[ss\] [**(equivalence between PMFCQ and SSC for differentiable convex systems).**]{} Assume that in all the functions $g_t$, $t\in T$, are convex and uniformly Fréchet differentiable at $\ox$ and that $h=A$ is a surjective continuous linear operator. Then the PMFCQ condition is equivalent to the following strong Slater condition $($SSC$)$: there is $\hat
x\in X$ such that $A\hat x=0$ and $$\begin{aligned}
\label{3.9}
\sup_{t\in T} g_t(\hat x)<0.\end{aligned}$$
[**Proof.**]{} Suppose first that the SSC holds at $\ox$, i.e., there are $\hat x\in X$ and $\delta>0$ such that $A\hat x=0$ and $g_t(\hat x)<-2\delta$ for all $t\in T$. By the assumptions made this implies that for each $\ve\in (0,\delta)$ and $t\in
T_\ve(\ox)$ we have $$\la\nabla g_t(\ox),\hat x-\ox\ra\le g_t(\hat x)-g_t(\ox)\le-2\delta+\ve\le-\delta.$$ Define further $\tilde x:=\hat x-\ox$ and get $A\tilde x=A\hat
x-A\ox=0$ with $\la\nabla g_t(\ox),\tilde x\ra\le-\delta$ for all $t\in T_\ve(\ox)$ and $\ve\in(0,\delta)$. This clearly implies the PMFCQ condition at $\ox$.
Conversely, assume that the PMFCQ condition holds at $\ox$. Then there are $\ve,\eta>0$ and $\Tilde x\in X$ such that $\la\nabla
g_t(\ox),\Tilde x\ra\le-\eta$ for all $t\in T_\ve(\ox)$ and that $A\Tilde x=0$. It follows from the assumed uniform Fréchet differentiability of $g_t$ at $\ox$ that for each $\lm>0$ we have $$\label{t}
g_t(\ox+\lm\Tilde x)\le g_t(\ox)+\lm\la\nabla g_t(\ox),\Tilde
x\ra+\lm\|\Tilde x\|s\big(\lm\|\Tilde x\|\big),$$ which readily implies that $g_t(\ox+\lm\Tilde
x)\le\lm\big(-\eta+\|\Tilde x\|s(\lm\|\Tilde x\|)\big)$ for all $t\in T_\ve(\ox)$. For $t\notin T_\ve(\ox)$ we observe from that $$g_t(\ox+\lm\Tilde x)\le-\ve+\lm\sup_{\tau\in T}\|\nabla
g_\tau(\ox)\|\cdot\|\Tilde x\|+\lm \|\Tilde x\|s\big(\lm\|\Tilde x\|\big),$$ which gives, combining with the above, that $$\sup_{t\in T} g_t(\ox+\lm\Tilde x)\le\max\big\{\lm\big(-\eta+\|\Tilde x\|s(\lm\|\Tilde x\|)\big),-\ve+\lm\|\Tilde x\|\big(\sup_{\tau\in T}\|\nabla g_\tau(\ox)\|+s(\lm\|\Tilde x\|)\big)\big\}.$$ The latter implies the existence of $\lm_0>0$ sufficiently small such that $\sup_{t\in T}g_t(\Hat x)<0$ with $\Hat
x:=\ox+\lm_0\Tilde x$. Furthermore, it is easy to see that $A\Hat
x=A\ox+\lm_0 A\Tilde x=0$. This concludes that the SSC holds at $\Hat x$ and thus completes the proof of the proposition. $\h$
Next we introduce another qualification condition of the [*closedness/Farkas-Minkowski type*]{} for infinite inequality constraints in .
\[cd3\] [**(Nonlinear Farkas-Minkowski Constraint Qualification).**]{} We say that system [(\[3.1\])]{} with $h(x)=0$ satisfies the [Nonlinear Farkas-Minkowski constraint qualification (NFMCQ)]{} at $\ox$ if the set $$\begin{aligned}
\label{fm}
\cone\big\{\big(\nabla g_t(\ox),\la\nabla
g_t(\ox),\ox\ra-g_t(\ox)\big)\big|\;t\in T\big\}\end{aligned}$$ is weak$^*$ closed in the product space $X^*\times\R$.
In the linear case of $g_t(x)=\la a^*_t,x\ra-b_t$ for some $(a^*_t,b_t)\in X^*\times \R$, $t\in T$, the NFMCQ condition above reduces to the classical Farkas-Minkowski qualification condition meaning that the set $\cone\{(a^*_t,b_t)|\; t\in T\}$ is weak$^*$ closed in $X^*\times\R$. It is well recognized that the latter condition plays an important role in linear semi-infinite and infinite optimization; see, e.g., [@BGW; @CLMP2; @DGL; @DMN1; @DMN2; @GL] for more details and references. Observe that the NFMCQ condition can be represented in the following equivalent form: the set $$\begin{aligned}
\cone\big\{\big(\nabla g_t(\ox),g_t(\ox)\big)\big|\;t\in
T\big\}\;\mbox{ is weak$^*$ closed in $X^*\times \R$}.\end{aligned}$$
Let us compare the new NFMCQ condition with the other qualification conditions discussed in this section in the case of infinite inequality constraints.
\[sfm\] [**(sufficient conditions for NFMCQ).**]{} Consider the constraint inequality system with $h=0$ therein. Then the NFMCQ condition is satisfied at $\ox\in\O$ in each of the following settings:
[**(i)**]{} The index $T$ is finite and the MFCQ condition holds at $\ox$.
[**(ii)**]{} $\dim X<\infty$, the set $\{(\nabla g_t(\ox),\la\nabla
g_t(\ox),\ox\ra-g_t(\ox))|\;t\in T\}$ is compact, and the PMFCQ condition holds at $\ox$.
[**(iii)**]{} The index $T$ is a compact metric space, $\dim
X<\infty$, the mappings $t\in T\mapsto g_t(\ox)$ and $t\in
T\mapsto\nabla g_t(\ox)$ are continuous, and the EMFCQ condition holds at $\ox$.
[**Proof.**]{} Define $\tilde g_t(x):=\la \nabla
g_t(\ox),x-\ox\ra+g_t(\ox)$ for all $x\in X$. To justify (i), suppose that $T$ is finite and that the MFCQ condition holds at $\ox$ for the inequality system in (\[3.1\]). It is clear that $\tilde g_t$ also satisfy the MFCQ at $\ox$. Since the functions $\tilde g_t$ are linear, we observe from Proposition \[ss\] that there is $\hat x\in X$ such that $\tilde g_t(\hat x)=\la \nabla
g_t(\ox),\hat x-\ox\ra+g_t(\ox)< 0$ for all $t\in T$. Thus it follows from [@DMN1 Proposition 6.1] that the NFMCQ condition holds.
Next we consider case (ii) with $X=\R^d$ therein. Suppose that the PMFCQ condition holds at $\ox$ and that the set $\{(\nabla
g_t(\ox),\la \nabla g_t(\ox),\ox\ra-g_t(\ox))|\;t\in T\}$ is compact in $\R^d\times\R$. Noting that the functions $\tilde g_t$ also satisfy the PMFCQ at $\ox$, we apply Proposition \[ss\] to these functions and find $\hat x\in X$ such that $\nabla
h(\ox)\hat x=0$ and that $$\begin{aligned}
\label{3.11}
\sup_{t\in T} \tilde g_t(\hat x)=\sup_{t\in T} \la \nabla g_t(\ox),
\hat x-\ox\ra+g_t(\ox)<0.\end{aligned}$$ Let us check that $(0,0)\not \in \co\{(\nabla g_t(\ox),\la\nabla
g_t(\ox),\ox\ra-g_t(\ox))|\; t\in T\}$. Indeed, otherwise ensures the existence of $\lm\in\tR$ with $\sum_{t\in T}\lm_t=1$ such that $$\begin{aligned}
(0,0)=\sum_{t\in T}\lm_t\big(\nabla g_t(\ox),\la\nabla
g_t(\ox),\ox\ra-g_t(\ox)\big).\end{aligned}$$ Combining the latter with (\[3.11\]) gives us that $$0=\sum_{t\in T}\lm_t\la\nabla g_t(\ox),\hat x\ra-\sum_{t\in
T}\lm_t\big(\la\nabla g_t(\ox),\ox\ra-g_t(\ox)\big)=\sum_{t\in
T}\lm_t\tilde g_t(\hat x)\le\sup_{t\in T}\tilde g_t(\hat x)<0,$$ which is a contradiction. Hence employing [@HL Theorem 1.4.7] in this setting, we have that the conic hull $\cone\{(\nabla
g_t(\ox),\la \nabla g_t(\ox),\ox\ra-g_t(\ox))|\;t\in T\}$ is closed in $\R^{d+1}$. This fully justifies (ii). Observing finally that (iii) follows from (ii) and Proposition \[p4\], we complete the proof of the proposition. $\h$
To conclude this section, let us show that the NFMCQ and PMFCQ conditions are independent for infinite inequality systems in finite dimensions.
\[exa2\] [**(independence of NFMCQ and PMFCQ).**]{}
It is easy to check that for the constraint inequality system from Example \[exa1\] the NFMCQ is satisfied at $\ox=(-1,0)$, since the corresponding conic hull $$\begin{aligned}
\cone\big\{(\nabla g_t(\ox),\la\nabla
g_t(\ox),\ox\ra-g_t(\ox))\big|\;t\in
T\big\}&=&\cone\Big((1,0,-1)\cup
\{(t,0,0)\big|t\in(0,1]\}\Big)\\
&=&\big\{x\in\R^3\big|\;x_1+x_3\ge 0,\;x_1\ge 0\ge
x_3,\;x_2=0\big\}\end{aligned}$$ is closed in $\R^3$. On the other hand, Example \[exa1\] demonstrates that the PMFCQ does not hold for this system at $\ox$.
To show that the NFMCQ does not generally follow from the PMFCQ (and even from the EMFCQ), consider the countable system of inequality constraints in $\R^2$ discussed in Example \[ex1\]. When $\ox=(-1,0)$, we get $T_\ve(\ox)=\{n\in
\IN\setminus\{1\}|\;n\le \frac{1}{\ve}\}\cup\{1\}$ for the the perturbed active index set in . It shows that the PMFCQ (and hence the EMFCQ) hold at $\ox$. On the other hand, the conic hull $$\begin{aligned}
\mbox{cone}\big\{(\nabla g_t(\ox),\la\nabla
g_t(\ox),\ox\ra-g_t(\ox))\big|\;t\in
T\big\}=\mbox{cone}\Big[(1,0,-1)\cup\Big\{\Big(\frac{1}{n},-1,
\frac{-2}{3n}\Big)\Big|\;n\in\IN\setminus\{1\}\Big\}\Big]\end{aligned}$$ is not closed in $\R^3$, i.e., the NFMCQ condition is not satisfies at $\ox$.
Normal Cones to Feasible Sets of Infinite Constraints
=====================================================
This section is devoted to computing both normal cones to the feasible solution sets for the class of nonconvex semi-infinite/infinite programs under consideration in the paper. These calculus results are certainly of independent interest while they play a crucial role in deriving necessary optimality conditions for in Section 5.
The first main theorem gives precise calculations of both Fréchet and limiting normal cones to the set $\O$ of feasible solutions in under the new Perturbed Mangasarian-Fromovitz Constraint Qualification of Definition \[cd2\]. Preliminary we present a known result from functional analysis whose simple proof is given for the reader’s convenience.
\[cr\] [**(weak$^*$ closed images of adjoint operators).**]{} Let $A:X\to Y$ be a surjective continuous linear operator. Then the image of its adjoint operator $A^*(Y^*)$ is a weak$^*$ closed subspace of $X^*$.
[**Proof.**]{} Define $C:=A^*(Y^*)\subset X^*$ and pick any $n\in\IN$. We claim that the set $A_n:=C\cap n\B_{X^*}$ is weak$^*$ closed in $X^*$. Considering a net $\{x^*_\nu\}_{\nu\in\N}\subset A_n$ weak$^*$ converging to $x^*\in X^*$ and taking into account that the ball $\B_{X^*}$ is weak$^*$ compact in $X^*$, we get $x^*\in n\B_{X^*}$. By construction there is a net $\{y^*_\nu\}_{\nu\in \N}\subset Y^*$ satisfying $x^*_\nu=A^*y^*_\nu$ whenever $\nu\in\N$. It follows from the surjectivity of $A$ that $$\|x^*_\nu\|=\|A^*y^*_\nu\|\ge\kappa\|y^*_\nu\|\;\mbox{ for all }\;\nu\in\N,$$ where $\kappa:=\inf\{\|A^*y^*\|\;\mbox{over}\;\|y^*\|=1\}\in(0,\infty)$; see, e.g., [@M1 Lemma 1.18]. Hence $\|y^*_\nu\|\le n\kappa^{-1}$ for all $\nu\in\N$. By passing to a subnet, suppose that $y^*_\nu$ weak$^*$ converges to some $y^*\in Y^*$ for which $x^*=A^*y^*\in A_n$. Thus we have that the set $A_n=C\cap n\B_{X^*}$ is weak$^*$ closed for all $n\in\IN$. The classical Banach-Dieudonné-Krein-Šmulian theorem yields therefore that the set $C$ is weak$^*$ closed in $X^*$. $\h$
Now we are ready to establish the main result of this section.
\[thm41\] [**(Fréchet and limiting normals to infinite constraint systems).**]{} Let $\ox\in\O$ for the set of feasible solutions to the infinite system satisfying the PMFCQ at $\ox$. Assume in addition that the inequality constraint functions $g_t$, $t\in T$, are uniformly Fréchet differentiable at $\ox$. Then the Fréchet normal cone to $\O$ at $\ox$ is computed by $$\begin{aligned}
\label{4.1}
\Hat N(\ox;\O)=\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla
g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}+\nabla h(\ox)^*(Y^*).\end{aligned}$$ If furthermore the functions $g_t$, $t\in T$, are uniformly strictly differentiable at $\ox$, then the limiting normal cone to $\O$ at $\ox$ is also computed by $$\begin{aligned}
\label{4.2}
N(\ox;\O)=\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla
g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}+\nabla h(\ox)^*(Y^*),\end{aligned}$$ and thus the set $\O$ of feasible solutions is normally regular at $\ox$.
[**Proof.**]{} First we justify under the assumptions made. It follows from the PMFCQ and the uniform Fréchet differentiability of $g_t$ at $\ox$ that there are $\tilde\ve>0$, $\delta>0$, and $\tilde x\in X$ such that $\nabla h(\ox)\tilde
x=0$ and $$\begin{aligned}
\label{4.3}
\disp\sup_{t\in T_\ve(\ox)}\la \nabla g_t(\ox),\tilde
x\ra<-\delta\quad {\rm for\ all}\ \ve\leq \tilde\ve.\end{aligned}$$ Let us prove the inclusion “$\supset$” in (\[4.1\]). To proceed, fix any $\ve\in(0,\tilde \ve)$ and pick an arbitrary element $x^*$ belonging to the right-hand side of (\[4.1\]). Then there exist a net $(\lm_\nu)_{\nu\in \N}\subset\tR$ and a dual element $y^*\in Y^*$ satisfying $$\begin{aligned}
\label{4.4}
x^*=w^*-\lim_\nu\sum_{t\in T_\ve(\ox)} \lm_{t\nu}\nabla g_t(\ox)+\nabla h(\ox)^*y^*.\end{aligned}$$ Combining the latter with (\[4.3\]) gives us $$\begin{aligned}
\begin{array}{ll}\label{4.5}
\la x^*,\tilde x\ra&=\disp\lim_\nu\sum_{t\in
T_\ve(\ox)}\lm_{t\nu}\la \nabla g_t(\ox),
\tilde x\ra+\la\nabla h(\ox)^*y^*,\tilde x\ra\\
&\disp\le\liminf_\nu\sum_{t\in T_\ve(\ox)} \lm_{t\nu}(-\delta)+\la
y^*,\nabla h(\ox)\tilde x\ra= -\delta\limsup_\nu\sum_{t\in
T_\ve(\ox)}\lm_{t\nu}.
\end{array}\end{aligned}$$ It follows further that for each $\eta>0$ and $x\in\O\cap
\B_\eta(\ox)$ we have $$\begin{aligned}
\la x^*,x-\ox\ra&=& \lim_\nu\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\la
\nabla g_t(\ox),x-\ox\ra+\la\nabla h(\ox)^*y^*,x-\ox\ra\\
&\le&\limsup_\nu\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\Big(
g_t(x)-g_t(\ox)+\|x-\ox\|s(\eta)\Big)+
\la y^*,\nabla h(\ox)(x-\ox)\ra\\
&\le&\limsup_\nu\sum_{t\in T_\ve(\ox)}
\lm_{t\nu}\Big(\ve+\|x-\ox\|s(\eta)\Big)+ \|y^*\|\Big
(\|h(x)-h(\ox)\|+o(\|x-\ox\|)\Big)\\
&\le&\Big(\ve+\|x-\ox\|s(\eta)\Big)\limsup_\nu\sum_{t\in
T_\ve(\ox)}\lm_{t\nu}+\|y^*\|o(\|x-\ox\|).\end{aligned}$$ Taking now the estimate (\[4.5\]) into account implies that $$\begin{aligned}
\la x^*,x-\ox\ra\le-\frac{\la x^*,\tilde
x\ra}{\delta}\Big(\ve+\|x-\ox\|s(\eta)\Big)+o(\|x-\ox\|)\|y^*\|,\end{aligned}$$ which yields in turn by $\ve\dn 0$ that $$\begin{aligned}
\la x^*,x-\ox\ra\le-\frac{\la x^*,\tilde
x\ra}{\delta}\|x-\ox\|s(\eta)+o(\|x-\ox\|)\|y^*\|.\end{aligned}$$ Since $s(\eta)\dn 0$ as $\eta\dn 0$, it follows from the latter inequality that $$\begin{aligned}
\disp\limsup_{x\st{\O}\to\ox}\frac{\la x^*,
x-\ox\ra}{\|x-\ox\|}\le 0,\end{aligned}$$ which means that $x^*\in\Hat N(\ox;\O)$ and thus justifies the inclusion “$\supset$” in (\[4.1\]).
Next we prove the inclusion “$\subset$” in (\[4.2\]) under the assumption that $g_t$ are uniformly strictly differentiable at $\ox$. This immediately implies the inclusion “$\subset$” in (\[4.1\]) under the latter assumption, while we note that similar arguments justify the inclusion “$\subset$” in (\[4.1\]) under merely the uniform Fréchet differentiability of $g_t$ at $\ox$.
To proceed with proving the inclusion “$\subset$” in (\[4.2\]), define the set $$\label{A}
A_\ve:=\cl{\rm cone}\big\{\nabla g_t(\ox)\big|\;t\in
T_\ve(\ox)\big\}+\nabla h(\ox)^*(Y^*)\;\mbox{ for }\;\ve>0.$$ Arguing by contradiction, pick an arbitrary element $x^*\in N(\ox;\O)\setminus\{0\}$ and suppose that $
x^*\notin A_\ve$ for some $\ve\in(0,\tilde\ve)$. We first claim that the set $A_\ve$ is weak$^*$ closed in $X^*$ for all $\ve\leq \tilde\ve$. To justify, take an arbitrary net $(u_\nu^*)_{\nu\in \N}\subset A_\ve$ weak$^*$ converging to some $u^*\in X^*$. Hence there are nets $(\lm_\nu)_{\nu\in \N}\subset \tR$, $(y^*_\nu)_{\nu\in \N} \subset Y^*$ such that $$u^*_\nu=\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\nabla g_t(\ox)+\nabla h(\ox)^*y^*_\nu\quad \st{w^*}\to u^*.$$ Similarly to the proof of (\[4.5\]) we derive the inequality $$\begin{aligned}
\la u^*,\tilde x\ra\leq-\delta\limsup_\nu\sum_{t\in T_\ve(\ox)} \lm_{t\nu}.\end{aligned}$$ Moreover, we have $$\|u^*_\nu-\nabla h(\ox)^*y^*_\nu\|=\|\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\nabla g_t(\ox)\|\leq \sup_{\tau\in T_\ve(\ox)}\|\nabla g_\tau(\ox)\|\sum_{t\in T_\ve(\ox)}\lm_{t\nu}.$$ It follows from two inequalities above that the net $\{u^*_\nu-\nabla h(\ox)^*y^*_\nu\}_{\nu\in \N}$ is bounded in $X^*$. By the classical Alaoglu-Bourbaki theorem, there is a subnet of $\{u^*_\nu-\nabla h(\ox)^*y^*_\nu\}$ (without relabeling) weak$^*$ converging to some $v^*\in \cl{\rm cone}\big\{\nabla g_t(\ox)\big|\;t\in
T_\ve(\ox)\big\}$. Thus the net $\{\nabla h(\ox)^*y^*_\nu\}$ weak$^*$ converges to $u^*-v^*$. Due to Lemma \[cr\], there is $y^*\in Y^*$ such that $u^*-v^*=\nabla h(\ox)^*y^*$. This implies that $u^*=v^*+\nabla h(\ox)^*y^*\in A_\ve$ and ensures that $A_\ve$ is weak$^*$ closed in $X^*$. Since $x^*\notin A_\ve$, we conclude from the classical separation theorem that there are $x_0\in X$ and $c>0$ satisfying $$\begin{aligned}
\label{4.6}
\la x^*,x_0\ra\ge 2c>0\ge\la\nabla g_t(\ox),x_0\ra+\la\nabla h(\ox)^*y^*,x_0\ra=\la\nabla g_t(\ox),x_0\ra+\la y^*,\nabla h(\ox)x_0\ra\end{aligned}$$ for all $t\in T_\ve(\ox)$ and $y^*\in Y^*$; hence $\nabla h(\ox)x_0=0$. Define further $$\disp\hat x:=x_0+\frac{c}{\|x^*\|\cdot\|\tilde x\|}
\tilde x$$ and observe that $\nabla h(\ox)\hat x=0$. Moreover, it follows from (\[4.6\]) and the PMFCQ that $$\begin{aligned}
\label{4.7}
\la x^*,\hat x\ra=\disp\la x^*,x_0+\frac{c}{\|x^*\|\cdot\|\tilde x\|}\tilde x\ra\ge 2c+\frac{c}{\|x^*\|\cdot\|\tilde x\|}\la x^*,
\tilde x\ra\ge 2c-c=c\;\mbox{ and}\end{aligned}$$ $$\begin{aligned}
\label{4.8}
\la\nabla g_t(\ox),\hat x\ra=\disp\la\nabla g_t(\ox),x_0\ra+\frac{c}{\|x^*\|\cdot\|\tilde x\|}\la\nabla g_t(\ox),
\tilde x\ra\le-\disp\frac{\delta c}{\|x^*\|\cdot\|\tilde x\|}=-\tilde \delta\end{aligned}$$ for all $t\in T_\ve(\ox)$ with $\tilde\delta:=\disp\frac{\delta c}{\|x^*\|\cdot\|\tilde x\|}>0$. Observing that $\hat x\ne 0$ by (\[4.8\]), suppose without loss of generality that $\|\hat x\|=1$. Furthermore, we get from definition of the limiting normal cone that there are sequences $\ve_n\dn 0$, $\eta_n\dn 0$, $x_n\st{\O}\to \ox$, and $x^*_n\st{w^*}\to x^*$ as $n\to\infty$ with $$\begin{aligned}
\label{4.9}
\la x^*_n,x-x_n\ra\le\ve_n\|x-x_n\|\;\mbox{ for all }\;x\in\B_{\eta_n}(x_n)\cap\O,\quad n\in\IN.\end{aligned}$$ Since the mapping $h$ is strictly differentiable at $\ox$ with the surjective derivative $\nabla h(\ox)$, it follows from the Lyusternik-Graves theorem (see, e.g., [@M1 Theorem 1.57]) that $h$ is [*metrically regular*]{} around $\ox$, i.e., there are neighborhoods $U$ of $\ox$ and $V$ of $0=h(\ox)$ and a constant $\mu>0$ such that $$\label{mr}
{\rm dist}\big(x; h^{-1}(y)\big)\le\mu\|y-h(x)\|\;\mbox{ for any }\;x\in U\;\mbox{ and }\;y\in V.$$ Since $h(x_n)=0$ and $\nabla h(\ox)\hat x=0$, we have $$\begin{aligned}
\|h(x_n+t\hat x)\|=\|h(x_n+t\hat x)-h(x_n)-\nabla h(\ox)(t\hat x)\|=o(t)\;\mbox{ for each small }\ t>0.\end{aligned}$$ Thus the metric regularity (\[mr\]) implies that for any small $t>0$ there is $ x_t\in h^{-1}(0)$ with $\|x_n+t\hat x-x_t\|=o(t)$ when $x_n\in U$. This allows us to find $\tilde\eta_n<\eta_n$ and $\tilde x_n:=x_{\tilde\eta_n}\in h^{-1}(0)$ satisfying $\tilde\eta_n+o(\tilde \eta_n)\le\eta_n$ and $\|x_n+\tilde\eta_n\hat x-\tilde x_n\|=o(\tilde\eta_n)$. Note that $$\|x_n-\tilde x_n\|\le\tilde\eta_n\|\hat x\|+\|x_n+\tilde\eta_n\hat x-\tilde x_n\|=\tilde\eta_n+o(\tilde\eta_n)\le\eta_n,$$ i.e., $\tilde x_n\in\B_{\eta_n}(x_n)$. Observe further that $$\|x_n-\tilde x_n\|\ge\tilde\eta_n\|\hat x\|-\|x_n+\tilde\eta_n\hat x-\tilde x_n\|=\tilde\eta_n-o(\tilde\eta_n).$$ By the classical uniform boundedness principle there is a constant $M$ such that $M>\|x^*_n\|$ for all $n\in\IN$ due to $x_n^*\st{w^*}\to x^*$ as $n\to\infty$. It follows from (\[4.7\]) that $\la x^*_n,\hat x\ra>0$ for $n\in\IN$ sufficiently large. Then we have $$\begin{aligned}
\disp\frac{\la x^*_n,\tilde x_n-x_n\ra}{\|\tilde x_n-x_n\|}&=&\disp\frac{\la x^*_n,\tilde x_n-\tilde\eta_n\hat x-x_n\ra}{\|\tilde x_n-x_n\|}
+\frac{\la x^*_n,\tilde\eta_n \hat x\ra}{\|\tilde x_n-x_n\|}\\
&\ge&\disp-M\frac{\|\tilde x_n-\tilde\eta_n\hat x-x_n\|}{\|\tilde x_n-x_n\|}+\tilde\eta_n\frac{\la x^*_n,\hat x\ra}{\|\tilde x_n-x_n\|}\\
&\ge&\disp-M\frac{o(\tilde\eta_n)}{\tilde\eta_n-o(\tilde\eta_n)}+\frac{\tilde\eta_n}{\tilde\eta_n+o(\tilde\eta_n)}\la x_n^*,\hat x\ra.\end{aligned}$$ Since $o(\tilde\eta_n)/\tilde\eta_n\to 0$ when $n\to\infty$, the latter inequalities yield that $$\begin{aligned}
\liminf_{n\to\infty}\disp\frac{\la x^*_n,\tilde x_n-x_n\ra}{\|\tilde x_n-x_n\|}\ge\la x^*,\hat x\ra.\end{aligned}$$ Combining this with (\[4.7\]) and (\[4.9\]) gives us that $\tilde x_n\notin\O$ for all large $n\in\IN$.
Now define $u_n:=x_n+\tilde\eta_n\hat x-\tilde x_n$ and get $\|u_n\|=o(\tilde\eta_n)$ and $\|\tilde x_n+u_n-x_n\|=\tilde\eta_n$ by the arguments above. It follows from our standing assumptions (SA), condition (\[3.4\]), and inequality (\[4.8\]) that for each $t\in T_\ve(\ox)$ we have $$\begin{aligned}
-\disp\tilde \delta&\geq&\disp\frac{\la\nabla g_t(\ox),\tilde\eta_n\hat x\ra}{\tilde\eta_n}=\disp\frac{\la\nabla g_t(\ox),\tilde x_n+u_n-x_n\ra}{\|\tilde x_n+u_n-x_n\|}=\disp\frac{\la \nabla g_t(\ox),\tilde x_n-x_n\ra}{\|\tilde x_n+u_n-x_n\|}
+\disp\frac{\la\nabla g_t(\ox),u_n\ra}{\|\tilde x_n+u_n-x_n\|}\\
&\ge&\disp\frac{\la\nabla g_t(\ox),\tilde x_n-x_n\ra}{\|\tilde x_n-x_n\|}\frac{\|\tilde x_n-x_n\|}{\|\tilde x_n+u_n-x_n\|}+\disp
\frac{\la\nabla g_t(\ox),u_n\ra}{\|\tilde x_n+u_n-x_n\|}\\
&\ge&\disp\Big(\frac{g_t(\tilde x_n)-g_t(x_n)}{\|\tilde x_n-x_n\|}-r(\hat\eta_n)\Big)\frac{\|\tilde x_n-x_n\|}{\|\tilde x_n+u_n-x_n\|}
-\sup_{\tau\in T_\ve(\ox)}\|\nabla g_\tau(\ox)\|\frac{o(\tilde\eta_n)}{\tilde\eta_n}\\
&\geq& \disp \Big(\frac{g_t(\tilde x_n)}{\|\tilde x_n-x_n\|}-r(\hat \eta_n)\Big)\frac{\|\tilde x_n-x_n\|}{\|\tilde x_n+u_n-x_n\|}- \sup_{\tau\in T}\|\nabla g_\tau(\ox)\|\frac{o(\tilde\eta_n)}{\tilde\eta_n},\end{aligned}$$ where $\hat\eta_n:=\max\{\|x_n-\ox\|$ and $\|\tilde x_n-\ox\|\}\to 0$ as $n\to\infty$. Note that $$\begin{aligned}
\frac{\tilde\eta_n-o(\tilde\eta_n)}{\tilde\eta_n}\le\frac{\|\tilde x_n-x_n\|}{\|\tilde x_n+u_n-x_n\|}\le\frac{\tilde\eta_n+o(\tilde \eta_n)}{\tilde\eta_n},\end{aligned}$$ which implies that $\disp\frac{\|\tilde x_n-x_n\|}{\|\tilde x_n+u_n-x_n\|}\to 1$ as $n\to\infty$. Furthermore, since $r(\hat\eta_n)\to 0$ and $\disp \frac{o(\tilde \eta_n)}{\tilde\eta_n}\to 0$ as $n\to \infty$, we have ${g_t(\tilde x_n)}\le-\disp\frac{\tilde\delta}{2}\|\tilde x_n-x_n\|\le 0$ for each $t\in T_\ve(\ox)$ when $n\in\IN$ is sufficiently large. Indeed, assuming otherwise that $t\notin T_\ve(\ox)$ gives us $$\begin{aligned}
g_t(\tilde x_n)&\le&g_t(\ox)+\la\nabla g_t(\ox),x_n-\ox\ra+\|x_n-\ox\|r(\hat\eta_n)\\
&\le&-\ve+\sup_{\tau\in T}\|\nabla g_\tau(\ox)\|\hat\eta_n+\hat\eta_n r(\hat\eta_n)\le 0\;\mbox{ for all large }\;n\in\IN.\end{aligned}$$ Thus $g_t(\tilde x_n)\le 0$ for all $t\in T$ and also $h(\tilde x_n)=0$ when $n\in\IN$ is sufficiently large, i.e., $\tilde x_n\in\O$, a contradiction. Hence we conclude that $N(\ox;\O)\subset A_\ve$ for all $\ve\in(0,\tilde\ve)$, which implies the inclusion “$\subset$” in (\[4.2\]) and completes the proof of the theorem. $\h$
Let us show now that the PMFCQ condition is essential for the validity of both normal cone representations in and ; moreover, this condition cannot be replaced by its weaker EMFCQ version.
\[exa3\] [**(violation of the normal cone representations with no PMFCQ).**]{} [Consider the infinite inequality system in $\R^2$ given in Example \[exa1\]. It is shown therein that the EMFCQ holds at $\ox=(-1,0)$ while the PMFCQ does not. It is easy to check that in this case $\Hat N(\ox;\O)=N(\ox;\O)=\R_+\times\R_-$ while $${\rm cl}\;{\rm cone}\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}={\rm cl}\;{\rm cone}\big\{(1,0)\cup\{(t,0)|\;t\in(0,\ve)\big\}\subset\R_+\times\{0\}.$$ i.e., the inclusions “$\subset$" in and are violated.]{}
The next example shows that the perturbed active index set $T_\ve(\ox)$ cannot be replaced by its unperturbed counterpart $T(\ox)$ in the normal cone representations and .
\[ex3\] [**(perturbation of the active index set is essential for the normal cone representations).**]{} [Let us reconsider the nonlinear infinite system in problem (\[3.7\]): $$\begin{aligned}
\left\{\begin{array}{ll}
g_1(x)=x_1+1\le 0,\\
g_n(x)=\disp\frac{1}{3n}x^3_1-x_2\le 0,\;\;n\in \IN\setminus\{1\},
\end{array}\right.\end{aligned}$$ where $x=(x_1,x_2)\in\R^2$ and $T:=\IN$. It is easy to check this inequality system satisfies our standing assumptions and that the functions $g_t$ are uniformly strictly differentiable at $\ox=(-1,0)$. Observe further that $\O=\{(x_1,x_2)\in \R^2|\;x_1\le-1,
\;x_2\ge 0\}$ and hence $N(\ox;\O)=\R_+\times\R_-$. As shown above, both PMFCQ and EMFCQ conditions hold at $\ox$. However, we have $T(\ox)=\{1\}$ and $$N(\ox;\O)\not=\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}=\cone\big\{\nabla g_1(\ox)\big\}=\cone\{(1,0)\}=\R_+\times \{0\},$$ which shows the violation of the unperturbed counterparts of and . Observe that $$\begin{aligned}
\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}&=&\cone\Big\{(1,0)\cup\big\{\big(\frac{1}{n},-1\big)\big|\; n\in\IN\setminus\{1\},\;n
\ge\frac{1}{\ve}\Big\}\\
&=&\big\{(x_1,x_2)\in\R^2\big|\;x_1\ge 0,\;x_2<0\big\},\end{aligned}$$ which is not a closed subset. On the other hand, we have $$N(\ox;\O)=\bigcap_{\ve>0}{\rm cl}\;\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\},$$ which illustrates the validity of the normal cone representations in Theorem \[thm41\].]{}
Now we derive several consequences of Theorem \[thm41\], which are of their independent interest. The first one concerns the case when the $\{\nabla g_t(\ox)|\;t\in T\}$ may not be bounded in $X^*$ as in our standing assumptions. It follows that the latter case can be reduced to the basic case of Theorem \[thm41\] with some modifications.
\[mc\] [**(normal cone representation for infinite systems with unbounded gradients).**]{} Considering the constraint system , assume the following:
[**(a)**]{} The functions
$g_t$, $t\in T$, are Fréchet differentiable at the point $\ox$ with $\|\nabla g_t(\ox)\|>0$ for all $t\in T$ and the mapping $h$ is strictly differentiable at $\ox$.
[**(b)**]{} We have that $\disp\lim_{\eta\dn0}\tilde r(\eta)=0$, where $\tilde r(\eta)$ is defined by $$\begin{aligned}
\label{r1}
\tilde r(\eta):=\sup_{t\in T}\sup_\substack{x,x^\prime\in\B_{\eta}(\ox)}{x\ne x'}\frac{|g_t(x)-g_t(x^\prime)-\la\nabla g_t(\ox),x-x'\ra|}{\|\nabla g_t(\ox)\|\cdot\|x-x'\|}\;\mbox{ for all }\;\eta>0.\end{aligned}$$
[**(c)**]{} The operator $\nabla h(\ox)\colon X\to Y$ is surjective and for some $\ve>0$ there are $\tilde x\in X$ and $\sigma>0$ such that $\nabla h(\ox)\tilde x=0$ and that $$\begin{aligned}
\label{4.11}
\la\nabla g_t(\ox),\tilde x+x\ra\le 0\;\mbox{ whenever }\;\|x\|\le\sigma\end{aligned}$$ for each $t\in\Tilde T_\ve(\ox):=\{t\in T|\ g_t(\ox)\ge-\ve\|\nabla g_t(\ox)\|\}$. Then the limiting normal cone to $\O$ at $\ox$ is computed by formula .
[**Proof.**]{} Define $\tilde g_t(x):=g_t(x)\|\nabla g_t(\ox)\|^{-1}$ for all $x\in X$ and $t\in T$ and observe that the feasible set $\O$ from admits the representation $$\O=\big\{x\in X\big|\;\tilde g_t(x)\le 0,\;h(x)=0\big\}.$$ Replacing $g_t$ by $\tilde g_t$ in Theorem \[thm41\], we have that the functions $\{\tilde g_t\}$ and $h$ satisfy the standing assumptions (SA) as well as condition (\[3.4\]) with the function instead of $r(\eta)$. Furthermore, it follows from (\[4.11\]) that for some $\ve>0$ there are $\tilde x\in X$ and $\sigma>0$ satisfying $\nabla h(\ox)\tilde x=0$ and such that $$\begin{aligned}
\la\nabla\tilde g_t(\ox),\tilde x\ra\le-\sup_{x\in\B_\sigma(\ox)}\la\nabla\tilde g_t(\ox),x\ra=-\sigma\|\nabla\tilde g_t(\ox)\|\;\mbox{ whenever }\; t\in \Tilde T_\ve(\ox),\end{aligned}$$ which turns into $\la\nabla\tilde g_t(\ox),\tilde x\ra\le-\sigma$ for all $t\in\Tilde T_\ve(\ox)=\{t\in T|\;\tilde g_t(\ox)\ge-\ve\}$. Hence the PMFCQ condition holds for the functions $\tilde g_t$ and $h$ at $\ox$. It follows from Theorem \[thm41\] that $$\begin{aligned}
\begin{array}{ll}
N(\ox;\O)&=\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla \tilde g_t(\ox)\big|\; t\in\Tilde T_\ve(\ox)\big\}+\nabla h(\ox)^*(Y^*)\\
&=\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\,\|\nabla g_t(\ox)\|^{-1}\big|\;t\in \Tilde T_\ve(\ox)\big\}+\nabla h(\ox)^*(Y^*)\\
&=\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\big|\;t\in\Tilde T_\ve(\ox)\big\}+\nabla h(\ox)^*(Y^*),
\end{array}\end{aligned}$$ which gives and completes the proof of the corollary. $\h$
Now we compare the result of Corollary \[mc\] with the recent one obtained in [@S Theorem 3.1 and Corollary 4.1] for inequality constraint systems, i.e., with $h=0$ in . The latter result is given by the inclusion form $$\begin{aligned}
N(\ox;\O)\subset\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}\end{aligned}$$ in the case of $\|\nabla g_t(\ox)\|=1$ for all $t\in T$ under the Fréchet differentiability of $g_t$ [*around*]{} $\ox$ (in (as) we need it merely [*at*]{} $\ox$) and the replacement of (b) of Corollary \[mc\] by the following [*equicontinuity*]{} requirement on $g_t$ at $\ox$: for each $\gamma>0$ there is $\eta>0$ such that $$\begin{aligned}
\label{4.12}
\|\nabla g_t(x)-\nabla g_t(\ox)\|\le\gamma\;\mbox{ for all }\;x\in\B_\eta(\ox),\;t\in T.\end{aligned}$$ Let us check that the latter assumption together with the Fréchet differentiability of $g_t$ around $\ox$ imply (b) in Corollary \[mc\]. Indeed, suppose that (\[4.12\]) holds and then pick any $x,x^\prime\in\B_\eta(\ox)$. Employing the classical Mean Value Theorem, find $\hat x\in [x,x^\prime]\subset\B_\eta(\ox)$ such that $g_t(x)-g_t(x^\prime)=\la\nabla g_t(\hat x),x-x^\prime\ra$. This gives $$\begin{aligned}
\disp\frac{|g_t(x)-g_t(x')-\la\nabla g_t(\ox),x-x'\ra|}{\|\nabla g_t(\ox)\|\cdot\|x-x'\|}&=&\disp\frac{|\la\nabla g_t(\hat x),x-x^\prime\ra-\la\nabla g_t(\ox),x-x'\ra|}{\|x-x'\|}\\
&\le&\disp\frac{|\la\nabla g_t(\hat x)-\nabla g_t(\ox),x-x^\prime\ra|}{\|x-x'\|}\\
&\le&\disp \|\nabla g_t(\hat x)-\nabla g_t(\ox)\|\le\gamma\end{aligned}$$ and yields $\disp\lim_{\eta\dn 0}\tilde r(\eta)\le\gamma$ for all $\gamma>0$, which ensures the validity of (b) in Corollary \[mc\].
The next consequence of Theorem \[thm41\] concerns problems of semi-infinite programming and presents sufficient conditions for the fulfillment of simplified representations of the normal cones to feasible constraints with no closure operations in and and with the replacement of the perturbed index set $T_\ve(\ox)$ by that of active constraints $T(\ox)$.
\[si\] [**(normal cones for semi-infinite constraints).**]{} Let $X$ and $Y$ be finite-dimensional spaces with $\dim Y<\dim X$. Assume that $T$ is a compact metric space, that the function $t\in T\mapsto g_t(\ox)$ is u.s.c., and the mapping $t\in T\mapsto\nabla g_t(\ox)$ is continuous. Suppose further that system [(\[3.1\])]{} satisfies the PMFCQ at $\ox$. Then we have $$\begin{aligned}
\label{4.13}
\Tilde N(\ox;\O)=\disp\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}+\nabla h(\ox)^*(Y^*),\end{aligned}$$ where $\Tilde N(\ox;\O)=\Hat N(\ox;\O)$ when the functions $g_t$ are uniformly Fréchet differentiable at $\ox$ and $\Tilde N(\ox;\O)= N(\ox;\O)$ when $g_t$ are uniformly strictly differentiable at $\ox$.\
In particular, if we assume in addition that both $t\in T\mapsto g_t(\ox)$ and $(x,t)\in X\times T\mapsto\nabla g_t(x)$ are continuous, then we also have [(\[4.13\])]{} for $\Tilde N(\ox;\O)= N(\ox;\O)$ provided that merely the EMFCQ condition holds at $\ox$.
[**Proof.**]{} Let $X=\R^d$ for some $d\in\IN$. It follows from Proposition \[p3\] that $g_t$, $t\in T$, and $h$ satisfy our standing assumptions $(SA)$. Since system (\[3.1\]) satisfies the PMFCQ at $\ox$, there are $\tilde\ve>0$, $\delta>0$, and $\tilde x\in X$ such that $\la\nabla g_t(\ox),\tilde x\ra<-\delta$ for all $t\in T_\ve(\ox)$ and $\ve\in (0,\tilde \ve)$. Observe that the perturbed active index set $T_\ve(\ox)$ is compact in $T$ for all $\ve>0$ due to the u.s.c. assumption on $t\in T\mapsto g_t(\ox)$. It follows from the continuity of $t\in T\mapsto\nabla g_t (\ox)$ that $\{\nabla g_t(\ox)|\; t\in T_\ve(\ox)\}$ is a compact subset of $\R^d$.
We now claim that $0\notin\co\{\nabla g_t(\ox)|\;t\in T_\ve(\ox)\}$. Indeed, it follows for any $\lm\in\Tilde \R_+^{T_\ve(\ox)}$ with $\sum_{t\in T_\ve(\ox)}\lm_t=1$ that $$\begin{aligned}
\sum_{t\in T_\ve(\ox)}\lm_t\la\nabla g_t(\ox),\tilde x\ra\le-\sum_{t\in T_\ve(\ox)}\lm_t\delta=-\delta<0,\end{aligned}$$ which yields that $0\ne\sum_{t\in T_\ve(\ox)}\lm_t\nabla g_t(\ox)$, i.e., $0\notin\co\{\nabla g_t(\ox)|\;t\in T_\ve(\ox)\}$.
Hence it follows from [@HL Proposition 1.4.7] that the conic hull cone$\{\nabla g_t(\ox)|\;t\in T_\ve(\ox)\}$ is closed in $\R^d$. Combining this with Theorem \[thm41\], it suffices to show that $$\begin{aligned}
\label{4.14}
\bigcap_{\ve>0}\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}=\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}.\end{aligned}$$ Observe that the inclusion “$\supset$” in (\[4.14\]) is obvious due to $T(\ox)\subset T_\ve(\ox)$ as $\ve>0$. To justify the converse inclusion, pick an arbitrary element $x^*$ from the set on the left-hand side of (\[4.14\]). By the classical Carathéodory theorem, for all large $n\in\IN$ we find $\lm_n\in \R^{d+1}_+$ and $$\nabla g_{t_{n_1}}(\ox),\ldots,\nabla g_{t_{n_{d+1}}}(\ox)\in\big\{\nabla g_t(\ox)\big|\;t\in T_{\frac{1}{n}}(\ox)\big\}\subset\R^d$$ satisfying the relationship $$\begin{aligned}
\label{4.15}
x^*=\sum_{k=1}^{d+1}\lm_{n_k}\nabla g_{t_{n_k}}(\ox),\end{aligned}$$ which implies in turn that $$\begin{aligned}
\la x^*,\tilde x\ra=\sum_{k=1}^{d+1}\lm_{n_k}\la\nabla g_{t_{n_k}}(\ox),\tilde x\ra\le-\sum_{k=1}^{d+1}\lm_{n_k}\delta.\end{aligned}$$ Hence the sequence $\{\lm_n\}$ is bounded in $\R^{d+1}$, and so is $$\big\{\lm_n\times(\nabla g_{t_{n_1}}(\ox),\ldots,\nabla g_{t_{n_{d+1}}})\big\}\subset\R^{d+1}\times\R^{d(d+1)}.$$ By the classical Bolzano-Weierstrass theorem and the compactness of $T$, we assume without loss of generality that the sequence $\{t_{n_k}\}$ converges to some $\bar t_k\in T$ for each $1\le k\le d+1$ and that $\{\lm_n\}$ converges to some $\bar\lm\in \R^{d+1}$ as $n\to \infty$. Note that $0\ge g_{t_{n_k}}(\ox)\geq -\frac{1}{n}$ for all $n\in\IN$ sufficiently large, which gives us $$0\ge g_{\bar t_k}(\ox)\ge\disp \limsup_{n\to\infty}g_{t_{n_k}}(\ox)\ge\limsup_{n\to\infty}-\frac{1}{n}=0$$ for all $1\le k\le d+1$. Combining the latter with (\[4.15\]) ensures that $$\begin{aligned}
x^*=\sum_{k=1}^{d+1}\bar\lm_{k}\nabla g_{\bar t_k}(\ox)\in\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\},\end{aligned}$$ which yields the inclusion “$\subset$” in (\[4.14\]). Thus we arrive at formula $(\ref{4.13})$.
The second part of the corollary follows from the first part, Proposition \[p3\], and Proposition \[p4\]. This completes the proof of the claimed result. $\h$
The results obtained in Corollary \[si\] can be compared with [@CHY Theorem 3.4], where “$\subset$” in (\[4.13\]) was obtained for $h=0$ under the following conditions: $T$ is scattered compact (meaning that every subset $S\subset T$ has an isolated point), $g_t$ are Fréchet differentiable for all $t\in T$, the mappings $(x,t)\in X\times T\mapsto g_t(x)$ and $(x,t)\in X\times T\mapsto\nabla g_t(x)$ are continuous, and the EMFCQ condition holds at $\ox$. We can see that these assumptions are significantly stronger than those Corollary \[si\]. Note, in particular, that the scattering compactness requirement on the index set $T$ is not different in applications from $T$ being finite.
The next question we address in this section is about the possibility to obtain normal cone representations of the “unperturbed" type as in Corollary \[si\] while in infinite programming settings with no finite dimensionality, compactness, and continuity assumptions made above. The following theorem shows that this can be done when the PMFCQ is accompanied by the NFMCQ condition of Definition \[cd3\].
\[thm42\] [**(unperturbed representations of normal cones for infinite constraint systems).**]{} Let the functions $g_t$, $t\in T$, be uniformly Fréchet differentiable at $\ox$, and let that system [(\[3.1\])]{} satisfy the PMFCQ and NFMCQ conditions at $\ox$. Then $$\begin{aligned}
\label{4.16}
\Hat N(\ox;\O)= \disp\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}+\nabla h(\ox)^*(Y^*).\end{aligned}$$ If in addition the functions $g_t$, $t\in T$, are uniformly strictly differentiable at $\ox$, then $$\begin{aligned}
\label{4.17}
N(\ox;\O)=\disp\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}+\nabla h(\ox)^*(Y^*).\end{aligned}$$
[**Proof.**]{} First we claim that the set $\disp\bigcap_{\ve>0}\cl\cone\{\nabla g_t(\ox)|\; t\in T_\ve(\ox)\}$ belongs to the set $$\begin{aligned}
\label{4.18}
\Big\{x^*\in X^*\Big|\;(x^*,\la x^*,\ox\ra)\in\cl\cone\big\{(\nabla g_t(\ox),\la\nabla g_t(\ox),\ox\ra-g_t(\ox))\big|\; t\in T\big\}\Big\}.\end{aligned}$$ Indeed, it follows from the PMFCQ for (\[3.1\]) at $\ox$ that $\nabla h(\ox)$ is surjective and there are $\tilde \ve>0$, $\delta>0$, and $\tilde x\in X$ such that $\nabla h(\ox)\tilde x=0$ and that $\la\nabla g_t(\ox),\tilde x\ra<-\delta$ for all $\ve\le\tilde\ve$ and $t\in T_\ve(\ox)$. To justify the claimed inclusion to (\[4.18\]), pick an arbitrary element $x^*\in\disp\bigcap_{\ve>0}\cl\cone\{\nabla g_t(\ox)|\;t\in T_\ve(\ox)\}$ and for any $\ve\in (0,\tilde\ve)$ find a net $(\lm_\nu)_{\nu\in\N}\subset\tR$ with $$\begin{aligned}
\label{4.19}
x^*=w^*-\lim_\nu\disp\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\nabla g_t(\ox).\end{aligned}$$ This implies the relationships $$\begin{aligned}
\label{4.20}
\la x^*,\tilde x\ra=\lim_\nu\disp\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\la \nabla g_t(\ox),\tilde x\ra\le-\delta\limsup_\nu\disp\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\;\mbox{ and}\end{aligned}$$ $$\begin{aligned}
\begin{array}{ll}
\la x^*,\ox\ra=\disp\lim_\nu\disp\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\la\nabla g_t(\ox),\ox\ra=\lim_\nu\disp\sum_{t\in T_\ve(\ox)}\lm_{t\nu}(\la\nabla g_t(\ox),\ox\ra-g_t(\ox)+g_t(\ox)).
\end{array}\end{aligned}$$ The later equality together with (\[4.20\]) give us that $$\begin{aligned}
0\ge\la x^*,\ox\ra-\limsup_\nu\sum_{t\in T_\ve(\ox)}\lm_{t\nu}(\la\nabla g_t(\ox),\ox\ra-g_t(\ox))\ge\liminf_\nu\sum_{t\in T_\ve(\ox)}\lm_{t\nu} g_t(\ox)\ge\disp\frac{\ve}{\delta}\la x^*,\tilde x\ra.\end{aligned}$$ By passing to a subnet and combining this with (\[4.19\]), we get $$(x^*,\la x^*,\ox\ra)\in\cl\cone\big\{(\nabla g_t(\ox),\la\nabla g_t(\ox),\ox\ra-g_t(\ox))\big|\;t\in T\big\}+\{0\}\times\disp[\frac{\ve}{\delta}\la x^*,\tilde x\ra,0]$$ for all $\ve\in(0,\tilde \ve)$, which implies that $x^*$ belongs to the set in (\[4.18\]) by taking $\ve\dn 0$.
Involving further the NFMCQ condition, we claim the equality $$\begin{aligned}
\label{4.21}
\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}=\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}.\end{aligned}$$ The inclusion “$\supset$” in is obvious since $T(\ox)\subset T_\ve(\ox)$ for all $\ve>0$. To justify the converse inclusion, pick any $x^*$ belonging to the left-hand side of (\[4.21\]). By the NFMCQ condition, it follows from (\[4.18\]) that there is $\lm\in\tR$ such that $$\begin{aligned}
\label{4.22}
(x^*,\la x^*,\ox\ra)=\sum_{t\in T}\lm_t\big(\nabla g_t(\ox),\la\nabla g_t(\ox),\ox\ra-g_t(\ox)\big),\end{aligned}$$ which readily yields the equalities $$\begin{aligned}
0=\sum_{t\in T}\lm_t\la\nabla g_t(\ox),\ox\ra-\sum_{t\in T}\lm_t\big(\la\nabla g_t(\ox),\ox\ra-g_t(\ox)\big)=\sum_{t\in T}\lm_t g_t(\ox).\end{aligned}$$ Since $g_t(\ox)\le 0$, we get $\lm_tg_t(\ox)=0$ for all $t\in T$. Combining this with (\[4.22\]) gives us $$\begin{aligned}
x^*\in\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\},\end{aligned}$$ which implies the inclusion “$\subset$” in (\[4.21\]). To complete the proof of the theorem, we combine the obtained equality with finally Theorem \[thm41\]. $\h$
Observe from Proposition \[fm\] that formula (\[4.16\]) holds under our standing assumptions (SA) and the MFCQ condition at $\ox$ when $T$ is a finite index set. Furthermore, the formula for the limiting normal cones (\[4.17\]) is also satisfied if all the functions $g_t$ are strictly differentiable at $\ox$. It follows from Proposition \[fm\] that Corollary \[si\] can be derived from a semi-infinite version of Theorem \[thm42\] in addition to the assumptions of this corollary we suppose that the function $t\in T\mapsto g_t(\ox)$ is continuous in $T$.
The next example shows that the PMFCQ condition cannot be replaced by the EMFCQ one in Theorem \[thm42\] to ensure the unperturbed normal cone representations and in the presence of the NFMCQ.
\[ex4\][**(EMFCQ combined with NFMCQ does not ensure the unperturbed normal cone representations).**]{} [We revisit the semi-infinite inequality constraint system in Example \[ex1\]. It is shown there that this system satisfied the EMFCQ but not PMFCQ at $\ox=(-1,0)$. It is easy to check that the set $$\begin{aligned}
\cone\big\{\big(\nabla g_t(\ox),\la\nabla g_t(\ox),\ox\ra-g_t(\ox)\big)\big|\;t\in T\big\}&=&\cone\big((1,0,-1)\cup\{(t,0,0)\big|t\in (0,1]\}\big)\\
&=&\big\{x\in\R^3\big|\;x_1+x_3\ge 0,\;x_1\ge 0\ge x_3,\;x_2=0\big\}\end{aligned}$$ is closed in $\R^3$, i.e., the NFMCQ condition holds at $\ox$. Observe however that both representations (\[4.16\]) and (\[4.17\]) are not satisfied for this system since we have $$\begin{aligned}
\Hat N(\ox;\O)=N(\ox;\O)\not=\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}=\cone\{(1,0)\}=\R_+\times\{0\}.\end{aligned}$$]{}
Now we present a consequence of Theorem \[thm42\] with the corresponding discussions.
\[cv\] [**(normal cone for infinite convex systems).**]{} Assume that all the functions $g_t$, $t\in T$, in are convex and uniformly Fréchet differentiable and that $h=A$ is a surjective continuous linear operator. Suppose further that system [(\[3.1\])]{} satisfies the PMFCQ $($equivalently the SSC$)$ at $\ox\in\O$. Then the normal cone to $\O$ at $\ox$ in sense of convex analysis is computed by $$\begin{aligned}
N(\ox;\O)=\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}+A^*(Y^*).\end{aligned}$$ If in addition the NFMCQ holds at $\ox$, then we have $$\begin{aligned}
\label{cnm}
N(\ox;\O)=\disp\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}+A^*(Y^*).\end{aligned}$$
[**Proof.**]{} It follows directly from Proposition \[ss\] and Theorem \[thm42\]. $\h$
For $h=0$ in the equality in (\[cnm\]) can be deduced from [@DMN2 Corollary 3.6] under another [*Farkas-Minkowski Constraint Qualification*]{} (FMCQ) defined as follows:
[**(FMCQ)**]{} [*The conic hull [cone]{}$\{\epi g_t^*|\; t\in T\}$ is weak$^*$ closed in $X^*\times\R$ under the additional assumption that the functions $g_t$ are l.s.c., where $$\ph^*(x^*):=\sup\big\{\la x^*,x\ra-\ph(x)\big|\;x\in X\big\},\quad x^*\in X^*,$$ stands for the Fenchel conjugate of a convex function.*]{}
It is worth noting that the above FMCQ condition is a [*global*]{} property, and hence formula (\[cnm\]) holds at every $\ox\in\O$. By the contrary, our new NFMCQ condition (\[fm\]) is constructed at a fixed point $\ox\in\O$. The next example shows that the combination of the PMFCQ (or the SSC) and the NFMCQ conditions for infinite convex inequality systems is not stronger than the FMCQ one.
\[ex5\] [**(PMFCQ combined with NFMCQ does not imply FMCQ for convex inequality systems).**]{} [Define a function $g_t:\R^2\to\R$ by $g_t(x_1,x_2):=tx_1^2-x_2$ for all $(x_1,x_2)\in\R^2$ and $t\in T:=(0,1)$, and let $\ox=(0,0)\in\R^2$. It is easy to see that all the functions $g_t$, $t\in T$, are convex and differentiable and that the standing assumptions are satisfied. For each $t\in T$ we have $$g_t^*(a_1,a_2)=\sup_{(x_1,x_2)\in\R^2}\left\{a_1x_1+a_2x_2-tx_1^2+x_2\right\}=\left\{
\begin{array}{ll}
\disp\frac{a_1^2}{4t} &\mbox{if $a_2=-1$},\\
\infty &\mbox{otherwise}.
\end{array}
\right.$$ This implies that $\epi g_t^*=\{(a,-1,\frac{a^2}{4t}+r)|\;a\in \R,\;r\ge 0\}$, which yields in turn that $$C:=\cone\big\{\epi g_t^*\big|\;t\in T\big\}=\cone\Big\{(a,-1,\frac{a^2}{4}+r)\Big|\;a\in \R,\;r\ge 0\Big\}.$$ The latter set is not closed in $\R^3$ since $\{0\}\times\{0\}\times\R_+\not\subset C$ while $\{0\}\times\{0\}\times \R_+\subset{\rm cl}C$. Moreover, we see that $\nabla g_t(\ox)=(0,-1)$ for all $t\in T$, and then the PMFCQ is satisfied. Furthermore, it follows that the set $$\cone\big\{(\nabla g_t(\ox),\la\nabla g_t(\ox),\ox\ra-g_t(\ox))\big|\;t\in T\big\}=\cone\{(0,-1, 0)\}=\{0\}\times\R_-\times\{0\}$$ is closed in $\R^3$. Hence the PMFCQ and NFMCQ conditions hold but the FMCQ does not.]{}
Finally in this section, we give specifications of obtained normal cone representations in the case linear infinite systems.
\[clm\] [**(normal cone representations for linear infinite constraint systems).**]{} Consider the constraint system with $g_t(x)=\la a^*_t,x\ra-b_t$ for all $t\in T$, and let $h=A:X\to Y$. Assume that $A$ is a surjective continuous linear operator and that the coefficient set $\{a^*_t|\;t\in T\}$ is bounded in $X^*$. If the SSC condition holds at $\ox$, then $$\begin{aligned}
N(\ox;\O)=\disp\bigcap_{\ve>0}\cl\cone\big\{a^*_t\big|\;t\in T_\ve(\ox)\big\}+A^*(Y^*)\end{aligned}$$ for the feasible set $\O:=\{x\in X|\;Ax=0,\;\la a^*_t,x\ra-b_t\le 0,\;t\in T\}$. On the other hand, assuming the weak$^*$ closedness of $\cone\{(a^*_t, b_t)|\;t\in T\}$ in $X^*\times\R$ and that $h=0$ gives us $$\begin{aligned}
N(\ox;\O)=\disp\cone\big\{a^*_t\big|\;t\in T(\ox)\big\}.\end{aligned}$$
[**Proof.**]{} The first statement is a specification of Corollary \[cv\]. The second one follows from the proofs given in [@CLMP1 Proposition 3.1] and [@CLMP2 Theorem 3.2] by using the classical Farkas Lemma for linear infinite systems. $\h$
Optimality Conditions in Nonlinear Infinite Programming
=======================================================
In this section we employ general principles in optimization and the calculus results on computing the normal cones to the infinite constraint sets in Section 4 to deriving necessary optimality conditions for problems of infinite and semi-infinite programming. We confine ourselves to optimality conditions of the “lower" subdifferential type conventional in minimization. Condition of the other (“upper" or superdifferential) type can be derived from the calculus results of Section 4 using an approach developed in [@M1 Chapter 5]; see also the recent paper [@CLMP2] for the implementation of the latter approach in the case of semi-infinite and infinite programs with linear constraints.
Our first theorem in this section concerns infinite programs of type in arbitrary Banach spaces involving Fréchet differentiable cost functions.
\[thm43\] [**(necessary optimality conditions for differentiable infinite programs in general Banach spaces).**]{} Let $\ox$ be a local minimizer of the infinite program [(\[1.1\])]{} under the PMFCQ condition imposed on the constraints at $\ox$. Suppose further that the inequality constraint functions $g_t$, $t\in T$, are uniformly Fréchet differentiable at $\ox$ and the cost function $f$ is Fréchet differentiable at this point. Then we have the inclusion $$\begin{aligned}
\label{5.1}
0\in\nabla f(\ox)+\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}+\nabla h(\ox)^*(Y^*).\end{aligned}$$ If in addition the NFMCQ holds at $\ox$, then there exist multipliers $\lm\in\tR$ and $y^*\in Y^*$ satisfying the differential KKT condition $$\begin{aligned}
\label{5.2}
0=\nabla f(\ox)+\disp\sum_{t\in T(\ox)}\lm_t\nabla g_t(\ox)+\nabla h(\ox)^*y^*.\end{aligned}$$
[**Proof.**]{} It is clear that $\ox$ is a local optimal solution to the following unconstrained optimization problem with the [*infinite penalty*]{}: $$\begin{aligned}
\label{5.3}
\mbox{minimize }\;f(x)+\delta(x;\O),\end{aligned}$$ where $\O$ is the feasible constraint set . Applying the generalized Fermat rule to the latter problem (see, e.g., [@M1 Proposition 1.114]), we have $$\begin{aligned}
\label{5.4}
0\in\Hat\partial\big(f+\delta(\cdot;\O)\big)(\ox).\end{aligned}$$ Since $f$ is Fréchet differentiable at $\ox$, it follows from the sum rule of [@M1 Theorem 1.107 ] applied to and from the first relationship in (\[2.4\]) that $$\label{5.4a}
0\in\nabla f(\ox)+\Hat\partial\delta(\ox;\O)(\ox)=\nabla f(\ox)+\Hat N(\ox;\O).$$ Now using the Fréchet normal cone representation of Theorem \[thm41\] in , we arrive at (\[5.1\]). The second part (\[5.2\]) of this theorem readily follows from Theorem \[thm42\]. $\h$
The next theorem establishes necessary conditions for local minimizers of infinite programs (\[1.1\]) with general nonsmooth cost functions in the framework of Asplund spaces.
\[thm44\] [**(necessary optimality conditions for nonconvex infinite programs defined on Asplund spaces, I).**]{} Let $\ox$ be a local minimizer of problem [(\[1.1\])]{}, where the domain space $X$ is Asplund while the image space $Y$ is arbitrary Banach. Suppose that the constraint functions $g_t$, $t\in T$, are uniformly strictly differentiable at $\ox$, that the cost function $f$ is l.s.c. around $\ox$ and SNEC at this point, and that the qualification condition $$\begin{aligned}
\label{5.5}
\partial^\infty f(\ox)\cap\Big[-\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}-\nabla h(\ox)^*(Y^*)\Big]=\{0\}\end{aligned}$$ is fulfilled; the latter two assumptions are automatic when $f$ is locally Lipschitzian around $\ox$. If the PMFCQ condition holds at $\ox$, then $$\begin{aligned}
\label{5.6}
0\in\partial f(\ox)+\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}+\nabla h(\ox)^*(Y^*).\end{aligned}$$ If in addition we assume that the NFMCQ holds at $\ox$ and replace by $$\begin{aligned}
\label{5.5a}
\partial^\infty f(\ox)\cap\Big[-\cone\big\{\nabla g_t(\ox)\big|\;t\in T(\ox)\big\}-\nabla h(\ox)^*(Y^*)\Big]=\{0\},\end{aligned}$$ then there exist multipliers $\lm\in\tR$ and $y^*\in Y^*$ such that the following subdifferential KKT condition is satisfied: $$\begin{aligned}
\label{5.6a}
0\in\partial f(\ox)+\disp\sum_{t\in T(\ox)}\lm_t\nabla g_t(\ox)+\nabla h(\ox)^*y^*.\end{aligned}$$
[**Proof.**]{} Observe first that the feasible set $\O$ is locally closed around $\ox$. Indeed, it follows from (\[3.4\]) that there are $\gg>0$ and $\eta>0$ sufficiently small such that $$\begin{aligned}
\|h(x)-h(x^\prime)\|\le(\|\nabla h(\ox)\|+\gg)\|x-x^\prime\|\;\mbox{ and }\;\|g_t(x)-g_t(x^\prime)\|\le\sup_{\tau\in T}(\|\nabla g_\tau(\ox)\|+\gg)\|x-x^\prime\|\end{aligned}$$ for all $x,x^\prime\in\B_\eta(\ox)$ and $t\in T$. Picking any sequence $\{x_n\}\subset\O\cap\B_\eta(\ox)$ converging to some $x_0$ as $n\to\infty$, we have $$\begin{aligned}
\|h(x_0)\|\le(\|\nabla h(\ox)\|+\gg)\|x_n-x_0\|\;\mbox{ and }\;g_t(x_0)\le\sup_{\tau\in T}(\|\nabla g_\tau(\ox)\|+\gg)\|x_n-x_0\|+g_t(x_n)\end{aligned}$$ for each $t\in T$ and $n\in\IN$. By passing to the limit as $n\to\infty$, the latter yields that $h(x_0)=0$ and $g_t(x_0)\le 0$ for all $t\in T$, i.e., $x_0\in\O\cap\B_\eta(\ox)$, which justifies the local closedness of the feasible set $\O$ around $\ox$.
Employing now the generalized Fermat rule to the solution $\ox$ of with the closed set $\O$ and using [@M1 Theorem 3.36] on the sum rule for basic/limiting subgradients in Asplund spaces when $f$ is SNEC at $\ox$ yield that $$\label{sr}
0\in\partial\big(f+\delta(\cdot;\O)\big)(\ox)\subset\partial f(\ox)+\partial\delta(\ox;\O)=\partial f(\ox)+N(\ox;\O)$$ provided that $\partial^\infty f(\ox)\cap\big(-N(\ox;\O)\big)=\{0\}$. We apply further to both latter conditions the limiting normal cone representation of Theorem \[thm41\]. This gives us the optimality condition under the fulfillment of and the PMFCQ at $\ox$. Applying finally Theorem \[thm42\] instead of Theorem \[thm41\] in the setting above, we arrive at the KKT condition under the assumed NFMCQ at $\ox$ and , which completes the proof of the theorem. $\h$
An important ingredient in the proof of Theorem \[thm44\] is applying the subdifferential sum rule from [@M1 Theorem 3.36 ] to the sum $f+\dd(\cdot;\O)$, which requires that either $f$ is SNEC at $\ox$ or $\O$ is SNC at this point. While the first possibility was used above, now we are going to explore the second alternative. The next proposition presents verifiable conditions ensuring the SNC property of the feasible set $\O$ at $\ox$.
\[asp\] [**(SNC property of feasible sets in infinite programming).**]{} Let $X$ be an Asplund space, and let $\dim Y<\infty$ in the framework of . Assume that all the functions $g_t$, $t\in T$, are Fréchet differentiable around some $\ox\in\O$ and that the corresponding derivative family $\{\nabla g_t\}_{t\in T}$ is equicontinuous around this point, i.e., there exists $\ve>0$ such that for each $x\in\B_\ve(\ox)$ and each $\gg>0$ there is $0<\tilde\ve<\ve$ with the property $$\begin{aligned}
\label{5.7}
\|\nabla g_t(x^\prime)-\nabla g_t(x)\|\le\gg\;\mbox{ whenever }\;x^\prime\in\B_{\tilde\ve}(x)\cap \O\;\mbox{ and }\;t\in T.\end{aligned}$$ Then the feasible set $\O$ in is locally closed around $\ox$ and SNC at this point provided that the PMFCQ condition holds at $\ox$.
[**Proof.**]{} Consider first the set $\O_1:=\{x\in X|\;g_t(x)\le 0,\;t\in T\}$. By using arguments similar to the proof of Theorem \[thm44\], we justify the local closedness of $\O_1$ around $\ox$. Now let us prove that $\O_1$ is SNC at this point. To proceed, pick any sequence $(x_n,x^*_n)\in\O_1\times X^*$, $n\in\IN$, satisfying $$\begin{aligned}
x_n\st{\O_1}\to\ox,\;x^*_n\in\Hat N(x_n;\O_1)\;\mbox{ and }\;x^*_n\st{w^*}\to 0\;\mbox{ as }\;n\to\infty.\end{aligned}$$ Taking into account, we see that the functions $g_t$, $t\in T$ satisfy the standing assumptions (SA) at $x_n$ for all $n\in\IN$ sufficiently large. Moreover, the proof showing that assumption (\[3.4\]) holds at $x_n$ follows from the discussions right after Corollary \[mc\]. Since the PMFCQ condition holds at $\ox$, there exist $\delta>0$, $\ve>0$, and $\tilde x\in X$ such that $\la\nabla g_t(\ox),\tilde x\ra\le-2\delta$ for all $t\in T_{2\ve}(\ox)$. Observe that $T_\ve(x_n)\subset T_{2\ve}(\ox)$ for all large $n\in\IN$. Indeed, whenever $t\in T_\ve(x_k)$ we have $$\begin{aligned}
0\ge g_t(\ox)&\ge& g_t( x_n)-\la \nabla g_t(\ox),x_n-\ox\ra-\|x_n-\ox\|s(\|x_k-\ox\|)\\
&\ge&-\ve-\sup_{\tau\in T}\|\nabla g_\tau(\ox)\|\cdot\|x_n-\ox\|-\|\tilde x_n-\ox\|s(\|x_n-\ox\|)\ge-2\ve\end{aligned}$$ for all large $n\in\IN$, where $s(\cdot)$ is defined in . Further, it follows from (\[5.7\]) that $$\begin{aligned}
\la\nabla g_t( x_n),\tilde x\ra\le\la\nabla g_t(\ox),\tilde x\ra+\|\nabla g_t( x_n)-\nabla g_t(\ox)\|\cdot\|\tilde x\|\le-2\delta+\|\nabla g_t( x_n)-\nabla g_t(\ox)\|\cdot\|\tilde x\|\le-\delta\end{aligned}$$ when $n\in\IN$ is sufficiently large. Hence we suppose without loss of generality that $$\begin{aligned}
\label{5.8}
T_\ve(x_n)\subset T_{2\ve}(\ox)\;\mbox{ and }\;\sup_{t\in T_\ve( x_n)}\la\nabla g_t(x_n),\tilde x\ra\le-\delta\;\mbox{ whenever }\;n\in\IN.\end{aligned}$$ Applying now Theorem \[thm41\] in this setting, we have that for each $n\in\IN$ there exists a net $\{\lm_{n_\nu}\}_{\nu\in\N}\subset\Tilde \R_+^{T_\ve( x_n)}$ such that $$\begin{aligned}
x^*_n=w^*-\lim_\nu\sum_{t\in T_\ve(x_n)}\lm_{tn_\nu}\nabla g_t(x_n).\end{aligned}$$ Combining this with (\[5.8\]) yields that $$\begin{aligned}
\la x^*_n,\tilde x\ra=\lim_\nu\sum_{t\in T_\ve(x_n)}\lm_{tn_\nu}\la\nabla g_t(x_n),\tilde x\ra\le-\delta\liminf_\nu\sum_{t\in T_\ve( x_n)}\lm_{tn_\nu}.\end{aligned}$$ Furthermore, for each $x\in X$ we get the relationships $$\begin{aligned}
\la x^*_n,x\ra&=&\liminf_\nu\sum_{t\in T_\ve( x_k)}\lm_{tn_\nu}\la\nabla g_t(x_n),x\ra\le\liminf_\nu\sum_{t\in T_\ve(x_n)}\lm_{tn_\nu}\sup_{\tau\in T}\|\nabla g_\tau(x_n)\|\cdot\|x\|\\
&\le&-\frac{\la x^*_n,\tilde x\ra}{\delta}\sup_{\tau\in T}\|\nabla g_\tau(x_n)\|\cdot\|x\|,\end{aligned}$$ which imply that $\|x^*_n\|\le-\frac{\la x^*_n,\tilde x\ra}{\delta}\sup_{\tau\in T}\|\nabla g_\tau(x_n)\|$ for all $n\in\IN$. Since $x^*_n\st{w^*}\to 0$, it follows from the latter that $\|x^*_n\|\to 0$ as $n\to\infty$ and thus the set $\O_1$ is SNC at $\ox$.
Consider now the set $\O_2:=\{x\in X|\;h(x)=0\}$, which is obviously closed around $\ox$. It follows from [@M1 Theorem 1.22] and finite dimensionality of $Y$ that $\O_2$ is SNC at $\ox$. Moreover, we get from [@M1 Theorem 1.17] that $N(\ox;\O_2)=\nabla h(\ox)^*(Y^*)$. Thus for any $x^*\in N(\ox;\O_1)\cap(-N(\ox;\O_2))$ there is $y^*\in Y^*$ such that $x^*+\nabla h(\ox)^*y^*=0$, and then $$\la x^*,\tilde x\ra=-\la\nabla h(\ox)^*y^*,\tilde x\ra=-\la y^*,\nabla h(\ox)\tilde x\ra=0.$$ Since $x^*\in N(\ox;\O_1)$, we find by Theorem \[thm41\] such a net $\{\lm_\nu\}_{\nu\in\N}\in\tR$ that $$\begin{aligned}
x^*=w^*-\lim_\nu\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\nabla g_t(\ox),\end{aligned}$$ which yields in turn that $$\begin{aligned}
0=\la x^*,\tilde x\ra=\lim_\nu\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\la \nabla g_t(\ox),\tilde x\ra\le-2\delta\liminf_{\nu}\sum_{t\in T_\ve(\ox)} \lm_{t\nu}.\end{aligned}$$ This ensures the relationships $$\la x^*,x\ra=\liminf_\nu\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\la\nabla g_t(\ox),x\ra\le\liminf_\nu\sum_{t\in T_\ve(\ox)}\lm_{t\nu}\sup_{\tau\in T}\|\la \nabla g_\tau(\ox)\|\|x\|=0$$ for all $x\in X$. Hence we have $x^*=0$, and so $N(\ox;\O_1)\cap(-N(\ox;\O_2))=\{0\}$. It finally follows from [@M1 Corollary 3.81] that the intersection $\O=\O_1\cap\O_2$ is SNC at $\ox$, which thus completes the proof of the proposition. $\h$
Observe that the assumption $\dim Y<\infty $ is essential in Proposition \[asp\]. To illustrate this, consider a particular case of when $T=\emp$. It follows from [@M1 Theorem 1.22] that the inverse image $\O=h^{-1}(0)$ is SNC at $\ox\in\O$ if and only if the set $\{0\}$ is SNC at $0\in Y$. Since $N(0;\{0\})=Y^*$, the latter holds if and only if the weak$^*$ topology in $Y^*$ agrees with the norm topology in $Y^*$, which is only the case of $\dim Y<\infty$ by the classical Josefson-Nissenzweig theorem from theory of Banach spaces.
Now we are ready to derive an aforementioned alternative counterpart of Theorem \[thm44\].
\[thm45\] [**(necessary optimality conditions for nonconvex infinite programs defined on Asplund spaces, II).**]{} Let $\ox$ be a local minimizer of infinite program [(\[3.1\])]{} under the assumptions of Proposition [\[asp\]]{}. Suppose also that $f$ is l.s.c. around $\ox$ and that the qualification condition is satisfied. Then we have the optimality condition . If in addition we assume that the NFMCQ holds at $\ox$ and replace by , then there exist multipliers $\lm\in\tR$ and $y^*\in Y^*$ such that the subdifferential KKT condition .
[**Proof.**]{} It is similar to the proof of Theorem \[thm44\] with applying Proposition \[asp\] on the SNC and closedness property of $\O$ in the sum rule of [@M1 Theorem 3.36]. $\h$
The next result provides necessary and sufficient optimality conditions for convex problems of infinite programming in general Banach spaces.
\[conv1\] [**(necessary and optimality conditions for convex infinite programs).**]{} Let both spaces $X$ and $Y$ be Banach. Assume that all the functions $g_t$, $t\in T$, are convex and uniformly Fréchet differentiable and that $h=A$ is a surjective continuous linear operator. Suppose further that the cost function $f$ is convex and continuous at some point in $\O$. If the PMFCQ condition $($equivalently the SSC condition) holds at $\ox$, then $\ox$ is a global minimizer of problem if and only if $$\begin{aligned}
0\in\partial f(\ox)+\disp\bigcap_{\ve>0}\cl\cone\big\{\nabla g_t(\ox)\big|\;t\in T_\ve(\ox)\big\}+A^*(Y^*).\end{aligned}$$ If in addition the NFMCQ condition holds, then $\ox$ is a global minimizer of problem if and only if there exist $\lm\in \tR$ and $y^*\in Y^*$ such that $$\begin{aligned}
\label{5.9}
0\in\partial f(\ox)+\disp\sum_{t\in T(\ox)}\lm_t\nabla g_t(\ox)+A^*y^*.\end{aligned}$$
[**Proof.**]{} Observe that $\ox$ is a global minimizer of problem (\[1.1\]) if and only if it is a global minimizer of the convex unconstrained problem (\[5.3\]), which is equivalent to the fact that $$\begin{aligned}
0\in\partial\big(f+\delta(\cdot;\O)\big)(\ox).\end{aligned}$$ Applying the convex subdifferential sum rule to the latter inclusion, we conclude that $\ox$ is a global minimizer of problem (\[1.1\]) if and only if $$\begin{aligned}
0\in\partial f(\ox)+\partial\delta(\ox;\O)=\partial f(\ox)+N(\ox;\O).\end{aligned}$$ The rest of the proof follows from Corollary \[cv\]. $\h$
Note that some versions of necessary optimality condition of the KKT type were derived in [@CLMP2 Theorems 3.1 and 3.2] for infinite problems with linear constraints but possibly nonconvex cost functions under the SSC and the linear counterpart of the FMCQ; see Example \[ex5\] and the corresponding discussions above.
Observe also that the results of Theorem \[thm45\] and Theorem \[conv1\] are formulated with no change in the case of semi-infinite programs, while in Theorem \[thm43\] we just drop the SNEC assumption on $f$, which holds automatically when $X$ is finite-dimensional.
In conclusion we present a consequence of our results for the classical framework of semi-infinite programming while involving nonsmooth cost functions.
\[cor6\] [**(necessary optimality conditions for semi-infinite programs with compact index sets).**]{} Let $\ox$ be a local minimizer of program [(\[1.1\])]{}, where both spaces $X$ and $Y$ are finite-dimensional with $\dim Y<\dim X$. Assume that the index set $T$ in is a compact metric space, that the mappings $(x,t):X\times T\mapsto g_t(x)$ and $(x,t):X\times T\mapsto\nabla g_t(x)$ are continuous, and that the cost function $f$ is l.s.c. around $\ox$ with the fulfillment of . If in addition the EMFCQ holds at $\ox$, then there exist multipliers $\lm\in\tR$ and $y^*\in Y^*$ satisfying the subdifferential KKT condition .
[**Proof.**]{} By Proposition \[sfm\] we have that the NFMCQ condition holds at $\ox$ under the assumptions made. Then this corollary follows directly from Theorem \[thm44\]. $\h$
When $f$ is smooth around $\ox$, assumption holds automatically while reduced to the differential KKT condition . Then Corollary \[cor6\] reduces to a well-known result in semi-infinite programming that can be found, e.g., in [@HK Theorem 3.3] and [@LS Theorem 2].
[99]{}
Anderson, E.J., Nash, P.: Linear Programming in Infinite-Dimensional Spaces. Wiley, Chicherster (1987)
Bonnans, J.F., Shapiro, A.: Perturbation Analysis of Optimization Problems. Springer, New York (2000)
Borwein, J.M., Zhu, Q.J.: Techniques of Variational Analysis. Springer, New York (2005)
Boţ, R.I., Grad, S.-M., Wanka, G.: On strong total Lagrange duality for convex optimization problems, J. Math. Anal. Appl. [**337**]{}, 1315–1325 (2008)
Cánovas, M.J., López, M.A., Mordukhovich, B.S., Parra J.: Variational analysis in semi-infinite and infinite programming, I: Stability of linear inequality systems of feasible solutions. SIAM J. Optim. [**20**]{}, 1504–1526 (2009)
Cánovas, M.J., López, M.A., Mordukhovich, B.S., Parra J.: Variational analysis in semi-infinite and infinite programming, II: Necessary optimality conditions. SIAM J. Optim. [**20**]{}, 2788–2806 (2010)
Chuong, T.D., Huy, N.Q., Yao, J.C.: Subdifferentials of marginal functions in semi-infinite programming. SIAM J. Optim., [**20**]{}, 1462–1477 (2009)
Dinh, N.; Goberna, M. A.; López, M. A.: From linear to convex systems: Consistency, Farkas’ lemma and applications. J. Convex Anal. [**13**]{}, 113–133 (2006)
Dinh, N., Goberna, M.A., López, M.A., Son, T.Q.: New Farkas-type results with applications to convex nfinite programming. ESAIM: Control Optim. Cal. Var. [**13**]{}, 580–597 (2007)
Dinh, N., Mordukhovich, B.S., Nghia, T.T.A.: Qualification and optimality conditions for DC programs with infinite programs. Acta Math. Vietnamica [**34**]{}, 125–155 (2009)
Dinh, N., Mordukhovich, B.S., Nghia, T.T.A.: Subdifferentials of value functions and optimality conditions for DC and bilevel infinite and semi-infinite programs. Math. Program. [**123**]{}, 101–138 (2010)
Fang, D. H., Li, C., Ng, K.F.: Constraint qualifications for extended Farkas’s lemmas and Lagrangian dualities in convex infinite programming. SIAM J. Optim. [**20**]{}, 1311–1332 (2009)
Goberna, M.A., López, M.A.: Linear Semi-Infinite Optimization. Wiley, Chichester (1998)
Goberna, M.A., López, M.A. (eds.): Semi-Infinite Programming: Recent Advances. Kluwer, Dordrecht (2001)
Hettich, R., Kortanek, K.O.: Semi-infinite programming: Theory, methods and applications. SIAM Review [**35**]{}, 380–429 (1993)
Hiriart-Urruty, J.B., Lemaréchal, C.: Fundamentals of Convex Analysis. Springer, Berlin (2001)
Jongen, H.Th., J.-J. Rückmann, J.-J., Stein, O.:, Generalized semi-infinite programming: A first-order optimality conditions and examples, Math. Program., [bf 83]{}, 145–158 (1998)
Jongen, H.Th., Twilt, F., Weber, G.W.: Semi-infinite optimization: Structure and stability of the feasible set. J. Optim. Theory Appl. [**72**]{}, 529–552 (1992)
Klatte, D., Henrion, R.: Regularity and stability in nonlinear semi-infinite optimization, in Semi-Infinite Programming, Reemstsen R. and Ruckmann, J.J., eds., pp. 69–102. Kluwer, Dordrecht (1998)
Li, C., Ng, K.F., Pong, T.K.: Constraint qualifications for convex inequality systems with applications in constrained optimization. SIAM J. Optim. [**19**]{}, 163–187 (2008)
López, M., Still, G.: Semi-infinite programming. Europ. J. Oper. Res. [**180**]{}, 491–518 (2007)
Mordukhovich, B.S.: Variational Analysis and Generalized Differentiation, I: Basic Theory. Springer, Berlin (2006)
Mordukhovich, B.S.: Variational Analysis and Generalized Differentiation, II: Applications. Springer, Berlin (2006)
Rockafellar, R.T., Wets, R. J-B.: Variational Analysis, Springer, Berlin (1998)
Schirotzek, W.: Nonsmooth Analysis, Springer, Berlin (2007)
Seidman, T.I.: Normal cones to infinite intersections, Nonlinear Anal. [**72**]{}, 3911–3917 (2010)
Shapiro, A.: Semi-infinite programming: Duality, discretization and optimality conditions. Optimization [**58**]{}, 133–161 (2009)
Stein, O.: Bilevel Strategies in Semi-Infinite Programming. Kluwer, Boston, Massachusetts (2003)
Zheng, X.Y., Yang, X.Q.: Lagrange multipliers in nonsmooth semi-infinite optimization, Math. Oper. Res., [**32**]{}, 168–181 (2007)
[^1]: Research was partially supported by the USA National Science Foundation under grants DMS-0603846 and DMS-1007132 and by the Australian Research Council under grant DP-12092508.
[^2]: Department of Mathematics, Wayne State University, Detroit, Michigan 48202; email: [email protected].
[^3]: Department of Mathematics, Wayne State University, Detroit, Michigan 48202; email: [email protected].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a group, a generalized torsion element is a non-identity element whose some non-empty finite product of its conjugates yields the identity. Such an element is an obstruction for a group to be bi-orderable. We show that the Weeks manifold, the figure-eight sister manifold, and the complement of Whitehead sister link admit generalized torsion elements in their fundamental groups. In particular, the Whitehead sister link, which is the pretzel link of type $(-2,3,8)$, can be generalized to hyperbolic pretzel links of type $(-2,3,2n)\ (n\ge 4)$. These give the first examples of hyperbolic links whose link groups admit generalized torsion elements.'
address: 'Department of Mathematic Education, Hiroshima University, 1-1-1 Kagamiyama, Higashi-hiroshima, 7398524, Japan'
author:
- Masakazu Teragaito
title: Generalized torsion elements and hyperbolic links
---
[^1]
Introduction
============
In a group $G$, a non-trivial element $g$ is called a *generalized torsion element* if some non-empty finite product of its conjugates is equal to the identity. That is, $g^{a_1}g^{a_2}\dots g^{a_k}=1$ for some $a_1,a_2,\dots, a_k\in G$. Here $g^a$ denotes a conjugate of $g$ by $a\in G$.
Every knot or link group is torsion free. However, it may contain a generalized torsion element. For example, all torus knot groups satisfy it. Naylor and Rolfsen [@NR] gave the first example of hyperbolic knot, which is the $(-2)$-twist knot, whose knot group contains a generalized torsion element. Then we showed that any negative twist knot enjoys the same property [@T].
The existence of generalized torsion element is an obstruction for a group $G$ to be bi-orderable. Recall that $G$ is said to be *bi-orderable* if it admits a strict total ordering which is invariant under multiplication from left and right sides; if $a<b$, then $gah<gbh$ for any $g$ and $h$ in $G$. If $G$ is bi-orderable, then any non-identity element $g$ is bigger or smaller than the identity. Let $g>1$. Since conjugation preserves the order, $g^a>1$ for any $a\in G$. Then any product of such elements is still bigger than $1$. Similarly for the case $g<1$. Thus any bi-orderable group has no generalized torsion element.
Conversely, even if $G$ has no generalized torsion element, we cannot claim that $G$ is bi-orderable [@B; @BL; @MR]. However, we expect that such phenomenon does not occur among $3$-manifold groups [@MT]. Many $3$-manifold groups are know to be not bi-orderable. For, any finitely generated bi-orderable group surjects on the infinite cyclic group (see [@CR]). In particular, if a $3$-manifold $M$ has finite first homology group, then $\pi_1(M)$ is not bi-orderable. In [@MT], we propose a conjecture that if $\pi_1(M)$ is not bi-orderable for a $3$-manifold $M$, then it contains a generalized torsion element. This is solved affirmatively for Seifert fibered manifolds, Solvable manifolds, and a few infinite families of hyperbolic manifolds.
In this paper, we first examine small volume cusped hyperbolic $3$-manifolds. The Weeks manifold is the unique closed orientable hyperbolic $3$-manifold of smallest volume [@GMM]. By [@CM], the minimal volume orientable hyperbolic $3$-manifold with one cusp is homeomorphic to either the figure-eight knot complement or the figure-eight sister manifold, which is also called the sibling manifold. Furthermore, the minimal volume orientable hyperbolic $3$-manifold with two cusps is homeomorphic to either the Whitehead link complement or the Whitehead sister manifold, which is the $(-2,3,8)$-pretzel link complement. The figure-eight knot complement and the Whitehead link complement have bi-orderable fundamental groups [@PR; @KR], so they do not admit a generalized torsion element in their fundamental groups.
Our first result claims that the others admit generalized torsion elements in the fundamental groups.
\[thm:main1\] The Weeks manifold, the figure-eight sister manifold and the Whitehead sister manifold admit generalized torsion elements in their fundamental groups.
As remarked above, the Whitehead sister manifold is the complement of the $(-2,3,8)$-pretzel link. Second, we generalize this to an infinite family of pretzel links.
\[thm:main2\] Let $L$ be the two-component pretzel link $P(-2,3,2n)$. If $n\ge 1$, then the link group $\pi_1(S^3-L)$ contains a generalized torsion element.
This theorem immediately gives the first examples of hyperbolic links whose link groups admit generalized torsion elements.
There are infinitely many hyperbolic two-component links whose link groups admit generalized torsion elements.
If $n\ge 4$, then $P(-2,3,2n)$ is hyperbolic [@MP]. The conclusion immediately follows from Theorem \[thm:main2\].
Throughout the paper, we use the notation $\bar{g}=g^{-1}$, $[g,h]=g^{-1}h^{-1}gh$ and $g^a=a^{-1}ga$ for the inverse, the commutator and the conjugate in a group.
The Weeks manifold and the figure-eight sister manifold
=======================================================
We start from the exterior $W$ of the Whitehead link. Then $\pi_1(W)$ has the following Wirtinger presentation $$\pi_1(W)=\langle a, b \mid aba\bar{b}\bar{a}bab=bab\bar{a}\bar{b}aba\rangle,$$ where $a$ and $b$ are meridians of the components as shown in Fig. \[fig:link\].
![The Whitehead link and the meridians $a$, $b$[]{data-label="fig:link"}](link.eps){width="4cm"}
The figure-eight sister manifold is described as the resulting manifold by performing $5/1$-Dehn filling on one boundary component of $W$ (see [@G]). Two components of the Whitehead link are interchangeable by an isotopy, so there is no ambiguity for the choice of boundary component.
\[thm:w\] For a slope $m/n\ (n\ge 1)$, let $W(m/n)$ be the resulting manifold by $m/n$-Dehn filling on one boundary component of the Whitehead link exterior $W$. If $m\ge 2n$, then $\pi_1(W(m/n))$ contains a generalized torsion element. In particular, the figure-eight sister manifold $W(5)$ satisfies this.
Perform $m/n$-surgery along the component with meridian $a$. Then the surgery yields a relation $$(\bar{b}^{ab\bar{a}\bar{b}}b)^n a^m=1,$$ because the longitude is $\bar{b}^{ab\bar{a}\bar{b}}b$. This gives $$\label{eq1}
(ba\bar{b}\bar{a}\bar{b}ab\bar{a})^na^m=1.$$ Hence $$\pi_1(W(m/n))=\langle a,b \mid
aba\bar{b}\bar{a}bab=bab\bar{a}\bar{b}aba,
(ba\bar{b}\bar{a}\bar{b}ab\bar{a})^na^m=1\rangle.$$
The first relation gives $$\label{eq2}
ba\bar{b}\bar{a}\bar{b}ab=a^{\bar{b}\bar{a}}a^b\bar{a}.$$ Hence (\[eq1\]) and (\[eq2\]) yield $(a^{\bar{b}\bar{a}}a^b\bar{a}^2)^na^m=1$, so we have $$\label{eq3}
UU^{a^2}U^{a^4}\cdots U^{a^{2(n-1)}}a^{m-2n}=1,$$ where $U=a^{\bar{b}\bar{a}}a^b$. Thus if $m\ge 2n$, then the left hand side of (\[eq3\]) is a product of conjugates of only $a$.
Since $H_1(W(m/n))=\mathbb{Z}\oplus \mathbb{Z}_{|m|}$ and the element $a$ goes to a generator of $\mathbb{Z}_{|m|}$ summand, $a$ is non-trivial in $\pi_1(W(m/n))$. Hence if $m\ge 2n$, then (\[eq3\]) shows that the element $a$ is a generalized torsion element in $\pi_1(W(m/n))$.
1. Since $W(1)$ is the exterior of the trefoil, $\pi_1(W(1))$ contains a generalized torsion element (see [@NR]).
2. If $m=-1$, then $W(-1/n)$ gives the exterior of the $n$-twist knot. For example, $W(-1)$ is the figure-eight knot exterior. Since the knot group of any positive twist knot is known to be bi-orderable [@C], $\pi_1(W(-1/n))$ does not contain a generalized torsion element for any $n\ge 1$. On the other hand, if $m=1$, then $\pi_1(W(1/n))$ contains a generalized torsion element, because the knot group of any negative twist knot admits a generalized torsion element [@T].
Let $M=W(m/n)$, and let $M(r)$ denote the $r$-Dehn filling on $M$. The Weeks manifold is $M(5/2)$ for $M=W(5)$ (see [@CD]).
\[cor:weeks\] If $m\ge 2n$. $\pi_1(M(r))$ contains a generalized torsion element for any slope $r\in \mathbb{Q}\cup \{1/0\}$. In particular, the Weeks manifold $M(5/2)$ satisfies this.
Since $\pi_1(M(r))$ is the quotient of $\pi_1(W(m/n))$, the relation (\[eq3\]) still holds in $\pi_1(M(r))$. Set $r=p/q$. The element $a$ projects to a generator of $\mathbb{Z}_{|m|}$ summand of $H_1(M(r)))=\mathbb{Z}_{|p|}\oplus \mathbb{Z}_{|m|}$, so it is nontrivial in $\pi_1(M(r))$. Hence $a$ remains to be a generalized torsion element in $\pi_1(M(r))$.
The Whitehead sister link and pretzel links
===========================================
Tunnel number one pretzel links
-------------------------------
Let $L$ be the pretzel link $P(-2,3,2n)$ for $n\ne 0$. In particular, $P(-2,3,8)$ is the Whitehead sister link. Let $G=\pi_1(S^3-L)$ be the link group. The purpose of this subsection is to obtain a presentation of $G$ with two generators and a single relation based on the fact that $L$ has tunnel number one.
Figure \[fig:tunnel\] shows the unknotting tunnel $\gamma$ for $L$. This means that the exterior of the regular neighborhood $N$ of $L\cup \gamma$ is a genus two handlebody $H$. Hence the exterior of $L$ is obtained from $H$ by attaching a $2$-handle along the co-core $\ell$ of the regular neighborhood of $\gamma$, which is regarded as a $1$-handle attached on the regular neighborhood of $L$. This implies that the link group $G$ has two generators and a single relation which comes from the $2$-handle addition.
![The pretzel link $P(-2,3,2n)$, where $n=3$, and its unknotting tunnel $\gamma$[]{data-label="fig:tunnel"}](pretzel.eps){width="10cm"}
To get a rank two presentation of $G$, we will trace the co-core $\ell$ on $\partial N$ during the unknotting transformation of $N$.
First, we replace the $2n$-twists on $L$ with $(-1/n)$-surgery on the additional unknotted circle as shown in Fig. \[fig:tunnel\]. Then Fig. \[fig:start\] shows $N$ and the co-core $\ell$.
![The regular neightborhood $N$ of $L\cup \gamma$ and the co-core $\ell$ on $\partial N$[]{data-label="fig:start"}](handle0.eps){width="6cm"}
As illustrated in Fig. \[fig:h1\], \[fig:h2\] and \[fig:h3\], we transform $N$. Here, the loop $\ell$ in Fig. \[fig:h2\] and \[fig:h3\] is described as a band sum of two circles for simplicity.
![The first move of $N$[]{data-label="fig:h1"}](handle1.eps){width="7cm"}
![The second move of $N$. Here, the loop $\ell$ is described as a band sum of two circles.[]{data-label="fig:h2"}](handle2.eps){width="10cm"}
![The third move of $N$[]{data-label="fig:h3"}](handle3.eps){width="8cm"}
Figure \[fig:h4\] shows the final form of $N$ with $\ell$ on $\partial N$. It is easy to see that the outside of $N$ is also a genus two handlebody $H$. Here, the loops $\alpha$ and $\beta$ bound mutually disjoint non-separating meridian disks of $H$. Hence if we take the generators $a$ and $b$ of $\pi_1(H)$ as the duals of $\alpha$ and $\beta$, then we can easily express $\ell$ as a word of $a$ and $b$ in $\pi_1(H)$ by following the intersection points between $\ell$ and $\alpha$ and $\beta$.
![The outsider of $N$ is a genus two handlebody $H$. The loops $\alpha$ and $\beta$ bound disjoint meridian disks of $H$. Here, $n=3$.[]{data-label="fig:h4"}](handle4.eps){width="10cm"}
\[prop:p\] The link group $G$ has the presentation $$\label{eq:p}
G=\langle a, b \mid
a\bar{b}^{n-1}a\bar{b}ab=ba\bar{b}a\bar{b}^{n-1}a\rangle.$$
The exterior of $L$ is obtained from the genus two handlebody $H$ by attaching a $2$-handle along the loop $\ell$. Let $a$ and $b$ be the generators of $\pi_1(H)$, which are duals of $\alpha$ and $\beta$ as shown in Fig. \[fig:h4\]. Hence the link group $G$ has a presentation with generators $a$ and $b$ and a single relation coming from $\ell$. We orient $\alpha$, $\beta$ and $\ell$ as in Fig. \[fig:h4\], and follow $\ell$ from the dot. Then we get the relation as in the statement.
Generalized torsion elements
----------------------------
\[lem:rel\] In $G$, $[a,b\bar{a}b^{n-1}\bar{a}b]=1$.
The relation of (\[eq:p\]) gives $U a^b = a^{\bar{b}} U$, where $U=a \bar{b} ^{n-1} a$. Hence we have $U^b a^{b^2}=aU^b$, so $a=a^{b^2 \bar{U}^b}$. This gives $[a,b^2\bar{U}^b]=1$, which yields the conclusion.
\[lem:decomp\] Let $w(\bar{a},b)$ be a word containing only $\bar{a}$ and $b$. Then the commutator $[a,w(\bar{a},b)]$ can be expressed as a product of conjugates of the commutator $[a,b]$.
In general, we have an equation $[a,uv]=[a,v][a,u]^v$. Since $[a,\bar{a}]=1$, $[a,w(\bar{a},b)]$ is decomposed into a product of conjugates of only $[a,b]$.
\[thm:pretzel\] If $n\ge 1$, then $G$ admits a generalized torsion element.
By Lemma \[lem:decomp\], $[a,b\bar{a}b^{n-1}\bar{a}b]$ is expressed as a product of conjugates of $[a,b]$ if $n\ge 1$. We know that $[a,b]\ne 1$ in $G$, because $G$ is not abelian. (The only knots and links whose groups are abelian are the unknot and the Hopf link.) Then Lemma \[lem:rel\] implies that $[a,b]$ is a generalized torsion element in $G$.
When $n=0$, then link $L$ is the connected sum of the trefoil and the Hopf link. Since the knot group of the trefoil contains a generalized torsion element, so does $G$. For $n<0$, our argument in the proof of Theorem \[thm:pretzel\] does not work.
**Proof of Theorem \[thm:main1\].** This follows from Theorems \[thm:w\] and \[thm:pretzel\] and Corollary \[cor:weeks\]. $\Box$
**Proof of Theorem \[thm:main2\].** This immediately follows from Theorem \[thm:pretzel\]. $\Box$
Problems
--------
Kin and Rolfsen [@KR] show that the pretzel links $P(-2,2k+1,2n)$ and $P(-2,-2k+1,2n)$ for $k\ge 1$ and $n\ge 2$ do not have bi-orderable link groups. Hence their link groups are expected to admit generalized torsion elements beyond our Theorem \[thm:main2\]. Fortunately, these pretzel links still have tunnel number one. Thus it is possible to apply our procedure to get a presentation of link group with two generators and a single relation. We tried this, but we could not find a generalized torsion element, because of the complicated relation.
The exterior of pretzel links $P(-2,3,2n)$ is obtained from the magic manifold by suitable Dehn filling. If the magic manifold contains a generalized torsion element in its fundamental group, then there would be a chance for the element to give a generalized torsion element for $P(-2,3,2n)$ as in the proof of Corollary \[cor:weeks\]. However, the result of [@KR] seems to suggest that the magic manifold would have bi-orderable fundamental group, so there is no generalized torsion element.
[16]{}
I. Agol, [*The minimal volume orientable hyperbolic $2$-cusped $3$-manifolds*]{}, Proc. Amer. Math. Soc. [**138**]{} (2010), no. 10, 3723–3732.
V. V. Bludov, [*An example of an unordered group with strictly isolated identity element*]{}, Algebra and Logic [**11**]{} (1972), 341–349.
V. V. Bludov and E. S. Lapshina, [*On ordering groups with a nilpotent commutant*]{} (Russian), Sibirsk. Mat. Zh. [**44**]{} (2003), no. 3, 513–520; translation in Siberian Math. J. [**44**]{} (2003), no. 3, 405–410.
D. Calegari and N. M. Dunfield, [*Laminations and groups of homeomorphisms of the circle*]{}, Invent. Math. [**152**]{} (2003), no. 1, 149–204.
C. Cao and G. R. Meyerhoff, [*The orientable cusped hyperbolic $3$-manifolds of minimum volume*]{}, Invent. Math. [**146**]{} (2001), no. 3, 451–478.
A. Clay, C. Desmarais and P. Naylor, [*Testing bi-orderability of knot groups*]{}, Canad. Math. Bull. [**59**]{} (2016), no. 3, 472–482.
A. Clay and D. Rolfsen, [*Ordered groups and topology*]{}, Graduate Studies in Mathematics, 176. American Mathematical Society, Providence, RI, 2016.
D. Gabai, R. Meyerhoff and P. Milley, [*Minimum volume cusped hyperbolic threemanifolds*]{}, J. Amer. Math. Soc. [**22**]{} (2009), no. 4, 1157–1215.
C. McA. Gordon, [*Dehn filling: a survey*]{}, Knot theory (Warsaw, 1995), 129–144, Banach Center Publ., 42, Polish Acad. Sci. Inst. Math., Warsaw, 1998.
E. Kin and D. Rolfsen, [*Braids, orderings, and minimal volume cusped hyperbolic $3$-manifolds*]{}, Groups Geom. Dyn. [**12**]{} (2018), no. 3, 961–1004.
B. Martelli and C. Petronio, [*Dehn filling of the “magic” $3$-manifold*]{}, Comm. Anal. Geom. [**14**]{} (2006), no. 5, 969–1026.
K. Motegi and M. Teragaito, [*Generalized torsion elements and bi-orderability of $3$-manifold groups*]{}, Canad. Math. Bull. [**60**]{} (2017), no. 4, 830–844.
R. Mura and A. Rhemtulla, [*Orderable groups*]{}, Lecture Notes in Pure and Applied Mathematics, 27. Marcel Dekker, Inc., New York-Basel, 1977.
G. Naylor and D. Rolfsen, [*Generalized torsion in knot groups*]{}, Canad. Math. Bull. [**59**]{} (2016), no. 1, 182–189.
B. Perron and D. Rolfsen, [*On orderability of fibred knot groups*]{}, Math. Proc. Cambridge Philos. Soc. [**135**]{} (2003), no. 1, 147–153.
M. Teragaito, [*Generalized torsion elements in the knot groups of twist knots*]{}, Proc. Amer. Math. Soc. [**144**]{} (2016), no. 6, 2677–2682.
[^1]: This work was supported by JSPS KAKENHI Grant Number 20K03587.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the question of complete gauge independence of the fermion pole mass by choosing a general class of gauge fixing which interpolates between the covariant, the axial and the Coulomb gauges for different values of the gauge fixing parameters. We derive the Nielsen identity describing the gauge parameter variation of the fermion two point function in this general class of gauges. Furthermore, we relate the denominator of the fermion propagator to the two point function. This then allows us to study directly the gauge parameter dependence of the denominator of the propagator using the Nielsen identity for the two point function. This leads to a simple proof that, when infrared divergences and mass shell singularities are not present at the pole, the fermion pole mass is gauge independent, in the complete sense, to all orders in perturbation theory. Namely, the pole is not only independent of the gauge fixing parameters, but has also the same value in both covariant and non-covariant gauges.'
author:
- 'Ashok K. Das,$^{a,b}$ R. R. Francisco$^{c}$ and J. Frenkel$^{c}$ [^1]'
title: On the gauge independence of the fermion pole mass
---
Introduction
============
In a relativistic field theory the mass of a particle is given by the Casimir (operator) relation $$P_{\mu}P^{\mu} - M_{p}^{2} \mathbbm{1} = 0,\label{intro_1}$$ acting on the one particle state. In an interacting theory, this is computationally quite involved. Instead, therefore, one defines the physical mass of the particle to be given by the pole of its propagator. This follows because the dynamical equation for a relativistic field (free equation after taking quantum corrections into account) is expected to encode the Casimir relation in some form and, therefore, the zero of the two point function or the pole of the propagator defines the physical mass of the particle. This is computationally much simpler and the mass of a particle can be calculated in a straightforward manner order by order in perturbation theory.
In a theory of fermions interacting with a gauge field, to do any calculaion in perturbation theory, one needs to choose a gauge. Normally one chooses a class of covariant gauge fixing by adding a term to the Lagrangian density of the form (say, in an Abelian theory) $${\cal L}_{\rm\sc GF} = - \frac{1}{2\xi}\, (\partial\cdot A)^{2},\label{intro_2}$$ which maintains manifest covariance in the intermediate steps of any calculation. Here $\xi$ denotes a gauge fixing parameter. For $\xi=1$ the gauge is known as the Feynman gauge while, for $\xi=0$, it is called the Landau gauge and so on. However, there are also other possible choices of gauge fixing, generically known as non-covariant gauges. Here the choice of gauge fixing depends on a direction conventionally denoted by $n^{\mu}$ (we will discuss this in more detail in sections [**2**]{} and [**3**]{}). The class of generalized axial gauges and Coulomb gauges belong to this class and are also known as physical gauges. In these gauges (for example, in the axial gauges) properties such as asymptotic freedom (in a non-Abelian gauge theory) can be seen in a very simple manner since ghost particles decouple.
As a result of gauge fixing in a gauge theory, the two point function and the propagator (which is a Green’s function) of a fermion interacting with a gauge field become gauge dependent. The gauge dependence has two distinct sources, namely, the propagator depends on the gauge fixing parameter (for example, $\xi$ in covariant gauges) and second, its actual form depends on the class of gauge fixing chosen (for example, whether it is the class of covariant or axial or Coulomb gauges). Nevertheless, one expects that the pole of the propagator should be independent of gauge since it defines a physical quantity (mass). This gauge independence should be complete in the sense that the pole mass should not only be gauge parameter independent within a class of chosen gauges (say, in covariant gauges), but should also have the same value in different classes of gauges (covariant or non-covariant).
It is indeed important to demonstrate that this expectation is true and this question has been studied within the class of covariant gauges from several points of view [@tarrach; @johnston; @reinders; @gray; @brown; @brecken; @kronfeld]. For example, in [@tarrach] it was shown in covariant gauges, both in QED and QCD, that the fermion mass is gauge parameter ($\xi$) independent up to two loops in perturbation theory. In [@kronfeld] this argument was extended in covariant gauges to show that, in these theories, the fermion pole mass is infrared finite and gauge parameter independent to all orders in perturbation theory. The most direct way to study the gauge dependence of any Green’s function or amplitude is through the Nielsen identities [@nielsen; @das] which follow from the BRST invariance [@brst] of the theory. References [@johnston; @brecken], in particular, used this approach (again in covariant gauges) to address the question of the gauge parameter independence of the fermion pole mass to all orders in perturbation theory by studying the gauge parameter variation of the fermion two point function.
These studies are all very important even though all of them are carried out within the given class of covariant gauges in . The gauge parameter independence shows, for example, that the fermion pole mass is the same in the Feynman gauge ($\xi=1$) as in the Landau gauge ($\xi=0$). However, these studies do not address the second source of gauge dependence, namely, one cannot conclude from these studies that the pole mass in the covariant gauges is the same as in non-covariant gauges like the axial gauges or the Coulomb gauges. This still remained to be demonstrated.
On the other side, there was very little systematic study of the question of gauge independence of the fermion pole mass in non-covariant gauges [@kummer; @frenkel; @leibbrandt; @fradkin; @bernstein; @andrasi]. This is a consequence of the fact that when there is an additional structure ($n^{\mu}$) present, the fermion two point function has a more complicated structure (than in covariant gauges) and extracting the mass from a study of the fermion two point function is nontrivial. In fact, since in this case amplitudes depend both on the momentum as well as the longitudinal component of the momentum (with respect to $n^{\mu}$) of the particle independently, it was an open question as to whether a consistent mass can even be defined in such theories (we will discuss these issues in more detail in the next section).
In a recent brief communication [@plb], using Nielsen identities we gave a simple proof of the complete gauge independence of the fermion pole mass in theories without any infrared divergence and mass shell singularities. This proof shows that the pole mass is gauge parameter independent in both covariant as well as non-covariant gauges and also shows that it is the same in different classes of gauges. This is achieved with three basic ingredients. First, we choose a general class of gauges which interpolate between the covariant and non-covariant gauges (axial and Coulomb) depending on the choice of the gauge fixing parameters [@taylor]. We derive the Nielsen identity for the gauge parameter variation of the fermion two point function in such a theory. Finally, since extracting the mass from the two point function is problematic in such a gauge, we study directly the gauge parameter variation of the denominator of the propagator by relating it to the fermion two point function (and using the Nielsen identity).
In this long paper, we give details of our analysis and discuss various other aspects associated with this question. In section [**2**]{}, we describe the generalized axial gauge fixing (in QED) in detail as well as the structure of the fermion self-energy (in the presence of an additional direction $n^{\mu}$) commonly assumed for such a theory. Using charge conjugation invariance in gauge theories, we show that this structure simplifies. The conventional definition of the fermion mass, in such theories, is taken directly from the properties of the fermion two point function in covariant gauges. We show that such a definition becomes gauge parameter dependent and does not coincide with the pole of the propagator starting at two loops. This is shown in two ways, first in a completely algebraic manner starting from the definition of the conventional mass taken for such theories and comparing it with the zero of the denominator of the propagator and second from an explicit two loop calculation. This makes clear that the standard definition of the mass taken from studies in covariant gauges does not work in non-covariant gauges and one should really look at the zero of the denominator (of the propagator) to determine the mass. In this section, we also relate the denominator of the fermion propagator to the two point function for later use in showing the gauge independence of the pole of the propagator. In section [**3**]{}, we discuss in detail the choice of an interpolating gauge (still in QED) and derive the Nielsen identity describing the gauge parameter variation of the fermion two point function. In section [**4**]{} we study directly the gauge parameter variation of the denominator of the propagator using the Nielsen identity for the fermion two point function as well as the relation between the denominator and the two point function. The proof of complete gauge independence then follows in a simple manner. In this section, we also show explicitly the gauge parameter independence of the zero of the denominator (pole of the propagator) up to two loops in the generalized axial gauges. The paper concludes with a short summary of our results and the derivation of the Nielsen identities in QCD in the interpolating gauge is described in the appendix.
Axial gauges
============
As we have already mentioned, non-covariant gauges are gauges where the gauge fixing uses an additional direction, $n^{\mu}$, to fix the gauge. (We recall that covariant gauges use only the covariant gradient vector to fix the gauge.) For example, one may fix the gauge using the longitudinal component of the gradient along the given direction. Such gauges are known as generalized axial gauges. One may alternatively use the transverse component of the gradient (to $n^{\mu}$) to fix the gauge. In particular, if the direction $n^{\mu}$ is timelike, then such gauges are known as generalized Coulomb gauges. Non-covariant gauges are also known as physical gauges. In generalized axial gauges, the gauge fixing Lagrangian density for QED is given by $${\cal L}_{\sc\rm GF} = - \frac{\beta^{2}}{2} \left(\partial_{\sc L}\cdot A\right)^{2},\label{axial_0}$$ where $\beta$ is the (constant) gauge fixing parameter and the longitudinal component of the gradient along $n^{\mu}$ is given by $$\partial^{\mu}_{\sc L} = \frac{(n\cdot \partial) n^{\mu}}{n^{2}}.\label{axial_0a}$$ For simplicity of discussions we assume that $n^{2}\neq 0$. With such a gauge fixing, the photon propagator has the form [@frenkel]
$$D_{\mu\nu} (p) = - \frac{1}{p^{2}}\left[\eta_{\mu\nu} - \frac{n_{\mu}p_{\nu} + n_{\nu}p_{\mu}}{n\cdot p} + \frac{p_{\mu}p_{\nu}n^{2}}{(n\cdot p)^{2}}\left(1 + \frac{n^{2} p^{2}}{\beta^{2} (n\cdot p)^{2}}\right)\right].\label{axial_0b}$$
It is worth noting here that the terms inside the square brackets are dimensionless and of order zero in both $n$ and $p$. Furthermore, the propagator, in addition to being symmetric in $p^{\mu}$, is also invariant under $n^{\mu}\leftrightarrow - n^{\mu}$. This symmetry is quite important as we will see in a short while. (The homogeneous axial gauge is obtained in the limit $\beta\rightarrow \infty$.) In axial gauges, the pole at $p^{2} = 0$ is treated, as usual, with the Feynman prescription while the poles at $n\cdot p =0$ are handled with the principal value (PV) prescription. The other thing to emphasize is that the photon propagator has a part (the first term) that coincides with the propagator in the Feynman gauge and is independent of the gauge fixing parameters $\beta, n^{\mu}$ while the rest of the terms depend manifestly on these parameters. So, in calculations studying gauge independence, sometimes it is useful to write the photon propagator in the generalized axial gauge as $$D_{\mu\nu} = D_{\mu\nu}^{(Feynman)} + \widetilde{D}_{\mu\nu} (n, \beta).\label{axial_0c}$$
In this section, we will start by recapitulating what is already known about the fermion mass in axial gauges. Let us recall that in covariant gauges the general structure of the fermion self-energy (in a massive theory) is simpler and has the form $$\Sigma^{(c)} (p) = m A + B p\sl,\label{axial_1}$$ where the coefficients $A, B$ are dimensionless functions of the Lorentz invariants $(p^{2},m)$ (in addition to the gauge fixing parameter in a gauge theory). In non-covariant gauges, however, because of the presence of a nontrivial direction vector $n^{\mu}$, the structure of the fermion self-energy becomes a bit more involved and is conventionally parameterized in the form [@pagels; @konetschny] $$\Sigma^{\rm (nc)} (p) = m A + B p\sl + C p\sl_{\sc L} + \frac{m D}{p^{2}_{\sc L}}\left(p\sl_{\sc L} p\sl - p\sl p\sl_{\sc L}\right),\label{axial_2}$$ where $p^{\mu}_{\sc L}$ denotes the component of the four momentum $p^{\mu}$ along the given direction $n^{\mu}$. We note here that given a direction vector $n^{\mu}$, any other vector can be decomposed into components parallel and perpendicular to this vector. For example, the momentum vector of the fermion can now have components $$p^{\mu}_{\sc L} = \frac{(n\cdot p)}{n^{2}}\,n^{\mu},\quad p^{\mu}_{\sc T} = p^{\mu} - p^{\mu}_{\sc L}.\label{axial_3}$$ Therefore, we can now construct three Lorentz invariants $p^{2}, p^{2}_{\sc L}$ and $p^{2}_{\sc T}$. However, not all three of these will be independent, rather they will be related as $p^{2} = p^{2}_{\sc L} + p^{2}_{\sc T}$. As a result, only two of them are independent and conventionally, one chooses $p^{\mu}, p^{\mu}_{\sc L}$ as independent vectors which is reflected in . The coefficients $A, B, C$ and $D$ in are, in general, dimensionless functions of $(p^{2},p^{2}_{\sc L}, n^{2}, m)$.
One can understand the rationale behind the structures in and simply as follows. A complete basis for matrices in the Dirac space is given by the matrices (generators of the Clifford algebra) $\mathbbm{1}, \gamma^{\mu}, \gamma^{[\mu}\gamma^{\nu]}, \gamma^{[\mu}\gamma^{\nu}\gamma^{\lambda]},\cdots$. Here the square bracket implies anti-symmetrization in the indices. When there is only one independent vector $p^{\mu}$ (as in covariant gauges), the most general structure for a Dirac matrix (without any Lorentz index), in a parity conserving theory, is given by . On the other hand, if there are two nontrivial independent vectors $p^{\mu}, p^{\mu}_{\sc L}$, it allows for more terms in the expansion in the basis leading to (in a parity conserving theory). It is clear that not having an additional direction $n^{\mu}$ (or $p^{\mu}_{\sc L}$) is equivalent to having the coefficients $C=D=0$ in and in this case, the structure reduces to .
The general structure of the self-energy in follows only from the simple structures of the basis of the Clifford algebra and this form has been conventionally used in all earlier discussions. However, in a gauge theory like QED, there are other symmetries which may limit the number of terms in . One such symmetry is the discrete charge conjugation symmetry which requires that the fermion two point function satisfies $${\cal C} \left(S^{-1} (p)\right)^{T} {\cal C}^{-1} = S^{-1} (-p),\label{axial_4}$$ where $T$ denotes the transpose of a matrix and ${\cal C}$ denotes the charge conjugation matrix defined by (we distinguish this from the coefficient $C$ in with a curly letter) $${\cal C}^{-1} \gamma^{\mu} {\cal C} = - \left(\gamma^{\mu}\right)^{T},\quad {\cal C} \left(\gamma^{\mu}\right)^{T} {\cal C}^{-1} = - \gamma^{\mu}.\label{axial_5}$$ Requiring charge conjugation invariance, namely, for $$S^{-1}_{(nc)} (p) = p\sl - m - \Sigma^{(nc)} (p),\label{axial_6}$$ leads to $$\begin{aligned}
A (-p) & = A(p),&\quad &B (-p) = B (p),\notag\\
C(-p) & = C(p),&\quad &D(-p) = - D(p).\label{axial_7}\end{aligned}$$ The last condition on the coefficient $D$ can be satisfied only if it depends on $p^{\mu}$ through an odd power of the Lorentz invariant $n\cdot p$. However, this would violate the symmetry under $n^{\mu}\leftrightarrow -n^{\mu}$ present in the photon propagator in (the $n^{\mu}$ dependence in the self-energy arises through the photon propagator). Therefore, we see that charge conjugation invariance in gauge theories like QED requires that $$D = 0,\label{axial_8}$$ We have verified explicitly that is true at one loop in $n$-dimensions. However, the charge conjugation symmetry restricts this coefficient to vanish identically to all orders. As a result, the general structure of the fermion self-energy is, in fact, simpler than conventionally assumed and has the form $$\Sigma^{\rm (nc)} (p) = m A + B p\sl + C p\sl_{\sc L},\label{axial_9}$$ with the coefficients $A, B, C$ depending only on $(p^{2}, p_{\sc L}^{2}, m, n^{2})$ (as well as the gauge fixing parameter $\beta$). The coefficients $A, B, C$ get contributions from loop diagrams at every order and represents the self-energy to all orders.
Determining the mass
--------------------
Let us recapitulate what is done in covariant gauges where things are simpler. The fermion two point function satisfies the equation $$\begin{aligned}
S^{-1}_{(c)}\, u\Big|_{p^{2}=M_{(c)}^{2}} & = (p\sl -m - \Sigma^{(c)}) u\Big|_{p^{2}=M_{(c)}^{2}}\notag\\
& = (p\sl - M_{(c)}) u\Big|_{p^{2}=M_{(c)}^{2}} = 0,\label{axial_10}\end{aligned}$$ where $M_{(c)}$ corresponds to the pole of the propagator (or the vanishing of the determinant). It follows now that $$\overline{u}\, S^{-1}_{(c)}\, u\Big|_{p^{2}=M_{(c)}^{2}} = \overline{u}\left(p\sl -M_{(c)}\right)u\Big|_{p^{2}= M_{(c)}^{2}} = 0,\label{axial_11}$$ which allows us to identify $$M_{(c)} = m + \overline{u}\,\Sigma^{(c)}\, u\Big|_{p^{2} = M_{(c)}^{2}}.\label{axial_12}$$
The dependence on an additional direction $n^{\mu}$, on the other hand, makes the extraction of the physical mass more complicated. Conventionally, one assumes that the mass, $p^{2} = \widetilde{M}^{2}$, can also be obtained, as in covariant gauges (see ), from the relation $$\begin{aligned}
\overline{u} S^{-1}_{\rm (nc)} (p) u\Big|_{p^{2}=\widetilde{M}^{2}} & = \overline{u}(p\sl - m - \Sigma^{(nc)}) u\Big|_{p^{2}=\widetilde{M}^{2}}\notag\\
& = \overline{u} (p\sl -\widetilde{M})u\Big|_{p^{2}=\widetilde{M}^{2}}= 0.\label{axial_13}\end{aligned}$$ This would then determine, as in , that the mass $\widetilde{M}$ is given by $$\widetilde{M} = m + \overline{u}\,\Sigma^{\rm (nc)}\, u\Big|_{p^{2}=\widetilde{M}^{2}}.\label{axial_14}$$ In fact, this mass can be evaluated from the fermion self-energy order by order and at one loop it turns out [@konetschny] that $\widetilde{M}$ is gauge parameter independent (independent of $\beta, n^{\mu}$). This one loop calculation has led to the expectation that the mass $\widetilde{M}$ determined from is gauge parameter independent to all orders as in covariant gauges and should correspond to the pole of the propagator. However, this expectation fails beginning at two loops and we will show this, in the following, in two different ways. First, we will show algebraically that $\widetilde{M}$ does not correspond to the pole of the fermion propagator and the difference between the two, starting at two loops, is manifestly gauge parameter dependent ($n^{\mu}$ dependent). Second, we will demonstrate through an explicit calculation of the fermion self-energy at two loops that $\overline{u}\,\Sigma^{\rm (nc)}\,u$ is manifestly gauge dependent at $p^{2} = \widetilde{M}^{2}$ (consistent with the algebraic result) so that $\widetilde{M} = \widetilde{M} (n)$ if holds. (The explicit calculation shows that $\widetilde{M}$ does not depend on the parameter $\beta$ at two loops.) The reason for this unexpected behavior can be traced to the fact that in non-covariant gauges where there is an additional structure ($n^{\mu}$) present, $\overline{u}\, S^{-1} (p)u\big|=0$ does not imply that $S^{-1} (p)u\big|=0$ where the restriction implies evaluating this at $p^{2}=\widetilde{M}^{2}$ (of course, the converse is always true). As a result $\widetilde{M}$ determined by relation does not represent the pole of the propagator beyond the one loop order. To demonstrate all this, we need the form of the fermion propagator which we derive in the next subsection.
Fermion propagator
------------------
Let us recall from the form of the fermion self-energy in that in the (non-covariant) axial gauge the fermion two point function can be written as $$S^{-1}(p) = {\cal A} - {\cal B} \mathbbm{1},\label{axial_15}$$ where ${\cal A}$ is a nontrivial matrix while ${\cal B}$ is a scalar of the forms $${\cal A} = (1-B) p\sl - C p\sl_{\sc L},\quad {\cal B} = m (1+A).\label{axial_16}$$ It is straightforward to check from (and the properties of the Dirac gamma matrices as well as the definition of the longitudinal momentum) that $${\cal A}^{2} = \left( (1-B)^{2} p^{2} + (C^{2} - 2C (1-B)) p_{\sc L}^{2}\right) \mathbbm{1}.\label{axial_17}$$ In such a case, the propagator (up to a factor $i$) can be obtained in a simple manner to have the form $$S (p) = \frac{\cal N}{\cal D},\label{axial_18}$$ where (the identity matrix in the denominator is not relevant and we have used )
$$\begin{aligned}
{\cal N} & = {\cal A} + {\cal B} \mathbbm{1} = - {\cal C} (S^{-1} (p))^{T} {\cal C}^{-1} = - S^{-1} (-p),\label{axial_19}\\
{\cal D} & = {\cal A}^{2} - {\cal B}^{2} = (1-B)^{2} p^{2} + (C^{2} - 2C (1-B))p_{\sc L}^{2} - m^{2} (1+A)^{2}.\label{axial_20}\end{aligned}$$
The pole of the propagator is determined from the zero of the denominator ${\cal D}$ and since the denominator depends on both $p^{2}$ and $p_{\sc L}^{2}$, it is not clear [*a priori*]{} whether a mass can even be defined. (This is quite different from covariant gauges.) For example, suppose the zero of the denominator occurs at $p^{2} = M_{p}^{2}$, then from we obtain $$M_{p}^{2} = \frac{[m^{2}(1+A)^{2} - (C^{2} - 2C(1-B))p_{\sc L}^{2}]}{(1-B)^{2}}\Big|_{p^{2}=M_{p}^{2}},\label{axial_21}$$ so that, in general, the location of the pole of the propagator seems to depend on the longitudinal momentum $p_{\sc L}$ and, therefore, cannot define a mass (unless all the dependence on the longitudinal momentum $p_{\sc L}$ cancels out). As we will show in section [**4**]{} using the Nielsen identity, this is indeed the case and $M_{p}$ is, in fact, independent of the gauge parameters $\beta$ as well as $n^{\mu}$. For later use, we note from - that we can write $$\begin{aligned}
{\cal D} \mathbbm{1} & = - S^{-1}(p) {\cal C} (S^{-1}(p))^{T} {\cal C}^{-1}\notag\\
& = - S^{-1} (p) S^{-1} (-p).\label{axial_22}\end{aligned}$$ Alternatively, taking the matrix trace in we determine the scalar denominator of the propagator to be given by $$\begin{aligned}
{\cal D} & = - \frac{1}{2^{[n/2]}}\, \text{Tr} \left(S^{-1}(p){\cal C} (S^{-1} (p))^{T} {\cal C}^{-1}\right)\notag\\
& = - \frac{1}{2^{[n/2]}}\,\text{Tr}\left(S^{-1}(p) S^{-1} (-p)\right).\label{axial_23}\end{aligned}$$ Here $n$ denotes the dimensionality of space-time, $[n/2]$ represents the integer part of $n/2$ and we have used the fact that, in $n$-dimensions, the Dirac matrices are $2^{[n/2]}\times 2^{[n/2]}$ matrices. (For completeness, note from and that $S^{-1} (p)$ and $S^{-1} (-p)$ commute.) We will show in section [**4**]{} that the pole of the propagator given in is gauge independent. For the moment, we only note that since $A,B,C$ receive contributions at various loops, the pole mass" $M_{p}$ can also be expanded in powers of loops as $$M_{p} = M_{p}^{(0)} + M_{p}^{(1)} + M_{p}^{(2)} + \cdots,\label{axial_24}$$ with the tree level mass identified with $M_{p}^{(0)} = m$ (namely, at tree level $A^{(0)}=B^{(0)}=C^{(0)}=0$).
On the other hand, $\widetilde{M}$ is defined in and to evaluate this we need an expression for $\overline{u} \gamma^{\mu}u$. In covariant theories with momentum as the only available four vector, this is normally determined to be $\overline{u} \gamma^{\mu} u\big|_{p^{2}=M^{2}} = \frac{p^{\mu}}{M}$ (so that we have $\overline{u}\,p\sl\, u\big|_{p^{2}=M^{2}} = M$). When there is an additional structure ($n^{\mu}$ or $p_{\sc L}^{\mu}$) present, this generalizes to $$\overline{u} \gamma^{\mu} u\Big|_{p^{2}=\widetilde{M}^{2}} = (1-a)\,\frac{p^{\mu}}{\widetilde{M}} + a \widetilde{M}\,\frac{p_{\sc L}^{\mu}}{p_{\sc L}^{2}}.\label{axial_25}$$ Here $a$ is a nontrivial parameter, in general, beginning at one loop and beyond since the additional structures arise in the self-energy only in that order. It follows from that (see also ) $$\begin{aligned}
& \overline{u}\, p\sl\, u\Big|_{p^{2}=\widetilde{M}^{2}} = \widetilde{M},\notag\\
& \overline{u}\, p\sl_{\sc L}\, u\Big|_{p^{2}=\widetilde{M}^{2}} = \frac{p_{\sc L}^{2}}{\widetilde{M}} - a \widetilde{M}\left(\frac{p_{\sc L}^{2}}{\widetilde{M}^{2}} -1\right).\label{axial_26}\end{aligned}$$ Using , we can now evaluate $\widetilde{M}$ from the definition in leading to
$$\widetilde{M}^{2} = \frac{1}{(1-B)^{2}}\left[m^{2}(1+A)^{2} - C^{2} \left(\frac{p_{\sc L}^{2}}{\widetilde{M}} - a\widetilde{M}\left(\frac{p_{\sc L}^{2}}{\widetilde{M}^{2}} -1\right)\right)^{2} + 2C(1-B) \left(p_{\sc L}^{2} - a \widetilde{M}^{2} \left(\frac{p_{\sc L}^{2}}{\widetilde{M}^{2}} -1\right)\right)\right]_{p^{2}=\widetilde{M}^{2}}.\label{axial_27}$$
Comparing and , it is clear that $M_{p}$ and $\widetilde{M}$ are, in general, different, but it is not clear at what order the difference arises. To that end, we note that we can also expand $$\widetilde{M} = \widetilde{M}^{(0)} + \widetilde{M}^{(1)} + \widetilde{M}^{(2)} + \cdots.\label{axial_28}$$ From and we can now determine $$\begin{aligned}
\widetilde{M}^{(0)} & = M_{p}^{(0)} = m,\notag\\
\widetilde{M}^{(1)} & = M_{p}^{(1)} = m (A^{(1)} + B^{(1)}) + C^{(1)}\frac{p_{\sc L}^{2}}{m}\Big|_{p^{2}=m^{2}},\notag\\
\widetilde{M}^{(2)} & = M_{p}^{(2)} - \left(\!(C^{(1)})^{2}\frac{p_{\sc L}^{2}}{2m} + ma^{(1)}C^{(1)}\!\right)\!\!\left(\frac{p_{\sc L}^{2}}{m^{2}}-1\right)_{p^{2}=m^{2}},\label{axial_29}\end{aligned}$$ and so on. Namely, $\widetilde{M}$ defined by coincides with the pole of the propagator only up to one loop, but this equivalence fails beginning at two loops. If $M_{p}$ is gauge independent, as we will see in section [**4**]{}, then shows that $\widetilde{M}$ becomes manifestly gauge dependent beginning at two loops. However, note from that in the special gauge $n^{\mu}\parallel p^{\mu}$ or $p_{\sc L}^{\mu} = p^{\mu}$, the difference vanishes. This is, in fact, not just an accident at two loops, rather it holds to all orders which can be seen as follows. In the gauge $p_{\sc L}^{\mu} = p_{\mu}$, we note that and reduce to the same equation and, therefore, coincide to all orders. Another way of seeing this is to note that in the special gauge $p_{\sc L}^{\mu} = p^{\mu}$, there is no additional structure and the fermion two point function has the same structure as in a covariant gauge for which the two equations and are equivalent.
We can also explicitly evaluate the self-energy through Feynman diagrams and determine $\widetilde{M}$ from there. It is already known from such a calculation that $\widetilde{M}$ is gauge parameter independent (independent of $\beta, n^{\mu}$) at one loop. So, let us concentrate on the contributions at two loops. At this order the relevant Feynman diagrams are given by Fig. \[fig1\].
![Feynman diagrams for the fermion self-energy in QED at two loops.[]{data-label="fig1"}](twoloop)
Here the diagrams $(a), (b), (c)$ denote the standard two loop graphs while the diagram in $(d)$ represents the two loop contribution coming from the one loop mass correction. We note that since the photon self-energy in $(c)$ is transverse, the last term in the photon propagator in gives vanishing contribution so that this diagram is independent of $\beta$. In fact, the $\beta$ dependent terms in all the other diagrams identically vanish because of the PV prescription. There are several other cancellations in the $n^{\mu}$ dependent terms involving diagrams $(a), (b), (c), (d)$ and the final result for $\widetilde{M}$ from this explicit calculation takes the factorized form (in $n$-dimensions)
$$\widetilde{M} = M^{(Feynman)} + \frac{e^{4}}{(2\pi)^{2n}}\,\overline{u} (p) \int \frac{d^{n}q_{1}}{q_{1}^{2} (n\cdot q_{1})}\, n\sl \frac{1}{p\sl + q\sl_{1} - m}\, (p\sl - m) \int \frac{d^{n}q_{2}}{q_{2}^{2} (n\cdot q_{2})}\,\frac{1}{p\sl + q\sl _{2} - m}\, n\sl\, u (p)\Big|_{p^{2}=m^{2}},\label{axial_30}$$
where $q_{1}, q_{2}$ denote the two internal loop momenta and $M^{(Feynman)}$ is the mass obtained from the first term ($\beta, n^{\mu}$ independent term) in the propagator in or . Since $M^{(Feynman)}$ is independent of the gauge parameters, it shows explicitly that $\widetilde{M}$ is manifestly gauge dependent at two loops consistent with the algebraic result in . The reason for this gauge dependence is clear from the structure in , namely, that it is not the factor $(p\sl - m)$, rather $n\sl$ which occurs at the two ends. Since $p\sl$ and $n\sl$ do not commute, we cannot move $(p\sl -m)$ move past $n\sl$ to obtain a vanishing result for the second term. However, note that if $n^{\mu}$ is parallel to $p^{\mu}$, then each of the (factorized) integrals can be evaluated to have the form $(a p\sl + b)$ which will commute with $(p\sl -m)$ and, in this case, we can take the factor $(p\sl - m)$ to one end to annihilate the spinor. So, only for this special gauge will $\widetilde{M}$ coincide with $M^{(Feynman)}$ as we have also seen in the algebraic method.
As we have tried to emphasize in this section, the conventionally accepted definition of mass of the fermion $\widetilde{M}$ in non-covariant gauges, , does not coincide with the pole of the propagator $M_{p}$ beyond one loop and becomes manifestly gauge parameter dependent. However, we are yet to show that the actual pole defines a gauge independent mass. As we have already pointed out in , it is not clear that we can even define a meaningful mass in non-covariant gauges. In the next two sections, we will show through Nielsen identities that the pole $M_{p}$ indeed defines a mass and is gauge independent in the complete sense.
Nielsen identity in an interpolating gauge
==========================================
We know that the fermion self-energy and, therefore, the two point function become gauge dependent once a gauge fixing condition has been chosen and the gauge parameter variation of the fermion two point function is best studied through the Nielsen identities. However, we wish to study the question of gauge independence of the fermion pole mass completely in the sense that we wish to show that the pole mass is not only independent of the gauge fixing parameter within a given class of gauges such as covariant or Coulomb or axial gauges, but also has the same value in all of these gauges. We achieve this in three basic steps, (i) choose a gauge fixing which will interpolate between covariant, Coulomb and axial gauges for different values of the gauge fixing parameters, (ii) derive the Nielsen identity for the gauge parameter variation of the fermion two point function for such a gauge fixing and (iii) study the gauge parameter variation of the zero of the denominator (of the propagator) , using the Nielsen identity for the two point function as well as the relation between the denominator and the two point function already discussed in . In this section, we only take up the first two steps leaving the last to the next section.
Choice of an interpolating gauge
--------------------------------
Let us consider massive QED in $n$ space-time dimensions (the generalization to QCD is discussed in the appendix) described by the Lagrangian density $${\cal L}_{\rm inv} = - \frac{1}{4}\, F_{\mu\nu}F^{\mu\nu} + \overline{\psi} (i D\!\sl - m) \psi.\label{nielsen_1}$$ As the first step, our goal is to study this theory in a general class of gauges which can interpolate between the covariant, the Coulomb and the axial gauges. To that end, we choose a gauge fixing Lagrangian density of the form [@taylor] $${\cal L}_{\rm\sc GF} = - \frac{1}{2}\left(\Lambda^{\mu} (\partial) A_{\mu}\right)^{2},\label{nielsen_2}$$ where $$\Lambda^{\mu} (\partial) = \alpha \partial^{\mu} + \beta \partial^{\mu}_{\sc L},\quad \partial^{\mu}_{\sc L} = \frac{(n\cdot \partial)}{n^{2}}\,n^{\mu},\label{nielsen_3}$$ and $\alpha,\beta$ are arbitrary constant parameters. Unlike usual gauge fixing Lagrangian densities which depend only on one gauge fixing parameter, here the dependence is on three independent parameters $\alpha, \beta$ and $n^{\mu}$ which we denote collectively by $$\phi_{(a)} = (\alpha, \beta, n^{\mu}).\label{nielsen_4}$$ The reason for this, as we have already emphasized earlier, is that this is an interpolating gauge between various classes of gauges in the sense that $$\begin{aligned}
\beta & = 0,& & \text{covariant gauges}\ (\text{with}\ \xi=\alpha^{-2}),\notag\\
\alpha & = 0,& & \text{generalized axial gauges},\notag\\
\beta & = -\alpha, n^{2}>0,& & \text{generalized Coulomb gauges}.\label{nielsen_4a}\end{aligned}$$ So, for different choices of the gauge fixing parameters one can easily go from covariant to axial to Coulomb gauges (which is our purpose in choosing such a gauge fixing term). Let us also note that, for the purposes of manifest BRST invariance (which is essential for deriving Ward identities and Nielsen identities), the gauge fixing Lagrangian density can also be written with an auxiliary (nondynamical) field $F$ as $${\cal L}_{\sc\rm GF} = \frac{1}{2}\, F^{2} + \left(\Lambda^{\mu} (\partial) F\right) A_{\mu}.\label{nielsen_5}$$
The ghost Lagrangian density corresponding to this general class of gauge choice, , is given by $${\cal L}_{\rm ghost} = \left(\Lambda^{\mu} (\partial) \overline{c}\right) \partial_{\mu}c,\label{nielsen_6}$$ and the combined Lagrangian density $${\cal L} = {\cal L}_{\rm inv} + {\cal L}_{\rm\sc GF} + {\cal L}_{\rm ghost},\label{nielsen_6a}$$ is manifestly invariant (with gauge fixing described by ) under the standard (nilpotent) BRST transformations of QED [@brst; @das], $$\begin{aligned}
& \delta A_{\mu} = \omega \partial_{\mu}c, & & \delta F = 0,\notag\\
& \delta \psi = - ie\omega c \psi, & & \delta \overline{\psi} = -ie\omega \overline{\psi} c,\notag\\
& \delta c = 0, & & \delta \overline{c} = - \omega F,\label{nielsen_7}\end{aligned}$$ where $\omega$ represents an arbitrary constant Grassmann parameter. For completeness we note here that the photon propagator for this generalized gauge fixing can be easily obtained to have the form [@taylor]
$$D_{\mu\nu} (p) = - \frac{1}{p^{2}}\left[\eta_{\mu\nu} - \frac{\Lambda_{\mu} (p) p_{\nu} + \Lambda_{\nu}(p) p_{\mu}}{\Lambda(p)\cdot p} + \frac{p_{\mu}p_{\nu} (\Lambda(p))^{2}}{(\Lambda(p)\cdot p)^{2}}\left(1 + \frac{p^{2}}{(\Lambda(p))^{2}}\right)\right],\label{nielsen_7a}$$
which can be compared with . Furthermore, it is straightforward to check that for different choices of the gauge fixing parameters, as in , this propagator reduces to the familiar propagators in the three different classes of gauges.
Nielsen identities
------------------
To derive the Green’s functions of the theory we need to introduce sources for the fundamental fields and to determine the gauge parameter variations of the effective action and the Green’s functions (Nielsen identities), we also need to introduce some other sources and these are described by the Lagrangian density $$\begin{aligned}
{\cal L}_{\rm source} & = J^{\mu}A_{\mu} + JF + i\left(\overline{\chi}\psi - \overline{\psi}\chi\right) + i\left(\overline{\eta} c - \overline{c}\eta\right)\notag\\
& + ie\left(\overline{M} c\psi - \overline{\psi}c M\right) + \left(H_{(\alpha)} (\partial^{\mu}\overline{c}) + H_{(\beta)} (\partial^{\mu}_{\sc L}\overline{c})\right)A_{\mu}\notag\\
& + \beta H_{(n)\,\mu} (N^{\mu\nu}\overline{c})A_{\nu}.\label{nielsen_8}\end{aligned}$$ Here we have identified $$N^{\mu\nu} = \frac{\partial \partial^{\nu}_{\sc L}}{\partial n_{\mu}} = \frac{(n\cdot \partial)}{n^{2}} \left(\eta^{\mu\nu} + \frac{\partial^{\mu}n^{\nu}}{n\cdot\partial} - \frac{2n^{\mu}n^{\nu}}{n^{2}}\right).\label{nielsen_9}$$ Therefore, the total Lagrangian density for the gauge fixed theory (with sources) is given by $$\begin{aligned}
{\cal L}_{\rm\sc TOT} & = {\cal L}_{\rm inv} + {\cal L}_{\rm\sc GF} + {\cal L}_{\rm ghost} + {\cal L}_{\rm source},\notag\\
S_{\rm\sc TOT} & = \int d^{n}x\, {\cal L}_{\rm\sc TOT}.\label{nielsen_10}\end{aligned}$$ As mentioned above, we note from that $(J^{\mu}, J, \overline{\chi}, \chi, \overline{\eta}, \eta)$ correspond respectively to the standard sources for the fields $(A_{\mu}, F, \psi, \overline{\psi}, c, \overline{c})$; the sources $(\overline{M}, M)$ generate respectively the composite BRST variations of $(\psi, \overline{\psi})$ in . In addition, we have introduced three sources $$H_{(a)} = \left(H_{(\alpha)}, H_{(\beta)}, H_{(n)}^{\mu}\right),\label{nielsen_11}$$ such that the BRST variations of these three source terms lead to the gauge variations of ${\cal L}_{\rm inv} + {\cal L}_{\rm\sc GF} + {\cal L}_{\rm ghost}$ with respect to the three parameters $\phi_{(a)}$ defined in . Note here that the additional sources $(\overline{M}, M)$ are commuting sources which carry spinor indices and ghost quantum numbers while the three sources $H_{(a)}$ defined in are anti-commuting and carry ghost quantum numbers (but have no spinor index).
The generating functional for this gauge fixed theory is now given by $$Z [{\cal J}] = e^{iW[{\cal J}]} = N \int {\cal D}\varphi\, e^{iS_{\rm\sc TOT}},\label{nielsen_12}$$ where we have denoted all the field variables and sources generically by $\varphi$ and ${\cal J}$ respectively and $N$ denotes the normalization constant for the path integral. If we make a field redefinition corresponding to the BRST transformations , namely, $$\varphi \rightarrow \varphi + \delta_{\sc BRST}\varphi,\label{nielsen_13}$$ inside the path integral, the generating functional does not change since the path integral involves integration over all field configurations (alternatively, the generating functional depends only on the sources and not on the fields). As a result, we obtain $$\delta Z [{\cal J}] = 0 = iN\int {\cal D}\varphi\, (\delta_{\sc BRST} S_{\rm\sc TOT})\, e^{iS_{\rm\sc TOT}}.\label{nielsen_14}$$ Only the source terms in $S_{\rm\sc TOT}$ in are not BRST invariant and substituting the variations of these terms in leads to an identity involving the generating functional $W[{\cal J}]$ of the form (we use the convention of left derivatives for anti-commuting variables) $$\begin{aligned}
& N e^{-iW}\int {\cal D}\varphi \int \!\!d^{n}x\left(\!H_{(a)}\frac{\partial {\cal L}_{\rm\sc TOT}}{\partial\phi_{(a)}} + O ((H_{(a)})^{2})\right) e^{iS_{\rm\sc TOT}}\notag\\
& = i\int d^{n}x\left(J^{\mu}\partial_{\mu} \frac{\delta W}{\delta\overline{\eta}} - \overline{\chi} \frac{\delta W}{\delta\overline{M}} + \frac{\delta W}{\delta M} \chi - \frac{\delta W}{\delta J} \eta\right).\label{nielsen_15}\end{aligned}$$ Furthermore, taking the functional derivative with respect to $H_{(a)}(x)$ and setting the sources $H_{(a)}=0$ and then integrating over $\int d^{n}x$, we obtain the Master identity for the generating functional for connected Green’s functions
$$\frac{\partial W}{\partial \phi_{(a)}} = i \int d^{n}x\,d^{n}y\left(J^{\mu}(y)\partial^{(y)}_{\mu} \frac{\delta^{2} W}{\delta H_{(a)}(x)\delta\overline{\eta}(y)} + \overline{\chi}(y) \frac{\delta^{2} W}{\delta\overline{M}(y) \delta H_{(a)}(x)} + \frac{\delta^{2} W}{\delta H_{(a)}(x) \delta M(y)} \chi (y) - \frac{\delta^{2} W}{\delta H_{(a)}(x) \delta J(y)} \eta(y)\right).\label{nielsen_16}$$
Upon differentiation, this equation leads to the gauge parameter variation of the connected Green’s functions of the theory. We emphasize that all the sources $H_{(a)}$ have been set to zero in .
To obtain the gauge parameter variation of the 1PI amplitudes of the theory, we make a Legendre transformation with respect to the standard sources for the dynamical fields of the theory, namely, (here the field variables correspond to the classical fields) $$\Gamma = W - \int d^{n}x \left(J^{\mu}A_{\mu} + JF + i (\overline{\chi}\psi - \overline{\psi} \chi) + i (\overline{\eta}c - \overline{c}\eta)\right).\label{nielsen_17}$$ This leads (from ) to the Master identity for the gauge parameter variation of the effective action in the form $$\begin{aligned}
\frac{\partial \Gamma}{\partial\phi_{(a)}} & = \int d^{n}z d^{n}w\left(\frac{\delta\Gamma}{\delta\psi_{\gamma}(w)} \frac{\delta^{2}\Gamma}{\delta \overline{M}_{\gamma} (w) \delta H_{(a)}(z)}\right.\notag\\
&\qquad\left. + \frac{\delta^{2}\Gamma}{\delta H_{(a)}(z) \delta M_{\gamma}(w)} \frac{\delta\Gamma}{\delta \overline{\psi}_{\gamma}(w)}\right),\label{nielsen_18}\end{aligned}$$ where, again, the sources $H_{(a)}$ have been set to zero. By taking derivatives of this identity with respect to various fields and setting all the fields (including $M, \overline{M}$) to zero, we can obtain the gauge parameter variation of any 1PI amplitude in the theory. For example, taking the functional derivative with respect to $\frac{\delta^{2}}{\delta\psi_{\beta}(y)\delta\overline{\psi}_{\alpha}(x)}$ and setting all fields to zero, we obtain the gauge parameter variation of the fermion two point function $$\begin{aligned}
\frac{\partial S^{-1}_{\alpha\beta}(x-y)}{\partial\phi_{(a)}} & = \int d^{n}z d^{n}w\left({\cal F}^{(a)}_{\alpha\gamma} (x,z,w) S^{-1}_{\gamma\beta} (w-y)\right.\notag\\
&\quad \left. + S^{-1}_{\alpha\gamma}(x-w) {\cal G}^{(a)}_{\gamma\beta} (w,z,y)\right).\label{nielsen_19}\end{aligned}$$ Here we have identified $$\begin{aligned}
S^{-1}_{\alpha\beta} (x-y) & = \frac{\delta^{2}\Gamma}{\delta\psi_{\beta} (y)\delta\overline{\psi}_{\alpha} (x)},\notag\\
{\cal F}^{(a)}_{\alpha\gamma}(x,z,w) & = \frac{\delta^{3}\Gamma}{\delta\overline{\psi}_{\alpha}(x)\delta H_{(a)}(z)\delta M_{\gamma}(w)},\notag\\
{\cal G}^{(a)}_{\gamma\beta} (w,z,y) & = \frac{\delta^{3}\Gamma}{\delta\overline{M}_{\gamma}(w)\delta H_{(a)}(z)\delta\psi_{\beta}(y)},\label{nielsen_20}\end{aligned}$$ with all field variables (including $M, \overline{M}$) set to zero. In momentum space, the Nielsen identity takes the simple form $$\frac{\partial S^{-1}_{\alpha\beta} (p)}{\partial \phi_{(a)}} = {\cal F}^{(a)}_{\alpha\gamma} (p) S^{-1}_{\gamma\beta} (p) + S^{-1}_{\alpha\gamma} (p) {\cal G}^{(a)}_{\gamma\beta} (p),\label{nielsen_21}$$ where we have identified the three point amplitudes $$\begin{aligned}
{\cal F}^{(a)}_{\alpha\gamma} (p) & = {\cal F}^{(a)}_{\alpha\gamma} (-p, 0,p),\notag\\
{\cal G}^{(a)}_{\gamma\beta} (p) & = {\cal G}^{(a)}_{\gamma\beta} (-p.0.p).\label{nielsen_22}\end{aligned}$$ Equation shows how the fermion two point function changes with respect to the three independent parameters $(\alpha,\beta,n^{\mu})$. We note here that since there are no vertices corresponding to ${\cal F}^{(a)}_{\alpha\beta} (p), {\cal G}^{(a)}_{\alpha\beta} (p)$ at the tree level Lagrangian density in , these amplitudes are nontrivial only at one loop and beyond. Therefore, the Nielsen identity implies that the dependence of the fermion two point function on the gauge fixing parameter arises only at one loop and beyond, as we expect.
Gauge independence of the pole of the fermion propagator
========================================================
We are now ready to show the gauge independence of the pole of the fermion propagator. We note that studying the pole of the propagator is equivalent to studying the zero of the denominator ${\cal D}$ of the propagator defined in and . Studying directly the gauge parameter independence of the zero of the denominator in the form (in a generalized interpolating gauge) is more involved than in the covariant gauge because of the presence of an additional structure. There is a much simpler way to do this which we discuss in the following. Let us recall from that the denominator of the fermion propagator can be related to the fermion two point function as $${\cal D} = - \frac{1}{2^{[n/2]}}\, \text{Tr} \left( S^{-1} (p) {\cal C} (S^{-1} (p))^{T} {\cal C}^{-1}\right).$$ Since the Nielsen identity describes how the fermion two point function changes with the change in any of the three gauge parameters, we can now use this to determine the gauge parameter variation of the denominator of the fermion propagator $$\begin{aligned}
& \frac{\partial {\cal D}}{\partial\phi_{(a)}} = - \frac{1}{2^{[n/2]}}\,{\rm Tr}\left(\frac{\partial S^{-1} (p)}{\partial\phi_{(a)}} {\cal C} (S^{-1}(p))^{T} {\cal C}^{-1}\right.\notag\\
& \qquad\qquad\qquad\qquad\left. + S^{-1}(p) {\cal C} \frac{\partial (S^{-1}(p))^{T}}{\partial\phi_{(a)}} {\cal C}^{-1}\right)\notag\\
& = {\cal D}\, {\rm Tr}\left(\!{\cal F}^{(a)}(p) + {\cal G}^{(a)}(p) + {\cal C} \left({\cal F}^{(a)}(p) + {\cal G}^{(a)}(p)\right)^{T}\!{\cal C}^{-1}\!\right)\notag\\
& = 2 {\cal D}\, {\rm Tr} ({\cal F}^{(a)}(p) + {\cal G}^{(a)}(p)).\label{gaugeind_1}\end{aligned}$$ Here we have used , cyclicity of trace as well as the fact that the trace of the transpose of a matrix coincides with that of the matrix itself.
Equation shows explicitly that the denominator depends on the gauge parameters (is gauge dependent) and the gauge parameter variation of the denominator with respect to the three independent parameters is proportional to the denominator itself. As a result, it follows that, if the amplitudes $({\cal F}^{(a)} (p), {\cal G}^{(a)}(p))$ are well behaved, the zero of the denominator ${\cal D}$ (which corresponds to the pole of the propagator) is, in fact, gauge parameter independent, namely, $$\frac{\partial {\cal D}}{\partial \phi_{(a)}}\Big|_{{\cal D}=0} = 0.\label{gaugeind_2}$$ This would, in fact, be the case when the theory does not have any infrared divergences or mass shell singularities at the pole of the propagator. Actually, this turns out to be the case in perturbative QED (and in QCD) in $4$ space-time dimensions [@kronfeld]. On the other hand, in lower dimensions, $n < 4$, such singularities can be present in the amplitudes $({\cal F}^{(a)} (p), {\cal G}^{(a)} (p))$ so that the pole of the propagator may become gauge dependent. This has been explicitly studied in the massive Schwinger model (massive QED in $1+1$ dimensions) in [@schubert]. Noting that near the zero of ${\cal D}$ (or the pole of the propagator), we can write $${\cal D}\xrightarrow{p^{2}\rightarrow M^{2}_{p}} Z (p^{2} - M^{2}_{p}),\label{gaugeind_3}$$ where the coefficient $Z$ is related to the wave function normalization $Z_{2}^{-1}$, equation leads to $$\frac{\partial M_{p}}{\partial \phi_{(a)}} = 0.\label{gaugeind_4}$$ Namely, the pole of the propagator or the physical mass of the fermion is independent of the three gauge fixing parameters $(\alpha, \beta, n^{\mu})$ in theories without infrared divergence and mass shell singularities at the pole. Furthermore, since the covariant, the axial and the Coulomb gauges correspond to specific values of these parameters which $M_{p}$ is independent of, the physical mass (or the location of the pole of the propagator) is the same in all three classes of gauges. This demonstrates the complete gauge independence of the fermion pole mass and it is worth emphasizing that this direct and simple demonstration of gauge invariance of the pole of the fermion propagator involves only three basic elements: choice of a general class of interpolating gauges, the Nielsen identity for the gauge variation of the fermion two point function for this gauge choice and the relation of the two point function to the denominator of the propagator.
We have explicitly verified the gauge independence of the pole of the fermion propagator, up to two loops, in the generalized axial gauge as follows. It is already known that at one loop, the pole is gauge parameter independent (namely, $\widetilde{M} = M_{p}$ is gauge parameter independent up to one loop). Therefore, we need to concentrate only on the contributions at two loops. At two loops, the contributions to the self-energy are given by the diagrams Fig. \[fig1\]. Using this in the expression for the denominator in the form, $${\cal D} = - \frac{1}{2^{[n/2]}}\,{\rm Tr} \left(S^{-1}(p) S^{-1}(-p)\right),$$ we can separate out the two loop contributions explicitly. Due to the transversality of the photon self-energy in diagram $(c)$ of Fig. \[fig1\], the gauge parameter ($(\beta, n^{\mu})$) dependent terms turn out to be proportional to $(p^{2}-m^{2})$ with a well behaved coefficient so that they vanish as $p^{2}\rightarrow m^{2}$ (note that since the amplitude is already of order two loops, $M_{p}^{2}$ can be set equal to $m^{2}$ in this order). Similarly the $\beta$ dependent terms in diagram $(d)$ as well as in the sum of the diagrams $(a)+(b)$ also have the same form with well behaved coefficients. Therefore, these also vanish at the pole and there is no $\beta$ dependence at all in the amplitude. The remaining $n^{\mu}$ dependent terms in the sum of the contributions from diagrams $(a)+(b)+(d)$ combine under the trace to be proportional to $(p^{2}-m^{2})$ (it is worth emphasizing that individual diagrams do not have this form). As a consequence, the complete two loop self-energy leads to $${\cal D} = {\cal D}^{(Feynman)} + Q(p,m,n) (p^{2}-m^{2}),\label{gaugeind_5}$$ where ${\cal D}^{(Feynman)}$ is manifestly gauge independent (calculated using $D_{\mu\nu}^{(Feynman)}$ in ) and $Q(p,m,n)$ is a complicated integral which is manifestly $n^{\mu}$ dependent (but has no $\beta$ dependence). It is clear from that if there are no mass shell singularities in $Q(p,m,n)$ at the pole, then the gauge dependent terms in vanish at the zero of the denominator, namely, as $p^{2}\rightarrow m^{2}$ (to this order) completely consistent with the conclusion following from the Nielsen identity in . One can estimate the behavior of the coefficient function and near the mass shell it behaves like $Q(p,m,n) \sim (p^{2}-m^{2})^{(n-4)/2}$ so that for $n\geq 4$ it is well behaved. However, for $n\leq 3$ mass shell singularities develop at the pole and one cannot conclude (as in ) that the pole of the propagator is gauge independent. We have seen this explicitly in an earlier study involving the massive Schwinger model [@schubert].
In fact, even in the Schwinger model (massless), where it is known that there is no infrared divergence, mass shell singularities do develop. For example, in the covariant gauge the fermion self-energy at one loop has the form (here we use the standard gauge fixing parameter $\xi = \alpha^{-2}$) $$\Sigma^{(1)}_{(c)} (p) = \frac{\xi e^{2}}{2\pi p^{2}}\, p\sl,\label{gaugeind_6}$$ and the singularity as $p^{2}\rightarrow 0$ is manifest in . In this case, the pole of the fermion propagator is easily determined to be $$M^{2}_{(c)} = \frac{\xi e^{2}}{2\pi},\label{gaugeind_7}$$ and is manifestly gauge parameter dependent. Furthermore, one can calculate the one loop fermion self-energy in the generalized axial gauge which has the form $$\Sigma^{(1)}_{(axial)} (p) = - \frac{e^{2}}{2\pi p_{\sc L}^{2}}\,p\sl_{\sc L},\label{gaugeind_8}$$ leading to a fermion pole mass (in the perturbative regime $e^{2}\ll |p_{\sc L}^{2}|$) $$M^{2}_{(axial)} = - \frac{e^{2}}{\pi}.\label{gaugeind_9}$$ There are two things to note from and . First, the mass in the axial gauge is independent of the gauge fixing parameters $(\beta, n^{\mu})$ while that in the covariant gauge is manifestly gauge parameter dependent so that the two do not coincide in general. This is a simple example of how a gauge parameter independent pole within one class of gauges (in this case the axial gauge) does not automatically imply that it will be the same in all classes of gauges. Second, because the poles in the two gauges do not coincide, even if the pole in the axial gauge is gauge parameter independent, it will be incorrect to conclude that it represents a physical mass. In fact, we see from that this mass is purely imaginary and, therefore, completely unphysical.
Conclusion
==========
In this paper we have studied the question of complete gauge independence of the fermion pole mass. This involves showing the gauge parameter independence of the pole within a given class of gauges (such as covariant or axial or Coulomb) as well as showing that the pole has the same value in all of these three classes of gauges. The demonstration of complete gauge independence is achieved in a simple manner by using three basic ingredients. First, we choose a general class of gauges which interpolate between the covariant, the axial and the Coulomb gauges for different values of the gauge parameters. We derive the Nielsen identity describing the gauge parameter variation of the fermion two point function in this class of interpolating gauges. And we relate the denominator of the fermion propagator to the two point function so that the gauge parameter variation of the denominator (or the pole of the propagator) can be studied directly using the Nielsen identity for the fermion two point function. With these three basic ingredients we are able to show in a simple manner that, when there are no infrared divergences and mass shell singularities at the pole, the pole of the fermion propagator is gauge independent in the complete sense. The presence of mass shell singularities can invalidate such a proof and this is pointed out with a simple example of the Schwinger model where the pole mass is manifestly gauge parameter dependent in the covariant gauge while it is independent of gauge parameters in the generalized axial gauges. The pole masses in different classes of gauges do not coincide in this case and, therefore, it would be wrong to conclude from a study of the pole in the axial gauge (where it is gauge parameter independent) that the pole represents a physical mass. In fact, the pole mass in this case (in the axial gauge) turns out to be purely imaginary and, therefore, unphysical. Physically, of course, this can be understood from the fact that in $1+1$ dimensions, the Coulomb potential increases linearly with distance leading to confinement and preventing any physical fermions in the asymptotic states.
[**Acknowledgments**]{}
A. D. would like to thank the Departamento de Física Matematica for hospitality where this work was done. This work was supported in part by USP, by CNPq and by FAPESP.
Quantum Chromodynamics
======================
In this appendix, we will briefly indicate how the derivation of the Nielsen identity generalizes to Quantum Chromodynamics in a straightforward manner. For a non-Abelian theory based on the gauge group $SU(N)$ the invariant Lagrangian density is given by $${\cal L}_{\rm inv} = - \frac{1}{4}\, F_{\mu\nu}^{a} F^{\mu\nu, a} + \overline{\psi}_{i} (i D\!\sl -m)\psi_{i},\label{app_1}$$ where $a=1,2,\cdots , N^{2}-1$ and $i=1,2,\cdots , N$ denote the color indices. The covariant derivative for the fermion is defined to be $$D_{\mu}\psi_{i} = \partial_{\mu}\psi_{i} + i gA_{\mu}^{a} (T^{a})_{ij} \psi_{j},\label{app_2}$$ where $T^{a}$ denotes the (Hermitian) generators of the group in the fundamental representation and $g$ denotes the coupling constant. One can generalize the interpolating gauge fixing in this case to be given by $${\cal L}_{\rm\sc GF} = -\frac{1}{2} \left(\Lambda\cdot A^{a}\right)^{2} = \frac{1}{2}\,F^{a}F^{a} + (\Lambda^{\mu}F^{a}) A_{\mu}^{a}.\label{app_3}$$ where $\Lambda^{\mu} = \Lambda^{\mu}(\partial)$ is given in .
The ghost Lagrangian density similarly generalizes to $${\cal L}_{\rm ghost} = (\Lambda^{\mu}\overline{c}^{a}) D_{\mu}c^{a},\label{app_4}$$ where the covariant derivative in the adjoint representation is defined as $$D_{\mu}c^{a} = \partial_{\mu}c^{a} - g f^{abc} A_{\mu}^{b} c^{c},\label{app_5}$$ with $f^{abc}$ representing the structure constants of the group. The Lagrangian density ${\cal L}_{\rm inv} + {\cal L}_{\rm\sc GF} + {\cal L}_{\rm ghost}$ is manifestly invariant under the standard (nilpotent) BRST transformations (see also ) $$\begin{aligned}
\delta A_{\mu}^{a} & = \omega (D_{\mu}c)^{a},\notag\\
\delta \psi_{i} & = -ig\omega c^{a} (T^{a})_{ij}\psi_{j},\notag\\
\delta \overline{\psi}_{i} & = - ig\omega \overline{\psi}_{j} (T^{a})_{ji} c^{a},\notag\\
\delta c^{a} & = \frac{g\omega}{2}\, f^{abc} c^{b}c^{c},\notag\\
\delta\overline{c}^{a} & = -\omega F^{a},\notag\\
\delta F^{a} & = 0,\label{app_6}\end{aligned}$$ where $\omega$ denotes a constant Grassmann parameter. The important thing to note here is that in the non-Abelian theory, the quadratic part of the Lagrangian density is diagonal in the color space so that the two point functions have the same structure as in the Abelian theory except for diagonal color factors such as $\delta^{ab}$ or $\delta_{ij}$. Therefore, much of our discussion goes through with minimal change in this case. For example, the denominator for the fermion propagator can now be related to the fermion two point function as in with the Tr" representing a trace over the spinor as well as the color indices and the normalization factor will be correspondingly different. (These are minor generalizations which have no effect on the final result.)
Since the BRST transformations are somewhat different (there are more composite variations), so are the source terms necessary to derive the Nielsen identity. In this case, the source Lagrangian density is given by
$$\begin{aligned}
{\cal L}_{\rm source} & = J^{\mu, a}A_{\mu}^{a} + J^{a} F^{a} + i (\overline{\chi}_{i}\psi_{i} - \overline{\psi}_{i}\chi_{i} )+ i(\overline{\eta}^{a} c^{a} - \overline{c}^{a}\eta^{a})
+ K^{\mu a} D_{\mu}c^{a} + ig (\overline{M}_{i} c^{a}(T^{a})_{ij}\psi_{j} - \overline{\psi}_{i}(T^{a})_{ij}c^{a}M_{j})\notag\\
& \ \ + K^{a} \left(\frac{g}{2} f^{abc} c^{b}c^{c}\right) + H_{(\alpha)} (\partial^{\mu}\overline{c}^{a}) A_{\mu}^{a} + H_{(\beta)} (\partial_{\sc L}^{\mu}\overline{c}^{a}) A_{\mu}^{a} + \beta H_{(n) \mu} (N^{\mu\nu}\overline{c}^{a}) A_{\nu}^{a},\label{app_7}\end{aligned}$$
where $N^{\mu\nu}$ is defined in . With this we can define the total Lagrangian density for the theory to be $${\cal L}_{\rm\sc TOT} = {\cal L}_{\rm inv} + {\cal L}_{\rm\sc GF} + {\cal L}_{\rm ghost} + {\cal L}_{\rm source}.\label{app_8}$$ One can now follow through the steps in - to obtain the Master identity for the gauge parameter variation of the generating functional for the connected Green’s functions (the Dirac spinor indices have been suppressed on the right hand side for simplicity)
$$\begin{aligned}
\frac{\partial W}{\partial\phi_{(a)}} & = i\int d^{n}x d^{n}y \left(i J^{\mu a} (y) \frac{\delta^{2}W}{\delta H_{(a)} (x)\delta K^{\mu a} (y)} + \overline{\chi}_{i}(y) \frac{\delta^{2}W}{\delta\overline{M}_{i}(y) \delta H_{(a)}(x)} + \frac{\delta^{2}W}{\delta H_{(a)}(x)\delta M_{i}(y)} \chi_{i}(y)\right.\notag\\
& \qquad\qquad\qquad \left. - \frac{\delta^{2}W}{\delta H_{(a)}(x)\delta J^{a}(y)} \eta^{a}(y) - \overline{\eta}^{a}(y) \frac{\delta^{2}W}{\delta H_{(a)}(x)\delta K^{a}(y)}\right).\label{app_9}\end{aligned}$$
This can be compared with the identity for the Abelian case in .
The Legendre transformation with respect to the standard sources for the field variables (see ) takes us to the Master identity for the gauge parameter variation of the effective action which has the form (we are suppressing the Dirac spinor indices on the right hand side for simplicity)
$$\begin{aligned}
\frac{\partial\Gamma}{\partial\phi_{(a)}} & = \int d^{n}x d^{n} y\left[\frac{\delta\Gamma}{\delta A_{\mu}^{a} (y)} \frac{\delta^{2}\Gamma}{\delta H_{(a)}(x) \delta K^{\mu a} (y)} + \frac{\delta\Gamma}{\delta \psi_{i}(y)}\frac{\delta^{2}\Gamma}{\delta \overline{M}_{i} (y) \delta H_{(a)} (x)} + \frac{\delta^{2}\Gamma}{\delta H_{(a)}(x) \delta M_{i}(y)} \frac{\delta\Gamma}{\delta \overline{\psi}_{i}(y)}\right.\notag\\
&\qquad\qquad\qquad \left. - \frac{\delta\Gamma}{\delta c^{a} (y)} \frac{\delta^{2}\Gamma}{\delta H_{(a)} (x) \delta K^{a}(y)}\right],\label{app_10}\end{aligned}$$
where all the fields (and the sources $M_{i}, \overline{M}_{i}$) have been set to zero. (This can be compared with for the Abelian theory.) Taking the functional derivative with respect $\frac{\delta^{2}}{\delta\psi_{j,\beta} (y) \delta\overline{\psi}_{i,\alpha} (x)}$, we can now obtain the gauge parameter variation of the fermion two point function which, in the matrix form (for simplicity), is given by $$\begin{aligned}
\frac{\partial S^{-1} (x-y)}{\partial\phi_{(a)}} & = \int d^{n}z d^{n} w\left({\cal F}^{(a)} (x,z,w) S^{-1} (w-y)\right.\notag\\
&\qquad\quad \left. + S^{-1}(x-w) {\cal G}^{(a)} (w,z,y)\right),\label{app_11}\end{aligned}$$ where we have identified (here we put back all the indices) $$\begin{aligned}
& S^{-1}_{ij,\alpha\beta} (x-y) = \frac{\delta^{2}\Gamma}{\delta\psi_{j,\beta}(y) \delta\overline{\psi}_{i,\alpha}(x)},\notag\\
& {\cal F}^{(a)}_{ik,\alpha\gamma}(x,z,w) = \frac{\delta^{3}\Gamma}{\delta\overline{\psi}_{i,\alpha}(x)\delta H_{(a)}(z)\delta M_{k,\gamma}(w)},\notag\\
& {\cal G}^{(a)}_{kj,\gamma\beta} (w,z,y) = \frac{\delta^{3}\Gamma}{\delta\overline{M}_{k,\gamma}(w) \delta H_{(a)}(z)\delta \psi_{j,\beta}(y)}.\label{app_12}\end{aligned}$$ In momentum space this relation takes the simple (matrix) form $$\frac{\partial S^{-1} (p)}{\partial\phi_{(a)}} = {\cal F}^{(a)} (p) S^{-1}(p) + S^{-1} (p) {\cal G}^{(a)}(p),\label{app_13}$$ where, as in , we have identified the matrices $$\begin{aligned}
{\cal F}^{(a)} (p) & = {\cal F}^{(a)} (-p,0,p),\notag\\
{\cal G}^{(a)} (p) & = {\cal G}^{(a)} (-p,0,p).\label{app_14}\end{aligned}$$ We note that in the matrix form the gauge parameter variation of the fermion two point function has the same form as in , the difference being that in the non-Abelian case, the matrices are matrices in the color as well as spinor space.
As we have mentioned earlier, the fermion two point function as well as the propagator are diagonal in the color space. So, still holds with the understanding that the identity matrix is a matrix in the color as well as the spinor space. This determines (see ) the denominator of the propagator to be given by $$\begin{aligned}
{\cal D} & = - \frac{1}{2^{[n/2]} N}\,{\rm Tr} \left(S^{-1}(p) ({\cal C} (S^{-1}(p))^{T} {\cal C}^{-1})\right)\notag\\
& = - \frac{1}{2^{[n/2]} N}\, {\rm Tr} \left(S^{-1}(p) S^{-1}(-p)\right),\label{app_15}\end{aligned}$$ where the trace is over color as well as spinor indices. We can now determine the variation of the denominator with respect to the gauge parameters using (exactly as was done in ) to be given by $$\frac{\partial{\cal D}}{\partial\phi_{(a)}} = 2 {\cal D}\,{\rm Tr} \left({\cal F}^{(a)} (p) + {\cal G}^{(a)} (p)\right),\label{app_16}$$ and the gauge independence of the pole of the propagator follows exactly as discussed in section [**4**]{}.
[10]{}
R. Tarrach, Nucl. Phys. [**B183**]{}, 384 (1981).
D. Johnston, preprint LPTHE Orsay 86/49 (unpublished).
L. J. Reinders and K. Stam, Phys. Lett. [**B195**]{}, 465 (1987).
N. Gray, D. J. Broadhurst, W. Grafe and K. Schilcher, Z. Phys. [**C48**]{}, 673 (1990).
L. S. Brown, [*Quantum Field Theory*]{}, Cambridge University Press, Cambridge (1992).
J. C. Breckenridge, M. J. Lavelle and T. G. Steele, Z. Phys. [**C65**]{}, 155 (1995).
A. S. Kronfeld, Phys. Rev. [**D58**]{}, 051501 (1999).
N. K. Nielsen, Nucl. Phys. [**101**]{}, 173 (1975).
A. Das, [*Finite Temperature Field Theory*]{}, World Scientific, Singapore (1997).
C. A. Becchi, A. Rouet and R. Stora, Comm. Math. Phys. [**42**]{}, 127 (1975); Ann. Phys. (N.Y.) [**98**]{}, 287 (1976).
W. Kummer, Acta Physica Austriaca [**41**]{}, 316 (1975).
J. Frenkel, Phys. Rev. [**D13**]{}, 2325 (1976).
G. Leibbrandt, Rev. Mod. Phys. [**59**]{}, 1067 (1987).
E. S. Fradkin and I. V. Tyutin, Phys. Rev. [**D2**]{}, 2841 (1970).
J. Bernstein, Nucl. Phys. [**B99**]{}, 541 (1975).
A. Andrasi and J. C. Taylor, Ann. Phys. [**326**]{}, 1053 (2011).
A. Das and J. Frenkel, arXiv:1306.6307.
J. Frenkel and J. C. Taylor, Nucl. Phys. [**B109**]{}, 439 (1976).
H. Pagels, Phys. Rev. [**D15**]{}, 2091 (1977).
W. Konetschny, Nuovo Cim. [**44A**]{}, 465 (1978).
A. Das, J. Frenkel and C. Schubert, Phys. Lett. [**B720**]{}, 414 (2013).
[^1]: $\ $ e-mail: [email protected], [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Monitoring with the Ryle Telescope at 15GHz of the Galactic X-ray transient source GRS1915+105 has revealed a remarkable range of rapid and extended flares which appear to be related to the X-ray emission as recorded by the [*RXTE*]{} all-sky monitor. Quasi-periodic oscillations in the range 20 – 40 min have been found and are probably related to oscillations in the soft X-ray flux.'
author:
- 'G. G. Pooley'
- 'R. P. Fender'
title: 'GRS1915+105: Flares, QPOs and other events at 15GHz'
---
\[poole\]
Introduction
============
The Galactic X-ray transient GRS1915+105 has proved to have a rich structure in its high- and low-energy X-ray emission. In the radio regime, it is no less remarkable: Mirabel & Rodríguez (1995) discovered a double-sided relativistic ejection for which they derive a velocity of 0.92$c$. The distance is estimated, from H[i]{} absorption measurements, to be 12.5 kpc. Radio monitoring had already shown the emission to be highly variable (Rodríguez et al. (1995), Foster et al. (1996)), and we started monitoring the source at 15 GHz in mid-1995. The observational details and further results are described by Pooley & Fender (1997).
Results
=======
Figure 1 shows the data for some 17 months. Individual observations lasted from less than 1 hour to about 6 hours. It can be seen that the variations are unpredictable and frequently very rapid; the flux density can increase from less than 1 mJy to 100 mJy and back in less than a day. The major flare event starting in 1996 July was characterised by relatively smooth variations during the first month or so; subsequently the source became highly variable and showed frequent examples of quasi-periodic variations.
Quasi-periodic oscillations
---------------------------
A characteristic form of the outbursts observed at 15 GHz is an event with a rise-time of less than 5 min, followed by a roughly exponential decay with time constant between 12 and 25 min. These may be isolated, they may recur with apparently random intervals, or they may repeat with some regularity; the QPOs seem to favour intervals near 25 and 40 minutes, although not exclusively so. Some examples are shown in Figure 2.
Comparison with other wavebands
-------------------------------
The pattern of radio emission is related in a rather complex way to that of the soft X-ray, as recorded by the [*RXTE*]{} all-sky monitor. Active (rapidly-varying) X-ray emission is usually, but not always, accompanied by radio emission. During the 1996 July radio flare, the X-ray emission was unusually constant.
Obtaining simultaneous observations of the rapid QPO events is difficult, since they are very unpredictable, but we believe that these observations will be necessary for a better understanding of the source. Perhaps the most intriguing simultaneous observation is that on 1996 Oct 24, when the PCA on the [*RXTE*]{} satellite detected oscillations which appear, on the basis of only a few cycles, to be related in phase with the 15-GHz data.
One observation with the VLBA at 8 GHz, on 1996 May 24, overlapped with a Ryle Telescope observation; the variations at 15 GHz appear about 4 minutes earlier than those at 8 GHz (Dhawan, , ).
Fender et al. (1997) have observed outbursts in the infrared with time-scales and flux densities similar to those of the radio events. We do not yet have simultaneous IR and radio data for these flares.
Fender, R. P. et al., 1997, submitted to
Foster, R. S. et al., 1996, , [**467**]{}, L81–84.
Mirabel, I. F. & Rodríguez, L. F., 1995, , [**92**]{} 11390–11392
Pooley, G. G. & Fender, R. P., 1997, submitted to
Rodríguez, L. F. et al., 1995, , [**101**]{} 173–179
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Hideki Omori[^1]\
Tokyo University of Science
- |
Yoshiaki Maeda[^2]\
Keio University
- |
Naoya Miyazaki[^3]\
Keio University
- |
Akira Yoshioka [^4]\
Tokyo University of Science
title: |
Deformation Expression for Elements of Algebras (III)\
–Generic product formula for ${*}$-exponentials\
of quadratic forms–
---
[**Keywords**]{}: Weyl algebra, Heisenberg Lie algebra, meta-plectic groups, spinor, polar elements.
[**Mathematics Subject Classification**]{}(2000): Primary 53D55, Secondary 53D17, 53D10
In a noncommutative algebra there is no canonical way to express elements in univalent way, which is often called “ordering problem”. In this note we give product formula of the Weyl algebra in generic ordered expression. In particular, the generic product formula of $*$-exponential functions of quadratic forms will be given.
In differential geometry, it is widely accepted that geometrical notion should have coordinate free expression. Obviously, algebraic structure of $({\mathbb C}[{\pmb u}], {*}_{\Lambda})$ depends only on the skew part of ${\Lambda}$. It seems reasonable to accept the independence of ordering principle as a basic principle that the physical implication should be independent of ordered expressions.
In the last section, we mention the independence of ordering principle (IOP), and how this principle breaks down in the system containing $*$-exponential functions of quadratic forms.
As a result, we obtain a kind of “double covering group” of $Sp(m;{\mathbb C})$ which is simply connected, but this contains the double covering group (meta-plectic group) of $Sp(m;{\mathbb R})$. Several extraordinary properties of $*$-exponential functions of quadratic forms will be given.
In these calculus, we found peculiar elements, called polar elements, each of which has infinitely many square roots.
General product formula and intertwiners {#PBW-theorem}
========================================
We start with a little general setting as follows: Let ${\mathfrak S}(n)$ and ${\mathfrak A}(n)$ be the spaces of complex symmetric matrices and skew-symmetric matrices respectively, and ${\mathfrak M}(n){=}
{\mathfrak S}(n)\oplus{\mathfrak A}(n)$. For an arbitrary fixed $n{\times}n$-complex matrix $\Lambda{\in}{\mathfrak M}(n)$, we define a product ${*}_{_{\Lambda}}$ on the space of polynomials ${\mathbb C}[\pmb u]$ by the formula $$\label{eq:KK}
f*_{_{\Lambda}}g=fe^{\frac{i{\hbar}}{2}
(\sum\overleftarrow{\partial_{u_i}}
{\Lambda}{}^{ij}\overrightarrow{\partial_{u_j}})}g
=\sum_{k}\frac{(i{\hbar})^k}{k!2^k}
{\Lambda}^{i_1j_1}\!{\cdots}{\Lambda}^{i_kj_k}
\partial_{u_{i_1}}\!{\cdots}\partial_{u_{i_k}}f\,\,
\partial_{u_{j_1}}\!{\cdots}\partial_{u_{j_k}}g.$$ It is known and not hard to prove that $({\mathbb C}[\pmb u],*_{_{\Lambda}})$ is an associative algebra. Clearly, if $\Lambda$ is symmetric, then the algebra obtained is commutative and is isomorphic to the standard polynomial algebra with ${\hbar}$.
For every $\Lambda$, $\partial_{u_{i}}$ acts as a derivation of the algebra $({\mathbb C}[\pmb u],*_{_{\Lambda}})$. Noting this, we define for any other constant symmetric matrix $K$ a new product $*_{_{\Lambda,K}}$ by the formula $$\begin{aligned}
f*_{_{\Lambda{,}K}}g=&f
e^{\frac{i{\hbar}}{2}
(\sum\overleftarrow{\partial_{u_i}}
{K}^{ij}{*_{_{\Lambda}}}\overrightarrow{\partial_{u_j}})}g\\
=&\sum_{k}\frac{(i{\hbar})^k}{k!2^k}
{K}^{i_1j_1}\cdots{K}^{i_kj_k}
(\partial_{u_{i_1}}\cdots\partial_{u_{i_k}}f){*_{_{\Lambda}}}
(\partial_{u_{j_1}}\cdots\partial_{u_{j_k}}g).
\end{aligned}$$ This is also an associative algebra $({\mathbb C}[\pmb u],*_{_{\Lambda,K}})$. Since $\Lambda$, $K$ are constant matrices and the non-commutativity of matrix algebra is not used in the calculation of the product formula, the new product formula can be rewritten as $$f*_{_{\Lambda,K}}g=
\sum_{k}\frac{(i{\hbar})^k}{k!2^k}
{(\Lambda{+}K)}^{i_1j_1}
\cdots{(\Lambda{+}K)}^{i_kj_k}
\partial_{u_{i_1}}\cdots\partial_{u_{i_k}}f
\partial_{u_{j_1}}\cdots\partial_{u_{j_k}}g$$ by noting that the exchanging indexes of $\partial_{u_{i_1}\cdots u_{i_k}}$ is permitted. That is, ${*}_{_{\Lambda,K}}={*}_{_{\Lambda+K}}$.
This formula may be written as $$\label{eq:prprod}
fe^{\frac{i{\hbar}}{2}(\sum\overleftarrow{\partial_{u_i}}
{(\Lambda{+}K)}^{ij}\overrightarrow{\partial_{u_j}})}g
=
fe^{\frac{i{\hbar}}{2}(\sum\overleftarrow{\partial_{u_i}}
{K}^{ij}
{e^{\frac{i{\hbar}}{2}(\sum\overleftarrow{\partial_{u_k}}
{\Lambda}^{kl}\overrightarrow{\partial_{u_k}})}}
\overrightarrow{\partial_{u_j}})}g.$$
Using a symmetric matrix $K$, we compute $\frac{1}{k!}
(\frac{i{\hbar}}{4}\sum{K}^{ij}
\partial_{u_i}\partial_{u_j})^k(f{*}_{_K}g)$ by noting that this is written as follows: $$\begin{aligned}
\sum_{p{+}q{+}r=k}
\frac{(i{\hbar})^r}{r!2^r}
{K}^{i_1j_1}\cdots&{K}^{i_rj_r}
\partial_{u_{i_1}}\cdots\partial_{u_{i_r}}
\frac{1}{p!}(\frac{i{\hbar}}{4}
\sum{K}^{ij}\partial_{u_i}\partial_{u_j})^pf\\
&\quad\times\partial_{u_{j_1}}\cdots\partial_{u_{j_r}}
\frac{1}{q!}
(\frac{i{\hbar}}{4}\sum{K}^{ij}\partial_{u_i}\partial_{u_j})^qg.
\end{aligned}$$ Using this formula, we have the following formula: $$\label{eq:Hochsch}
\begin{aligned}
e^{\frac{i{\hbar}}{4}\sum{K}^{ij}\partial_{u_i}\partial_{u_j}}
\Big(&\big(e^{-\frac{i{\hbar}}{4}\sum{K}^{ij}
\partial_{u_i}\partial_{u_j}}f\big)
{*_{_\Lambda}}
\big(e^{-\frac{i{\hbar}}{4}\sum{K}^{ij}\partial_{u_i}\partial_{u_j}}g\big)\Big)\\
=&fe^{\frac{i{\hbar}}{2}(\sum\overleftarrow{\partial_{u_i}}
{*_{\Lambda}}{K}^{ij}{*_{_\Lambda}}
\overrightarrow{\partial_{u_j}})}g= f{*}_{_{\Lambda{+}K}}g.
\end{aligned}$$
Set $\Lambda=K{+}J$ where $K$, $J$ are the symmetric part and the skew-part of $\Lambda$ respectively. Since the commutator $[u_i,u_j]={i{\hbar}}J^{ij}$ is given by the skew-part of $\Lambda$, the algebraic structure of $({\mathbb C}[\pmb u], *_{_\Lambda})$ depends only on $J$, whose isomorphism class may be denoted by $({\mathbb C}[\pmb u], *_{_J})$ or simply by $({\mathbb C}[\pmb u], *)$ by noticing this class consists of a [*single*]{} algebra.
This is confirmed directly by the formula . Namely, we see the following:
Let $I_0^{^K}(f)=
e^{\frac{i{\hbar}}{4}\sum{K}^{ij}\partial_{u_i}\partial_{u_j}}$, and $I_{_K}^0(f)=
e^{-\frac{i{\hbar}}{4}\sum{K}^{ij}\partial_{u_i}\partial_{u_j}}$. Then $I_0^{^K}$ is an isomorphism of $({\mathbb C}[\pmb u];{*}_{_{\Lambda}})$ onto $({\mathbb C}[\pmb u];{*}_{_{\Lambda+K}})$.
It is clear that the product $f{*_{_{\Lambda}}}g$ is defined if one of $f, g$ is a polynomial and another is a smooth function.
Let $H{\!o}l({\mathbb C}^n)$ be the space of all holomorphic functions on the complex $n$-plane ${\mathbb C}^n$ with the uniform convergence topology on each compact domain. The next one gives a useful remark:
\[frechet\] $H{\!o}l({\mathbb C}^n)$ with the topology above is a Fr[é]{}chet space defined by a countable family of seminorms.
\[extholom\] For every $p(\pmb u)\in{\mathbb C}[\pmb u]$, the left-multiplication $f\to p(\pmb u)*_{_{\Lambda}}f$ and the right-multiplication $f\to f*_{_{\Lambda}}p(\pmb u)$ are both continuous linear mapping of $H\!ol({\mathbb C}^n)$ into itself.
If two of $f, g, h$ are polynomials, then associativity $(f{*_{_{\Lambda}}}g{*_{_{\Lambda}}})h=
f{*_{_{\Lambda}}}(g{*_{_{\Lambda}}}h)$ holds.
Expression parameters and intertwiners {#Expinter}
--------------------------------------
In what follows we treat the case of $2m$ variables, and we use notations $$\label{Weyl}
{\pmb u}=(u_1,u_2,\cdots,u_{2m})=
(\tilde{\pmb u},\tilde{\pmb v}),\quad
\tilde{\pmb u}=(\tilde{u}_1,\cdots,\tilde{u}_m),\,\,
\tilde{\pmb v}=(\tilde{v}_1,\cdots,\tilde{v}_m).$$ The skew part $J$ is fixed to be the standard skew-symmetric matrix $J=\left[
{\footnotesize
{\begin{matrix}
0 & {-}I\\
I & 0
\end{matrix}}}
\right]$. The algebra is called the [**Weyl algebra**]{} and the isomorphism class is denoted by $W_{{\hbar}}(2m)$.
We use sometimes notations $(u_1,\cdots,u_m, v_1,\cdots,v_m)$ instead of $(\tilde{u}_1,\cdots,\tilde{u}_m,\tilde{v}_1,\cdots,\tilde{v}_m)$ when no confusion is suspected.
For the case of a universal enveloping algebra of a Lie algebra, Poincar[é]{}-Birkhoff-Witt theorem ensures that this is realized on the space of ordinary polynomials by giving a new associative product. However, there is no standard way of unique expressing elements of algebra.
Note that if the generator system is fixed, then Proposition\[extholom\] gives a representation of the algebra. The product formula gives also the unique expression of elements of this algebra by the usual polynomials. For instance, computing $u^i{*}u^j{*}u^k$ by using gives the expression of $u^i{*}u^j{*}u^k$ as a polynomial. Thus, the product formula will be referred to $K$-[*ordered expression*]{} (or ${K}$-[*ordering*]{}), i.e. if generators are fixed, giving an ordering expression is nothing but giving a product formula on the space of polynomials which defines the Weyl algebra $W_{{\hbar}}$.
By this formulation of orderings, the intertwiner between $K$-ordered expression and $K'$-ordered expression is explicitly given as follows:
\[intwn\] For every $K, K'\in{\mathfrak S}(n)$, the intertwiner is defined by $$\label{intertwiner}
I_{_K}^{^{K'}}(f)=
\exp\Big(\frac{i{\hbar}}{4}\sum_{i,j}(K^{'ij}{-}K^{ij})
\partial_{u_i}\partial_{u_j}\Big)f \,\,
(=I_{0}^{^{K'}}(I_{0}^{^{K}})^{-1}(f)),$$ and by it gives an isomorphism $I_{_K}^{^{K'}}:({\mathbb C}[{\pmb u}]; *_{_{K+J}})\rightarrow
({\mathbb C}[{\pmb u}]; *_{_{K'+J}})$. Namely, the following identity holds for any $f,g \in {\mathbb C}[{\pmb u}]:$ $$\label{intertwiner2}
I_{_K}^{^{K'}}(f*_{_\Lambda}g)=
I_{_K}^{^{K'}}(f)*_{_{\Lambda'}}I_{_K}^{^{K'}}(g),$$ where $\Lambda=K{+}J$, $\Lambda'=K'{+}J$.
Intertwiners do not change the algebraic structure $*$, but do change the expression of elements by the ordinary commutative structure.
In what follows, we use the notation $*_{_K}$ instead of $*_{_\Lambda}$, since the skew-part $J$ is fixed as the standard skew-matrix. We use notations $${\pmb u}=(u_1,u_2,\cdots,u_{2m})=
(\tilde{\pmb u},\tilde{\pmb v}),\quad
\tilde{\pmb u}=(\tilde{u}_1,\cdots,\tilde{u}_m),\,\,
\tilde{\pmb v}=(\tilde{v}_1,\cdots,\tilde{v}_m).$$
As in the case of one variable, infinitesimal intertwiner $$dI_{_K}(K')=\frac{d}{dt}\Big|_{t=0}I_{_K}^{^{K{+}tK'}}=
\frac{i{\hbar}}{4} {K'}_{ij}\partial_{u_i}\partial_{u_j}$$ is viewed as a flat connection on the trivial bundle $\coprod_{K\in{\mathfrak S}(n)}
H{\!o}l({\Bbb C}^{n})$. The equation of parallel translation along a curve $K(t)$ is given by $$\label{parallel}
\frac{d}{dt}f_t=
dI_{_{\dot K}(t)}(\dot K(t))f_t, \quad
\dot K(t)=\frac{d}{dt}K(t),$$ but this may not have a solution for some initial function.
Note that according to the choice of $K=0, K_0, {-}K_0$, $I$, where $$(0,\,\, K_0, {-}K_0, I)=
\left(
\left[
{\footnotesize
{\begin{matrix}
0 & 0\\
0 & 0
\end{matrix}}}
\right],\,\,
\left[
{\footnotesize
{\begin{matrix}
0 & I\\
I & 0
\end{matrix}}}
\right],\,\,
\left[
{\footnotesize
{\begin{matrix}
0 &\!\!{-}I\\
{-}I&\!\!0
\end{matrix}}}\right],\,\,
\left[
{\footnotesize
{\begin{matrix}
I&\!\!0\\
0&\!\!I
\end{matrix}}}\right]
\right),$$
---------------------------------------------------------
Choice of $K$ (name of ordering)
------------------------ --------------------------------
$K=0$ Weyl ordered expression
$K_0=\left[ Normal ordered expression
{\footnotesize{
\begin{matrix}
0 & I\\
I & 0
\end{matrix}}}
\right]$
$-K_0$ Anti-normal ordered expression
$\left[{\footnotesize{ Unit ordered expression
\begin{matrix}
I& 0\\
0 & I
\end{matrix}}}
\right]$
General $K$ $K$-ordered expression
---------------------------------------------------------
$$\label{ppformula}
\begin{aligned}[c]
f({\pmb u}){*{_{_0}}}g({\pmb u})=&
f\exp
\frac{{\hbar}i}{2}\{\overleftarrow{\partial_{v}}
{\wedge}\overrightarrow{\partial_{u}}\}g,
\quad{\text{(Moyal product formula)}}\\
f({\pmb u}){*{_{_{K_0}}}}g({\pmb u})=&
f\exp {{\hbar}i}\{\overleftarrow{\partial_{v}}\,\,
\overrightarrow{\partial_{u}}\}g,
\qquad{\text{($\Psi$DO product formula)}} \\
f({\pmb u}){*{_{_{{-}K_0}}}}g({\pmb u})=&
f\exp{-{\hbar}i}\{\overleftarrow{\partial_{u}}\,\,
\overrightarrow{\partial_{v}}\}g,
\quad{\text{($\overline{\Psi}$DO product formula)}}
\end{aligned}$$ where $\overleftarrow{\partial_{v}}{\wedge}
\overrightarrow{\partial_{u}}
=\sum_i(\overleftarrow{\partial_{\tilde{v}_i}}
\overrightarrow{\partial_{\tilde{u}_i}}
-\overleftarrow{\partial_{\tilde{u}_i}}
\overrightarrow{\partial_{\tilde{v}_i}})$ and $\overleftarrow{\partial_{v}}\,\,
\overrightarrow{\partial_{u}}
=\sum_i\overleftarrow{\partial_{\tilde{v}_i}}\,\,
\overrightarrow{\partial_{\tilde{u}_i}}$.
The product formula for the unit ordered expression is a bit complicated to write down, but it is easy to obtain. For instance $$u_{*_{_I}}^2{=}u^2{+}\frac{i{\hbar}}{2}, \quad
u{*_{_I}}e^{-\frac{1}{i{\hbar}}u^2}{=}0{=}
e^{-\frac{1}{i{\hbar}}u^2}{*_{_I}}u
\quad e.t.c.$$ while the Weyl ordered expression gives $$v{*_{_0}}e^{-\frac{2}{i{\hbar}}uv}{=}
0{=}e^{-\frac{2}{i{\hbar}}uv}{*_{_0}}u.$$
The next one is trivial, but an important remark:
\[trivrmk\] Every entire function $f(u,v)=\sum a_{kl}u^kv^l$ can be viewed as a $K$-ordered expression of an element of extended Weyl algebra.
The relations between two different expressions are given by intertwiners, but computations in the algebra can be done by using only the associativity and the fundamental commutation relations. Note for instance that $u{*}v{-}v{*}u=-i{\hbar}$ give for every polynomial $p(v{*}v)$ of $v{*}u$ that $$u{*}p(v{*}u)=p(u{*}v){*}u,\quad (\text{bumping identity}).$$ Let $u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v=\frac{1}{2}(u{*}v{+}v{*}u)$; the symmetric product. The bumping identity gives $$u{*}(u{*}v){*}v=u{*}(u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v{+}\frac{1}{2}i{\hbar}){*}v
=(u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v{+}\frac{1}{2}i{\hbar}){*}(u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v{+}\frac{3}{2}i{\hbar}).$$ Throughout this series, we use notation ${:}{\bullet}{:}_{_K}$ to indicate the expression parameter for elements of $W_{{\hbar}}$. For instance, we write $${:}u_i{*}u_j{:}_{_K}{=}u^2{+}\frac{i{\hbar}}{2}(K{+}J)_{ij},\quad
{:}u_j{*}u_j{:}_{_I}=u_j^2{+}\frac{i{\hbar}}{2}\quad etc.$$ A remarkable feature of the first three formulas of is seen as follows: $$\begin{aligned}
&{:}{\tilde u}_j{*}{\tilde v}_j{:}_0={\tilde u}_j{\tilde v}_j{-}\frac{1}{2}i{\hbar}, \quad
{:}{\tilde v}_j{*}{\tilde u}_j{:}_0={\tilde u}_j{\tilde v}_j{+}\frac{1}{2}i{\hbar},\quad(\text{Weyl ordering})\\
&{:}\sum a_{kl}\tilde u_*^k{*}\tilde v_*^l{:}_{_{K_0}}=\sum a_{kl}\tilde u^k\tilde v^l, \qquad
(\text{normal ordering}), \\
&{:}\sum a_{kl}\tilde v_*^k{*}\tilde u_*^l{:}_{_{{-}K_0}}=\sum a_{kl}\tilde u^k\tilde v^l,\quad
(\text{anti-normal ordering}),\\
\end{aligned}$$ but concrete product formulas will be used to extend the algebra transcendentally.
[**Weyl ordered expression**]{}. In general, define $w_*(u_kv_l)$ by $\frac{1}{(k{+}l)!}\sum x_1{*}x_2{*}\cdots{*}x_{k{+}l}$, where $x_i$ is ${\tilde u}_k$ or ${\tilde v}_l$ and the summation runs through all possible rearrangement of $u^k{*}v^l$. $$\label{nikou}
(\tilde u{+}\tilde v)_*^n=\sum_k \,\,{}_nC_kw_*(\tilde u^k\tilde v^{n{-}k}),
\quad {:}w_*(\tilde u^k\tilde v^l){:}_{0}=\tilde u^k\tilde v^{l}.$$
[**Special expression parameter $K_s$**]{}. In [@OMMY3], we introduced the special ordered expression $K_s$ to control the distribution of singular points of $*$-exponential functions of quadratic forms.
By $K_s$-product formula, we see that $(H\!ol({\mathbb C}^n),*_{_{K_s}})$ contains a subalgebra which is isomorphic to the Clifford algebra $Clif\!f(m)$.
[**Siegel class of expression parameters**]{}. In [@OMMY3], we introduced the class ${\mathfrak S}_+({\mathbb R}^n)$ of expression parameters and gave several remarks. This is $${\mathfrak S}_+({\mathbb R}^n)=
\{K; {\rm{Re}}\,\frac{1}{{\hbar}}\langle \xi(iK),\xi\rangle\geq c_{K}|\xi|^2,
\quad \exists c_k>0,~\forall\xi\in {\mathbb R}^n\}$$ which will be called the [*imaginary positive definite class*]{} or the [*Siegel class*]{}. ${\mathfrak S}_+({\mathbb R}^n)$ is $G\!L(n,{\mathbb R})$-invariant. Expressions in this class is easy to treat up to $*$-exponential functions of linear forms and their integrals. Further remarks will be given in the last section.
### Linear change of generators
Next, we consider the effect of linear changes of generators such as $${u'}_i=\sum u_kS_i^k, \quad S\in G\!L(n,{\mathbb C}),\quad
({\pmb u}'={\pmb u}S).$$ Since $\partial_{u_i}=\sum S_i^k\partial_{u'_k}$, the product formula is rewritten by using new generators as $$\label{eq:KK2}
f*_{_{\Lambda}}g=fe^{\frac{i{\hbar}}{2}
(\sum\overleftarrow{\partial_{u'_i}}
({}^t\!S{\Lambda}S)^{ij}\overrightarrow{\partial_{u'_j}})}g.$$ Hence the notation $*_{_\Lambda}$ is better to be replaced $*_{_{\Lambda'}}$ where $\Lambda'{=}{}^t\!S{\Lambda}S$.
Therefore the algebraic structure of $({\mathbb C}[\pmb u], *_{_\Lambda})$ depends only on the conjugate class of the skew part $J$. If ${}^t\!SJS=J$, that is, $S$ is a symplectic linear change of generators such as $${u'}_i=\sum u_kS_i^k, \quad S\in Sp(m,{\mathbb C}),$$ the mapping ${\pmb u}\to {\pmb u}'$ does not change the algebraic structure. Change of generators are viewed often as coordinate transformations, but note here that $I_{_K}^{^{{}^t\!SKS}}$ is something like the “square root” of symplectic coordinate transformations.
Since $\det S=1$ for $S\in Sp(m,{\mathbb C})$, we see $\det{}^t\!SKS=\det K$, hence the isomorphic change by the intertwiner $I_{_K}^{^{K'}}$ can not be covered by a coordinate transformation if $\det K\not=\det K'$.
Star-exponential functions of linear functions
----------------------------------------------
For ${\pmb a}, {\pmb b}\in{\mathbb C}^{2m}$, we set $\langle{\pmb a}{\varGamma},{\pmb b}\rangle
=\sum_{ij=1}^{2m}{\varGamma}^{ij}a_ib_j$, $\langle{\pmb a},{\pmb u}\rangle=\sum_{i=1}^{2m} a_iu_i$. These will be denoted also by ${\pmb a}{\varGamma}\,{}^t{\pmb b}$ and $\langle{\pmb a},{\pmb u}\rangle={\pmb a}\,{}^t\!{\pmb u}$. For $f(\pmb u)\in H{\!o}l({\mathbb C}^{2m})$, the direct calculation via the product formula by using Taylor expansion gives the following: $$\label{extend}
\begin{aligned}
&e^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{*_{_K}}f({\pmb u})=
e^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
f({\pmb u}{+}\frac{s}{2}{\pmb a}(K{+}J)),\\
&
f({\pmb u}){*_{_K}}
e^{-s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}=
f({\pmb u}{+}\frac{s}{2}{\pmb a}(-K{+}J))
e^{-s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
\end{aligned}$$ as natural extension of the product formula. This gives also the associativity $$\label{adjointlike}
\Big(e^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{*_{_K}}f({\pmb u})\Big){*_{_K}}e^{t\frac{1}{i{\hbar}}\langle{\pmb b},{\pmb u}\rangle}
=
e^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{*_{_K}}\Big(f({\pmb u}){*_{_K}}e^{t\frac{1}{i{\hbar}}\langle{\pmb b},{\pmb u}\rangle}\Big),
\quad f(\pmb u)\in H{\!o}l({\mathbb C}^{2m}).$$
By a direct calculation of intertwiner, we see that $$\label{eq:intwin}
I_{_K}^{^{K'}}(e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle})
=e^{\frac{1}{4i{\hbar}}\langle{\pmb a}(K'{-}K),{\pmb a}\rangle}
e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}.$$ Hence, $\{e^{\frac{1}{4i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle}
e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle};
K\in{\mathfrak S}_{\mathbb C}(2m)\}$ is a parallel section of $\coprod_{K\in{\mathfrak S}_{\mathbb C}(2m)}
{H{\!o}l}({\Bbb C}^{2m}).$
We denote this element symbolically by $e_*^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$. Namely we denote $$\label{eq:tempexp}
:e_*^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}:_{_K}
=e^{\frac{1}{4i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle}
e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
=e^{\frac{1}{4i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle
{+}\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}.$$ By using the product formula in $K$-ordered expression, we have easily the exponential law $${:}e_*^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}}{*_{_{K}}}
{:}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}}=
{:}e_*^{(s{+}t)\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}},\,\,
\forall K\in {\mathfrak S}(2m).$$ The exponential law may be written by omitting the suffix $K$ as $$e_*^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{*}
e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}=
e_*^{(s{+}t)\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle},\quad
e^{s}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}=
e_*^{s{+}t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$$ together with the exponential law with the ordinary exponential functions.
Furthermore note also that ${:}\langle{\pmb a},{\pmb u}\rangle{:}_{_{K}}=
\langle{\pmb a},{\pmb u}\rangle$ for every $K$, and $e_*^{\frac{s}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$ is the solution of the evolution equation $$\frac{d}{dt}{:}e_*^{\frac{t}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}}
=\frac{1}{i{\hbar}}{:}\langle{\pmb a},{\pmb u}\rangle{:}_{_{K}}{*_{_K}}
{:}e_*^{\frac{t}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_{K}}
\,\,{\text{with initial data}}\,\, {:}e_*^{\frac{0}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_{K}}=1.$$
$e_*^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}=
\{e^{s^2\frac{1}{4i{\hbar}}\langle{\pmb a}K\,{\pmb a}\rangle}
e^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle};
K\in{\mathfrak S}(2m)\}$ forms a one parameter group of parallel sections.
By applying to ${:}e_*^{\pm s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}$, we have for every $f\in H\!ol({\mathbb C}^n)$ the associativity $$\label{adjointlike}
{:}(e_*^{s\frac{1}{i{\hbar}}
\langle{\pmb a},{\pmb u}\rangle}{*}f_*({\pmb u}))
{*}e_*^{-s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}
={:}f_*({\pmb u}{+}{s}{\pmb a}J){:}_{_K}=
{:}e_*^{s\frac{1}{i{\hbar}}
\langle{\pmb a},{\pmb u}\rangle}{*}(f_*({\pmb u})
{*}e_*^{-s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}){:}_{_K}.$$ This gives also the real analyticity of $e_*^{s\frac{1}{i{\hbar}}
\langle{\pmb a},{\pmb u}\rangle}{*}f_*({\pmb u})
{*}e_*^{-s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$ in $s$.
It is remarkable that if $K=0$, then $:e_*^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}:_{_K}=
e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$, that is, $*$-exponential functions of linear functions are ordinary exponential functions in Weyl ordered expression. On the other hand, if $K\in{\mathfrak S}_+({\mathbb R}^n)$ then ${:}e_*^{\pm s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}$ has a very strong property that $${:}e_*^{\pm s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}
=e^{\frac{s^2}{4i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle}
e^{\pm\frac{s}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$$ is rapidly decreasing in $s\in{\mathbb R}$.
### Extension of products {#Extprodpp}
For every positive real number $p$, we set $$\label{sysnorm1}
{\mathcal E}_p({\Bbb C}^{n})=
\{f \in H\!ol({\Bbb C}^{n})\, ;\,
\|f\|_{p,s}=\sup\, |f|\, e^{-s|\xi|^p}<\infty,\,\,\forall s>0\}$$ where $|\xi|=(\sum_i|u_i|^2)^{1/2}$. The family of seminorms $\{||\,\cdot\,||_{p,s}\}_{s>0}$ induces a topology on $ {\mathcal E}_p({\Bbb C}^{n} )$ and $({\mathcal E}_p({\Bbb C}^{n} ),\cdot)$ is an associative commutative Fréchet algebra, where the dot $\cdot$ is the ordinary product for functions in ${\mathcal E}_p({\Bbb C}^{n})$. It is easily seen that for $0<p<p'$, there is a continuous embedding $${\mathcal E}_p({\Bbb C}^{n} )\subset
{\mathcal E}_{p'}({\Bbb C}^{n} )$$ as commutative Fréchet algebras (cf.[@GS]), and that ${\mathcal E}_p({\Bbb C}^{n})$ is $G\!L(n,{\mathbb C})$-invariant.
We denote $${\mathcal E}_{p+}({\Bbb C}^{n})
=\bigcap_{p'>p}{\mathcal E}_{p'}({\Bbb C}^{n}),\quad
(\text{with the intersection topology})$$ It is obvious that every polynomial is contained in ${\mathcal E}_{p}({\Bbb C}^{n})$, that is $p({\pmb u})\in {\mathcal E}_{0+}({\Bbb C}^{n})$, and $\Bbb C[{\pmb u}]$ is dense in ${\mathcal E}_{p}({\Bbb C}^{n})$ for any $p>0$ in the Fr[é]{}chet topology defined by the family of seminorms $\{||\,\,||_{p,s}\}_{s>0}$.
We easily see that $e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
\in{\mathcal E}_{1+}({\Bbb C}^{n})$. Moreover, it is not difficult to show that an exponential function $e^{p({\pmb u})}$ of a polynomial of degree $d$ is contained in ${\mathcal E}_{d+}({\Bbb C}^{n})$, but not in ${\mathcal E}_{d}({\Bbb C}^{n})$.
Theorems \[main01\] and \[main02\] stated below give basic tools to study $*$-functions (cf. [@ommy] for the proof), although most of the concrete formulas can be obtained without these theorems.
\[main01\] For $0<p\leq 2$, the product formula extends to give the following[:]{}
$(1)$ The space $({\mathcal E}_{p}({\Bbb C}^{n}),*_{_{K}})$ forms a complete noncommutative topological associative algebra over $\mathbb C$. $(2)$ The intertwiner $I_{_K}^{^{K'}}$ extends to give an isomorphism of $({\mathcal E}_{p}({\Bbb C}^{n}),*_{_{K}})$ onto $({\mathcal E}_{p}({\Bbb C}^{n}),*_{_{K'}})$.
[**Remark**]{}For the second statement, it is enough to prove that $I_{_K}^{^{K'}}$ extends to give a linear isomorphism of ${\mathcal E}_{p}({\Bbb C}^{n})$ onto itself. The property (2) shows that if $p\leq 2$, $\coprod_{K\in{\mathfrak S}(n)}
{\mathcal E}_p({\Bbb C}^{n})$ is a trivial subbundle, and this is in fact an algebra bundle $$\coprod_{K\in{\mathfrak S}(n)}
({\mathcal E}_{p}({\Bbb C}^{n}),*_{_K}).
\quad (0<p\leq 2)$$ The equation of parallel translation has a unique solution for the initial function $f$ is in ${\mathcal E}_{p}({\Bbb C}^{n})$, $0<p\leq 2$.
It is easily seen that the following identities hold on ${\mathcal E}_{p}({\Bbb C}^{n})$, $p\leq 2$ $$\label{eq:intid}
I_{_{K'}}^{^{K}}I_{_{K}}^{^{K'}}=1,\quad
I_{_{K'}}^{^{K''}}I_{_K}^{^{K'}}=I_{_K}^{^{K''}}.$$ Hence, for every $f\in {\mathcal E}_{p}({\Bbb C}^{n})$, the set $f_*({\pmb u})=
\{I_{0}^{^{K}}(f); K\in{\mathfrak S}_{\mathbb C}(n)\}$ is a globally defined parallel section.
For $p>2$, we note the following:
\[main02\] For $p>2$, the product formula gives continuous bi-linear mappings of $${\mathcal E}_{p}({\Bbb C}^{n})\times
{\mathcal E}_{p'}({\Bbb C}^{n})\rightarrow
{\mathcal E}_{p}({\Bbb C}^{n}),\quad
{\mathcal E}_{p'}({\Bbb C}^{n})\times
{\mathcal E}_{p}({\Bbb C}^{n})\rightarrow
{\mathcal E}_{p}({\Bbb C}^{n}),$$ for $\forall p'$ such that $\frac{1}{p}+\frac{1}{p'}\geq 1$.
For $f, g, h\in{\mathcal E}_{p}({\Bbb C}^{n})$ $(p>2)$, the associativity $(f{*}_{_{K}}g){*}_{_{K}}h=f{*}_{_{K}}(g{*}_{_{K}}h)$ holds if two of $f, g, h$ are in ${\mathcal E}_{p'}({\Bbb C}^{n})$ such that $\frac{1}{p}+\frac{1}{p'}\geq 1$.
Note that the linear change of coordinate ${\pmb u}'={\pmb u}S$ by $S\in G\!L(n,{\mathbb C})$ gives naturally the topological linear isomorphism $\Phi_S:{\mathcal E}_{p}({\Bbb C}^{n})\to
{\mathcal E}_{p}({\Bbb C}^{n}),\,\, p \geq 0,$ and this is an isomorphism as ${\mathbb C}[{\pmb u}]$-bi-modules for every $p\geq 0$; $$\Phi_S:({\mathcal E}_{p}({\Bbb C}^{n}); *_{_\Lambda})\to
({\mathcal E}_{p}({\Bbb C}^{n}; *_{_{{}^t\!S\Lambda S}}).$$ This is not an automorphism, but an [*outer isomorphism*]{}.
### Remarks on elements obtained by integrals
Suppose $f(x)$ is a continuous mapping of a compact domain $D$ into ${\mathcal E}_{p}({\Bbb C}^{n})$. As $\|f(x)\|_{p,s}$ is bounded on $D$, its integral over $D$ is bounded. Hence we see $\int_Df(x)dx \in {\mathcal E}_{p}({\Bbb C}^{n}).$ This will be used to compute Fourier series.
\[safe\] For every compact interval $I$, the integral $\int_{I}
{:}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}}dt$ gives an element of ${\mathcal E}_{1+}({\mathbb C}^{n})$ for every $K\in {\mathfrak S}(n)$, and $$I_{_K}^{^{K'}}\Big(\int_{I}
{:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb a},{\pmb u}\rangle}{:}_{_{K}}dt\Big)
=\int_{I}
{:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb a},{\pmb u}\rangle}{:}_{_{K'}}dt$$ $\{\int_{I}{:}
e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_{K}}dt;
K\in {\mathfrak S}(n)\}$ is a parallel section which may be denoted by $\int_{I}
e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}dt$ without showing expression parameters.
Since $e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
\in{\mathcal E}_{1+}({\Bbb C}^{n})$, Theorem \[main02\] shows that $e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{*_{_K}}f$, $f{*_{K}}e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$ are defined for every $f\in \bigcup_{p>2}{\mathcal E}_{p}({\Bbb C}^{n})$, but in fact these are defined for every $f\in H\!ol({\mathbb C}^n)$ by .
As we have seen above, ${*}$-product integrals of exponential functions of linear functions are remained in the class. ${\mathcal E}_{1+}({\mathbb C}^{2m})$. Note that usual integral can be defined for elements of ${\mathcal E}_{1+}({\mathbb C}^{2m})$. However, we have to be careful to use the integral on noncompact domain, for such integrals often give elements outside the domain where the integrand is considered. Here, we give a typical example.
Suppose ${\rm{Re}}(\frac{1}{i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle)<0$, that is $K$ is in the Siegel class. Then, the integral $\int_{-\infty}^{\infty}
{:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}dt$ converges. The formula of Fourier transform gives $$\int_{-\infty}^{\infty}
{:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}dt=
\int_{\mathbb R}e^{\frac{t^2}{4i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle}e^{t\frac{1}{i{\hbar}}
\langle{\pmb a},{\pmb u}\rangle}dt=
2(\frac{-i{\hbar}\pi}{\langle{\pmb a}K,{\pmb a}\rangle})^{1/2}
\,\,e^{-\frac{1}{i{\hbar}}\frac{1}{\langle{\pmb a}K,{\pmb a}\rangle}
\langle{\pmb a},{\pmb u}\rangle^2}.$$ Since $\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle{*}:
H\!ol({\mathbb C}^{2m})\to H\!ol({\mathbb C}^{2m})$ is continuous, we see that $$\label{eq:confirm}
\begin{aligned}
\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle{*}
\int_{-\infty}^{\infty}
e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}dt=
\lim_{N,N'\to\infty}
\int_{-N}^{N'}\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle{*}
e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}dt
=\int_{-\infty}^{\infty}
\frac{d}{dt}
e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}dt=0.
\end{aligned}$$ Since $\langle{\pmb a},{\pmb u}\rangle{*_{_K}}
f(\langle{\pmb a},{\pmb u}\rangle)
=\langle{\pmb a},{\pmb u}\rangle
f(\langle{\pmb a},{\pmb u}\rangle)
{+}\frac{i{\hbar}}{2}\langle{\pmb a}K,{\pmb a}\rangle
f'(\langle{\pmb a},{\pmb u}\rangle)$, the direct calculation also gives $$\langle{\pmb a},{\pmb u}\rangle{*_{_K}}
e^{-\frac{1}{i{\hbar}}\frac{1}{\langle{\pmb a}K,{\pmb a}\rangle}
\langle{\pmb a},{\pmb u}\rangle^2}=0.$$ Under the condition ${\rm{Re}}(\frac{1}{i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle)<0$, we denote as in [@OMMY3] $$\frac{1}{2\pi{\hbar}}\int_{-\infty}^{\infty}
e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}dt
=\delta_*(\langle{\pmb a},{\pmb u}\rangle)$$ Moreover, we see that integrals $\int_{-\infty}^{0}
{:}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}dt,
\,\,
{-}\!\int_{0}^{\infty}
{:}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}dt$ are both inverses of $\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle$. We denote these by $$(\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle)_{*+}^{-1}
=\int_{-\infty}^{0}
{:}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}dt,
\quad
(\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle)_{*-}^{-1}=
-\int_{0}^{\infty}
{:}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_K}dt.$$ This apparently breaks associativity $$\Big((\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle)_{*+}^{-1}{*_{_K}}
\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle\Big){*_{_K}}
(\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle)_{*-}^{-1}
\not=
(\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle)_{*+}^{-1}{*_{_K}}
\Big(\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle{*_{_K}}
(\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle)_{*-}^{-1}\Big).$$
### Remarks on real analyticity and on associativity
A mapping $f: U\to F$ from an open subset $U$ of $\mathbb R$ into a Fr[é]{}chet space $F$ is called to be [**real analytic**]{}, if for every $a\in U$ there is ${\varepsilon}(a)>0$ such that $f$ is written in the form $$f(a+s)=\sum_k \frac{1}{k!}a_k s^k, \quad a_k\in F, \quad
|s|<{\varepsilon}(a).$$ $a_k$ is given by $a_k=\partial_s^kf|_{s=0}$.
If $F$ is a Banach space and $\sum_k\frac{1}{k!}\|a_k\||s|^k$ converges, then the power series $\sum_k a_k s^k$ is called to [converge absolutely]{} under the norm.
If a Fr[é]{}chet space $F$ is defined by a countable family of seminorms $\{\|f\|_m; m=1,2,3\cdots\}$, then replace this part by the absolute convergence of $\sum_k\frac{1}{k!}\|a_k\|_m|s|^k$ w.r.t. seminorms $\|\cdot\|_m$. A power series $\sum_k a_k s^k$ converges if this converges absolutely under every seminorms.
[**Radius of convergence**]{}Suppose a Fr[é]{}chet space $F$ is defined by a countable family of seminorms $\{\|f\|_m; m=1,2,3\cdots\}$.
\[powerser\] For a power series $\sum_k a_k s^k$, $a_k\in F$, there exists uniquely a real number $R$ $(0\leq R\leq \infty)$ satisfying $(1)$ and $(2)$ below:
(1)
: If $|s|<R$, then the power series $\sum_k a_k s^k$ converges absolutely under every seminorm $\|\cdot\|_m$.
(2)
: If $|s|>R$, then $\sum_k a_k s^k$ does not converge for some seminorm.
[**Proof**]{}Suppose $\sum_k a_k s_0^k$ converges at $s_0$. Then $a_k s_0^k$ is bounded under every seminorm $\|\cdot\|_m$. Set $\sup_k\|a_k s_0^k\|_m\leq M_m$. Then for every $s$ such that $|s|<|s_0|$ $$\sum_k\|a_ks^k\|_m\leq \sum_k M_m|s/s_0|^k
= M_m\frac{1}{1-|s/s_0|}<\infty.$$ It follows the convergence of $\sum_ka_ks^k$. ${}$$\square$
\[powbibun\] $\sum_{k\geq 0}a_k s^k$ and $\sum_{k\geq 1}ka_k s^{k-1}$ have same radius of convergence.
Real analyticity is left invariant under every continuous linear transformation:
\[linsbl\] Let $F, G$ be Fr[é]{}chet spaces and $\varphi : F\to G$ be a continuous linear mapping. If $f: U\to F$ is real analytic, then $\varphi f: U\to G$ is also real analytic.
Since $X\to p({\pmb u}){*}X{*}q({\pmb u})$ is a continuous linear mapping, Lemma\[frechet\] gives the following:
\[realanal\] Let $U$ be an connected open neighborhood of $0$ of ${\mathbb R}^{\ell}$ Suppose $\psi: U\to H{\!o}l({\mathbb C}^n)$ be a real analytic mapping. Then $x\to p({\pmb u}){*}\psi(x){*}q({\pmb u})$ is also a real analytic on $U$ for every polynomial $p({\pmb u}),\,q({\pmb u})$.
[**Remarks on the associativity**]{}
Products of exponential functions of quadratic forms may not be defined, and even if the product is defined the associativity may not hold, since these are elements of ${\mathcal E}_{2+}({\Bbb C}^{n})$. In general, we do not have the associativity even for a polynomial $p({\pmb u})$ $$(e^{H({\pmb u})}{*}p({\pmb u})){*}e^{K({\pmb u})},
\quad e^{H({\pmb u})}{*}(p({\pmb u}){*}e^{K({\pmb u})}),$$ since $p(u)$ has two different $*$-inverses in general.
However, if we can treat elements in $({\mathbb C}[\pmb u][[{\hbar}]], {*_{_K}})$, the space of formal power series of ${\hbar}$, then $*_{_\Lambda}$-product is always defined by the product formula and the associativity holds.
Elements of ${\mathcal E}_{2+}({\Bbb C}^{n})$ are often given as a real analytic function of ${\hbar}$ defined on certain interval containing ${\hbar}=0$. The following is easy to see:
\[assocthm\] Suppose $f({\hbar},{\pmb u})$, $g({\hbar},{\pmb u})$ and $h({\hbar},{\pmb u})$ are given as real analytic function of ${\hbar}$ in some interval $[0,H]$. If $$f({\hbar},{\pmb u}){*_{_K}}g({\hbar},{\pmb u}),\,\,
(f({\hbar},{\pmb u}){*_{_K}}g({\hbar},{\pmb u})){*_{_K}}h({\hbar},{\pmb u}),\,\,
g({\hbar},{\pmb u}){*_{_K}}h({\hbar},{\pmb u}),\,\,
f({\hbar},{\pmb u}){*_{_K}}(g({\hbar},{\pmb u}){*_{_K}}h({\hbar},{\pmb u}))$$ are defined as real analytic functions on $[0,H]$, then the associativity $$(f({\hbar},{\pmb u}){*_{_K}}g({\hbar},{\pmb u})){*_{_K}}h({\hbar},{\pmb u})
=f({\hbar},{\pmb u}){*_{_K}}(g({\hbar},{\pmb u}){*_{_K}}h({\hbar},{\pmb u}))$$ holds.
We refer this theorem to the [**formal associativity theorem**]{}.
[**Remark 1**]{}. In what follows, elements are often given in the form $f(\frac{1}{i{\hbar}}\varphi(t),{\pmb u})$ by using a real analytic function $f(t,{\pmb u})$, $t{\in}[0,T]$, where $\varphi(t)$ is a real analytic function such that $\varphi(0){=}0$. (Cf., In such a case, replacing $t$ by $s{\hbar}$ gives a real analytic function of ${\hbar}$, and such an element is embedded in $({\mathbb C}[\pmb u][[{\hbar}]], {*_{_K}})$. Thus, we can apply the above theorem. We call such elements [**classical elements**]{}. However, there are many elements in ${\mathcal E}_{2+}({\Bbb C}^{n})$ written in the form $f(\frac{1}{i{\hbar}}\varphi(t),{\pmb u})$ such that $\varphi(0){\not=}0$.
Blurred covering group of $Sp(m,{\mathbb C})$
=============================================
In this section we first treat the infinitesimal $*$-action of quadratic forms on the space of exponential functions of quadratic forms. We treat this in general expressions by using intertwiners. Since the space of quadratic forms is isomorphic to the Lie algebra of $Sp(m,\mathbb C)$, i.e. $$\{\langle{\pmb u}A,{\pmb u}\rangle; A\in{\mathfrak S}(2m)\}
\cong {\mathfrak{sp}}(m,{\mathbb C})=
\{\alpha; \alpha J{+}J\,{}^t\!\alpha=0\}$$ as Lie algebra, the natural ${*}$-action of quadratic forms can be viewed as the [**infinitesimal action**]{} of the Lie group $Sp(m,\mathbb C)$.
In contrast with that infinitesimal intertwiners are viewed as a flat connection on the trivial bundle $\coprod_{K{\in}{\mathfrak S}(2m)}{\mathbb C}
e^{{\mathfrak S}(2m)}$, whose fiber is the space of exponential functions of quadratic forms ${\mathbb C}e^{{\mathfrak S}(2m)}$, all possible infinitesimal actions of quadratic forms gives a tangential distribution on each fiber $\coprod_{K{\in}{\mathfrak S}(2m)}{\mathbb C}
e^{{\mathfrak S}(2m)}$.
Infinitesimal actions of quadratic forms {#infaction}
----------------------------------------
On every fiber at $K$, consider left multiplication $${:}\langle{\pmb u}A,{\pmb u}\rangle{:}_{_K}{*_{_K}} :
{\mathbb C}e^{{\mathfrak S}(2m)}
\to {\mathbb C}e^{{\mathfrak S}(2m)}$$ Since $\frac{1}{i\hbar}{:}\langle{\pmb u}A,{\pmb u}\rangle{:}_{_K}=
\frac{1}{i\hbar}\langle{\pmb u}A,{\pmb u}\rangle
+\frac{1}{2}{\rm Tr}(A\!K)$, we see $$\label{bulletone}
\frac{1}{i\hbar}{:}\langle{\pmb u}A,{\pmb u}\rangle{:}_{_K}{*_{_K}}
(ge^{\frac{1}{i\hbar}\langle{\pmb u}Q,{\pmb u}\rangle})
=\Big(\frac{1}{2}{\rm{Tr}}\big((K{-}J)A(K{+}J)Q+A\!K\big)
{+}\frac{1}{i\hbar}\langle{\pmb u}{Q'},{\pmb u}\rangle\Big)
ge^{\frac{1}{i\hbar}\langle{\pmb u}{Q}),{\pmb u}\rangle}
$$ where ${Q'}=
{A}+{A}(K{+}J){Q}+{Q}(K{-}J){A}
+{Q}(K{-}J){A}(K{+}J){Q}$, and $A\in{\mathfrak S}(2m)$. The term $\frac{1}{i\hbar}\langle{\pmb u}{Q'},{\pmb u}\rangle$ will be called the [*infinitesimal phase*]{} part, and $\frac{1}{2}{\rm{Tr}}\big((K{-}J)A(K{+}J)Q+A\!K\big)$ will be called the [*infinitesimal amplitude*]{} part.
Moving $A\in{\mathfrak S}(2m)$ at every fixed $ge^{\frac{1}{i\hbar}\langle{\pmb u}Q,{\pmb u}\rangle}$, we have a linear subspace of the tangent space of ${\mathbb C}e^{{\mathfrak S}(2m)}$ at $ge^{\frac{1}{i\hbar}\langle{\pmb u}Q,{\pmb u}\rangle}$. We call this the singular distribution of infinitesimal actions of quadratic forms.
On the other hand, there is a natural correspondence between ${\mathfrak{sp}}(m;\mathbb C)$ and ${\mathfrak S}(2m)$. $${\mathfrak{sp}}(m;\mathbb C)\cong {\mathfrak S}(2m)\quad
{\text{via}}\quad
\alpha\in {\mathfrak{sp}}(m;\mathbb C)
\Leftrightarrow \alpha J\in{\mathfrak S}(2m),
\quad
J=
\begin{bmatrix}
0&{-}1\\
1&0
\end{bmatrix}.$$ We make the correspondence as follows: $$A \Leftrightarrow\alpha{=}{-}AJ,\qquad
Q \Leftrightarrow\xi{=}{-}QJ.$$ We set also ${\kappa'}{=}J\!K'$, $\kappa{=}J\!K$ in ${\mathfrak{sp}}(m;\mathbb C)$. Through these, intertwiners $I_{_K}^{^{K'}}$ defined on $\coprod_{K\in{\mathfrak S}(2m)}{\mathbb C}
e^{{\mathfrak S}(2m)}$ in §3.1 of [@OMMY4] is easily translated on $\coprod_{\kappa\in{\mathfrak{sp}}(m;\mathbb C)}
{\mathbb C}e^{{\mathfrak{sp}}(m;\mathbb C)J}$ as $$I_{\kappa}^{\kappa'}(ge^{\langle{\pmb u}(\frac{1}{i{\hbar}}\alpha J),{\pmb u}\rangle})
=\frac{g}{\sqrt{\det(I{-}\alpha(\kappa'{-}\kappa))}}
e^{\langle{\pmb u}(\frac{1}{i{\hbar}}\frac{1}{I{-}\alpha(\kappa'{-}\kappa)}\alpha J),
{\pmb u}\rangle} .$$ These intertwiners may be viewed as coordinate transformations: $I_{\kappa}^{\kappa'}:{\mathbb C}
e^{{\mathfrak{sp}}(m;\mathbb C)J}\to
{\mathbb C}e^{{\mathfrak{sp}}(m;\mathbb C)J}$. For the precise treatment of patching by intertwiners, we set $$I_{\kappa}^{\kappa'}\frac{1}{\sqrt{\det(I{-}\alpha\kappa))}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}
(\frac{1}{I{-}\alpha\kappa}\alpha J),{\pmb u}\rangle}=
\frac{1}{\sqrt{\det(I{-}\alpha\kappa'))}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}
(\frac{1}{I{-}\alpha\kappa'}\alpha J),{\pmb u}\rangle}$$ $$\widetilde{\mathcal D}_{\kappa}
=\{
\frac{1}{\sqrt{\det(I{-}\alpha\kappa))}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}
(\frac{1}{I{-}\alpha\kappa}\alpha J),{\pmb u}\rangle}; \alpha\in
{\mathcal D}_{\kappa}\},\quad
{\mathcal D}_{\kappa}{=}\{\alpha; \det ({I{-}\alpha\kappa) }\not=0\}.$$ As $\kappa$ moves in the whole space ${sp}(m,{\mathbb C})$, and for any $\alpha$, we can find $\kappa$ such that $\det(I{-}\alpha\kappa){\not=}0$, we easily see that $$\label{spspsp}
\bigcup_{\kappa}{\mathcal D}_{\kappa}=
{\mathfrak{sp}}(m,{\mathbb C}), \quad
\bigcap_{\kappa}{\mathcal D}_{\kappa}=\{0\}.$$
$\widetilde{\mathcal D}_{\kappa}$ is a double cover of ${\mathcal D}_{\kappa}$. (Recall we set $\sqrt{1}=\{\pm 1\}$ in the case $\kappa=0$.) Let $\pi:\widetilde{\mathcal D}_{\kappa}\to{\mathcal D}_{\kappa}$ be the natural projection. As it was seen in in §3.1 of [@OMMY4] intertwiners $I_{\kappa}^{\kappa'}$ give isomorphisms $$\begin{matrix}
\widetilde{\mathcal D}_{\kappa}& \supset &
\pi^{-1}({\mathcal D}_{\kappa}{\cap}{\mathcal D}_{\kappa'})&
\overset{I_{\kappa}^{\kappa'}}{\longrightarrow}&
\pi^{-1}({\mathcal D}_{\kappa'}{\cap}{\mathcal D}_{\kappa})&
\subset &\widetilde{\mathcal D}_{\kappa'}\\
\downarrow\,\pi& &\downarrow\,\pi& &\downarrow\,\pi&
&\downarrow\,\pi \\
{\mathcal D}_{\kappa}& \supset &
{\mathcal D}_{\kappa}{\cap}{\mathcal D}_{\kappa'}&
=\!=&
{\mathcal D}_{\kappa'}{\cap}{\mathcal D}_{\kappa}&
\subset&{\mathcal D}_{\kappa'}
\end{matrix}$$ However intertwiners $I_{\kappa}^{\kappa'}$ are 2-to-2 mappings. Thus, the union $\bigcup_{\kappa}\widetilde{\mathcal D}_{\kappa}$ is a manifold-like object glued by 2-to-2 coordinate transformations.
Set ${\alpha'}{=}{-}Q'J$, ${\xi}{=}{-}QJ$, ${\alpha}{=}{-}AJ$, $\kappa{=}J\!K$. These are $\in {\mathfrak{sp}}(m;\mathbb C)$. We want to translate the equality $\eqref{bulletone}$ by these replacement. First, the infinitesimal phase part is rewritten as $$\begin{aligned}[t]
{\alpha'} =&{\alpha}-{\alpha}(I{-}\kappa){\xi}+{\xi}(I{+}\kappa){\alpha}
-{\xi}(I{+}\kappa){\alpha}(I{-}\kappa){\xi}\\
=&\big(I+{\xi}(I{+}\kappa)\big){\alpha}\big(I-(I{-}\kappa){\xi}\big),
\end{aligned}$$ and it is easy to see $\big(I+{\xi}(I{+}\kappa)\big){\alpha}\big(I-(I{-}\kappa){\xi}\big)\in
{\mathfrak{sp}}(m;\mathbb C)$ by moving $J$ in the l.h.s of\
$J\big(I+{\xi}(I{+}\kappa)\big){\alpha}\big(I-(I{-}\kappa){\xi}\big)$ to the r.h.s. Hence, the equality $\eqref{bulletone}$ is translated into $$\label{diamond}
\begin{aligned}
(\diamondsuit)\quad \frac{1}{i\hbar}
{:}\langle{\pmb u}(\alpha J),{\pmb u}\rangle{:}_{_K}{*_{_K}}
(ge^{\frac{1}{i\hbar}\langle{\pmb u}\,\xi J,{\pmb u}\rangle})
&{=}
\Big(\frac{1}{2}{\rm{Tr}}\big((\kappa{+}I)\alpha(\kappa{-}I)\xi{+}\alpha\kappa\big)
{+}\langle{\pmb u}(\alpha'J),{\pmb u}\rangle\Big)
ge^{\frac{1}{i\hbar}\langle{\pmb u}\,{{\xi}J}),{\pmb u}\rangle}\\
\alpha'&{=}(I+{\xi}(I{+}\kappa)){\alpha}(I-(I{-}\kappa){\xi}).
\end{aligned}$$
By moving $\alpha{\in}{\mathfrak{sp}}(m,{\mathbb C})$ we make a subspace $D_{(\kappa,ge^{\xi})}$ of the tangent space $T_{ge^{\xi}}{\mathbb C}e^{{\mathfrak S}(2m)}$ of ${\mathbb C}e^{{\mathfrak S}(2m)}$ at $ge^{\frac{1}{i\hbar}\langle{\pmb u}\,\xi J,{\pmb u}\rangle}$. We make also a distribution (a singular subbundle): $$\label{distr}
\begin{aligned}
D_{(\kappa,ge^{\xi})}=&
\Big\{\Big(\frac{1}{2}
{\rm{Tr}}\big((\kappa{+}I)\alpha(\kappa{-}I)\xi{+}\alpha\kappa\big)
{+}\frac{1}{i{\hbar}}
\langle{\pmb u}\,{{\alpha'}\!J},{\pmb u}\rangle,\,\,\alpha\Big);
\alpha\in {sp}(m,{\mathbb C})\Big\}\\
&\qquad\qquad{\text{where }} \,\,{\alpha'}=
\big(I{+}{\xi}(I{+}\kappa)\big){\alpha}
\big(I{-}(I{-}\kappa){\xi}\big).
\end{aligned}$$ on the space ${\mathbb C}e^{{\mathfrak S}(2m)}$. Note the following:
\[lemma0\] $\det(I{+}{\xi}(I{+}\kappa))
{=}\det(I{-}(I{-}\kappa)\xi)
{=}\det(I{-}{\xi}(I{-}\kappa))$.
[**Proof**]{} For the first equality, use $\det J{=}1$ and $$\det(J(I{+}{\xi}(I{+}\kappa))){=}
\det((I{-}{}^t\!{\xi}(I{-}{}^t\!\kappa))J)=
\det(I{-}(I{-}\kappa){\xi}).$$ For the second, we use the standard trick $$\det(I{-}(I{-}\kappa){\xi})=
\det({\xi}^{-1}{\xi}{-}(I{-}\kappa){\xi})
=\det({\xi}{\xi}^{-1}{-}{\xi}(I{-}\kappa))$$ via an appropriate approximation of $\xi$ by nonsingular element.$\Box$
First of all, we consider open subsets where the rank of distribution is constant:
$\alpha\to{\alpha'}$ is a bijection of ${\mathfrak{sp}}(m;\mathbb C)$ onto itself, if and only if $\det(I{+}{\xi}(I{+}\kappa))\not=0.$ In this case, Lemma\[lemma0\] gives $$\alpha=(I{+}{\xi}(I{+}\kappa))^{-1}\alpha'(I{-}(I{-}\kappa)\xi)^{-1}.$$ That is in $K$-ordered expression, the infinitesimal action $\{\langle{\pmb u}(\alpha J),{\pmb u}\rangle{*};\,\,
\alpha{\in}{sp}(m;\mathbb C)\}$ degenerates only at the point $\xi$ such that $\det(I{+}{\xi}(I{+}\kappa))=0$.
Let ${\mathcal O}_{\kappa}
=\{\xi\in {\mathfrak{sp}}(m;\mathbb C);
\det(I{+}\xi(I{+}\kappa))\not=0\}$ for every ${\kappa}{\in}{\mathfrak{sp}}(m;{\mathbb C})$. Since this distribution is given by the infinitesimal action of a Lie group, we have
The distribution $D_{(\kappa,ge^{\xi})}$ is constant corank one and involutive on ${\mathcal O}_{\kappa}$.
The goal of this section is as follows:
\[goal\] Maximal integral submanifold through $g\in {\mathbb C}_{\times}$ over ${\mathcal O}_{\kappa}$ is given by $$\label{eq:kyokumen1}
\{g\sqrt{\det(I{+}(I{{+}}\kappa){\alpha})}
e^{\langle{\pmb u}(\frac{1}{i\hbar}{\alpha}J), {\pmb u}\rangle};
\alpha\in{\mathcal O}_{\kappa}\}$$ This is a nontrivial double cover of ${\mathcal O}_{\kappa}$.
The proof is given in several steps. Note first that the maximal integral submanifold through $1$ must be closed under $*_{\kappa}$-product.
[**Step 1**]{} First note that the phase part of the distribution takes arbitrary element. Thus consider elements $g(t)e^{\frac{1}{i{\hbar}}\langle{\pmb u}(t\tilde\alpha J),{\pmb u}\rangle}$ by fixing $\tilde\alpha$. We want to make the tangent vectors of this curve are always in the distribution. Taking derivative, we have $$\Big(\frac{d}{dt}g(t){+}
\frac{1}{i{\hbar}}\langle{\pmb u}(\tilde\alpha J),{\pmb u}\rangle g(t)\Big)
e^{\frac{1}{i{\hbar}}\langle{\pmb u}(t\tilde\alpha J),{\pmb u}\rangle}.$$ Comparing this with at $\xi=t\tilde\alpha$, we take $\alpha(t)$ so that $$\tilde\alpha=
(I{+}t\tilde\alpha(I{+}\kappa))\alpha(t)(I{-}t\tilde\alpha(I{-}\kappa)).$$ Then, the infinitesimal action by $\langle{\pmb u}(\alpha(t)J),{\pmb u}\rangle$ satisfies $$\begin{aligned}
&{:}\frac{1}{i{\hbar}}\langle{\pmb u}(\alpha(t)J),{\pmb u}\rangle{:}_{_K}{*_{_K}}
(g(t)e^{\frac{1}{i{\hbar}}\langle{\pmb u}(t\tilde\alpha J,{\pmb u}\rangle})\\
&=
\left\{
\frac{1}{2}{\rm{Tr}}
\big((\kappa{+}I){\alpha(t)}(\kappa{-}I)t\tilde\alpha{+}\alpha(t)\kappa\big)
{+}\frac{1}{i{\hbar}}\langle{\pmb u}(\tilde\alpha J),{\pmb u}\rangle
\right\}g(t)e^{\frac{1}{i{\hbar}}\langle{\pmb u}(t\tilde\alpha J),{\pmb u}\rangle}\\
&=\Big(\frac{d}{dt}g(t){+}
\frac{1}{i{\hbar}}\langle{\pmb u}(\tilde\alpha J),{\pmb u}\rangle g(t)\Big)
e^{\frac{1}{i{\hbar}}\langle{\pmb u}(t\tilde\alpha J),{\pmb u}\rangle}
\end{aligned}.$$ Plugging in $\alpha(t)=(I{+}{t\tilde\alpha}(I{+}\kappa))^{-1}
\tilde\alpha(I{-}(I{-}\kappa)t\tilde\alpha)^{-1}$ into the above, $g(t)$ is obtained by solving $$\frac{d}{dt}g(t)=
\frac{1}{2}{\rm{Tr}}
\big((\kappa{+}I){\alpha(t)}(\kappa{-}I)t\tilde\alpha{+}\alpha(t)\kappa\big)g(t),\quad
g(0)=g.$$ [**Step 2**]{} To solve this equation, we first solve it in the case $\kappa{=}0$. The equation becomes $$\frac{d}{dt}\log g(t)=
\frac{1}{2}
{\rm{Tr}}\frac{t\tilde\alpha^2}{1-(t\tilde\alpha)^2}=
\frac{1}{4}\frac{d}{dt}{\rm{Tr}}\log(1{-}(t\tilde\alpha)^2).$$ It follows that $$g(t)=e^{{\rm{Tr}}\log(1{-}(t\tilde\alpha)^2)^{\frac{1}{4}}}
=\sqrt[4]{\det(1{-}(t\tilde\alpha)^2)}.$$
On the other hand, since $\det(1{-}t\tilde\alpha){=}\det(1{+}t\tilde\alpha)$, we have $g(t)=\sqrt{\det(1{+}t\tilde\alpha)},$ that is,
\[integsug\] $\sqrt{\det(1{+}t\tilde\alpha)}\,\,
e^{\frac{1}{i{\hbar}}\langle{\pmb u}(t\tilde\alpha J),{\pmb u}\rangle}$ is in an integral submanifold.
[**Step 3**]{} The integral manifold for the general $\kappa$ is obtained by the intertwiner $I_0^{\kappa}$. We have $$I_0^{\kappa}\Big(\sqrt{\det(1{+}\tilde\alpha)}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}(t\tilde\alpha J),{\pmb u}\rangle}\Big)
=\frac{\sqrt{\det(I{+}\tilde\alpha)}}{\sqrt{\det(I{-}\tilde\alpha\kappa)}}
e^{\frac{1}{i{\hbar}}
\langle{\pmb u}(\frac{1}{I{-}\tilde\alpha}\tilde\alpha J),{\pmb u}\rangle}.$$ Replacing $\frac{1}{I{-}\tilde\alpha}\tilde\alpha=\alpha$ gives $$\tilde\alpha=\alpha\frac{1}{1+\alpha\kappa}{=}\frac{1}{I{+}\alpha\kappa}\alpha.$$ Plugging this and using the algebraic calculation such that $\frac{\sqrt{x}}{\sqrt{x}}=1$, we have the following:
\[integmanif\] In $\kappa$-ordered expression, the maximal integral submanifold is given by $$c{\widetilde{\mathcal O}}_{\kappa}=
\{
c\sqrt{\det(I{+}\alpha(I{+}\kappa))}\,\,
e^{\frac{1}{i{\hbar}}\langle{\pmb u}(\alpha J),{\pmb u}\rangle}; \alpha\in
{\mathcal O}_{\kappa}\}$$ where ${\mathcal O}_{\kappa}=
\{\alpha\in {sp}(m,{\mathbb C}); \det(I{+}\alpha(I{+}\kappa))\not=0\}$.
Note that Proposition\[integmanif\] shows that we have only to know the phase part to know the integral submanifold. By definition ${\widetilde{\mathcal O}}_{\kappa}$ is the maximal integral submanifold through $(1,0)\in {\mathbb C}{\times}{sp}(m;\mathbb C)$. Setting $c=1$ in Proposition\[integmanif\], we see that $$\label{integralmanifmax}
\pi_{\kappa}: {\widetilde{\mathcal O}}_{\kappa}\to{\mathcal O}_{\kappa}$$ is a nontrivial double cover, which is just the forgetful mapping of the amplitude part. The significance of the set ${\mathcal O}_{\kappa}$ will be explained in the next section by the Cayley transform.
### Integral submanifolds and twisted Cayley transforms {#Cayley}
The Cayley transform $C_0(X)=\frac{I-X}{I+X}$ has following properties: For $X\in {\mathfrak{sp}}(m;\mathbb C)$ with $\det(I+X)\not=0$, we see $C_0(X)\in Sp(m;\mathbb C)$ and $\det(I+C_0(X))=(\det(I+X))^{-1}$. $$\label{eq:Cayley}
X\in {\mathfrak{sp}}(m;\mathbb C)\Leftrightarrow C_0(X)\in Sp(m;\mathbb C),
\quad C_0^2(X)=X.$$
Let ${\mathcal O}_0=\{X\in {\mathfrak{sp}}(m;\mathbb C); \det(I{+}X)\not=0\}$. $C_0:{\mathcal O}_0\to Sp(m;\mathbb C)$ is viewed as a local coordinate system $Sp(m;\mathbb C)$, which covers an open dense subset of $Sp(m;\mathbb C)$.
Let ${\mathcal O}_\kappa=
\{\alpha\in {\mathfrak{sp}}(m;\mathbb C); \det(I{+}(I{+}\kappa)\alpha)\not=0\}$, and define $$\label{eq:cayley22}
\begin{aligned}[c]
C_{\kappa}(\alpha)=
(I-(I{-}{\kappa})\alpha)&\frac{1}{I+(I{+}{\kappa})\alpha}
=\frac{1}{I{+}\alpha(I{+}{\kappa})}(I{-}\alpha(I{-}{\kappa})), \\
(C_{\kappa})^{-1}(Y)=
&\frac{1}{I{-}{\kappa}{+}Y(I{+}{\kappa})}(I{-}Y)=
(I{-}Y)\frac{1}{I{-}{\kappa}{+}(I{+}{\kappa})Y}.
\end{aligned}$$ $C_{\kappa}:{\mathcal O}_{\kappa}\to Sp(m;\mathbb C)$ gives also a local coordinate system of $Sp(m;\mathbb C)$. We call the [*twisted Cayley transform*]{}. Since $C_{\kappa}: {\mathcal O}_{\kappa}\to
C_{\kappa}({\mathcal O}_{\kappa})$ is a diffeomorphism, we often identify ${\mathcal O}_{\kappa}$ with $C_{\kappa}({\mathcal O}_{\kappa})$ through the twisted Cayley transform $C_{\kappa}$. The following Lemma is crucial to our purpose:
\[crucial00\] $\bigcup\{C_{\kappa}({\mathcal O}_{\kappa});
\kappa\in {\mathfrak{sp}}(m,{\mathbb C})\}
=Sp(m,{\mathbb C})$. On the other hand, if $\alpha\not\in {\mathcal O}_0$, then $\frac{1}{I{-}\alpha\kappa}\alpha
\not\in{\mathcal O}_{\kappa}$ for every $\kappa\in {\mathcal D}_{\alpha}{=}
\{\kappa{\in}{\mathfrak{sp}}(m;{\mathbb C});
\det(I-\alpha\kappa)\not=0\}$.
[**Proof**]{} Suppose there is a $Y{\in}Sp(m,{\mathbb C})$ such that $\det(I{-}{\kappa}{+}Y(I{+}{\kappa})){=}0$ for every $\kappa\in {\mathfrak{sp}}(m,{\mathbb C})$. Then such a $Y$ must satisfy $\det(\frac{1{-}\kappa}{I{+}\kappa}{+}Y){=}0$. Since $\frac{1{-}\kappa}{I{+}\kappa}$ moves in an open dense domain of $Sp(m,{\mathbb C})$, it follows $\det(X{+}Y){=}0$ for every $X\in Sp(m,{\mathbb C})$. Set $X{=}Y$ to get a contradiction. Thus we see that for every $Y$ there is $\kappa$ such that $\det(I{-}{\kappa}{+}Y(I{+}{\kappa})){\not=}0.$ Hence $C_{\kappa}^{-1}(Y)$ exists. The rest of Lemma follows easily. $\Box$
Define $T_{\kappa'{-}\kappa}(\alpha)=\frac{1}{I-\alpha(\kappa'{-}\kappa)}\alpha$. Then $T^{-1}_{\kappa'{-}\kappa}(\alpha)=\frac{1}{I{+}\alpha(\kappa'{-}\kappa)}\alpha$ $$T_{\kappa'{-}\kappa}(\alpha)\in {\mathfrak{sp}}(m;\mathbb C)
\Longleftrightarrow \alpha \in {\mathfrak{sp}}(m;\mathbb C).$$ It is easy to see $$\frac{1}{I{-}\alpha(\kappa'{-}\kappa)}(I{+}\alpha(I{+}{\kappa}))=
I{+}T_{\kappa'{-}\kappa}(\alpha)(I{+}{\kappa'}).$$ Hence, we have the following:
On $Sp(m,{\mathbb C})$, the coordinate transformations are given by the phase part of the intertwiners.
Hence, by setting ${\mathcal O}_{\kappa\kappa'}={\mathcal O}_{\kappa}\cap{\mathcal O}_{\kappa'}$, intertwiners give 2-to-2 mappings $$\label{eq:coodtrns}
I_{\kappa}^{\kappa'}: \pi_{\kappa}^{-1}({\mathcal O}_{\kappa\kappa'})\to
\pi_{\kappa'}^{-1}({\mathcal O}_{\kappa'\kappa})$$ just as in Proposition 3.1in §3.1 of [@OMMY4]. Since $0{\in}\bigcap_{\kappa}{\mathcal O}_{\kappa}$, and $C_{\kappa}(0)=I$, we can consider $\bigcup_{\kappa}{\widetilde{\mathcal O}}_{\kappa}$ as an object patched by the intertwiners $I_{\kappa}^{\kappa'}$ as a bundle-like object over $\bigcup_{\kappa}C_{\kappa}({\mathcal O}_{\kappa})=Sp(m;{\mathbb C})$.
Computing the derivative of the twisted Cayley transform, we have $$\begin{aligned}[t]
(dC_{\kappa})_{\xi}({\alpha}){=}
-(I{-}\kappa){\alpha}\frac{1}{I{+}(I{+}\kappa){\xi}}
{=}
-(I{-}(I{-}\kappa){\xi})\frac{1}{I{+}(I{+}\kappa){\xi}}
(I{+}\kappa){\alpha}\frac{1}{I{+}(I+\kappa){\xi}}.
\end{aligned}$$ Using the bumping identity $\frac{1}{I{+}(I{+}\kappa){\xi}}(I{+}\kappa)=
(I{+}\kappa)\frac{1}{I{+}{\xi}(I{+}\kappa)}$, we easily see $$\label{eq:keykey}
(dC_{\kappa})_{{\xi}}((I{+}{\xi}(I{+}\kappa))\alpha
(1{-}(I{-}\kappa){\xi}))=-2\alpha C_{\kappa}({\xi}).$$ Thus, the phase part of the distribution $D_{\kappa}$ is translated by $C_{\kappa}$ into the right invariant tangential distribution on $Sp(m;\mathbb C)$. Recall that $(I{+}{\xi}(I{+}\kappa))\alpha(1{-}(I{-}\kappa){\xi})$ appeared already in , $(\diamondsuit)$ as the infinitesimal phase part.
\[infactionquad\] The infinitesimal phase part of the infinitesimal action ${:}\frac{1}{i{\hbar}}\langle{\pmb u}\alpha J,{\pmb u}\rangle{:}_{\kappa}{*_\kappa}$ is translated by the twisted Cayley transform $C_{\kappa}$ into the right invariant distribution by $\alpha\in {\mathfrak{sp}}(m;\mathbb C)$ on $Sp(m;\mathbb C)$.
As the distribution in the previous section is defined by the infinitesimal action of $*$-exponential functions $e_*^{t\frac{1}{i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle}$, the maximal integral submanifold must be closed by the left multiplication $e_*^{t\frac{1}{i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle}{*}$.
Therefore, the joint object $\{\widetilde{\mathcal O}_{\kappa};\kappa\in {sp}(m,{\mathbb C})\}$ must have certain “Lie group-like” properties with manifold-like properties patched by $2$-to-$2$ coordinate transformations. A general product formula will be given in the next section.
On the other hand, recall the second statement of Lemma\[crucial00\] gives
\[vacpolar\] If $\det(I{+}\alpha)=0$, then the infinitesimal $*$-action of the quadratic forms to the parallel section $$\Big\{\frac{1}{\sqrt{\det(I{-}\alpha\kappa)}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}\frac{1}{I{-}\alpha\kappa}\alpha,{\pmb u}\rangle};
\kappa\in {\mathfrak{sp}}(m;{\mathbb C}) \Big\}$$ degenerates at every $\kappa$. Namely, $(\frac{1}{\sqrt{\det(I{-}\alpha\kappa)}};
\frac{1}{I{-}\alpha\kappa}\alpha)\not\in
{\widetilde{\mathcal O}}_\kappa$ for every $\kappa\in {\mathcal D}_{\alpha}$, where ${\mathcal D}_{\alpha}=$\
$\{\kappa{\in}{\mathfrak{sp}}(m;{\mathbb C});
\det(I-\alpha\kappa)\not=0\}$.
The most degenerate (minimal rank) orbit will be called the orbit of [**vacuums**]{}. An example is given by $\alpha=
\begin{bmatrix}
1&0\\
0&{-}1
\end{bmatrix}$ in the case $m=1$. These elements will be used to make matrix representations in the later chapter.
General product formula {#productformula}
-----------------------
It is rather hard to construct general manifold theory patched together by $2$-to-$2$ coordinate transformations, for such objects do not have underlying topological spaces. In spite of this, there is no difficulty forming a local/classical differential geometry. Hence a certain general theory is easy to construct for a Lie group-like object by using infinitesimal algebraic notions other than point set pictures. It is natural to think this gives an intuitive concrete object of “gerbes”.
Proposition\[infactionquad\] shows that if one concerns only the phase parts of the $*$-product, then one can compute these via the group structure of $Sp(m;{\mathbb C})$ through twisted Cayley transform.
By this observation, we first investigate the product $*_0$ defined on ${\mathbb C}_{\times}\times{\mathcal O}_{0}$ as follows:
$(g; a){*}_{0}(g';b)
=\left(gg'\Big(\frac{1}{\sqrt{\det(1{+}ab)}}\Big);
C_0^{-1}(C_0(a)C_0(b))\right)$\
$C_0^{-1}(C_0(a)C_0(b))=
\frac{1}{1{+}a}(a{+}b)\frac{1}{1{+}ab}(1{+}a)$
Note first the following general identity:
$\frac{1}{1{+}a}(a{+}b)\frac{1}{1{+}ab}(1{+}a)\sim
(1{+}b)\frac{1}{1{+}ab}(a{+}b)\frac{1}{1{+}b}$, where the reason of the notation $\sim$ instead of $=$ is that algebraic calculation such as $(1{+}a)\frac{1}{1{+}a}=1$ is used in the proof. Hence, one may replace $\frac{1}{1+a}(a+b)\frac{1}{1+ab}(1+a)$ by $(1{+}b)\frac{1}{1{+}ab}(a{+}b)\frac{1}{1{+}b}$.
[**Proof**]{} follows immediately by the identity $(a{+}b)\big(1{+}\frac{1}{1{+}ab}(a{+}b)\big)=
\big(1{+}(a{+}b)\frac{1}{1{+}ab}\big)(a{+}b)$. ${}$ $\Box$
As far as concerning the phase part $C_0(a)$, and forgetting about the singularity, this gives a group which is isomorphic to $Sp(m;{\mathbb C})$.
To consider the amplitude part, we define $$(g;\alpha){*}_{\kappa}(g';\beta)\sim
I_0^{\kappa}
\big(I^0_{\kappa}(g;\alpha){*}_0I^0_{\kappa}(g';\beta)\big).$$ Since $I^0_{\kappa}(g;\alpha)
=\big( g(\det(I{+}\alpha\kappa))^{-\frac{1}{2}};T_{-\kappa}(\alpha)\big)$, the definition of $C_{\kappa}$ gives that $$\label{prodkappa}
(g;\alpha)*_{{\kappa}}(g';\beta)
=\Big(gg'(\frac{\det(P{+}Q(I{+}\kappa))}
{\det(P(I{+}\alpha(I{+}\kappa))(I{+}\beta(I{+}\kappa)))})^{\frac{1}{2}}
;C_{\kappa}^{-1}(C_{\kappa}(\alpha)C_{\kappa}(\beta))\Big)
$$ where $P=I{+}\alpha(I{-}\kappa)\beta(I{+}\kappa)$, $Q=\alpha{+}\beta{+}2\alpha\kappa\beta$, and $$C_{\kappa}^{-1}(C_{\kappa}(\alpha)C_{\kappa}(\beta))
=(I{+}\beta(I{+}{\kappa}))\,\frac{1}{P}\,Q \,\,\frac{1}{I{+}(I{+}{\kappa})\beta}.$$ We easily see that $\det(P{+}Q(I{+}\kappa))
=\det(I{+}\alpha(I{+}\kappa))(I{+}\beta(I{+}\kappa))$. Hence, the first component of the r.h.s of is $gg'(\frac{1}{\det P})^{\frac{1}{2}}$. Hence, we obtain $$\label{eq:prodfmla}
(g;\alpha)*_{{\kappa}}(g';\beta){=}
\Big(gg'(\frac{1}{\det P})^{\frac{1}{2}};
(I{+}\beta(I{+}{\kappa}))\,\frac{1}{P}\,Q \,\,
\frac{1}{I{+}(I{+}{\kappa})\beta}\Big).$$ The product formula works only for $\alpha, \beta$ such that $\det P\not=0$, and $\det(I{+}(I{+}{\kappa})\beta)\not=0$. But, one can choose the expression parameter $\kappa$ so that these conditions are satisfied.
This follows from that we treat elements written in the form $e^{\frac{1}{i{\hbar}}Q({\pmb u})}$. Therefore, for elements written in the form $e^{Q({\pmb u})}$ the product must be written in the form $e^{(i{\hbar})^2R({\pmb u})}$, and hence the product formula is real analytic in ${\hbar}\geq 0$. Hence one can apply the formal associativity Theorem\[assocthm\].
\[blurred\] Associativity holds with $\pm$ ambiguity.
We call this object a [*blurred Lie group*]{}, and denote it by $(Sp^{(\frac{1}{2})}_{\mathbb C}(m);{*})$. This is not an object in which $\pm a$ is treated simply as a single point, since they can be locally distinguished.
For instance, we first note the following:
$(1;0)$ is the identity with respect to ${*}_{\kappa}$-product for $\forall\kappa$
In particular, in the Weyl ordered expression, the integral manifold through $(1;0)$ is $${\widetilde{\mathcal O}}_0=
\{\sqrt{(\det(1{+}a)};a); a\in {\mathcal O}_0\}.$$
Although the sign ambiguity remains, we obtain the following:
\[doublevalinv00\]
(a)
: If $A, B\in {\widetilde{\mathcal O}}_0$, and if $A{*}_0B$ is defined, then $A{*}_0B\in {\widetilde{\mathcal O}}_0$.
(b)
: $(1;0)$ is the identity.
(c)
: The inverse $(\sqrt{\det(1{+}a)};a)^{-1}$ is given by $(\sqrt{\det(1{-}a)}; -a)$.
$$(\sqrt{\det(1{+}a)}; a)*_{0}(\sqrt{\det(1{-}a)}; -a)
=(\sqrt{1}; 0).$$
In general, $\sqrt{1}$ must be treated as $\pm 1$, but concerning the inverse, this should be $1$ by [**continuous tracing**]{} from the identity $(1;0)$ to the point $(\sqrt{\det(1{+}a)}; a)$.
Since ${\widetilde{\mathcal O}}_{\kappa}$ is a local Lie group with the identity and ${\mathfrak{sp}}(m;{\mathbb C})$ as its tangent space, we have the following:
\[Liealgebra\] ${\mathfrak{sp}}(m;{\mathbb C})$ is the Lie algebra of $Sp^{(\frac{1}{2})}_{\mathbb C}(m)$.
By $A=\alpha J$, ${\mathfrak{sp}}(m;{\mathbb C})$ is naturally identified with the space of expression parameters ${\mathfrak S}(2m)$.
Since the element $(1; 0)$ may be viewed as the identity of the blurred Lie group $Sp^{(\frac{1}{2})}_{\mathbb C}(m)$, the tangent space of $Sp^{(\frac{1}{2})}_{\mathbb C}(m)$ at $(1;0)$ is naturally identified with ${\mathfrak{sp}}(m,{\mathbb C})$.
In the next section, we define one parameter subgroups of $(Sp^{(\frac{1}{2})}_{\mathbb C}(m);{*})$, and the $*$-exponential mapping $$\exp_*: {\mathfrak{sp}}(m;{\mathbb C})\to Sp^{(\frac{1}{2})}_{\mathbb C}(m),$$ and we show that every one parameter subgroup has discrete branched singular points in generic ordered expression.
Abstract definition of blurred Lie groups
-----------------------------------------
Here we give a [*tentative*]{} abstract definition of blurred Lie (covering) groups. As we do not have enough concrete examples, it seems to be too early to give the notion of isomorphisms or the general theory.
Let $G$ be a locally simply arcwise connected topological group and let $\{{\mathcal O}_{\alpha}; \alpha\in I\}$ be an open covering of $G$.
\(a) For every $\alpha\in I$, ${\mathcal O}_{\alpha}$ contains the identity $e$. ${\mathcal O}_{\alpha}$ is called an
**abstract expression**
space
, and $\alpha$ is called an expression parameter.
\(b) For every $\alpha\in I$, ${\mathcal O}_{\alpha}$ is open, dense and connected, but it may not be simply connected.
\(c) For every $\alpha, \beta\in I$, there is a homeomorphism $\phi_{\alpha}^{\beta}: {\mathcal O}_{\alpha}\to {\mathcal O}_{\beta}$.
\(d) For every $g, h\in G$, there is $\alpha\in I$ and continuous path $g(t), h(t)\in G$, $t\in [0,1]$, such that
$g(0)=h(0)=e$, $g(1)=g, h(1)=h$ and $g(t), h(t), g(t)h(t)$ are in ${\mathcal O}_{\alpha}$ for every $t\in [0,1]$.
An open covering $\{{\mathcal O}_{\alpha}; \alpha\in I\}$ is called [**natural covering**]{} of $G$ if it satisfies $(a){\sim}(d)$. The condition $(c)$ shows that there is an abstract topological space $X$ homeomorphic to every ${\mathcal O}_{\alpha}$. We consider a connected covering space $\pi:\tilde X\to X$. This is same to say we consider a connected covering $\pi_{\alpha}: \widetilde{\mathcal O}_{\alpha}\to
{\mathcal O}_{\alpha}$ for each $\alpha$. It is easy see that $\pi_{\alpha}^{-1}(e)$ is a group given as a quotient group of the fundamental group of ${\mathcal O}_{\alpha}$. As $G$ is locally simply connected, $\pi_{\alpha}^{-1}(e)$ forms a discrete group, and $\phi_{\alpha}^{\beta}$ lifts to an isomorphism $\tilde\phi_{\alpha}^{\beta}: \pi^{-1}_{\alpha}(e)\to\pi^{-1}_{\beta}(e)$. We denote $\pi^{-1}_{\alpha}(e)=\varGamma_{\alpha}$, and the isomorphism class is denoted by ${\varGamma}$.
Choose $\tilde{e}_{\alpha}\in\pi_{\alpha}^{-1}(e)$ and call $\tilde{e}_{\alpha}$ a tentative identity. For any continuous path $g(t)$ in ${\mathcal O}_{\alpha}$ such that $g(0)=g(1)=e$, the continuous tracing among the set $\pi^{-1}(g(t))$ starting at $\tilde{e}_{\alpha}$ gives a group element $\gamma \in \varGamma_{\alpha}$.
By a standard argument, it is easy to make $\widetilde{\mathcal O}_{\alpha}$ a local group such that $\pi_{\alpha}$ is a homomorphism: We define first that $\tilde e_{\alpha}\tilde e_{\alpha}=\tilde e_{\alpha}$. For paths $g(t), h(t), g(t)h(t)$ such that they are in ${\mathcal O}_{\alpha}$ for every $t\in [0,1]$ and $g(0)=h(0)=e$, we define the product by a continuous tracing among the set-to-set mapping $$\pi^{-1}_{\alpha}(g(t))\pi^{-1}_{\alpha}(h(t))=\pi^{-1}_{\alpha}(g(t)h(t)).$$ We set ${\mathcal O}_{\alpha\beta}=
{\mathcal O}_{\alpha}\cap {\mathcal O}_{\beta},\quad
{\mathcal O}_{\alpha\beta\gamma}=
{\mathcal O}_{\alpha}\cap {\mathcal O}_{\beta}\cap
{\mathcal O}_{\gamma}$ for simplicity.
As $G$ is locally simply connected, the full inverse $\pi_{\alpha}^{-1}V$ of a simply connected neighborhood $V\subset{\mathcal O}_{\alpha}$ of the identity $e\in G$ is the disjoint union $\coprod_{\lambda}{\tilde V}_{\lambda}$, each member ${\tilde X}_{\lambda}$ of which is homeomorphic to $V$. Moreover $\pi_{\alpha}^{-1}{\mathcal O}_{\alpha\beta}$ is also a local group for every $\beta$.
[**Isomorphisms modulo $\varGamma$, Controlled discontinuity**]{}
For every $\alpha,\beta$, we define the notion of “isomorphism” $I_{\alpha}^{\beta}$ of local groups, which corresponds to the notion of intertwiners in the previous section: $$\begin{matrix}
\widetilde{\mathcal O}_{\alpha}&\supset&
\pi_{\alpha}^{-1}{\mathcal O}_{\alpha\beta}&
\overset{I_{\alpha}^{\beta}}\longrightarrow&
\pi_{\beta}^{-1}{\mathcal O}_{\beta\alpha}&\subset&
\widetilde{\mathcal O}_{\alpha}\\
\downarrow\pi_{\alpha}&{ }&{ }&{ }&{ }&{ }&
\downarrow\pi_{\beta}\\
{\mathcal O}_{\alpha}&\supset&
{\mathcal O}_{\alpha\beta}&
=\!=\!=&
{\mathcal O}_{\beta\alpha}&\subset&{\mathcal O}_{\beta}
\end{matrix}$$ such that $I_{\beta}^{\alpha}=(I_{\alpha}^{\beta})^{-1}$, but the cocycle condition $I_{\alpha}^{\beta}I_{\beta}^{\gamma}I_{\gamma}^{\alpha}=1$ is not required for ${\mathcal O}_{\alpha\beta\gamma}$.
Since the correspondence $I_{\alpha}^{\beta}$ does not make sense as a point set mapping, we should be careful for the definition.
Note that $I_{\alpha}^{\beta}$ is a collection of $1$-to-$1$ mapping $I_{\alpha}^{\beta}(g): \pi_{\alpha}^{-1}(g)\to
\pi_{\beta}^{-1}(g)$ for every $g\in{\mathcal O}_{\alpha\beta}=
{\mathcal O}_{\beta\alpha}$, which may not be continuous in $g$.
For each $g$ there is a neighborhood $V_g$ of the identity $e$ such that $V_gg\subset {\mathcal O}_{\alpha\beta}$ and the local trivialization $\pi_{\alpha}^{-1}(V_gg)=V_gg{\times}\pi_{\alpha}^{-1}(g)$. Thus $I_{\alpha}^{\beta}(g)$ extends to the correspondence $${\tilde I}_{\alpha}^{\beta}(h,g):
\pi_{\alpha}^{-1}(hg)\to
\pi_{\beta}^{-1}(hg),\quad h\in V_g$$ which commutes with the local deck transformations.
The collection $I_{\alpha}^{\beta}{=}\{I_{\alpha}^{\beta}(g); g\in{\mathcal O}_{\alpha\beta}\}$ is called an isomorphism modulo $\varGamma$, if the product $I_{\beta}^{\alpha}(hg){\tilde I}_{\alpha}^{\beta}(h,g)$ is in the group $\varGamma$ for every $g\in {\mathcal O}_{\alpha\beta}$ and $h\in V_g$. $($It follows the continuity of $I_{\alpha}^{\beta}(hg)$ w.r.t. $h$.$)$
The condition given by this definition means roughly that $I_{\alpha}^{\beta}(g)$ has discontinuity in $g$ only in the group $\varGamma$.
${\widetilde G}=\{\widetilde{\mathcal O}_{\alpha},
\pi_{\alpha}, I_{\alpha}^{\beta}; \alpha,\beta\in I\}$ is called a [**blurred covering group**]{} of $G$ if each $\widetilde{\mathcal O}_{\alpha}$ is a covering local group of ${\mathcal O}_{\alpha}$, where $\{{\mathcal O}_{\alpha};\alpha\in I\}$ is a natural open covering of a locally simply arcwise connected topological group $G$ and $I_{\alpha}^{\beta}$ are isomorphisms modulo $\varGamma$.
Because of the failure of the cocycle condition, this object does neither form a covering group, nor a topological point set. However, this object looks like a covering group.
For $g$, let $I_g$ be the set of expression parameters involving $g$; $I_g=\{\alpha\in I; {\mathcal O}_{\alpha}\ni g\}$. For every $\alpha\in I(g,h,gh)=I_g\cap I_h\cap I_{gh}$, we easily see that $\pi_{\alpha}^{-1}(g)\pi_{\alpha}^{-1}(h)=
\pi_{\alpha}^{-1}(gh)$. In general, this is viewed as set-to-set correspondence, but if $g$ or $h$ is in a small neighborhood of the identity, we can make these correspondence a genuine point set mapping. Hence, we have the notion of indefinite small action or “infinitesimal left/right action” of small elements to the object. This corresponds to the infinitesimal action $w_*^2{*}$ or ${*}w_*^2$ in the previous section.
Next, we choose an element $\tilde{e}_{\alpha}\in\pi_{\alpha}^{-1}(e)$, and call it a local identity. On the other hand, $\pi_{\alpha}^{-1}(e)$ is called [*the set of local identities*]{} of ${\widetilde{G}}$. The failure of the cocycle condition gives that ${\mathfrak M}_{\alpha}\tilde{e}_{\alpha}$ may not be a single point set, but forms a discrete abelian group. Hence an identity of our object is always a [*local identity*]{}.
Since $G$ is a locally simply connected, there is an open simply connected neighborhood $V_{\beta}$ of $e$ contained in ${\mathcal O}_{\beta}$. Hence, there is the unique lift $\tilde{V}_{\beta}$ through $\tilde{e}_{\beta}$. Setting $\tilde{V}_{\beta\gamma}=
\tilde{V}_{\beta}\cap\tilde{V}_{\gamma}$ e.t.c., we see easily $I_{\beta}^{\gamma}(\tilde{V}_{\beta\gamma})
=\tilde{V}_{\gamma\beta}.$
The $\{{\tilde g}_{\alpha}\in \tilde{\mathcal O}_{\alpha};\alpha\in I\}$ may be viewed as an element of ${\widetilde{G}}$ if $I_{\alpha}^{\beta}{\tilde g}_{\alpha}={\tilde g}_{\beta}$, but this is not a single point set by the same reason. In spite of this, one can distinguish individual points within a small local area.
Star-exponential functions of quadratic forms {#star-exponential2}
=============================================
For an element $H_*$ of the algebra, we define the $*$-exponential function $e_*^{tH_*}$ as the real analytic solution of $$\label{leftevol}
\frac{d}{dt}f_*(t)=H_*{*}f_*(t), \quad f(0){=}1,$$ provided the solution exists. More precisely, we define $e_*^{tH_*}$ as the family $\{f_t(K)\}$ of univalent solutions of the evolution equation $$\label{starexp}
\frac{d}{dt}f_t(K)={:}H_*{:}_{_K}{*_{_K}}f_t(K),$$ with the initial condition $f_0(K)=1$. We think of $f_t(K)$ as the $K$-ordered expression of $e_*^{tH_*}$, and denote it by ${:}e_*^{tH_*}{:}_{_K}=f_t(K)$. Uniqueness is ensured if we consider only real analytic solutions. is called the [*left evolution equation*]{}. The [*right evolution equation*]{} is defined similarly, but this is not used except when otherwise mentioned.
If $H_*$ is a $*$-polynomial, can be rewritten as a partial differential evolution equation. If the equation $\frac{d}{dt}{:}f_*(t){:}_{_K}={:}H_*{:}{*}f_*(t){:}_{_K}$ has a unique solution for the initial element $f_*(0){=}g_*$, then the solution will be denoted by ${:}e_*^{tH_*}{*}g_*{:}_{_K}$.
As it was seen in §\[Expinter\], a star exponential function $e_*^{\frac{1}{i{\hbar}}\langle{\pmb\xi}, {\pmb u}\rangle}$ of a linear form $\langle{\pmb\xi}, {\pmb u}\rangle$, was welldefined as the family $\{e^{\frac{1}{4i{\hbar}}
\langle{\pmb\xi}K,{\pmb\xi}\rangle}
e^{\frac{1}{i{\hbar}}\langle{\pmb\xi}, {\pmb u}\rangle}\}$ for all $K{\in}{\mathfrak S}(n)$. Provided ${:}e_*^{sH_*}{:}_{_{K}}$ exists for every $s{\in}{\mathbb C}$, they form a complex one parameter subgroup, for the exponential law holds by the uniqueness of real analytic solutions.
Here we give several general remarks on $*$-exponential functions of quadratic forms.
\(1) If $H_*$ is a quadratic form, ${:}e_*^{sH_*}{:}_{_{K}}$ is defined with double branched singularities on a discrete set
(c.f. ). Thus, we have to prepare two sheets to consider ${:}e_*^{sH_*}{:}_{_{K}}$ for $s\in {\mathbb C}$. But, the origin $0$ of
another sheet does not correspond to $1$, but $-1$.
(1.1) In general, there is no reflection symmetry in $s$ for the domain of existence of the solution of
. That is, the existence of ${:}e_*^{sH_*}{:}_{_{K}}$ does not necessarily imply that ${:}e_*^{-sH_*}{:}_{_{K}}$ exists: e.g. $${:}e_*^{t\frac{1}{i{\hbar}}(u^2+v^2)}{:}_{_I}=
\frac{1}{\cos t{-}\sin t}
e^{\frac{1}{i{\hbar}}\frac{\sin t}{\cos t{-}\sin t}(u^2+v^2)}\quad cf.
\eqref{Delta}.$$
(1.2) Moreover ${:}e_*^{sH_*}{:}_{_{K}}$ is double-valued holomorphic function in $K$ on an open connected dense
domain, i.e. double-valued holomorphic parallel section.
\(2) If $H_*$, $G_*$ are quadratic forms, then the product ${:}e_*^{tH_*}{*}e_*^{G_*}{:}_{_K}$ is defined as a double-valued
holomorphic function of $(t,K)$ defined on an open connected dense domain containing $(0,0)$.
For a given $K$, suppose that has real analytic solutions in $t$ on some domain $D(K)$ including $0$ for the initial functions $1$ and $g$. We denote the solution of with initial function $g$ by $${:}e_*^{tH_*}{:}_{_K}{*_{_K}}g,\quad t{\in}D(K).$$
\[exppoly\] If $H_*$ is a polynomial and ${:}e_*^{tH_*}{:}_{_K}$ is defined on a domain $D(K)$, then ${:}e_*^{tH_*}{:}_{_K}{*_{_K}}p(\pmb u)$ is defined for every polynomial $p(\pmb u)$ on the same domain $D(K)$.
If $p(\pmb u)=\sum A_\alpha(s){\pmb u}^{\alpha}$ is a polynomial whose coefficients depend smoothly on $s$, then the formula $$\partial_s^{\ell}{:}e_*^{tH_*}{:}_{_K}{*_{_K}}p(\pmb u)=
{:}e_*^{tH_*}{:}_{_K}{*_{_K}}\partial_{s}^{\ell}p(\pmb u)$$ holds for every $\ell$.
[**Proof**]{} Multiplying the defining equation by ${*}p(\pmb u)$ and applying the associativity in Proposition\[extholom\], we have $$\label{eq-11}
\frac{d}{dt}f_t(K){*}p(\pmb u)=
{:}H_*{:}_{_K}{*_{_K}}(f_t(K){*}p(\pmb u)),
\quad f_0(K)= 1.$$ Since $f_t(K){*}p(\pmb u)$ is a real analytic solution, this is written in our notation as $e_*^{tH_*}{*}p(\pmb u)$. Applying $\partial^{\ell}_s$ to gives the second assertion by a similar argument. $\Box$
For a quadratic form $\langle{\pmb u}A, {\pmb u}\rangle_*$, the $*$-exponential function $e_*^{\frac{t}{i{\hbar}}\langle{\pmb u}A, {\pmb u}\rangle_*}$ is given in a concrete form. For every $\alpha\in {\mathfrak{sp}}(m,{\mathbb C})$, we consider first the one parameter subgroup $e^{-2t\alpha}$ of $Sp(m,{\mathbb C})$, and consider the inverse image of twisted Cayley transform $C^{-1}_{\kappa}(e^{-2t\alpha})$: We set $$\label{eq:kyakukei}
C_{\kappa}^{-1}(e^{-2t\alpha})=
\frac{1}{(I{-}{\kappa}){+}
e^{-2t\alpha}(I{+}{\kappa})}(I{-}e^{-2t\alpha})\\
=\frac{1}{\cosh t\alpha{-}(\sinh t\alpha){\kappa}}\sinh t\alpha.$$ where $\frac{1}{X}$ stands for $X^{-1}$.
The exponential function must lie on the integral manifold $\widetilde{\mathcal O}_{\kappa}$ through $(1;0)$, and the point of the integral manifold is determined by its phase part. Hence we have $$\label{eq:starexp}
\exp_{*_{{\kappa}}}\frac{1}{i{\hbar}}t{\alpha}=
\Big(\big(\det(I+C_{\kappa}^{-1}(e^{-2t{\alpha}})
(I{{+}}{\kappa}))\big)^{\frac{1}{2}}
; C_{\kappa}^{-1}(e^{-2t{\alpha}})\Big).$$ In the original notation, we see $e_*^{s\frac{1}{i{\hbar}}\langle{\pmb u}
({\alpha}J),{\pmb u}\rangle_*}$ as follows by setting $\kappa{=}J\!K$: $${:}e_*^{s\frac{1}{i{\hbar}}\langle{\pmb u}
({\alpha}J),{\pmb u}\rangle_*}{:}_{_{\kappa}}=
\big(\det(I{+}C_{\kappa}^{-1}
(e^{-2s{\alpha}})(I{{+}}{\kappa}))\big)^{\frac{1}{2}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}
(C_{\kappa}^{-1}(e^{-2s{\alpha}})J),{\pmb u}\rangle}.$$ More precisely, for every $\alpha{\in}{\mathfrak{sp}}(m,{\mathbb C})$, the $K$-ordered expression of the $*$-exponential function is given as follows: (Cf.[@ommy] [@OMMY6] [@OMMY7] for special cases.) $$\label{fundamenta1}
\begin{aligned}
{:}e_*^{\frac{t}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}
{=}
\frac{2^m}{\sqrt{\det(I{-}{\kappa}{+}e^{-2t\alpha}(I{+}\kappa))}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}
\frac{1}{I{-}{\kappa}{+}e^{-2t\alpha}(I{+}\kappa)}
(I{-}e^{-2t\alpha})J,{\pmb u}\rangle}
\end{aligned}$$ where $\kappa{=}J\!K$. It is not hard to see that this is the real analytic solution of . By this concrete form we see this is an element of ${\mathcal E}_{2+}({\mathbb C}^{2m})$ whenever this is defined. But it is remarkable that ${:}e_*^{\frac{t}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}$ remains in the space ${\mathbb C}e^{Q(u,v)}$ given in Theorem\[meromuseful\].
Adjoint action to $V_{2m}$. {#adjoint}
---------------------------
$Sp_{\mathbb C}^{(\frac{1}{2})}(m)$ is not a genuine Lie group, as elements have double-valued nature in general, and it looks something like a double covering group of $Sp(m,{\mathbb C})$. But, because of this reason, $Sp_{\mathbb C}^{(\frac{1}{2})}(m)$ contains several genuine groups such as the metaplectic group which is not contained in $Sp(m,{\mathbb C})$. Moreover, $Sp_{\mathbb C}^{(\frac{1}{2})}(m)$ contains $Spin(m)$ under the special ordered expression $K_s$ (cf. [@OMMY4].) In the case $m=1$, we have seen in [@OMMY3] some basic properties of Jacobi’s $\theta$-functions by means of $*$-exponential functions of quadratic forms.
To avoid the vague issue of sign ambiguity, we first consider adjoint representations of $Sp_{\mathbb C}^{(\frac{1}{2})}(m)$ on the linear space of generators, for the sign ambiguity disappears in adjoint representations, and it is independent of the expression parameter $K$.
For $\alpha\in {\mathfrak{sp}}(m;\mathbb C)$, the quadratic form $\langle{\pmb u}(\alpha J),{\pmb u}\rangle$ acts on the space of linear functions: $$[\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle,
\langle{\pmb a},{\pmb u}\rangle]={-}\langle{\pmb a\alpha},{\pmb u}\rangle.$$
Hence, the Lie algebra ${\mathfrak{sp}}(m;\mathbb C)$ is obtained by the adjoint representation of quadratic forms $${\rm{ad}}(\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle)
={-}\alpha \in {\mathfrak{sp}}(m;\mathbb C).$$ It follows that for every $*$-function such as $*$-polynomials or $f_*(\pmb u)=\int_{{\mathbb R}^n}{\hat f}(\xi)
e_*^{\frac{1}{i{\hbar}}\langle\xi,{\pmb u}\rangle}{d\!\!\!{\lower-0.6ex\hbox{$-$}}\!}\xi$, $$e^{t{\rm{ad}}
(\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle)}f_*({\pmb u})
=f_*(e^{{-}t\alpha}{\pmb u}),$$ where $e^{{-}t\alpha}$ is a linear transformation $e^{-t\alpha}\in Sp(m;\mathbb C)$.
A concrete form for the case $m=1$ is given by using the transposed matrices as follows, $$\label{eq:ad-quad}
{\text{ad}}(\frac{i}{2\hbar}(au^2+bv^2+2cuv))
\begin{bmatrix}
u\\v
\end{bmatrix}
=
\begin{bmatrix}
-c& -b\\
a& c
\end{bmatrix}
\begin{bmatrix}
u\\v
\end{bmatrix}$$
Let $V_{2m}{=}
\{\langle{\pmb\xi},{\pmb u}\rangle;
{\pmb\xi}{\in}{\mathbb C}^{2m}\}$. For every quadratic form $\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*$, ${\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*)$ is welldefined as a linear mapping independent of expression parameters. $${\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*):
V_{2m}\to V_{2m},\quad
{\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*):
H\!ol({\mathbb C}^{2m})\to H\!ol({\mathbb C}^{2m})$$ It is easy to see that ${\text{ad}}(\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle)
{=}{-}\alpha{\in}{\mathfrak{sp}}(m,{\mathbb C})$, hence it extends as a ${*}$-derivation $${\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*):
({\mathcal E}_2({\mathbb C}^{2m}),{*})\to
({\mathcal E}_2({\mathbb C}^{2m}),{*}).$$ Linear algebra on finite dimensional vector space gives linear isomorphisms $$e^{{\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*)}:
V_{2m}\to V_{2m},\quad
e^{{\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*)}:
H\!ol({\mathbb C}^{2m})\to H\!ol({\mathbb C}^{2m})$$ and a ${*}$-isomorphism $$e^{{\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*)}:
({\mathcal E}_2({\mathbb C}^{2m}),{*})\to
({\mathcal E}_2({\mathbb C}^{2m}),{*}).$$
Set $A{=}\alpha J$. Since ${:}e_*^{\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}{:}_{_K}$ is defined as a multi-valued holomorphic mapping from an open connected dense domain $D$ containing the origin into ${\mathcal E}_{2+}({\mathbb C}^{2m})$. and the first associativity Theorem\[assocthm\] applied to $t{=}{\hbar}s$ shows the following:
\[lem15\] Both sides are well-defined and associativity $${:}(e_*^{\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}
{*}\langle{\pmb\xi},{\pmb u}\rangle)
{*}e_*^{{-}\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}{:}_{_K}{=}
{:}e_*^{\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}
{*}(\langle{\pmb\xi},{\pmb u}\rangle
{*}e_*^{{-}\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*})
{:}_{_K}$$ holds for every $t{\in}D$.
Differentiating the identity of Lemma\[lem15\] by using Theorem\[assocthm\] several times, gives that $$\frac{d}{dt}{:}e_*^{\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}
{*}\langle{\pmb\xi},{\pmb u}\rangle
{*}e_*^{{-}\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}{:}_{_K}{=}
{:}{\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*)
(e_*^{\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}
{*}\langle{\pmb\xi},{\pmb u}\rangle
{*}e_*^{{-}\frac{t}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*})
{:}_{_K}.$$
Uniqueness of the real analytic solution gives that the matrix obtained is independent of expression parameters: $${:}e_*^{\frac{t}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*}
{*}\langle{\pmb\xi},{\pmb u}\rangle
{*}e_*^{{-}\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}{:}_{_K}{=}
e^{t{\rm{ad}}(\frac{1}{2i{\hbar}}\langle{\pmb u}A,{\pmb u}\rangle_*)}
\langle{\pmb\xi},{\pmb u}\rangle{=}
\langle{\pmb\xi}e^{{-}t\alpha},{\pmb u}\rangle
\,\,({=}\langle{\pmb\xi}, e^{{-}t\alpha}{\pmb u}\rangle),$$ where $A{=}\alpha J, \alpha{\in}{sp}(m,{\mathbb C})$.
\[The basic formula\] If ${:}e_*^{\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}{:}_{_K}$ is defined, then $$e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
*{\langle{\pmb a},{\pmb u}\rangle}{*}
e_*^{-t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
=\langle {\pmb a}e^{-t\alpha},{\pmb u}\rangle.$$
The proof is based on the fact that $e_*^{\frac{t}{2i{\hbar}}
\langle{\pmb u}A,{\pmb u}\rangle_*}{*}
{\langle{\pmb a},{\pmb u}\rangle}{*}
e_*^{-t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}$ is defined and real analytic on an open dense connected domain of $t$ containing $0$. Hence, one may replace ${\langle{\pmb a},{\pmb u}\rangle}$ by any polynomial.
Since $\{e^{\alpha}, \alpha{\in}\mathfrak{sp}(m,{\mathbb C})\}$ generates $Sp(m,{\mathbb C})$, the following is easy to see:
\[ajoint222\] As linear transformation of $V_{2m}$, we have ${\rm{Ad}}(e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle})
=e^{t{\rm{ad}}(\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle)}.$ Hence, ${\rm{Ad}}(e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle})$ has no singular point and generates the group $Sp(m,{\mathbb C})$.
This identity holds in spite of the ambiguity of the amplitude of $e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}$, because the ambiguity of amplitude disappears in the adjoint formula. Hence,
The blurred Lie group $Sp^{(\frac{1}{2})}_{\mathbb C}(m)$ generated by $e_*^{t\langle{\pmb u}(\frac{1}{i{\hbar}}\alpha J),{\pmb u}\rangle}$ looks like a double covering group of $Sp(m,{\mathbb C})$ which is known to be simply connected.
Several point set pictures for blurred subgroups
------------------------------------------------
Recall the surjective “homomorphism” $${\rm{Ad}}: Sp^{(\frac{1}{2})}_{\mathbb C}(m)\to Sp(m,\mathbb C).$$ For every subgroup $G$ of $Sp(m,{\mathbb C})$, the full inverse ${\rm{Ad}}^{-1}G$ may be viewed as a [*blurred*]{} covering of $G$. However, it is often possible that ${\rm{Ad}}^{-1}G$ is a genuine Lie group under a suitable expression parameter.
Suppose we have a subgroup $G$ of $Sp(m,{\mathbb C})$. Take a simple open covering $\{V_{\alpha}\}_\alpha$ of $Sp(m,{\mathbb C})$ such that $\{V_{\alpha}{\cap}\,G\}_\alpha$ is also a simple open covering of $G$, and each $V_\alpha$ is contained in some $C_{\kappa}({\mathcal O}_{\kappa})$. (Cf. Lemma\[crucial00\].) For every $\alpha, \beta, \gamma$ we denote $$V_{\alpha\beta}{\cap}\,G{=}
V_{\alpha}{\cap}V_{\beta}{\cap}G, \quad
V_{\alpha\beta\gamma}{\cap}\,G{=}
V_{\alpha}{\cap}V_{\beta}{\cap}V_{\gamma}
{\cap}\,G,\quad e.t.c.$$ Although $Sp^{(\frac{1}{2})}_{\mathbb C}(m)$ is a blurred double covering, the simplicity of $V_{\alpha}{\cap}\,G$ ensures that\
${\rm{Ad}}^{-1}(V_{\alpha}{\cap}\,G)
{=}(V_{\alpha}{\cap}\,G, {\mathbb Z}_2)$, and the patching diffeomorphisms $\phi_{\alpha\beta}: V_{\alpha\beta}{\cap}\,G\to{\mathbb Z}_2 $ satisfies the cocycle condition $$\phi_{\alpha\beta}\phi_{\beta\gamma}\phi_{\gamma\alpha}{=}\pm 1$$ as $2$-to-$2$ mappings. These $2$-to-$2$ patching diffeomorphisms give on each $(\alpha,\beta)$ two choices of patching diffeomorphisms, say $\pm\phi_{\alpha\beta}$. In a certain case, we can select one of these sign to clear the cocycle condition to obtain a genuine subset.
\[dblecover\] For a connected subgroup $G$ of $Sp(m,{\mathbb C})$, if we can select patching diffeomorphisms so that they satisfies the cocycle condition, then there is a group $\tilde G$ contained in ${\rm{Ad}}^{-1}(G)$ such that ${\rm{Ad}}:
\tilde G \to G$ is a surjective homomorphism.
[**Proof**]{}. Since patching diffeomorphisms are so adjusted that the cocycle condition is satisfied, we have a genuine point set. But it is easy to see that these satisfies the condition of covering group of $G$. Note that such a point set picture may not be unique. ${\square}$
We have already in [@OMMY4] an example that $Sp^{(\frac{1}{2})}_{\mathbb C}(m)$ contains $Spin(m)$ under a special ordered expression $K_s$. Here, we give a simplest example. Note that $$\frac{i}{2{\hbar}}[\sum(u_i^2+v_i^2),
\binom{\pmb u}{\pmb v}]=
\begin{bmatrix}
0&-I_m\\
I_m& 0\\
\end{bmatrix}\binom{\pmb u}{\pmb v}.$$ We see that $Sp(m, \mathbb C)$ contains $U(1)$ in the form $$\label{eq:U(1)}
U(1)=
\left\{\begin{bmatrix}
\cos\theta I_m&-\sin\theta I_m\\
\sin\theta I_m&\,\,\,\cos\theta I_m
\end{bmatrix}; \theta \in\mathbb R \right\}.$$ Hence we see that $\{{\rm{Ad}}(e_{*}^{\frac{i\theta}{2{\hbar}}\sum_m(u_i^2+v_i^2)})\}=U(1)$ and the full inverse ${\tilde U}(1)={\rm{Ad}}^{-1}(U(1))$ is a double covering group of $U(1)\subset Sp(m,{\mathbb C})$. In the next section, we see that there are open subsets ${\mathfrak K}_1$, ${\mathfrak K}_2$ of expression parameters such that $${:}{\tilde U}(1){:}_{_K}=
\left\{
\begin{matrix}
U(1){\times}{\mathbb Z}_2 & K\in {\mathfrak K}_1\\
\text{the connected double cover of }\,\,U(1)&K\in {\mathfrak K}_2
\end{matrix}
\right.$$ Then, noting that $Sp(m,\mathbb R)\supset U(1)$, the full inverse ${\rm{Ad}}^{-1}(Sp(m,\mathbb R))$ is a genuine connected double covering group of $Sp(m,\mathbb R)$ under the $K$-ordered expression such that $K\in {\mathfrak K}_2$. This is called the [**metaplectic group**]{} and denoted by $M\!p(m)$. The metaplectic group is the connected double covering group of $Sp(m,{\mathbb R})$, which appears naturally as patching diffeomorphisms of the symbols of the group of invertible Fourier integral operators. It is known that $M\!p(m)$ has no complexification as Lie groups. Thus $Sp^{(\frac{1}{2})}_{\mathbb C}(m)$ is viewed as its complexification as blurred Lie groups.
For concrete computation, note that the adjoint mapping ${\rm{Ad}}$ gives $$e_*^{\frac{r}{{\hbar}}u^2}\rightarrow
\left[\begin{array}{cc}
1,& 0\\
-ri,& 1
\end{array}\right], \quad
e_*^{\frac{s}{{\hbar}}iuv}\rightarrow
\left[\begin{array}{cc}
e^{-s},& 0\\
0,& e^{s}
\end{array}\right], \quad
e_*^{\frac{t}{{\hbar}}v^2}\rightarrow
\left[\begin{array}{cc}
1,& ti\\
0,& 1
\end{array}\right]$$ $$e_*^{\frac{\theta}{2{\hbar}}(u^2+v^2)}\rightarrow
\left[\begin{array}{cc}
\cosh\theta,& i\sinh\theta\\
-i\sinh\theta,& \cosh\theta
\end{array}\right], \quad
e_*^{\frac{s}{2{\hbar}}(u^2-v^2)}\rightarrow
\left[\begin{array}{cc}
\cos s, & i\sin s\\
i\sin s,& \cos s
\end{array}\right]$$ In particular, $Sp(1,\mathbb C)=S\!L(2,\mathbb C)$ contains $S\!L(2,\mathbb R)$ and $$\begin{aligned}
SU(2)=&\Big\{
\begin{bmatrix}
\alpha & \beta\\
-\bar\beta &\bar\alpha\\
\end{bmatrix};
|\alpha|^2{+}|\beta|^2=1\Big\}\cong
S^3, \\
SU(1,1)=&\Big\{
\begin{bmatrix}
\alpha & \beta\\
\bar\beta &\bar\alpha\\
\end{bmatrix};
|\alpha|^2{-}|\beta|^2=1\Big\}.
\end{aligned}$$ Through these subgroups we take the full inverse ${\rm{Ad}}^{-1}(G)$. Hence, for $S\!L(2,\mathbb R)$ we see $$\begin{aligned}
\{{\rm{Ad}}&(e_*^{\frac{r}{2{\hbar}}i(u^2+v^2)}
*e_*^{\frac{s}{{\hbar}}iuv}*e_*^{\frac{ti}{{\hbar}}u^2}); r, s, t\in \Bbb R\}\\
=&\left\{
\left[\begin{array}{cc}
\cos r, & -\sin r\\
\sin r,& \cos r
\end{array}\right]
\left[\begin{array}{cc}
e^{-s},& 0\\
0,& e^{s}
\end{array}\right]
\left[\begin{array}{cc}
1,& 0\\
t,& 1
\end{array}\right]; r,s,t\in \Bbb R \right\}
\end{aligned}$$ Hence, ${\rm{Ad}}^{-1}(S\!L(2,{\mathbb R}))$ is the connected double covering of $S\!L(2,{\mathbb R} )$ under the $K$-ordered expression such that $K{\in}{\mathfrak K}_2$.
Similarly, under the $K$-ordered expression such that $K{\in}{\mathfrak K}_2$, we see ${\widetilde{SU}}(1,1)={\rm{Ad}}^{-1}(SU(1,1))$ is the connected double covering group of $SU(1,1)$.
Next, consider $${\widetilde{SU}}(2)={\rm{Ad}}^{-1}(SU(2)).$$ Indeed, this is the simplest toy model of blurred covering group. More precisely, decompose $SU(2)$ as $$\left\{
\left[\begin{array}{cc}
\cos\theta, & -\sin\theta e^{i\psi}\\
\sin\theta e^{-i\psi},& \cos\theta
\end{array}\right]
\left[\begin{array}{cc}
e^{i\rho},& 0\\
0,& e^{-i\rho}
\end{array}\right]
; \theta, \psi, \rho\in \Bbb R, \,\,|\theta|<\frac{\pi}{2}\right\}$$ with singular points at $\cos\theta=0$, where $\theta, \psi$ may be viewed as the latitude and longitude respectively. Under a suitable expression parameter, we have a double covering group of the group $\{e^{i\rho}\}$. Hence, we have a covering space by replacing $\rho$ by $\rho/2$ for each decomposition.
By this observation we see also
\[byprod22\] There is no expression parameter $K$ under which all one parameter subgroup are not $2\pi$-periodic but $4\pi$-periodic.
Several remarks on $*$-exponential functions
--------------------------------------------
By noting that $\det(e^{t\alpha}){=}1$ for every $\alpha{\in}{\mathfrak{sp}}(m,{\mathbb C})$, is rewritten as $$\label{fundamenta2}
\begin{aligned}
{:}e_*^{\frac{t}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}
{=}
\frac{2^m}{\sqrt{\det(e^{t\alpha}(I{-}{\kappa})
{+}e^{-t\alpha}(I{+}\kappa))}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}
\frac{1}{e^{t\alpha}(I{-}{\kappa}){+}e^{-t\alpha}(I{+}\kappa)}
(e^{t\alpha}{-}e^{-t\alpha})J,{\pmb u}\rangle}
\end{aligned}$$ In spite of the sign ambiguity of $\sqrt{\,\,}$, the exponential law $$\label{explaw11}
\begin{aligned}
{:}e_*^{(s{+}t)\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}=&
{:}e_*^{s\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}
{*_{_K}}{:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}\\
{:}e_*^{s(a+\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*)}{:}_{_K}=&
{:}e^{as}e_*^{s\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}
\end{aligned}$$ holds under computations such as $\sqrt{a}\sqrt{b}{=}\sqrt{ab}$. This is because that the exponential law and associativity holds on the group $Sp(m,{\mathbb C})$. Note however that $\sqrt{1}{=}\pm 1$.
By this observation we have the following:
\[oneparam\] For every fixed $\alpha$, $\kappa$, a suitable choice of angle $\theta$ gives various real one parameter subgroups ${:}e_*^{se^{i\theta}\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}$, $s{\in}{\mathbb R}$. Moreover, we can find many complex semi-groups on various sectors.
By the concrete formula , we have also the following:
\[hbaranalytic\] Replacing $t$ by $t{\hbar},$ ${:}e_*^{t
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{_K}$ is real analytic in ${\hbar}$ in an open connected domain containing ${\hbar}=0$.
As has double branched singular points, we have to use two sheets by setting [*slits*]{} in the complex plane to treat ${:}e_*^{tH_*}{:}_{_K}$ univalent way. Although there is no general rule to set the slits, it is natural to set [**the slits periodically, since the singular points are distributed periodically**]{}. We adopt this rule throughout this series.
Note that $J\in {\mathfrak{sp}}(m,{\mathbb C})$ and also $J\in {Sp}(m,{\mathbb C})=
\{g\in G\!L(2m,{\mathbb C}); gJ\,{}^t\!g=J\}$. For every $g\in Sp(m,{\mathbb C})$, $\tilde J=gJg^{-1}$ is both an element of Lie algebra and a group element satisfies $\tilde J^2{=}{-}I$ and $e^{t\tilde J}{=}\cos tI{+}(\sin t)\tilde J$. Recall the formula , which is rewritten as $$\begin{aligned}
{:}e_*^{\frac{t}{i{\hbar}}
\langle{\pmb u}(\alpha J),
{\pmb u}\rangle_*}{:}_{_{({-}\!J\kappa)}}
{=}
\frac{2^m}{\sqrt{\det(I{-}{\kappa}{+}e^{-2t\alpha}(I{+}\kappa))}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}
\frac{1}{I{-}{\kappa}{+}e^{-2t\alpha}(I{+}\kappa)}
(I{-}e^{-2t\alpha})J,{\pmb u}\rangle}, \quad \kappa{=}J\!K.
\end{aligned}$$ Setting $\alpha=\tilde J$ and noting $\alpha J=gJg^{-1}J={-}g\,{}^t\!g$, we see first $$\begin{aligned}
I{-}{\kappa}&{+}e^{-2t\alpha}(I{+}\kappa)
=I{-}{\kappa}{+}(\cos 2t I{-}(\sin2t)\tilde J)(I{+}\kappa)\\
&=2(\cos t I{-}(\sin t)\tilde J)
(\cos t I{-}(\sin t){\tilde J}\kappa)\\
&=2g(\cos t I{-}(\sin t)J)
(\cos t I{-}(\sin t){J}\tilde\kappa)g^{-1},\quad
(\tilde\kappa=g^{-1}\kappa g).
\end{aligned}$$ We have also that $$(I-e^{-2t\alpha})J{=}J{-}(\cos 2t I{-}(\sin2t)\tilde J)J
{=}-2g\sin t(\cos t I-(\sin t)J){}^t\!g.$$
Since $\det(\cos t I{-}(\sin t)J)=1$, it follows $$\det(I{-}{\kappa}{+}e^{-2t\alpha}(I{+}\kappa))
=2^{2m}\det(\cos t I{-}(\sin t)J\tilde{\kappa}).$$ Recalling that $K{=}{-}J\kappa, \,\,\kappa{=}g{\tilde\kappa}g^{-1}$, and plugging these, we have $$\label{tildeKK}
{:}e_*^{-\frac{t}{i{\hbar}}
\langle{\pmb u}g,
{\pmb u}g\rangle_*}{:}_{_{({-}\!J\kappa)}}
{=}
{:}e_*^{\frac{t}{i{\hbar}}
u\langle{\pmb u}(\alpha J),
{\pmb u}\rangle_*}{:}_{_{({-}\!J\kappa)}}\\
{=}
\frac{1}{\sqrt{\det(\cos t I{-}(\sin t)J\tilde\kappa)}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}g
\frac{-\sin t}{\cos t I{-}(\sin t){J}\tilde\kappa},{\pmb u}g\rangle}$$ where $\cos t I{-}(\sin t){J}\tilde\kappa$ is a symmetric matrix.
Now, one may assume in generic ordered expressions, $-J\tilde\kappa$ has disjoint $2m$ simple eigenvalues. Considering the diagonalization of $J\tilde\kappa$ in , we easily see that
\[singularpts00\] In a generic $($open dense$)$ ordered expression, the singular points distributed $\pi$-periodically along $2m$ lines parallel to the real axis, and the singular points are all simple double branched singular points. Moreover, ${:}e_*^{-\frac{t}{i{\hbar}}
\langle{\pmb u}g,
{\pmb u}g\rangle_*}{:}_{_{({-}\!J\kappa)}}$ is rapidly decreasing along lines parallel to the pure imaginary axis of the growth order $e^{-|t|^m}$, where $2m=n$.
[**Generic assumption**]{}Throughout this series, we suppose above properties for generic ordered expressions except otherwise stated.
In addition to generic assumption, we may suppose the following:
\[1parameter\] In generic ordered expression $K$, one may assume that ${:}e_*^{\frac{t}{i{\hbar}}
\langle{\pmb u}g,
{\pmb u}g\rangle_*}{:}_{_{({-}\!J\kappa)}}$ has no singular point on the real line. Hence, the exponential law proved by the uniqueness in the left evolution equation gives that ${:}e_*^{\frac{t}{i{\hbar}}
\langle{\pmb u}g,
{\pmb u}g\rangle_*}{:}_{_{({-}\!J\kappa)}},$ $t\in{\mathbb R}$ forms a one parameter subgroup of period $\pi$, or $2\pi$ depending on the expression parameter $K$.
One of the remarkable feature of this concrete formula (\[tildeKK\]) is that it shows several extraordinary properties of $*$-exponential functions. For instance, we will see in the next section the following (cf. , Lemma\[singularpts00\]):
\[remakable\] If $\alpha=gJg^{-1}$ for some $g\in Sp(m,{\mathbb C})$, then the $*$-exponential function of quadratic form $\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*$ in a [**generic**]{} $($open dense$)$ ordered expression $\kappa$ is $2\pi$-periodic along real line $($in precise, $\pi$-periodic or alternating $\pi$-periodic$)$, and rapidly decreasing in both sides along the imaginary axis $i{\mathbb R}$ in the growth order $e^{-|t|^m}$. Hence such a $*$-exponential function must have singular points by Liouville’s theorem.
By , we see also that $
{:}e_*^{s\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{\kappa}
$ has in general discrete branched singularities on the complex space $s{\in}{\mathbb C}$ with some periodicity depending on the parameter $\kappa{=}JK$. To obtain the value without sign ambiguity, we have to fix the path from $0$. To stress this, we use sometimes the notation $$\label{uninotation}
{:}e_*^{[0\sim s]\frac{1}{i{\hbar}}
\langle{\pmb u}(\alpha J), {\pmb u}\rangle_*}{:}_{\kappa}$$ where $[0\sim s]$ indicates a path joining $0$ to $s$ avoiding singular points.
Replacing $-J\kappa$ by $K$ in , we have $J\tilde\kappa= -{}^tgKg$, and replacing $t$ by $-t$ we have the formula again: $$\label{tildeKKK}
\begin{aligned}
{:}e_*^{\frac{t}{i{\hbar}}
\langle{\pmb u}g,
{\pmb u}g\rangle_*}{:}_{_{K}}
&{=}
\frac{1}{\sqrt{\det(\cos t I{-}(\sin t){}^t\!gKg)}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}g
\frac{\sin t}{\cos tI-\sin t \,{}^t\!g Kg},{\pmb u}g\rangle}
\end{aligned}$$
By requiring $1$ at $t=0$ and by using $\det g=1$, we have by setting $t=\pm\pi$, and $t=\pm\frac{\pi}{2}$, that $$\label{defpolar00}
{:}e_*^{\pi\frac{\pm 1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}
{:}_{_{K}}=\sqrt{(-1)^{2m}}{=}\sqrt{1}, \quad
{:}e_*^{\pi\frac{\pm 1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}
=
\frac{1}{\sqrt{\det{K}}}
e^{{-}\frac{1}{i{\hbar}}\langle{\pmb u}\frac{1}{K},{\pmb u}\rangle}.$$ Note that the r.h.s. of the first equality looks independent of $g$ and the expression parameters, and that the r.h.s. of the second equality looks independent of $g$. Since $Sp(m,{\mathbb C})$ is connected, it looks the sign of $\sqrt{1}$ and $\sqrt{\det K}$ can be fixed. However, the sign of $\sqrt{1}$ depends both on the expression $K$ and on the path from $0$ to $\pi$ by which we choose the sign of ${:}e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}$ under the condition $e_*^{0\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}=1$, where the path should be so chosen that there is no singular point on the path.
In the case $K{=}0$ (the Weyl ordered expression), the r.h.s. of the second identity diverges and the first identity gives $${:}e_*^{[0\sim t]\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{0}=
\sqrt{(\cos t)^{2m}}
e^{\frac{1}{i{\hbar}}\langle{\pmb u}g
\frac{\sin t}{\cos t}I,{\pmb u}g\rangle}
=(\cos t)^me^{\frac{1}{i{\hbar}}\langle{\pmb u}g
\frac{\sin t}{\cos t}I,{\pmb u}g\rangle}$$ by requesting $1$ at $t=0$. Hence ${:}e_*^{[0\sim\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{0}=(-1)^m$.
In general the $\pm$-sign depends on the path from $0$ to $\pi$ or $\pi/2$. It depends on which sheet the end point of the path is sitting. By this observation, we see that $$\label{signsheets}
{:}e_*^{[0\sim\pi]\frac{\pm 1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_K}=(-1)^m, \,\,\,(\text{resp.} -(-1)^m)$$ if ${:}e_*^{\pi\frac{\pm 1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_K}$ is sitting in the same (resp. opposite) sheet as in ${:}e_*^{0\frac{\pm 1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_K}$.
On the other hand, for a fixed $K$, the r.h.s. of the second equality is independent of $g$. Since $Sp(m,{\mathbb C})$ is connected, it looks that one can fix the sign of $\sqrt{\det K}$ in the r.h.s. of the second equality. Here, we meet the strange phenomenon that we have already met in [@OMMY4]. We call $e_*^{\pi\frac{1}{2i{\hbar}}\langle{\pmb u}g, {\pmb u}g\rangle_*}$ the (total) [**polar element**]{} and denote this by ${{\varepsilon}}_{00}$. The polar element will be discussed in the next section more closely.
### The case $m=1$
In this section, we treat the case of two variables $u,v$ (i.e. the case $m{=}1$). Note first that $\{\langle{\pmb u}g, {\pmb u}g\rangle_*; g{\in}S\!L(2,{\mathbb C})\}$ is spanned by quadratic forms given by $$\begin{bmatrix}
\cosh r& \sinh r\\
\sinh r&\cosh r
\end{bmatrix},\quad
\begin{bmatrix}
\cos r& i\sin r\\
i\sin r&\cos r
\end{bmatrix},\quad
\begin{bmatrix}
e^{is}& 0\\
0&e^{-is}
\end{bmatrix},\quad
\begin{bmatrix}
e^{s}& 0\\
0&e^{-s}
\end{bmatrix},\quad r, s\in{\mathbb R}.$$ In particular, we treat $*$-exponential functions $e_*^{t\frac{1}{i{\hbar}}(u_*^2{+}v_*^2)}$, $e_*^{t\frac{1}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}$ more closely.
In , we set the expression parameter $K= -J\!\kappa{=}
\left[\begin{smallmatrix}
a&c\\
c&b
\end{smallmatrix}\right]$, and we set the amplitude part of $\frac{1}{\sqrt{\Delta_{K}(t)}}$ where $$\label{Delta}
\Delta_{K}(t){=}
\det((\cos t)I{+}(\sin t)K){=}
\cos^2t{-}(a{+}b)\sin t\cos t{+}(ab{-}c^2)\sin^2t$$ Note that ${a{+}b}$ and ${ab{-}c^2}$ can be arbitrary complex numbers.
$\Delta_{K}(t)$ and the phase part of are both $\pi$-periodic, but the sign of $\sqrt{\Delta_{K}(t)}$ depends on the expression parameter $K$ and the path from $0$ to $t$ in the complex plane. The sign ambiguity is removed by putting the initial condition $e_*^{0\frac{1}{i{\hbar}}H_*}{=}1$ at $t{=}0$ only in the case that $a=-b$ and $c^2{+}a^2=1$, i.e. $\Delta_{K}(t)=1$, or the case that $(a{-}b)^2{+}4c^2{=}0$, i.e. $\Delta_{K}(t){=}\frac{1}{4}(2\cos{t}{+}(a{+}b)\sin{t})^2$.
Moreover, singular points depend on expression parameters (cf.[@OMMY4]). The case $c{=}0$ where $a, b$ are arbitrary in $\mathbb C$ gives an overview how the singular points are moving: $${:}e_*^{t\frac{1}{i{\hbar}}(u_*^2{+}v_*^2)}{:}_{_K}{=}
\frac{1}{\sqrt{(\cos t{-}a\sin t)(\cos t{-}b\sin t)}}
\exp\frac{1}{i{\hbar}}\Big(\frac{\sin t}
{\cos t{-}b\sin t}u^2{+}\frac{\sin t}{\cos t{-}a\sin t}v^2\Big).$$ By these observations, we see that the singular points appear $\pi$-periodically in general on two lines parallel to the real axis and the ${*}$-exponential functions have $e^{-|t|}$-growth with the exponential decay on the line parallel to the pure imaginary axis when these do not hit singular points.
The observation here gives in addition the following:
\[reflem\] Choosing the expression parameter $K$, we can make both ${:}e_*^{[0{\to}\pi]\frac{1}{i{\hbar}}
(u_*^2{+}v_*^2)}{:}_{_{K}}=1,
\,\,\text{and}\,\,-1$. Moreover, multiplying $e^{t}$, we have an extremal point, called [*vacuum*]{}, $$\lim_{t\to\infty}e^t
{:}e_*^{it\frac{1}{i{\hbar}}
(u_*^2{+}v_*^2)}{:}_{_{K}}=
\frac{2}{\sqrt{(1{-}a)(1{-}b)}}
\exp\frac{1}{i{\hbar}}
\Big(\frac{1}{1{-}b}u^2{+}\frac{1}{1{-}a}v^2\Big).$$ depending on the expression parameters.
We fix the expression parameter $K$ as follows: $$K_{re}=
\begin{bmatrix}
\rho&ic'\\
ic'&\rho\\
\end{bmatrix},\,\,\,{or}\,\,\,
K_{im}=
\begin{bmatrix}
i\rho&c\\
c&i\rho\\
\end{bmatrix},\,\,\rho,\,c,\,c'\,{\in}{\mathbb R}.$$ The formula is rewritten in this case as $$\begin{aligned}
\Delta_{K_{re}}(t)&{=}
\det((\cos t)I{+}(\sin t)K_{re}){=}
\cos^2t{+}2\rho\sin t\cos t{+}(\rho^2{+}{c'}^2)\sin^2t,\\
\Delta_{K_{im}}(t)&{=}
\det((\cos t)I{+}(\sin t)K_{im}){=}
\cos^2t{+}2i\rho\sin t\cos t{-}({\rho}^2{+}c^2)\sin^2t.
\end{aligned}$$ The first one is obviously positive definite if $c'\not=0$ (i.e. Siegel ordered expression in the case $m=1$) and hence $\sqrt{\Delta_{K_{re}}(t)}$ does not change sign when $t$ moves $0$ to $\pi$ along real line.
On the other hand, $$\sqrt{\Delta_{K_{im}}(t)}
{=}
\frac{1}{2}e^{-it}
\sqrt{(1{+}\rho)^2{+}c^2}\sqrt
{(e^{2it}{+}\alpha)(e^{2it}{+}\bar\alpha)}, \quad
\alpha=\frac{1{-}(\rho{+}ic)}{1{+}(\rho{+}ic)}.$$ One may assume generically that $|\alpha|\not=1$. Hence, $\sqrt{\Delta_{K_{im}}(t)}$ changes sign when $t$ moves from $0$ to $\pi$. Thus, we have
\[111\] ${:}e_*^{t\frac{1}{i{\hbar}}(u^2{+}v^2)}{:}_{_{K_{re}}}$ is $\pi$-periodic, and the two lines of singular points are sitting in both upper and lower half plane. The real line is between these.
On the other hand, ${:}e_*^{t\frac{1}{i{\hbar}}(u^2{+}v^2)}{:}_{_{K_{im}}}$ is alternating $\pi$-periodic, and the two lines of singular points are sitting in upper or lower half plane depending on the sign of $\rho$.
The expression parameter $K_{im}$ is the case $m=1$ of the special expression parameter $K_s$ used in [@OMMY4].
Next, we take our attention to the quadratic form $2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v$, but we take a general expression parameter $K{=}
\footnotesize
{\begin{bmatrix}
\delta&c\\
c&\delta'
\end{bmatrix}}$. A little complicated calculation via intertwiner $I_{_{K_0}}^{^{K}}$ from the normal ordered expression gives by setting $\Delta{=}e^t{+}e^{-t}{-}c(e^t{-}e^{-t})$ that $$\label{genericparam00}
{:}e_*^{t\frac{1}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_{K}}{=}
\frac{2}{\sqrt{\Delta^2{-}(e^t{-}e^{-t})^2\delta\delta'}}
\,\,e^{\frac{1}{i{\hbar}}
\frac{e^t-e^{-t}}{\Delta^2{-}(e^t{-}e^{-t})^2\delta\delta'}
\big((e^t-e^{-t})(\delta' u^2{+}\delta v^2){+}2\Delta uv\big)}.$$ The $q$-scalar and the polar element are obtained by setting $t=\pm{\pi i}$ and $t=\pm\frac{\pi i}{2}$ respectively.
For the simplest case in , that the case $c=\delta=\delta'=0$ is the Weyl ordered expression. This is not a generic ordered expression having singular points on the imaginary axis, and this is $\pi i$-alternating periodic.
On the contrary, the unit ordered expression is given by $K=I$, i.e. $\delta=1, \delta'=1$, $c=0$. By , we have $${:}e_*^{t\frac{1}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_{I}}=
\frac{2}{\sqrt{4}}e^{\frac{1}{4i{\hbar}}
(e^{2t}{+}e^{-2t}{+}2)(u^2{+}v^2){+}2(e^{2t}{-}e^{-2t})uv}.$$ This is $\pi i$-periodic and there is no singular point.
For the case $\delta=\delta'=0$ but $c\not=0$ which involves the normal ordered expression, we see that $$\label{normalantinormal}
{:}e_*^{t\frac{1}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_{K}}{=}
\frac{2}{\sqrt{\Delta^2}}
\,\,e^{\frac{1}{i{\hbar}}
\frac{e^t-e^{-t}}{\Delta^2}\big(2\Delta uv\big)}
=
\frac{2}{\Delta}
\,\,e^{\frac{1}{i{\hbar}}
\frac{e^t-e^{-t}}{\Delta}2uv}.$$ This is the case where the singular points are not branching ones and they are sitting $\pi i$ periodically on a single line parallel to the imaginary axis whose real part are given by $\log\big|\frac{c+1}{c-1}\big|$. We see also that ${:}e_*^{\frac{t}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_K}$ is alternating $\pi i$-periodic along the imaginary axis.
Suppose in that $K{=}K_{re}$; $$\label{genkkk}
\delta{=}\delta'{=}\rho,\,\,c=ic' \quad
\rho,\,\,c'\in{\mathbb R},\,\,
c'{\not=}0.$$ By setting $\beta{=}\rho+ic'$, we have that $$\label{valueatpi77}
\frac{1}{2}\sqrt{\Delta^2{-}(e^t{-}e^{-t})^2\delta\delta'}{=}
e^{-t}\sqrt{(1{-}\beta)(1{+}\bar\beta)}
\sqrt{e^{2t}{-}\frac{1{+}\beta}{1{-}\beta}}
\sqrt{e^{2t}{-}\frac{1{-}\bar\beta}{1{+}\bar\beta}}.$$ Obviously $|\frac{1{+}\beta}{1{-}\beta}|
|\frac{1{-}\bar\beta}{1{+}\bar\beta}|=1$, but one may assume in generic ordered expression that $|\frac{1{+}\beta}{1{-}\beta}|{\not=}1$. Hence, $\sqrt{e^{2t}{-}\frac{1{+}\beta}{1{-}\beta}}
\sqrt{e^{2t}{-}\frac{1{-}\bar\beta}{1{+}\bar\beta}}$ changes the sign when $t$ moves $0$ to $\pi i$. Thus we see $\frac{1}{2}\sqrt{\Delta^2{-}(e^t{-}e^{-t})^2\delta\delta'}$ does not change the sign on the interval $[0,\pi i]$. Hence ${:}e_*^{t\frac{1}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_{K_{re}}}$ is $\pi i$-periodic. Remark now this is the case in ${:}e_*^{\frac{it}{i{\hbar}}\langle{\pmb u}g,
{\pmb u}g\rangle}{:}_{_K}$ where $$g=
\frac{1}{\sqrt{2}}
\begin{bmatrix}
1&i\\
i&1
\end{bmatrix}, \quad
K_{re}=
\begin{bmatrix}
\rho&ic'\\
ic'&\rho
\end{bmatrix}.$$ The concrete expression of polar element is $$\label{polar400}
{:}{{\varepsilon}}_{00}{:}_{_{K_{re}}}=
{:}e_*^{\pi i\frac{1}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_{K_{re}}}{=}
\frac{1}{\sqrt{(\rho^2{+}{c'}^2)}}
e^{-\frac{1}{i{\hbar}}\frac{1}{\rho^2{+}{c'}^2}
\rho(u^2{+}v^2){-}2c'iuv}.$$
Note that the quadratic form $u_*^2{+}v_*^2$ is a representative of general quadratic forms $au_*^2{+}bv_*^2{+}2cu{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v$ with the discriminant $c^2{-}ab={-}1$ via $S\!L(2;{\mathbb C})$-linear change of generators.
Since $S\!L(2;{\mathbb C})=Sp(1;{\mathbb C})$, such a linear change is covered by a change of expression parameters by the formula . Thus, even if an expression parameter $K$ is fixed generically, these patterns for the quadratic form $u_*^2{+}v_*^2$ must appear for $au_*^2{+}bv_*^2{+}2cu{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v$ via changing coefficients. We shall show that this appears slightly different, more delicate shape.
In general, we set $$e_*^{itH_*}{=}
e_*^{\frac{it}{i{\hbar}}(au_*^2{+}bv_*^2{+}2cu{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)},
\quad c^2{-}ab=1.$$
We have then three patterns as follows:
$(Q(1))$: $e_*^{itH_*}$ is alternating $\pi$-periodic and the 2 lines of singular points are in the upper
half-plane.
$(Q(2))$: $e_*^{itH_*}$ is alternating $\pi$-periodic and the 2 lines of singular points are in the lower
half-plane.
$(Q(3))$: $e_*^{itH_*}$ is $\pi$-periodic and the real line are between 2 lines of singular points.
$(Q(k))$ are open subsets of $\{(a,b,c); c^2{-}ab{=}1\}$ such that $(Q(1))\cup(Q(2))\cup(Q(3))$ is dense. Since the time reversing sends the line of singularities to the opposite side, we see that $(Q(k))$ has the property $$(Q(1))^{-1}=(Q(2)),\quad (Q(3))^{-1}=(Q(3)).$$ This means that if $e_*^{itH_*}\in (Q(1))$, then $e_*^{it(-H_*)}\in (Q(2))$.
[**Remark**]{}Alternating $\pi$-periodicity appears when no sheet changing occurs. Thus, $${:}e_*^{\pi\frac{1}{i{\hbar}}(u^2{+}v^2)}{:}_{_{K}}=-1$$ always on the positive sheet, as far as requesting $e_*^{0\frac{1}{i{\hbar}}(u^2{+}v^2)}=1$. On the other hand , $${:}e_*^{\pi\frac{1}{i{\hbar}}(u^2{+}v^2)}{:}_{_{K}}=1,$$ when the sheet changing occurs on a path from $0$ to $\pi$. It is very easy to make a mistake.
Recall first the anomalous phenomena mentioned in [@OMMY4] that a polar element is obtained not only by $e_*^{\frac{\pi i}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}$ but also by $e_*^{\frac{\pi i}{2i{\hbar}}(au_*^2{+}bv_*^2{+}c2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}$, $c^2{-}ab=1$. This shows that a polar element is sitting on various one parameter subgroups. This is just like the longitude lines starting at the north pole meet again at the South Pole. We show in the next section this is a generic phenomena of $e_*^{\frac{\pi i}{2i{\hbar}}(au_*^2{+}bv_*^2{+}2cu{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}$, $c^2{-}ab=1$. Thus, a polar element has infinitely many [*square roots*]{} sitting on the equator.
Beyond the south pole the longitude lines come back again to the north pole, where we give the initial value $1$ to every one parameter subgroup parameterized by longitude. However, it is a little surprising that the periodicity of these periodic movement depends on expression parameters.
### Product structure {#prodstr}
The product formula shows that the space ${\mathbb C}e^{Q(u,v)}$ of exponential functions of polynomial of degree 2 forms a very special subclass in the space ${\mathcal E}_{2+}({\mathbb C}^{2m})$. It is useful to memorize the next theorem:
\[meromuseful\] In a generic ordered expression $K$, the $*_{_K}$-product $${\pi}_{_K}: {\mathbb C}e^{Q(u,v)}\times {\mathbb C}e^{Q(u,v)}\to
{\mathbb C}e^{Q(u,v)}$$ is a mapping given in the form ${\pi}_{_K}(ae^{Q}, be^{R})= ab\sqrt{f(Q,R,K)}\,\,e^{g(Q,R,K)}$ where $f$ and $g$ are meromorphic functions of $Q, R, K$. Hence the continuity $$\lim_{(k,\ell)}\pi_{_K}(a_ke^{Q_k}, b_{\ell}e^{R_{\ell}})=
\pi_{_K}(\lim_ka_ke^{Q_k}, \lim_{\ell}b_{\ell}e^{R_{\ell}})$$ holds whenever $\lim_ka_ke^{Q_k}$, $\lim_{\ell}b_{\ell}e^{R_{\ell}}$ are defined in the space ${\mathbb C}e^{Q(u,v)}$, and $$\pi_{_K}(a_ke^{Q_k}, b_{\ell}e^{R_{\ell}}),\quad
\pi_{_K}(\lim_ka_ke^{Q_k}, \lim_{\ell}b_{\ell}e^{R_{\ell}})$$ are defined.
As for products, we know already that associativity holds always with sign ambiguity. However the following theorem is useful as a corollary of Proposition\[hbaranalytic\] and the formal associativity theorem (cf. Theorem 1.3),
\[2ndassociative\] For quadratic forms $K_*$, $L_*$, $M_*$, associativity $$(e_{*}^{[0\sim r]K_*}{*}e_*^{[0\sim s]L_*})
{*}e_*^{[0\sim t]M_*}{=}
e_{*}^{[0\sim r]K_*}{*}(e_*^{[0\sim s]L_*}
{*}e_*^{[0\sim t]M_*})$$ holds without sign ambiguity whenever both sides are defined, where paths in both left/right hand sides with same symbol should be same path $($synchronized path selecting$)$.
We next consider the product $e_{*}^{sH_*}{*}e_*^{tK_*}$ for two quadratic forms $H_*, K_*$ such that $[H_*,K_*]=0$. First of all, we show the following
\[prodcomm\] If $e_{*}^{sH_*}{*}e_*^{tK_*}$ are defined on $(s,t){\in}[0,a]^2$, then $e_{*}^{sH_*}{*}e_*^{tK_*}{=}e_{*}^{tK_*}{*}e_*^{sH_*}$.
[**Proof**]{}Since $K_*{*}e_{*}^{sH_*}$ and $e_{*}^{sH_*}{*}K_*$ satisfies the equation $\frac{d}{ds}f_s=H_*{*}f_s$ with $f_0=K_*$, we have $K_*{*}e_{*}^{sH_*}=e_{*}^{sH_*}{*}K_*$. Hence, we have $$\frac{d}{dt}e_{*}^{sH_*}{*}e_*^{tK_*}{=}K_*{*}e_{*}^{sH_*}{*}e_*^{tK_*},
\quad
\frac{d}{dt}e_{*}^{tK_*}{*}e_*^{tH_*}{=}K_*{*}e_{*}^{tK_*}{*}e_*^{sH_*}$$ with the same initial condition $e_{*}^{sH_*}$. The uniqueness gives the proof. $\Box$
If $(s,t){\in}{\mathbb C}^2$, then we have in general $e_{*}^{sH_*}{*}e_*^{tK_*}=\pm e_*^{tK_*}{*}e_{*}^{sH_*}$ with the sign ambiguity by the product formula . This means in particular and the phase parts of both sides coincides (the sign ambiguity appears only in the amplitude parts). In general, $e_{*}^{sH_*}{*}e_*^{tK_*}$ has a singular set $S$ of complex codimension 1. We see that the origin $(0,0)$ is not contained in $S$. Since $S$ is a branched singularity, we have to prepare two sheets ${\mathbb C}_+^2$, ${\mathbb C}_-^2$ and “slit” $\Sigma$ of real codimension 1 to connect these two sheets. $\Sigma$ is set so that ${\mathbb C}^2{\setminus}\Sigma$ is locally simply connected and there is no singular point.
Now, restrict the parameter $(s,t)\in {\mathbb R}^2$ in $e_{*}^{sH_*}{*}e_*^{tK_*}$ and suppose $e_{*}^{sH_*}{*}e_*^{tK_*}$ has a singular point in $(s,t){\in}(0,a){\times}(0,b)$. One may assume that ${\mathbb R}^2$ is transversal to $S$ in generic ordered expression. Hence if $S\cap{\mathbb R}^2\not=\emptyset$, then this is a discrete set and $\Sigma\cap{\mathbb R}^2$ is a collection of (real one dimensional) curves starting at a singular point ending another singular point or $\infty$. Hence one may assume that the boundary $\partial([0,a]{\times}[0,b])$ cuts the slit just once for all.
\[singularpt\] Under the assumption as above, we have $e_{*}^{sH_*}{*}e_*^{tK_*}={-}e_{*}^{sK_*}{*}e_*^{tH_*}$
[**Proof**]{} As we have two sheets, there are two “origin”, $(0,0){\in}{\mathbb C}_+^2$ and $(0,0){\in}{\mathbb C}_-^2$. Since $e_{*}^{0H_*}{*}e_*^{0K_*}$ is $1$ in the positive sheet ${\mathbb C}_+^2$, the origin in the negative sheet must be treated as $-1$. Now, consider $e_{*}^{aH_*}{*}e_*^{bK_*}$ and $e_*^{bK_*}{*}e_{*}^{aH_*}$. The first one is defined by the solution of the evolution equation $$\frac{d}{dt}f_t=H_*{*}f_t,\quad f_0=e_*^{bK_*}.$$ We indicate this by the notation $e_{*}^{[0\to a]H_*}{*}e_*^{bK_*}$. This is the clockwise tracing from the origin. On the contrary, $e_*^{[0\to b]K_*}{*}e_{*}^{aH_*}$ means the anti-clockwise tracing from the origin. Now suppose there is a singular point $(s_0,t_0)$, then one of the paths $e_{*}^{[0\to a]H_*}{*}e_*^{bK_*}$ and $e_*^{[0\to b]K_*}{*}e_{*}^{aH_*}$ is crossing the slit hence they are sitting mutually in the opposite sheet. By this way, the sign changes around a singular point.$\Box$
Rule of setting slits and polar elements {#GenPat}
========================================
If it is an absolute scalar, then $(\sqrt{1})^2{=}1$ is trivial. Recall first
\[polar111\] If $e^{2\pi\alpha}{=}I$ such as $\alpha{=}J$ $($e.g. $\alpha=gJg^{-1}$, $\forall g{\in}Sp(m,{\mathbb C})$$)$, then ${:}e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}
=\sqrt{1}$ independent of $K$.
Note that l.h.s. is not a [*classic*]{} element, for this identity does not hold for ${\hbar}=0$.
Hence, the strict exponential law might be failed, that is, ${:}e_*^{2\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}
=1$ or $${:}e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}
{*_{\kappa}}{:}e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}
=1$$ may not hold automatically. If ${:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}$ has a singular point on the interval $[0,2\pi]$, then it may occur $(e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*})^2{\not=}
e_*^{2\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}$, although the equality holds modulo $\pm$ sign.
To avoid such a strange nature, we give a general rule to set [*slits*]{}. Because of the double branching singular points, we have to use two sheets by setting [*slits*]{} in the complex plane to treat these $*$-exponential functions ${:}e_*^{tH_*}{:}_{_K}$ univalent way.
By virtue of this rule, we have
\[period\] If $e^{2\pi\alpha}{=}I$ $($e.g. $\alpha=gJg^{-1}$, $\forall g{\in}Sp(m,{\mathbb C})$$)$, then ${:}\big(e_*^{[0{\to}\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*})^2{:}_{\kappa}{=}1$ for every $\kappa$-ordered expression such that ${:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}$ has no singular point on the interval $[0,\pi]$. Moreover, we have $${:}e_*^{2\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}{=}
{:}(e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*})^2{:}_{\kappa}.$$
[**Proof**]{}. Note first that this is by no means trivial. It is crucial that the assumption and the $\pi$-periodicity of singular points shows that ${:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}$ has no singular point on the interval $[\pi,2\pi]$, but if there is no rule to set the slit, it may happen that path $[0{\to}2\pi]$ cross the slit only once.
By the rule of setting slits $\clubsuit$, we see that the slits are set $\pi$-periodically. Thus, the line segment $[0,2\pi]$ must cross the slits even (possibly $0$) times. It follows ${:}e_*^{2\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}{=}1$, since this is sitting in the positive sheet.
To confirm ${:}(e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*})^2{:}_{\kappa}
={:}e_*^{2\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa},$ we have to recall how the $*$-product $e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{*}g$ is defined. We use the definition which is given by the evolution equation .
Since ${:}e_*^{\pi\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}
={\pm 1}$, one can define $${:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}
{*}_{\kappa}{:}1{:}_{\kappa}, \quad
{\text{or}}\quad
{:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}
{*}_{\kappa}{:}({-}1){:}_{\kappa}$$ by the solution of the evolution equation with the initial condition ${\pm 1}$. By Proposition \[exppoly\], the solution is ${:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}$ or $-{:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}{\alpha}J,{\pmb u}\rangle_*}{:}_{\kappa}$ respectively. This gives the result. $\square$
General polar element as $q$-scalars
------------------------------------
It is interesting that polar element ${{\varepsilon}}_{00}$ behaves just like a scalar, but it behaves various ways. Sometimes, it behaves as if it were $-1$, and sometimes it looks as if $i$ depending on $K$. We call such elements $q$-scalars. But, to treat this as a univalent element, we have to distinguish more strictly.
The strange double-valued nature of the polar element ${{\varepsilon}}_{00}$ is caused by that $e_*^{\pi\frac{1}{2i{\hbar}}\langle{\pmb u}g, {\pmb u}g\rangle_*}$ is moving discontinuously in both positive and negative sheets when $g$ moves in $Sp(m,{\mathbb C})$.
In this section, we analyze this phenomenon more clearly. In particular, we investigate the generic patterns of periodicity and singularities of $*$-exponential functions of quadratic forms under the assumption $\clubsuit$. In particular, we are interested the behaviour of polar element. In what follows, we use several notions for the path as follows:
$[0{\to}a]$ : the path starting from the origin $0$ ending at $a$ along the line segment, but the $*$-exponential is evaluated at $t{=}a$ by the continuous chase from $0$ to $a$ along the path $[0{\to}a]$.
$[0{\sim}a]$ : a path starting from the origin $0$ ending at $a$ avoiding singular points, but evaluated at $a$.
$[0{\approx}a]$ : a path starting from the origin $0$ ending at $a$ avoiding singular points and slits so that the end point is sitting in the same sheet as the origin.
For a fixed $g$, ${{\varepsilon}}_{00}={:}e_*^{\pi\frac{1}{2i{\hbar}}\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_K}$ is always viewed as a double-valued single parallel section. If $K$ is fixed, ${{\varepsilon}}_{00}$ looks independent of $g$ with $\pm$ ambiguity. To distinguish the sign, we use the notation $$\label{strictnotion}
{:}{{\varepsilon}}_{00}[g]{:}_{_K}=
{:}e_*^{[0{\to}\pi]
\frac{1}{2i{\hbar}}\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_K}=
\frac{1}{\sqrt{\det(\cos([0{\to}1]\frac{\pi}{2})I{-}(\sin([0{\to}1]\frac{\pi}{2}){}^t\!gKg)}}
e^{-\frac{1}{i{\hbar}}\langle{\pmb u}\frac{1}{K}, {\pmb u}\rangle}$$ to fix the sign of ${{\varepsilon}}_{00}$, where $[0{\to}a]$ is the path along the straight line segment. Note that ${:}{{\varepsilon}}_{00}[g]{:}_{_K}$ may not be defined at some $g$, when a singular point appears in the interval $(0,\pi/2]$. Although ${{\varepsilon}}_{00}=\pm{{\varepsilon}}_{00}[g]$ and ${{\varepsilon}}_{00}$ is independent of $g$, ${{\varepsilon}}_{00}[g]$ may not be continuous w.r.t. $g$. The sign changes discontinuously at some $g$. For a generic $K$, there is $g{\in}Sp(m,{\mathbb C})$ such that ${}^t\!gKg$ is a real diagonal matrix. Hence ${:}e_*^{t\frac{1}{2i{\hbar}}\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_K}$ has a singular point.
Note that for every $g{\in}Sp(m,{\mathbb C})$ there is $k{\in}Sp(m,{\mathbb C})$ such that $-\langle{\pmb u}g,{\pmb u}g\rangle_*{=}
\langle{\pmb u}k,{\pmb u}k\rangle_*$. This is shown for instance $$\label{sp-inverse}
g
\begin{bmatrix}
iI&0\\
0&{-i}I
\end{bmatrix}
\begin{bmatrix}
iI&0\\
0&{-i}I
\end{bmatrix}
\,{}^t\!g= -g\,{}^t\!g.$$
Recall the rule $\clubsuit$ of setting slits. As sheets are set $\pi$-periodically we see the next result:
In generic $K$-expression, ${:}e_*^{[0\to\pi]\frac{1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_K}$ and ${:}e_*^{[0\to\pi]\frac{-1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_K}$ belong to the same sheet, and $${:}e_*^{[0\to\pi]\frac{1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_K}=
{:}e_*^{[0\to\pi]\frac{-1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_K}
= 1\quad {\text{or}}\quad -1.$$ However, this may not belong to the same (positive) sheet as ${:}e_*^{0\frac{1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_K}$.
[**Proof**]{}If the path $[0{\to}\pi]$ crosses the slit $\ell$-times, then the end point ${:}e_*^{[0{\to}\pi]\frac{1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_K}$ is sitting on the $(-1)^{\ell}$-sheet. Since sheets are set $\pi$-periodically, the path $[0{\to}\pi]$ for ${:}e_*^{[0\to\pi]\frac{-1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_K}$ also crosses the slit $\ell$-times. $\Box$
On the other hand, the second equality of does not necessarily imply that $${:}e_*^{[0{\to}\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}=
{:}e_*^{[0{\to}\pi]\frac{-1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}.$$ The sheet change may occur in the continuous tracing of $\sqrt{\det(\cos t I{-}(\sin t){}^t\!gKg)}$ from $-\frac{\pi}{2}$ to $\frac{\pi}{2}$ if the path from $-\frac{\pi}{2}$ to $\frac{\pi}{2}$ crosses the slit odd-times. shows
\[Rm b\] ${:}{{\varepsilon}}_{00}[g]{:}_{_{K}}={:}{{\varepsilon}}_{00}[g]^{-1}{:}_{_K}$ if and only if $$\sqrt{\det(\cos([0{\to}\pi])I{-}(\sin([0{\to}\pi]){}^t\!gKg)}=1.$$
If there is no singular point on ${:}e_*^{t\frac{1}{2i{\hbar}}\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}$, $t\in{\mathbb R}$, then this forms a one parameter group, and thus the equality above is equivalent with ${:}e_*^{[0{\to}\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}=1$ by the exponential law.
Lemma\[reflem\] in the this section shows that for a certain $K$ there are $g$, $g'$ such that $\sqrt{(-1)^{2m}}=1$ and $-1$ respectively. Thus, even if $K$ is fixed, the sign may depend on $g$ and the path from $0$ to $\pi$. Since $Sp(m,{\mathbb C})$ is connected, the sign changes discontinuously when the path from $0$ to $\pi$ hits a singular point. The sign changes by the changing sheet caused when the path crossing the slit drawn from the set of the singular points.
In the argument above, paths were restricted in line segment to fix the ambiguous sign. In fact, we can relax this condition. The next lemma shows that the sign-changing is caused only when the path moves across the set $S$ of singular points. Take an open connected subset $U$ of $Sp(m,{\mathbb C})$ which may be $U\not=-U$. Suppose we can fix path ${:}e_*^{[0\sim\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_{K}}$ from $t{=}0$ to $t{=}\pi$ avoiding singular points but depending continuously in $g\in U$. By setting $t=\pi$, and $t=\frac{\pi}{2}$, we have the following :
\[madamada\] Under the assumption for $U$ mentioned above, the $*$-exponential function $${:}e_*^{[0{\sim}t]\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}
{:}_{_{K}}$$ is defined uniquely without sign ambiguity by the continuous tracing from the identity, and we see $$\label{defpolar}
{:}e_*^{[0{\sim}\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}
{:}_{_{K}}=\sqrt{(-1)^{2m}},$$ where $\sqrt{(-1)^{2m}}=(-1)^m$, when the end point of path is sitting in the same (positive) sheet as $0$, and $-(-1)^m$, when the end point of path is sitting in the opposite (negative) sheet.
On the other hand for the polar element, we have $$\label{defpolar22}
{:}e_*^{[0\sim\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}
=
\frac{1}{\sqrt{\det{K}}}
e^{{-}\frac{1}{i{\hbar}}\langle{\pmb u}K^{{-}1},{\pmb u}\rangle}.$$ The sign of $\sqrt{\det{K}}$ is determined by the sheet on which the end point of the path $[0{\sim}\pi]$ is sitting.
Note that $(-1)^m$ in Lemma\[madamada\] is $-1$ if $m=$odd, and $1$ if $m=$even. Thus, the mathematical context depends on $(-1)^m=\pm1$ in the next Proposition.
\[strange2\] Suppose there is $g\in Sp(m,{\mathbb C})$ such that ${:}e_*^{[0\to\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_{K}}=-1$. Then, there must exist $h\in Sp(m,{\mathbb C})$ such that ${:}e_*^{[0\to\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}h, {\pmb u}h\rangle_*}{:}_{_{K}}=1$, and $\hat{h}\in Sp(m,{\mathbb C})$ such that the path ${:}e_*^{[0\to\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}\hat{h}, {\pmb u}\hat{h}\rangle_*}{:}_{_{K}}$ must hit a singular points.
[**Proof**]{} Suppose ${:}e_*^{[0\to\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_{K}}=-1$ for every $g{\in}Sp(m,{\mathbb C})$ and suppose there is no singular point on the path $[0{\to}\pi]$.
As $Sp(m,{\mathbb C})$ is connected, and the second equality of give that for the mid-point $$\label{contrdict}
{:}e_*^{[0{\to}\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_{K}}
{=}
{:}e_*^{[0{\to}\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}k,{\pmb u}k\rangle_*}{:}_{_{K}}
{=}{:}e_*^{[0{\to}\pi]\frac{-1}{2i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_{K}}.$$ The exponential law gives $${:}e_*^{[0{\to}\pi]\frac{-1}{2i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_{K}}=
{:}\big(e_*^{[0{\to}\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}\big)^{-1}{:}_{_{K}}$$ and therefor multiplying $e_*^{[0{\to}\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}$ to both sides of , we have the contradiction $$-1=e_*^{[0{\to}\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}=
e_*^{[0{\to}\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}
{*}\big(e_*^{[0{\to}\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g,{\pmb u}g\rangle_*}\big)^{-1}=1.$$ As a result $Sp(m,{\mathbb C})$ is divided into three parts $D_+, D_-, D_{sing}$ such that $${:}e_*^{[0{\to}\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}
=
\left\{
\begin{matrix}
-1& g{\in} D_+\\
{:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}
{\text{has a singular point on}}\,\,(0,\pi)&
g{\in} D_{sing}\\
1& g{\in} D_-
\end{matrix}
\right.$$ and $D_{+}\subsetneqq Sp(m,{\mathbb C})$. In particular this yields $D_{sing}\not=\emptyset$.
Now, we show that $D_-\not=\emptyset$. Since the points of $D_{sing}$ are branched singular points, the value of ${:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}$ changes sign around branched singular point. Since we assumed as a generic assumption that the singular points distributed $\pi$-periodically along $2m$ lines parallel to the real line, there is at most one singular point on $(0,\pi)$. Thus, ${:}e_*^{[0\to 1]\frac{\pi}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}$ must change sign at $g{\in}D_{sing}$. Hence we see $D_-{\not=}\emptyset$. $\Box$
We note that $Sp(k,{\mathbb C})$ is naturally included in $Sp(m,{\mathbb C})$ for $m{>}k$. Apparently, the result mentioned in [@OMMY4] is a special case for $m=1$, $g=
\frac{1}{\sqrt{2}}\begin{bmatrix}
1&i\\
i&1
\end{bmatrix}
$ and $K{=}K_0$ (normal ordered expression).
Proposition\[strange2\] gives in particular that if $D_+{\not=}\emptyset$, then $D_-{\not=}\emptyset$ and $D_{sing}{\not=}\emptyset$.
Consider now whether it is possible $D_{+}=\emptyset$ in Lemma\[madamada\]. First we note the following:
\[madamadada\] If $D_{sing}{\not=}\emptyset$, then $D_{\pm}{\not=}\emptyset$.
[**Proof**]{}For $(t,g){\in}{\mathbb C}{\times}Sp(m,{\mathbb C})$, the set $S$ of singular points of ${:}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_{K}}$ is a closed subset of complex codimension $1$. The slit $\Sigma$ is set so that $({\mathbb C}{\times}Sp(m,{\mathbb C})){\setminus}{\sigma}$ is locally simply connected. Hence, if ${:}e_*^{[0\to\pi]\frac{1}{i{\hbar}}\langle{\pmb u}g,{\pmb u}g\rangle_*}{:}_{_{K}}$ hits a singular point for some $g$, then there are $h, h'\in Sp(m,{\mathbb C})$ in a neighborhood of $g$ such that ${:}e_*^{[0\to\pi]\frac{1}{i{\hbar}}\langle{\pmb u}h,{\pmb u}h\rangle_*}{:}_{_{K}}$ hits $\Sigma$, but ${:}e_*^{[0\to\pi]\frac{1}{i{\hbar}}\langle{\pmb u}h',{\pmb u}h'\rangle_*}{:}_{_{K}}$ does not. Hence, these two must have different sign. $\Box$.
Now note that the comment following shows that $D_{sing}{\not=}\emptyset$. Thus, we have
\[doesnotexist\] Suppose $K$ is a generic expression parameter. Then, $Sp(m,{\mathbb C})$ is divided into three non empty subsets $D_+, D_-, D_{sing}$.
[**Remark 1**]{}As singular points are distributed $\pi$- periodically, if $g\in D_{sing}$, then ${:}e_*^{t\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}$ has singular points not only in the interval $(0,\frac{\pi}{2}]$ but also in the interval $(-\pi,-\frac{\pi}{2}]$.
Theorem\[doesnotexist\] shows a polar element ${{\varepsilon}}_{00}$ is a member of various one parameter subgroups with different periodicity ${{\varepsilon}}_{00}^2=1$, and ${{\varepsilon}}_{00}^2={-}1$.
[**Note**]{}Sometimes, $D_-$ contains a compact subgroup of $Sp(m,{\mathbb C})$. Indeed, we will show in the next section that such a case exists. That is, in the case $m=1$ there is a class $K_{re}$ of expression parameters such that $${:}e_*^{[0{\to}\pi]\frac{1}{i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}=1 \quad \text{for
every }\,\, g{\in}SU(2)\quad {\rm {cf.\,\,Proposition\,\ref{nicenice22}}}.$$
[**Remark 2**]{} A polar element is a double-valued single element. Thus even though ${:}e_*^{\frac{\pi}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}{:}_{_{K}}
=
{:}e_*^{\frac{\pi}{2i{\hbar}}
\langle{\pmb u}h, {\pmb u}h\rangle_*}{:}_{_{K}}$, square of these may be different $${:}(e_*^{\pm\frac{\pi}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*})^2{:}_{_{K}}{\not=}
{:}(e_*^{\pm\frac{\pi}{2i{\hbar}}
\langle{\pmb u}h, {\pmb u}h\rangle_*})^2{:}_{_{K}}$$ if the paths from $0$ to $\pi$ have different numbers of crossing slits.
It is quite difficult to control the $\pm$ sign in the product formula. We have always to chase continuously from the identity. Even though ${:}{{\varepsilon}}_{00}[g]{:}_{_K}=\pm {:}{{\varepsilon}}_{00}[h]{:}_{_K}$, it does not necessarily imply ${:}{{\varepsilon}}_{00}[g]^2{:}_{_K}={:}{{\varepsilon}}_{00}[h]^2{:}_{_K}$. Furthermore, we do not have enough information in order to determine ${:}{{\varepsilon}}_{00}[g]{*}{{\varepsilon}}_{00}[h]{:}_{_K}$, though this is $\{\pm 1\}$ by the product formula with sign ambiguity. In such a situation, we cannot use ${{\varepsilon}}_{00}[g], {{\varepsilon}}_{00}[h]$ as elements of a system with binary operations.
By these observation, it seems to be better to treat every element always together with a path from the origin, and products are defined always by path connecting. However, this is sometimes too much to treat the detail, for the object turns out to be a groupoid. We have to seek an amenable object to treat which gives informations what we want to know.
[**Strict polar element**]{}
Let $[0\approx\pi]$ be a path from $0$ to $\pi$ avoiding singular points and slits so that $e_*^{\pi\frac{1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}$ is sitting in the same sheet as in $e_*^{0\frac{1}{2i{\hbar}}\langle{\pmb u}g, {\pmb u}g\rangle_*}$. Then $e_*^{[0\approx\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}$ is determined without sign ambiguity.
$e_*^{[0\approx\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}$ is called the [**strict polar element**]{} by requesting that the path is so chosen that $e_*^{\pi\frac{1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*}$ is sitting in the same sheet as in $e_*^{0\frac{1}{2i{\hbar}}\langle{\pmb u}g, {\pmb u}g\rangle_*}$, and it will be denoted by $\hat{{\varepsilon}}_{00}$. In precise, $$\label{strictpolar}
\hat{{\varepsilon}}_{00}=
e_*^{[0\approx\pi]\frac{1}{2i{\hbar}}
\langle{\pmb u}g, {\pmb u}g\rangle_*},\quad
{:}\hat{{\varepsilon}}_{00}{:}_{_K}=
\frac{1}{\sqrt{\det(\cos([0{\approx}1]\frac{\pi}{2})I{-}(\sin([0{\approx}1]\frac{\pi}{2}){}^t\!gKg)}}
e^{-\frac{1}{i{\hbar}}\langle{\pmb u}\frac{1}{K}, {\pmb u}\rangle}$$ but a little care is required for the r.h.s., for the sheet is not distinguished by the notation itself.
Since singular points and slits are not sitting $\pi/2$-periodically but only $\pi$-periodically, the square ${\hat{{\varepsilon}}_{00}}^2=\hat{{\varepsilon}}_{00}{*}\hat{{\varepsilon}}_{00}$ is defined only with sign ambiguity (cf.). That is, ${\hat{{\varepsilon}}_{00}}^2=\pm 1$ and the sign depends on $g$ and $K$ discontinuously, while ${\hat{{\varepsilon}}_{00}}^4=1$ (cf. Proposition\[period\]).
But recall here that change of generators is covered by change of expression parameters. Hence the same phenomenon must occur in the change of expression parameters even when $g$ is fixed.
Sign-changing by the order of continuous tracing {#Adjoint}
------------------------------------------------
Recall §\[prodstr\]. We have discussed the product formula $e_*^{sH_*}{*}e_*^{tK_*}$ for the case $[H_*,K_*]=0$ in Propositions\[prodcomm\],\[singularpt\]. In this section, we consider the the case $[H_*,K_*]{\not=}0$ and we give the product formula corresponding to Propositions\[prodcomm\],\[singularpt\]. As it is mentioned before, the product $${:}e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}{*}f{:}_{_K},\quad
{:}f{*}e_*^{{-}t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}{:}_{_K}$$ are defined by the left/right evolution equations $$\frac{d}{dt}f_t=\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle{*}f_t, \quad
\frac{d}{dt}f_t=f_t{*}\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle,$$ with initial data $f$ Suppose $f$ is another $*$-exponential function ${:}e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\beta J),
{\pmb u}\rangle}{:}_{_K}$.
$${:}e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}(f{*}e_*^{{-}t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}){:}_{_K}
{\quad }
{:}(e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}
{*}f){*}e_*^{{-}t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}{:}_{_K}
\quad \text{e.t.c.}$$ are defined holomorphically, but multi-valued in $t$ on an open connected domain $D$ containing the origin $0\in{\mathbb C}$.
Even in such a case, we can fix the value by tracing along a real analytic path from $0$. We have then the following synchronized associativity:
\[synchronized\] Whenever the same path is used to fix the value in both sides, associativity $${:}e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}(f{*}e_*^{{-}t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}){:}_{_K}
={:}(e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}f){*}e_*^{{-}t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}{:}_{_K}$$ holds and differentiating this gives as in §\[adjoint\] $${:}e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}f{*}e_*^{{-}t\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}{:}_{_K}
=
{:}e^{t{\rm{ad}}(\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle)}f{:}_{_K}$$
Using Theorem\[assocthm\] and Theorem\[synchronized\], we see also the following:
\[quadexpquad\] Suppose $e_*^{s\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}$ and $e_*^{[0{\sim}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}$ are defined, where $[0{\sim}\pi]$ is a real analytic curve in ${\mathbb C}$ joining $0$ to $\pi$ avoiding singular points.
Since for every fixed $s$, $$e_*^{s\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}
{*}
e_*^{{-}s\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}$$ are defined as a multi-valued holomorphic element on an open connected neighbourhood of $[0{\sim}\pi]$, we see $${:}{\rm{Ad}}(e_*^{s\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle})
e_*^{[0{\sim}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{:}_{_{K}}{=}
{:}e_*^{[0{\sim}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}{:}_{_K}, \,\,
\tilde\beta(s){=}e^{s\alpha}\beta e^{-s\alpha}$$ hold without sign ambiguity, where $[0{\sim}t]$ in the r.h.s. is the path naturally given by the adjoint action for the path of the l.h.s.
In particular, ${:}e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{:}_{_{K}}$ and ${:}e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}{:}_{_K}$ must have the same periodicity.
Here we used the same notation as in previous section to stress that $e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s) J),{\pmb u}\rangle}$ is defined by solving the evolution equation along the real segment $[0,t]$: $$\label{adjoint020}
\frac{d}{dt}f_*(t)=
\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle{*}f_*(t),\quad
f_*(0){=}1.$$
Consider now ${\rm{Ad}}(e_*^{s\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle})
e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}$ of two variables $(s,t){\in}[0,\pi]{\times}[0,\pi]$. Note that Corollary\[quadexpquad\] holds even if there is a singular point $(s_0,t_0)$ in the open square $(0,\pi){\times}(0,\pi)$, but there happens another phenomenon of change sheets depending on the order of continuous tracing of values.
By the observation Proposition\[singularpt\] in the previous section, we see that the singular points in ${\mathbb C}^2$ forms a set $S$ of complex codimension 1, which is transversal to the real plane. One may assume that $S\cap{\mathbb R}^2$ is a discrete set. Suppose now there is a singular point $(s_0,t_0)$ in the open square $(0,\pi){\times}(0,\pi)$. Then, there must be a slit starting from $(s_0,t_0)$ going outside the square. In what follows, we see that $*$-exponential function $e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s_0)J),{\pmb u}\rangle}$ is discontinuous at $t=t_0$.
Hence fixing $t$ as $t_0{<}t{<}\pi$ and tracing $e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}$ by moving $s$ from $s{=}0$, the curve must hit the slit and changes the sheet.
As it mentioned in Proposition\[singularpt\], the sheet changing gives $$\label{2sheets}
e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta([0,s])J),{\pmb u}\rangle}
={-}e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}$$ where the l.h.s. is the element obtained by tracing continuously from $e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(0)J),{\pmb u}\rangle}$ ${=}$ $e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}$ to $e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\beta(s)J),{\pmb u}\rangle}$ under a fixed $t$.
One may understand how the sign changes by noting the difference $([0,t],s)$ and $(t,[0,s])$ in the next picture.
(300,130)(0,-10) (0,0)[(1,0)[100]{}]{} (0,0)[(0,1)[100]{}]{} (0,100)(100,100)(100,0) (45,45)[$\odot$]{} (49,47)[(1,1)[50]{}]{} (30,94)[$\bullet$]{} (30,100)[$e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}{=}
{-}e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta([0,s])J),{\pmb u}\rangle}$]{} (94,30)[$\bullet$]{} (100,30)[$e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(\sigma)J),{\pmb u}\rangle}{=}
e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta([0,\sigma])J),{\pmb u}\rangle}$]{} (85,75)[.]{} (87,66) (94,45)[(-1,2)[5]{}]{} (120,10)[$\tilde\beta(s){=}e^{s\alpha}\beta e^{-s\alpha}$]{}
Now the formal associativity Theorem\[assocthm\] gives the translation identity from the right evolution equation into the left evolution equation: $$\label{master01}
{:}e_*^{s\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}
e_*^{[0{\to}t]\langle{\pmb u}(\frac{1}{2i{\hbar}}\beta J),
{\pmb u}\rangle}{:}_{_K}
={:}e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}{*}
e_*^{s\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}{:}_{_K}.$$
On the other hand, note that $e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta([0,s])J),{\pmb u}\rangle}$ is the solution of $$\label{Adjoint111}
\frac{d}{d\eta}f_*(\eta)=
[\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle,f_*(\eta)],\quad f_*(0){=}
e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle},
\quad
{\text{$t$ is fixed.}}$$ Since this is real analytic, the formal associativity theorem gives for fixed $t$ that $$\label{master02}
{:}e_*^{[0{\to}s]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{:}_{_K}
={:}e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta([0,s])J),{\pmb u}\rangle}{*}
e_*^{[0{\to}s]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}{:}_{_K}.$$ If we use the tracing , then we have $$\begin{aligned}
{:}
e_*^{[0{\to}s]\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}
{:}_{_K}
=&-{:}e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}{*}
e_*^{[0{\to}s]\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}{:}_{_K}\\
=&-{:}e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}{*}
e_*^{s\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}{:}_{_K}
\end{aligned}$$ for $e_*^{[0{\to}s]\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}$ on the right hand side can be replaced simply by $e_*^{s\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}$ without changing meaning. It follows a little tricky result as follows:
\[master01\] If the square $[0,s]{\times}[0,t]$ contains no singular point, then the identify $${:}e_*^{[0{\to}s]\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}
{*}e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{:}_{_K}.
={:}e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}{*}
e_*^{s\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{:}_{_K}$$ holds, but if the square $[0,s]{\times}[0,t]$ contains a singular point $(s_0,t_0)$ in the interior, then $${:}e_*^{[0{\to}s]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}e_*^{t\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{:}_{_K}.
=-{:}e_*^{[0{\to}t]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\tilde\beta(s)J),{\pmb u}\rangle}{*}
e_*^{s\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{:}_{_K}$$ since the sheet is exchanged.
\[trickymaster\] Suppose $\tilde\beta({\pi}){=}
e^{{\pi}\alpha}\beta e^{-{\pi}\alpha}{=}{-}\beta$. If there is no singular point in $(0,\pi){\times}(0,\pi)$, then $${:}e_*^{[0{\to}\pi]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}e_*^{\pi\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{:}_{_K}.
={:}e_*^{[0{\to}\pi]\langle{\pmb u}
(\frac{-1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{*}
e_*^{\pi\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}{:}_{_K}$$ but if there is a singular point in $(0,\pi){\times}(0,\pi)$, then $${:}e_*^{[0{\to}\pi]\langle{\pmb u}
(\frac{1}{2i{\hbar}}\alpha J),{\pmb u}\rangle}
{*}e_*^{\pi\langle{\pmb u}
(\frac{1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{:}_{_K}.
=-{:}e_*^{[0{\to}\pi]\langle{\pmb u}
(\frac{-1}{2i{\hbar}}\beta J),{\pmb u}\rangle}{*}
e_*^{\pi\langle{\pmb u}(\frac{1}{2i{\hbar}}\alpha J),
{\pmb u}\rangle}{:}_{_K}$$
The relation such as $e^{{\pi}\alpha}\beta e^{-{\pi}\alpha}{=}{-}\beta$ appears naturally in the next section, but the relation in Corollary\[trickymaster\] is [*not*]{} a classical relation, for such a relation does not hold in the limit ${\hbar}\to 0$.
### Formula obtained by adjoint relations
In this subsection, we apply these results to the case $m=1$. First of all, we recall
\[polar\] In a generic ordered expression $K$, ${:}e_*^{\frac{t}{i{\hbar}}(au^2{+}bv^2{+}2cu{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}{:}_{_K}$ has no singular point on the real line and the pure imaginary line.
Providing $c^2{-}ab=1$, polar element ${:}e_*^{\frac{\pi i}{2i{\hbar}}(au^2{+}bv^2{+}2cu{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}{:}_{_K}$ depends only on $K$ and the path from $0$ to $\pi$.
Except otherwise stated, the path is chosen as the segment $[0{\to}\pi]$.
Note first that $\frac{1}{i{\hbar}}[u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v,
\begin{bmatrix}
u\\v
\end{bmatrix}]=
\begin{bmatrix}
1&0\\
0&{-}1
\end{bmatrix}
\begin{bmatrix}
u\\v
\end{bmatrix}$. It follows that $e^{\frac{it}{i{\hbar}}{\rm{ad}}(u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}
\begin{bmatrix}
u\\v
\end{bmatrix}=
\begin{bmatrix}
e^{it}&0\\
0&e^{-it}
\end{bmatrix}
\begin{bmatrix}
u\\v
\end{bmatrix}$, and hence for any $*$-function such as $f_*(u,v,{\hbar})=
\frac{1}{2\pi}\int_{{\mathbb R}^2}{\hat f}(s,t,{\hbar})
e_*^{\frac{1}{i{\hbar}}(su{+}tv)}dsdt$ depending real analytically on ${\hbar}$ in some interval involving ${\hbar}=0$, we have $$e_*^{\frac{is}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{*}f_*(u,v,{\hbar})
{*}e_*^{{-}\frac{is}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{=}
f_*(e^{is}u, e^{-is}v, {\hbar})$$ by the formal associativity theorem. Furthermore, we have by the same reason that $$\label{adjointuv}
e_*^{\frac{is}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{*}f_*(u,v,{\hbar}){=}
f_*(e^{is}u, e^{-is}v, {\hbar}){*}e_*^{\frac{is}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}.$$
$$\frac{1}{2i{\hbar}}[u^2{-}v^2,
\begin{bmatrix}
u\\v
\end{bmatrix}]=
\begin{bmatrix}
0&-1\\
-1&0
\end{bmatrix}
\begin{bmatrix}
u\\v
\end{bmatrix},\quad
\frac{1}{2i{\hbar}}[u^2{+}v^2,
\begin{bmatrix}
u\\v
\end{bmatrix}]=
\begin{bmatrix}
0&1\\
-1&0
\end{bmatrix}
\begin{bmatrix}
u\\v
\end{bmatrix}.$$ It follows that $$e^{\frac{it}{2i{\hbar}}{\rm{ad}}(u^2{-}v^2)}
\begin{bmatrix}
u\\v
\end{bmatrix}=
\begin{bmatrix}
\cos t&{-}i\sin t\\
{-}i\sin t&\cos t
\end{bmatrix}
\begin{bmatrix}
u\\v
\end{bmatrix},\quad
e^{\frac{t}{2i{\hbar}}{\rm{ad}}(u^2{+}v^2)}
\begin{bmatrix}
u\\v
\end{bmatrix}=
\begin{bmatrix}
\cos t&\sin t\\
{-}\sin t&\cos t
\end{bmatrix}
\begin{bmatrix}
u\\v
\end{bmatrix}.$$
Now, even if $f_*$ is a $*$-exponential function of quadratic form ${:}e_*^{t\langle{\pmb u}(\frac{1}{2i{\hbar}}\beta J),
{\pmb u}\rangle}{:}_{_K}$, we can make several commutation relations by using the product formula . But for that purpose, we have to use [*synchronized*]{} path in both sides.
[**Polar elements are splitting**]{}
We next compute the case $c{=}0$ and $\delta\delta'{\not=}\pm 1$ in i.e. $K=diag\{\delta,\delta'\}$. Then, $$\begin{aligned}
&\sqrt{\Delta^2{-}(e^t{-}e^{-t})^2\delta\delta'}\\
&{=}
\sqrt{1{-}\delta\delta'}e^{-t}
\sqrt{e^{4t}{+}2\frac{1{+}\delta\delta'}
{1{-}\delta\delta'}e^{2t}{+}1}
{=}\sqrt{1{-}\delta\delta'}e^{-t}
\sqrt{(e^{2t}{+}\beta)(e^{2t}{+}\beta^{-1})}
\end{aligned}$$ where $\beta=
\frac{1{+}\sqrt{\delta\delta'}}{1{-}\sqrt{\delta\delta'}}$. If $|\beta|\not=1$ i.e. $\delta\delta'\not\in {\mathbb R}_{<0}$ then only one of $\sqrt{e^{2t}{+}\beta}$ or $\sqrt{e^{2t}{+}\beta^{-1}}$ changes sign when $t$ moves from $0$ to $\pi i$. Thus, this is the case where the singular points are distributed $\pi i$-periodically along two lines parallel to the imaginary axis both positive and negative real parts, whose real parts are $\pm\log\big|
\frac{\sqrt{\delta\delta'}+1}
{\sqrt{\delta\delta'}-1}\big|$, and ${:}e_*^{\frac{t}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_K}$ is $\pi i$-periodic along the imaginary axis.
\[unitlike\] Suppose $
K{=}
\begin{bmatrix}
\delta&0\\
0&\delta'
\end{bmatrix}
$ such that $\delta\delta'\not=0, 1$, and $\delta\delta'$ is not a negative real. Then, ${:}e_*^{\frac{t}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_K}$ is $\pi i$-periodic along the pure imaginary axis and singular points distributed $\pi i$-periodically along two lines parallel to the imaginary axis both positive and negative real parts.
If $\delta\delta'=1$, then ${:}e_*^{\frac{t}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_K}$ is $\pi i$-periodic along the pure imaginary axis and there is no singular point.
However, if $|\beta|=1$ i.e. $\delta\delta'$ is negative real, then ${:}e_*^{t\frac{1}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_K}$, $K=diag\{\delta,\delta'\}$ has two branching singular points on the open interval $i(0,\pi)$. Even if this is the case, one may change $\delta, \delta'$ slightly so that $\delta\delta'$ avoids negative real and $1$. By this procedure, we have the same periodical nature and the pattern of singularities as above.
We next change the generator by $(u,v)
\frac{1}{\sqrt{2}}
\begin{bmatrix}
1&-1\\
1&1
\end{bmatrix}$, and the expression parameters by two different ways: $$\label{genchange}
\begin{aligned}
K_{re}&=
\frac{1}{2}
\begin{bmatrix}
1&1\\
-1&1
\end{bmatrix}
\begin{bmatrix}
\rho{-}ic'&0\\
\vspace{.3cm}
0&\rho{+}ic'
\end{bmatrix}
\begin{bmatrix}
1&-1\\
1&1
\end{bmatrix}=
\begin{bmatrix}
\rho&ic'\\
ic'&\rho
\end{bmatrix}\\
K_{im}&=
\frac{1}{2}
\begin{bmatrix}
1&1\\
-1&1
\end{bmatrix}
\begin{bmatrix}
i\rho{-}c&0\\
0&i\rho{+}c
\end{bmatrix}
\begin{bmatrix}
1&-1\\
1&1
\end{bmatrix}=
\begin{bmatrix}
i\rho&c\\
c&i\rho
\end{bmatrix}.
\end{aligned}$$ Then for both cases, we see by that $${:}e_*^{\frac{t}{i{\hbar}}(u^2{-}v^2)}{:}_{_{K_{re}}}=
{:}e_*^{\frac{t}{i{\hbar}}2u'v'}{:}_{_{\hat{K}_0}},\quad
{:}e_*^{\frac{t}{i{\hbar}}(u^2{-}v^2)}{:}_{_{K_{im}}}=
{:}e_*^{\frac{t}{i{\hbar}}2u'v'}{:}_{_{K_0'}},$$ where $u'=\frac{1}{\sqrt{2}}(u-v),\,v'=
\frac{1}{\sqrt{2}}(u+v)$, $\hat{K}_0=diag\{\rho{-}ic,\rho{+}ic\}$, $K_0'= diag\{i\rho{-}c,i\rho{+}c\}$.
Note now that $$\begin{bmatrix}
\cos r& i\sin r\\
i\sin r&\cos r
\end{bmatrix} \subset
\begin{bmatrix}
\rho& ic\\
ic&\rho
\end{bmatrix},\quad \rho, c\in{\mathbb R},$$ and $$\begin{bmatrix}
\cos r& i\sin r\\
i\sin r&\cos r
\end{bmatrix},\quad
\begin{bmatrix}
\cos s& -\sin s\\
sin s&\cos s
\end{bmatrix},\quad
\begin{bmatrix}
e^{i\theta}& 0\\
0&e^{-i\theta}
\end{bmatrix}$$ generate $SU(2)$.
By these observation, we have first the following:
\[nicenice22\] If $K_{re}{=}
\begin{bmatrix}
\rho&ic'\\
ic'&\rho
\end{bmatrix}$ with $c', \rho \in {\mathbb R}$, satisfies $|\frac{1{+}\rho{+}ic'}{1{-}\rho{-}ic'}|{\not=}1$, then $K_{re}$ ordered expressions of those three $*$-exponential functions $$e_*^{\frac{it}{i{\hbar}}2u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}, \quad
e_*^{\frac{t}{i{\hbar}}(u^2+v^2)},\quad
e_*^{\frac{it}{i{\hbar}}(u^2-v^2)},$$ have no singular point on the real axis and $\pi$-periodic, but each of them has singular points sitting $\pi$-periodically along two lines parallel to the real axis on both upper and lower half plane.
Hence, the polar element ${{\varepsilon}}_{00}$ may be written in the $K_{re}$-expression by $${:}{{\varepsilon}}_{00}{:}_{_{K_{re}}}=
{:}e_*^{\frac{\pi i}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{:}_{_{K_{re}}}=
{:}e_*^{\frac{\pi i}{2i{\hbar}}(u^2{-}v^2)}{:}_{_{K_{re}}}=
{:}e_*^{-\frac{\pi}{2i{\hbar}}(u^2+v^2)}{:}_{_{K_{re}}}.$$ and ${{\varepsilon}}_{00}^2=1$. Therefore, we have three square roots $$e_1=e_*^{\frac{\pi i}{2i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}, \quad
e_2=e_*^{\frac{\pi}{4i{\hbar}}(u^2{+}v^2)},\quad
e_3=e_*^{\frac{\pi i}{4i{\hbar}}(u^2-v^2)}$$ such that $e_i^2={{\varepsilon}}_{00}$.
To avoid possible confusion, we restrict the expression parameter in the class $K_{re}$ in what follows.
For every $s$, and Corollary\[quadexpquad\] gives in generic ordered expression $K$ that
$$\label{pmidenty}
\begin{aligned}
&{:}e_*^{\frac{\pi i}{4i{\hbar}}(u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}
{*}e_*^{[0{\sim}s]\frac{1}{i{\hbar}}(u_*^2+v_*^2)}
{*}e_*^{{-}\frac{\pi i}{4i{\hbar}}(u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}{:}_{_K}
{=}{:}e_*^{[0{\sim}s]\frac{i}{i{\hbar}}(u_*^2-v_*^2)}{:}_{_K},\\
&{:}e_*^{\frac{\pi i}{2i{\hbar}}(u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}
{*}e_*^{[0{\sim}s]\frac{1}{i{\hbar}}(u_*^2+v_*^2)}
{*}e_*^{-\frac{\pi i}{2i{\hbar}}(u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v)}{:}_{_K}
{=}{:}e_*^{[0{\sim}s]\frac{-1}{i{\hbar}}(u_*^2+v_*^2)}{:}_{_K}
\end{aligned}$$
without sign ambiguity, where $[0{\sim}s]$ in the l.h.s. is a path from $0$ to $s$ in a complex plane on which there is no singular point, and $[0{\sim}s]$ in the r.h.s. is the path given naturally by the adjoint transformation.
We have also $$\label{partI}
\begin{aligned}
&e_*^{\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{*}e_*^{\frac{i[0\sim s]}{i{\hbar}}(u_*^2-v_*^2)}{*}
e_*^{{-}\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{=}e_*^{-\frac{[0\sim s]}{i{\hbar}}(u_*^2+v_*^2)},\\
&e_*^{\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{*}
e_*^{\frac{i[0\sim s]}{i{\hbar}}(u_*^2-v_*^2)}
{=}e_*^{-\frac{[0\sim s]}{i{\hbar}}(u_*^2+v_*^2)}
{*}e_*^{\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
\end{aligned}$$
Taking the synchronized use of path, we have $$\label{partII}
\begin{aligned}
&e_*^{\frac{\pi}{4i{\hbar}}(u_*^2{+}v_*^2)}
{*}e_*^{\frac{i[0\sim s]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{=}
e_*^{-\frac{i[0\sim s]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{*}
e_*^{\frac{\pi}{4i{\hbar}}(u_*^2{+}v_*^2)},\\
&e_*^{\frac{\pi i}{4i{\hbar}}(u_*^2{-}v_*^2)}{*}
e_*^{\frac{i[0\sim s]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{=}e_*^{-\frac{i[0\sim s]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{*}e_*^{\frac{\pi i}{4i{\hbar}}(u_*^2{-}v_*^2)}.
\end{aligned}$$
In these notations, we have also $$e_*^{\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{*}e_*^{\frac{[0\sim\pi]}{4i{\hbar}}(u_*^2+v_*^2)}
{*}e_*^{{-}\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{=}e_*^{\frac{i[0\sim\pi]}{4i{\hbar}}(u_*^2-v_*^2)},$$ Applying the second equality of to the part $e_*^{\frac{[0\sim\pi]}{4i{\hbar}}(u^2+v^2)}
{*}e_*^{{-}\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}$, we have $$e_*^{\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{*}e_*^{\frac{[0\sim\pi]}{4i{\hbar}}(u^2+v^2)}
{*}e_*^{{-}\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{=}
e_*^{\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{*}e_*^{\frac{\pi i}{4i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{*}e_*^{\frac{[0\sim\pi]}{4i{\hbar}}(u^2+v^2)}
{=}e_*^{\frac{i[0\sim\pi]}{4i{\hbar}}(u^2-v^2)}$$ hold. This may be written simply by $$\label{123}
e_*^{\frac{\pi i}{2i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}
{*}e_*^{\frac{\pi}{4i{\hbar}}(u^2+v^2)}
{=}e_*^{\frac{\pi i}{4i{\hbar}}(u^2-v^2)}$$
Note also that yields a tricky result as follows:
\[rootsdouble\] The polar element $e_*^{\frac{i[0\sim \pi]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}$ satisfies the equality $$e_*^{\frac{\pi}{4i{\hbar}}(u_*^2{+}v_*^2)}
{*}e_*^{\frac{i[0\sim \pi]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{=}
e_*^{-\frac{i[0\sim\pi]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}{*}
e_*^{\frac{\pi}{4i{\hbar}}(u_*^2{+}v_*^2)}.$$ Hence, such a polar element commutes with another square root of a polar element, if and only if $e_*^{-\frac{i[0\sim\pi]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}=
e_*^{\frac{i[0\sim\pi]}{i{\hbar}}u{{\lower-0.2ex\hbox{${\scriptscriptstyle{\circ}}$}}}v}$.
Recall first . This gives $e_1{*}e_2=e_3$ in the $K_{re}$-ordered expression.
Generally, adjoint relations of quadratic forms give the following master relations for elements of square roots of the polar element.
\[master\] Let $H_*$ be a quadratic form with the discriminant $1$. Then, $e^{\pi i{\rm{ad}}(H_*)}e_j{=}e_j^{-1}$. This implies that $e_i{*}e_j{*}e_i^{-1}{=}e_j^{-1}$ by Theorem \[synchronized\]. These relations hold without sign ambiguity.
[**Proof**]{} The first equality is easy to see. The second identity is a special case of the identity which is proved by using formal associativity theorem. $\Box$
By the master relation, we have in general $$e_i{*}e_j=e_j^{-1}{*}e_i
={{\varepsilon}}_{00}{*}e_j{*}e_i.$$ By the identity $e_3=e_1{*}e_2$, we have $$e_2{*}e_3=
e_2{*}e_1{*}e_2=
e_2{*}e_2^{-1}{*}e_1=e_1.$$ Similarly, $$e_3{*}e_1=
e_3{*}e_2{*}e_3=
e_3{*}e_3^{-1}{*}e_2=e_2.$$
Note that all $e_i$ are elements of ${\mathcal E}_{2+}({\mathbb C}^{2})$. Hence, we have
\[surprise0\] In the $K_{re}$-ordered expression such that $|\frac{1{+}\rho{+}ic'}{1{-}\rho{-}ic'}|{\not=}1$, $\{{{\varepsilon}}_{00}, e_1, e_2, e_3\}$ generates an algebra ${\mathcal A}$ where exist two idempotent elements $\frac{1}{2}(1{+}{{\varepsilon}}_{00})$, $\frac{1}{2}(1{-}{{\varepsilon}}_{00})$ such that $$1=\frac{1}{2}(1{+}{{\varepsilon}}_{00}){+}\frac{1}{2}(1-{{\varepsilon}}_{00}), \quad
\frac{1}{2}(1{+}{{\varepsilon}}_{00}){*}\frac{1}{2}(1-{{\varepsilon}}_{00})=0.$$
The subalgebra $\frac{1}{2}(1{-}{{\varepsilon}}_{00}){*}{\mathcal A}$ is naturally isomorphic to the complexification ${\mathbb C}{\otimes}{\mathbb H}$ of the quaternion field $\mathbb H$ such that by denoting $\hat 1=\frac{1}{2}(1{-}{{\varepsilon}}_{00})$ $$\hat{{\varepsilon}}_{00}=\frac{1}{2}(1{-}{{\varepsilon}}_{00}){*}{{\varepsilon}}_{00}=-\hat 1, \quad
\hat{e}_i^2=-\hat 1, \quad
\hat{e}_i{*}\hat{e}_j={-}\hat 1{*}\hat{e}_j{*}\hat{e}_i,\quad 1\leq i,j\leq 3,$$ where $\hat{e}_i=\frac{1}{2}(1{-}{{\varepsilon}}_{00}){*}{e}_i$, and the subalgebra $\frac{1}{2}(1{+}{{\varepsilon}}_{00}){*}{\mathcal A}$ is the group ring over ${\mathbb C}$ of the Klein’s four group.
Independence of ordering principle and its failure
--------------------------------------------------
In differential geometry, it is widely accepted that geometrical notion should have coordinate free expression. Obviously, algebraic structure of $({\mathbb C}[{\pmb u}], {*}_{\Lambda})$ depends only on the skew part of ${\Lambda}$. It seems reasonable to accept the independence of ordering principle as a basic principle that the physical implication should be independent of ordered expressions. Theorem \[main01\] supports this principle for elements in a class ${\mathcal E}_{2}({\mathbb C}^{n})$.
By a direct calculation of intertwiner, we see that $$\label{eq:intwin}
I_{_K}^{^{K'}}(e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle})
=e^{\frac{1}{4i{\hbar}}\langle{\pmb a}(K'{-}K),{\pmb a}\rangle}
e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}.$$ Hence, $\{e^{\frac{1}{4i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle}
e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle};
K\in{\mathfrak S}_{\mathbb C}(2m)\}$ is a parallel section of $\coprod_{K\in{\mathfrak S}_{\mathbb C}(2m)}
{H{\!o}l}({\Bbb C}^{2m}).$
We denoted this collection symbolically by $e_*^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$ and we regard each member $$\label{eq:tempexp}
:e_*^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}:_{_K}
=e^{\frac{1}{4i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle}
e^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
=e^{\frac{1}{4i{\hbar}}\langle{\pmb a}K,{\pmb a}\rangle
{+}\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$$ as its $K$-expression. Furthermore for every $K$, $e_*^{\frac{s}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$ is the solution of the evolution equation $$\frac{d}{dt}{:}e_*^{\frac{s}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}}
=\frac{1}{i{\hbar}}{:}\langle{\pmb a},{\pmb u}\rangle{:}_{_{K}}{*_{_K}}
{:}e_*^{\frac{s}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{:}_{_{K}}
\,\,{\text{with initial data}}\,\, {:}1{:}_{_{K}}=1.$$ Note also that ${:}\langle{\pmb a},{\pmb u}\rangle{:}_{_{K}}
=\langle{\pmb a},{\pmb u}\rangle$. $e_*^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}=
\{e^{s^2\frac{1}{4i{\hbar}}\langle{\pmb a}K\,{\pmb a}\rangle}
e^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle};
K\in{\mathfrak S}(2m)\}$ forms a one parameter group of parallel sections. The product formula in $K$-ordered expression gives the exponential law ${:}e_*^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}}{*_{_{K}}}
{:}e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}}=
{:}e_*^{(s{+}t)\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}
{:}_{_{K}}$ for every $K{\in}{\mathfrak S}(2m)$. Hence, this may be written by omitting the suffix $K$ as $e_*^{s\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{*}
e_*^{t\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}=
e_*^{(s{+}t)\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}$. The product formula may be written as $$\label{prodformula}
e_*^{\frac{1}{i{\hbar}}\langle{\pmb a},{\pmb u}\rangle}{*}
e_*^{\frac{1}{i{\hbar}}\langle{\pmb b},{\pmb u}\rangle}
=e^{\frac{1}{2i{\hbar}}{\langle\pmb a}{J},{\pmb b}\rangle}
e_*^{\frac{1}{i{\hbar}}\langle({\pmb a}{+}{\pmb b}),
{\pmb u}\rangle}.$$
The main point is that we do not use operator theory, but instead various ordered expressions under the leading principle that a physical/mathematical object should be free from ordered expressions ([**the independence of ordering principle**]{}, (IOP) in short), just as geometrical objects are independent of local coordinate expressions.
Recall this principle in geometry forced to accept the absolute abstract notion “underlying topological space” before a collection of local coordinate system. However, we saw in [@OMMY4] that the topology of a set depends on expression parameters. That is, $${\mathfrak P}^{(2)}_{K_0}\cong SO(m,{\mathbb C}){\times}{\mathbb Z}_2,
\quad \text{but} \quad
{\mathfrak P}^{(2)}_{K_s}\cong Spin(m)\otimes{\mathbb C}.$$
Furthermore, if we apply this principle to our system, then it becomes a true nature that every linear form has two different inverses, for this holds for generic (open dense) expressions. In general, parallel sections of $\coprod_{K{\in}{\mathfrak S}(n)}{\mathcal E}_{2+}({\mathbb C}^{n})$ are multi-valued with branched singular points depending on expression parameters. It is difficult to explain multi-valued parallel section in a picture of point set topology. Thus we have to think twice about the role of expression parameters in geometry and physics.
### Philosophy of general dynamics
It was widely accepted in classical physics that every dynamical movement must be caused by some Hamiltonian $H$. Another word, this is the definition of “dynamical movement”. (IOP) is also widely accepted together with differential geometrical expressions, e.g. contact geometry, $G$-structures.
The philosophy was succeeded in non-relativistic quantum dynamics by replacing $H$ by a quantum Hamiltonian. This is given by the evolution equation of every quantum observable $f_t({\pmb u})$: $$\frac{d}{dt}f_t({\pmb u})=[H, f_t(\pmb u)]_*$$ and the solution is given by $f_t=e^{t{\rm{ad}}(H)}f_0$, where $e^{t{\rm{ad}}(H)}$ acts on the space of quantum observables.
In the relativity theory, “time” $t$ is never an absolute scalar, but a coordinate function of “space-time”. Thus, the Hamiltonian $H$ which governs a relativistic movement is given by $H=H(t,e(t,\pmb u))$ involving $t$ and the quantum canonical conjugate $e$ of $t$. $e$ is called the “energy” variable, relating each other by $e=e(t,\pmb u)$, or $t=t(e,{\pmb u})$.
The equation of the relativistic movement is written by similar differential equation by using “proper time” $\tau$ viewed as the individual time of observer: $$\frac{d}{d\tau}\phi_\tau(e,t,{\pmb u})
=[H, \phi_\tau(e,t,\pmb u)],
\quad
\frac{d}{d\tau}e{*}t=[H,e{*}t]=0,$$ where $\phi_{\tau}$ is any quantum observable, and the solution is given by $\phi_{\tau}=e^{\tau{\rm{ad}}(H)}\phi_0$, where $e^{\tau{\rm{ad}}(H)}$ acts on the space of quantum observables. If one forgets about physics by neglecting the positivity of energy, such equations can be treated as Fourier integral operators, and the principle (IOP) remains safe.
However in the field theory, quantum observables $\phi_0$ are regarded as operators acting on some pre-Hilbert space ${\mathbb H}$, and we are requested to have $e_*^{\tau H}$ acting on ${\mathbb H}$ with suitable associativity such that $\phi_{\tau}$ may be written as $$\phi_{\tau}=(e_*^{\tau H}{*}\phi_0){*}e_*^{-\tau H}
=e_*^{\tau H}{*}(\phi_0{*}e_*^{-\tau H}).$$ On the other hand, as it is seen throughout this series of papers, $*$-exponential functions such as $e_*^{\tau H}$ often has branched singular point and the periodicity depends on the expression parameters. We have delicate problems of failing associativity related to [**moving branched singular points**]{}, which depends on expression parameters. Stone’s theorem shows that there is no essential selfadjoint operator $H$ such that $\int_{\mathbb R}e^{tH}dt$ is finite.
At a first glance it is natural to replace $e_*^{\tau H}$ by ${\rm{Ad}}(e_*^{\tau H})$ and operator representation of ${\rm{Ad}}(e_*^{\tau H})$. But this can not be the mathematical Messiah, because such strange phenomena are already involved in the transcendently extended algebra of ordinary calculus. Hence phisysists are required always the mathematical consistency. Strictly speaking, this means that physical phenomenon in field theory depends on how the element is expressed. Nature of individual element, in particular the nature of periodicity depends on expression parameters.
[OM]{}
, ,
, ,
, Ann. Phys. 111, (1977), 61-151.
, Acad. Press, 1968.
, [Geometric Asymptotics]{}. A.M.S. Mathematical surveys, 14, 1977.
, ,
, AMS Trans. Mono.129, 1993.
,[*Toward geometric quantum theory*]{}, in From Geometry to Quantum Mechanics. Prog. Math. 252, Birkh[ä]{}user,2007, 213-251.
[H. Omori, Y. Maeda, N. Miyazaki, A. Yoshioka]{}, , in Conférence Moshé Flato 1999, Quantizations, Deformations, and Symmetries, Vol II, Math. Phys. Studies 22, Kluwer Academic Press, 2000, 233-246.
[H. Omori, Y. Maeda, N. Miyazaki, A. Yoshioka]{}, , Jour. Lie Theorey, Vol 13, no.2, (2003) 491-510.
, , , 483-492.
, [arXiv:1104.2109]{}
, [arXiv:1105.1218]{}
, Memoir. A.M.S. 106, 1993.
[^1]: Department of Mathematics, Faculty of Sciences and Technology, Tokyo University of Science, 2641, Noda, Chiba, 278-8510, Japan, email: [email protected]
[^2]: Department of Mathematics, Faculty of Science and Technology, Keio University, 3-14-1, Hiyoshi, Yokohama,223-8522, Japan, email: [email protected]
[^3]: Department of Mathematics, Faculty of Economics, Keio University, 4-1-1, Hiyoshi, Yokohama, 223-8521, Japan, email: [email protected]
[^4]: Department of Mathematics, Faculty of Science, Tokyo University of Science, 1-3, Kagurazaka, Tokyo, 102-8601, Japan, email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the dynamics of many charges interacting with the Maxwell field. The particles are modeled by means of non-negative distribution functions $f^+$ and $f^-$ representing two species of charged matter with positive and negative charge, respectively. If their initial velocities are small compared to the speed of light, $c$, then in lowest order, the Newtonian or classical limit, their motion is governed by the Vlasov-Poisson system. We investigate higher order corrections with an explicit control on the error terms. The Darwin order correction, order $|v/c|^2$, has been proved previously. In this contribution we obtain the dissipative corrections due to radiation damping, which are of order $|v/c|^3$ relative to the Newtonian limit. If all particles have the same charge-to-mass ratio, the dissipation would vanish at that order.'
author:
- |
[S. Bauer[^1]]{}\
Universität Duisburg-Essen, Fachbereich Mathematik,\
D-45117 Essen, Germany\
[**Key words:**]{} Vlasov-Maxwell system, post-Newtonian expansion,\
radiation reaction
title: 'Post-Newtonian Dynamics at Order 1.5 in the Vlasov-Maxwell System'
---
\[section\]
\[theorem\][Definition]{} \[theorem\][Definition and Lemma]{} \[theorem\][Lemma]{} \[theorem\][Example]{}
\[theorem\][Remark]{} \[theorem\][Remarks]{} \[theorem\][Corollary]{} \[theorem\][Proposition]{} \[theorem\][Proposition]{} \[theorem\][Assumption]{}
Introduction
============
In classical electrodynamics it is well known that accelerated charges loose energy by radiation and there is a large amount of literature concerning effective equations which include effects due to radiation damping without giving a completely relativistic description of the system of fields and charges.
A similar but more involved situation occurs in the theory of general relativity where accelerated matter emits gravitational radiation and is thus damped. The probably best studied example is the Hulse-Taylor binary pulsar consisting of a strongly self-gravitating system of two stars rotating about their common center of mass. Due to the difficulties as the non-linearity and the necessity of finding appropriate coodinates it seems to be out of reach to treat a system like the one already mentioned within the full theory. Hence, it is desirable to have effective equations valid in certain limits as in the electromagnetic case.
In many applications, as e.g. the Hulse-Taylor binary pulsar, the occurring velocities are small compared to the speed of light. Thus, it is a natural strategy to expand the metric in powers of $|v/c|$, ($c$ denotes the velocity of light). The contribution in order zero corresponds to the non-relativistic limit where gravity is governed by Newtonian theory. Therefore, higher order corrections are usually addressed as post-Newtonian approximations. For an overview concerning post-Newtonian expansions see [@bla]. Whereas it is relatively straightforward to establish relations between the full system and the equations of approximation, it is much more difficult to give the relation between the solutions of the two sets of equations. While order zero is done in [@ADR] for asymptotically flat solutions any further progress seems to be difficult at this point.
For this reason it seems to be useful to investigate the very similar but less involved system of charged matter coupled to electromagnetic fields. In [@KS1] and [@KS2] the first post-Newtonian approximations of the Abraham model, a model consisting of single charged particles coupled to the Maxwell fields which they create collectively, are considered yielding the Darwin corrections, order $|v/c|^2$, and radiation corrections, order $|v/c|^3$ with respect to the Newtonian limit. Explicit estimates of the error terms are given.
In the present paper we choose a model of many particles governed by a statistical approach. For sake of simplicity we assume that there are only two different species of matter with mass normalized to unity and charge normalized to plus unity and minus unity, respectively. These distributions of the large number of particles in phase-space are modeled through the non-negative distribution functions $f^+$ and ${f^-}$, $f^\pm=f^\pm(t, x, p)$, depending on time $t\in\R$, on position $x\in\R^3$ and momentum $p\in\R^3$. The dynamics is governed by the relativistic Vlasov-Maxwell system.
$$\label{RVMC}
\left.\begin{array}{lclcrcl}
\multicolumn{7}{c}{\partial_t
f^\pm+\hat{p}\cdot\nabla_x
f^\pm \pm(E+c^{-1}\hat{p}\times B)\cdot\nabla_p
f^\pm=0,} \\[2ex]
\;c\nabla\times E & = & -\partial_t B, & &
c\nabla\times B &
= & \partial_t E+4\pi j \\[2ex]
\quad\, \nabla\cdot E & = & 4\pi\rho, & \quad &
\nabla\cdot B
& = & 0, \\[0ex]
\qquad\quad\rho & := & \displaystyle
\int(f^+-f^-)\,dp, & & j & := & \displaystyle
\int\hat{p} (f^+-f^-)\,dp
\end{array}\right\}\tag{RVM${}_{c}$}$$
Here $$\label{hatv-def}
\hat{p}=(1+c^{-2} p^2)^{-1/2}p\in\R^3$$ is the relativistic velocity associated to $p$. The Lorentz force $E+c^{-1}\hat{p}\times B$ realizes the coupling of the Maxwell fields $E(t,
x)\in\R^3$ and $B(t, x)\in\R^3$ to the Vlasov equation, and conversely the density functions $f^\pm$ enter the field equations via the scalar charge density $\rho(t, x)$ and the current density $j(t, x)\in\R^3$, which act as source terms for the Maxwell equations. The parameter $c$ gives the speed of light for given units of time and space of the physical system represented. As usual we shall deal with the limit of small velocities by letting $c\to
\infty$. Some background on this procedure is given Section \[PN-sec\]. In order to give the Cauchy problem of (\[RVMC\]) one has to prescribe initial data for the densities and the fields at a certain time $t$, say $t=0$, $$\label{data-def}
f^\pm(0, x, p)=f^{\circ, \pm}_c(x, p), \quad
E(0, x)=E^\circ_c(x), \quad B(0,
x)=B^\circ_c(x).$$ Henceforward in our notation we will only express the initial data’s dependency upon the light velocity by the subscript $c$ while the dependency of the solution will be suppressed.
In the next section we shall describe the post-Newtonian expansions used in this paper. First we define a “naive” post-Newtonian expansion of (\[RVMC\]), see (\[exp-def\])-(\[fields-def\]); whereas this expansion is well defined for all orders in $c^{-1}$ and the solutions of the expansion equations up to order $ c^{-2}$ are good approximations of the solution of the Cauchy problem of (\[RVMC\]) if the initial data are well adapted, see [@schaeffer:86] and [@degond] for the Newtonian limit and [@bauku] and [@degondravi] for the corrections in order $c^{-2}$, it does not include damping effects due to radiation which occur in the order $c^{-3}$, see [@BKRR]. Hence, we introduce a more sophisticated expansion up to order $c^{-3}$ containing a radiation reaction term, see (\[dVPc\]) below. Assuming only one species of matter, say $f^-=0$, this term would vanish reflecting that there is no radiation in this order of $c^{-1}$ in that case. As in the case of single charges one has to circumvent the occurence of so-called run-away solutions in the resulting dynamics. Details are explained in Section \[cm-sec\]. We show that the resulting effective dynamics is well defined, at least locally in time, and we give some nice further properties of the solutions, Proposition \[f0-th\] and Lemma \[f2-le\]. Furthermore, we give an explicit expression of the fields as function of the values of the densities; in this expansion the fields are not a degree of freedom anymore but are enslaved by the densities, see formulas (\[radE-rep\]) and (\[radB-rep\]).
In Section \[comp-sec\] we state our main results concerning the comparison of solutions of our expansion and solutions of (\[RVMC\]).
The phase-space of the comparison dynamics defined in Section \[PN-sec\] consists of a pair $f^{\circ, \pm}_c$, of smooth non-negative functions with compact support defined on $\R^3\times \R^3$. The field quantities are to be computed from this densities e.g. by means of the formulas (\[radE-rep\]) and (\[radB-rep\]). On the other hand treating the Cauchy problem of (\[RVMC\]) we also have to specify the initial fields $E^\circ_c$ and $B^\circ_c $ and it is the question for which choice of initial data the dynamics of Section \[PN-sec\] is a good comparison dynamics and for which not. Surely, it is possible to choose initial data for the fields such that the densities of the two dynamics evolve in completely different ways. In Section \[comp-ad-sec\] we will adapt the initial fields for the Maxwell dynamics from the comparison dynamics, see formula (\[IC\]). That means for given data for the particle densities we compute special fields by means of the formulas (\[radE-rep\]) and (\[radB-rep\]) and impose these fields upon the initial values of the Maxwell fields. That has the advantage that, from a mathematical point of view, existence and uniqueness theorems of local in times solutions for both dynamics are at hand, see [@glstr], [@schaeffer:86], [@bau1]. We prove that the error between the solutions of the two systems is of order ${\cal
O}(c^{-4})$, see Theorem \[Hauptsatz\]. We want to mention that in [@schaeffer:86], [@degond] and [@bauku] the fields are adapted up to the relevant orders in the same way.
There are two drawbacks of this method. Post-Newtonian expansion is in essence an expansion of the relativistic velocity $\hat{p}$ and the retardet time $t-c^{-1}|x-y|$. It is clear that assuming localized sources the expansion of the retardet time is only a good approximation in the near zone of the source where $|x-y|\ll c$. This is reflected in the fact that the estimates of the fields in Theorem \[Hauptsatz\] and \[Hauptsatz2\] are only local in the space variable $x$. For this reason also the adapted initial fields are only reliable in the near zone, in particular they are not of finite energy. From a more physical point of view it also seem to be questionable to use the Cauchy problem at all. Recalling the motivation of post-Newtonian expansions one is more interested in localized systems, isolated from the rest of the world, which have already evolved for a long time with small velocities. Therefore the Cauchy problem might not be the right formulation since it is not clear how to incorporate these properties into the initial fields. In physics textbooks isolated systems are characterized by the absence of incoming radiation, that is energy coming into the system from past null infinity by means of electromagnetic fields, for a rigorous definition see [@cal1]. For given sources fields free of incoming radiation are usually calculated by means of the retardet potential. Past null infinity is that region of space-time which is reached in the direction of backward lightcones. In Section \[retRVMC-sec\] we consider a familiy of solutions, parametrized by $c$ of (\[RVMC\]), passing through $f^{\circ,
\pm}_c$ at time $t=0$ where in contrast to the Cauchy problem of (\[RVMC\]) the electromagnetic fields are computed by means of the retardet potentials alone, see (\[retRVMC\]). Because it is our goal to model slow systems we assume that the momenta are bounded uniformly in $c\ge 1$ and time $t\in\R$, see Assumption \[retRVMC-ass\](b). It is not the aim of this paper to investigate existence of such solutions with mathematical rigor, instead we will just assume their existence and some nice properties used in the sequel, see Assumption \[retRVMC-ass\]. Note however that in [@cal1] the existence of global solutions is proved for small $f^{\circ, \pm}_c$. Furhermore it is shown that such a solution is also unique in a certain class of “nearly free streaming solutions” and is free of incoming radiation. We also want to emphasize that the Larmor formula has been proved for this system, see (\[lam-for\]) and [@BKRR]. The underlying physical picture is that in the absence of incoming radiation from outside the system any solution of (\[RVMC\]) will approach a solution of (\[retRVMC\]). That means that solutions of (\[retRVMC\]) constitute a kind of initial layer.
We prove that the error between solutions of our comparsion dynamics and solutions of (\[retRVMC\]) is of order ${c^{-4}}$, see Theorem \[Hauptsatz2\].
In Section \[not-sec\] we collect some more notations used in the proofs of Theorem \[Hauptsatz\] and Theorem \[Hauptsatz2\]. The proof of Theorem \[Hauptsatz\] is elaborated in Section \[HS-bew\] while the somehow cumbersome computations of some representation formulas for the fields is outsourced to the Appendix \[append\]. The proof of Theorem \[Hauptsatz2\] is presented in Section \[main2-th\].
Post-Newtonian expansion {#PN-sec}
========================
We adopt the definition of a post-Newtonian approximation from [@KR2], see also [@ADR2] for the Einstein case. Therefore matter and fields are described by functions $(f^\pm(c), E(c), B(c))$ depending on a parameter $c\in [c_0, \infty)$ giving a one-parameter family of solutions of (\[RVMC\]). This means that $(f^\pm(c), E(c), B(c))$ describes a one-parameter family of solutions of physical systems which are represented in parameter-dependent units where the numerical value of the speed of light is given by $c$. A more conventional physical description of the post-Newtonian expansion would say that in a fixed system of units the occurring velocities are small compared to the speed of light. To be more precise, note that $(f^\pm, E, B)$ is a solution of (\[RVMC\]) with $c={\varepsilon}^{-1/2}$ if and only if $$\begin{aligned}
f^{\pm, {\varepsilon}}(t, x, p) & = & {\varepsilon}^{3/2} f^\pm(
{\varepsilon}^{3/2} t, {\varepsilon}x, {\varepsilon}^{-1/2} p),
\nonumber \\
E^{\varepsilon}(t, x) & = & {\varepsilon}^2 E({\varepsilon}^{3/2} t, {\varepsilon}x ),
\label{sca-def} \\
B^{\varepsilon}(t, x) & = & {\varepsilon}^2 B({\varepsilon}^{3/2} t, {\varepsilon}x)
\nonumber\end{aligned}$$ is a solution of (\[RVMC\]) with $c=1$. In this scaling the masses of the system remain unchanged $$\int\!\!\!\int f^{\pm, {\varepsilon}}\,dp\,dx =
\int\!\!\!\int f^\pm{\,dp}\,dx$$ while the average momenta $$\bar{p}^{\,{\varepsilon}}=\int\!\!\!\int p f^{\pm,
{\varepsilon}}\,dp\,dx=\sqrt{{\varepsilon}}\int\!\!\!\int p
f^\pm\,dp\,dx=\sqrt{{\varepsilon}}\bar{p}$$ are scaled by $\sqrt{{\varepsilon}}$. By definition of the rescaled fields these fields are slowly varying in their space and time variables. Thus, the limit $c\to \infty$ corresponds to an adiabatic limit.
In this work we treat the speed of light $c$ as a parameter and study the behavior of the system as $c\to \infty$, but note that all Theorems can also be formulated in a parameter independent fashion. In that case the value of $c$ is fixed, say $c=1$, and the initial data has to be scaled according to (\[sca-def\]), see [@bauku] for details.
We start with a formal expansion of all coefficients occurring in (\[RVMC\]) in powers of $c^{-1}$. $$\begin{aligned}
f^\pm & = &
f^\pm_0+c^{-1}f^\pm_1
+c^{-2}f^\pm_2+c^{-3}f^\pm_3+\ldots,
\nonumber \\
E & = &
E_0+c^{-1}E_1+c^{-2}E_2+c^{-3}E_3+\ldots,
\nonumber\\
B & = &
B_0+c^{-1}B_1+c^{-2}B_2+c^{-3}B_3+\ldots,
\label{exp-def}\\
\rho & = &
\rho_0+c^{-1}\rho_1+c^{-2}\rho_2
+c^{-3}\rho_3+\ldots,
\nonumber \\
j & = &
j_0+c^{-1}j_1+c^{-2}j_2+c^{-3}j_3+\ldots.
\nonumber \end{aligned}$$ Moreover, also the initial denities are assumed to be expandable $f^{\circ, \pm}_c= f^{\circ,
\pm}_0+c^{-1} f^{\circ, \pm}_1+...\,$. While considering the Cauchy problem the same is understood for the electromagnetic initial fields. Finally $\hat{p}=p-(c^{-2}/2)p^2p+\ldots$ by (\[hatv-def\]), where $p^2=|p|^2$. These expansions can be substituted into (\[RVMC\]). Comparing coefficients at every order gives a hierarchy of equations for these coefficients. The equations in order $k$ will be addressed as the $k/2$-PN equations and the solutions $(f^{\pm,
k/2PN}=\sum_{j=0}^kc^{-j}f^\pm_j, E^{k/2PN}=...,
B^{k/2PN}=...)$ as the $k/2$-PN approximation contributing to the fact that in the context of general relativity post-Newtonian approximations are usually counted in orders of $ c^{-2}$. In order zero the well known Vlasov-Poisson system of plasma physics appears. $$\label{VP}
\left.\begin{array}{rcl}
\multicolumn{3}{c}{\partial_t
f^\pm_0+p\cdot\nabla_x f^\pm_0
\pm E_0\cdot\nabla_p f^\pm_0=0,} \\[1.5ex]
E_0(t, x) &=&-\displaystyle\int
|z|^{-2}\bar{z}\,\rho_0(t, x+z)\,dz, \\[2ex]
\rho_0 & = & \displaystyle\int
(f^+_0-f^-_0)\,dp, \\[2ex]
f^\pm_0(0, x, p) & = & f^{\circ, \pm}_0(x,
p),\end{array}\right\}\tag{VP}$$ where $$\bar{z}=z{|z|^{-1}}.$$ Note that the degrees of freedom of the electromagnetic fields up to this order are lost, reflecting the fact that the limit $c\to \infty$ is singular and the hyperbolic field equations become elliptic. We recall that in [@schaeffer:86] it has been shown that as $c\to\infty$ the solutions of (\[RVMC\]) approach a solution of the Vlasov-Poisson system with the rate ${\cal O}(c^{-1})$; see [@asanouk; @degond] for similar results and [@lee] for the case of two spatial dimensions. The respective Newtonian limits of other related systems are derived in [@ADR; @calee].
Concerning a general $k$ we assume that the lower order coefficients are already computed. Then the fields in order $k$ have to solve $$\begin{array}{rclcrcl}
\nabla\times E_k & = & -\partial_t B_{k-1},
&\quad& \nabla\cdot E_k & = & 4\pi \rho_k, \\
\nabla\times B_k & = & \partial_t E_{k-1} +
4\pi j_{k-1}, &\quad & \nabla\cdot B_k & = & 0.
\end{array}$$ The corresponding Vlasov equation is given by $$\partial_t f^\pm_k+p\cdot \nabla_x f^\pm_k \pm
E_0\cdot \nabla_p f^\pm_k
= \mp E_k\cdot \nabla_p f^\pm_0+R_k$$ where $R_k$ is depending on $f^\pm_j,\;\nabla_x
f^\pm_j,\; \nabla_p f^\pm_,\; E_j,\,B_j, \;j=0,
\cdots,k-1$. If we assume that $f^{\circ, \pm}_k=0$ for all odd $k$, what we will do henceforward without mentioning, and employing the explicit form of $R_k$ it is easy to show that we consistently can set $$\label{shadow-def}
f^\pm_{2l+1}=0,\qquad E_{2l+1}=0,\qquad B_{2l}=0$$ for all $l=0, 1,2, \ldots\;$. Granted $E_k$ is known, we can easily solve for $f^\pm_k$ using characteristics. Note that for all orders $k$ the characteristic flow is given by the vectorfield $(p, \pm
E_0)$. Thus, for $k\ge 1$ $E_k$ only enters the Vlasov equation via the right hand side.
Using the vector identity $-\nabla\times \nabla
\times +\nabla\nabla\cdot=\Delta$ we can rewrite the field equations. $$\begin{aligned}
\label{fields-def}
E_{2k} & = & 4\pi\Delta^{-1} (\nabla\rho_{2k}
+\partial_t j_{2k-2}) +\Delta^{-1} (\partial_t^2
E_{2k-2})
\nonumber\\
& = & 4\pi \sum_{l=0}^k \Delta^{-1-l}
\Big[\partial_t^{2l}(\nabla \rho_{2k-2l}
+\partial_t j_{2k-2l-2})\Big]
\nonumber \\
B_{2k+1} & = & \Delta^{-1} (\partial_t^2
B_{2k-2}) -4\pi\Delta^{-1}(\nabla\times j_{2k})
\nonumber \\
& = & -4\pi\sum_{l=0}^k
\Delta^{-1-l} \Big[\partial_t^{2l} \nabla\times
j_{2k-2l}\Big], \end{aligned}$$ where quantities carrying a negative index are set to zero. For given $f^\pm_k$, thus $\rho_k$ and $j_k$ are given, and if we assume that all densities have compact support we can solve these equations using the convolution with the fundamental solution of $\Delta^{-1-l}$. Of course, without boundary conditions these solutions are not unique, and at least for higher orders these solutions will not vanish at infinity.
Nevertheless, if we choose these fields the coupled equtions are easily solved by a fix-point iteration for $E_k$. Thus, this (naive) PN approximation scheme is well defined.
According to this scheme $B_1$ is given by $$\label{B1-def}
B_1(t, x) = \int|z|^{-2}\bar{z}\times j_0(t,
x+z)\,dz$$ where $$\label{j0-def}
j_0=\int p(f^+_0-f^-_0)\,dp.$$ The couple $(f_2, E_2)$ is the solution of $$\label{LVP}
\left.\begin{array}{c}
\partial_t f_2^\pm+p\cdot\nabla_x
f^\pm_2-\frac{1}{2} p^2\,p\cdot\nabla_x
f^\pm_0
\pm E_0\cdot\nabla_p f^\pm_2\pm(E_2+p\times
B_1)\cdot\nabla_p f^\pm_0=0, \\[1ex]
E_2(t, x) = \displaystyle \frac{1}{2}
\int\bar{z}\partial_t^2\rho_0{(t, x+z)}{\,dz}-\int{|z|^{-1}}\partial_t j_0{(t, x+z)}{\,dz}-\int|z|^{-2}\bar{z}\rho_2{(t, x+z)}{\,dz}\\[2ex]
\rho_2 = \displaystyle \int(f^+_2-f^-_2)\,dp
\\[2ex]
f_2(0, x, p)=f^{\circ,\pm}_2(x,
p).\end{array}\right\}\tag{LVP}$$ In analogy to the particle model the 1-PN approximation $$\label{1pnapp-def}
f^{\pm, 1PN}=f^\pm_0+ c^{-2} f^\pm_2,\qquad
E^{1PN}=E_0+ c^{-2} E_2, \qquad B^{1PN}=c^{-1} B_1$$ is also called the Darwin approximation. It is Hamiltonian in the following sense. If the conserved energy $${\cal E}=\int\!\!\!\int \sqrt{1+ c^{-2} p^2/2}\,
(f^++f^-)\,dp\,dx +\frac{1}{8\pi}\int
(E^2+B^2)\,dx$$ of (\[RVMC\]) is expanded according to (\[exp-def\]) one obtains the Darwin energy defined by ${\cal E}_D={\cal E}_{D, kin}+{\cal
E}_{D, pot}$. The kinetic and potential energy are given by $$\begin{aligned}
{\cal E}_{D, kin} & = & \int\!\!\!\int
(p^2/2- c^{-2} p^4/8) (f^+_0+f^-_0)+
c^{-2}p^2/2(f^+_2+f^-_2)\,dp\,dx \quad\text{and}\\
{\cal E}_{D, pot} & = & \frac{1}{8\pi} \int
E_0^2 +2 c^{-2} E_0\cdot E_2+ c^{-2} B_1^2\,dx\end{aligned}$$ respectively. One can check that ${\cal E}_D$ is conserved along solutions of the 1-PN approximation. If in the Cauchy problem of (\[RVMC\]) we adapt the initial data (\[data-def\]) to suit the data of the 1PN approximation the solutions are tracked down with an error of order $c^{-3}$, see [@bauku]. Hence, the naive post-Newtonian expansion is valid up to this order.
Radiation damping in the 1.5 PN approximation
---------------------------------------------
Using the naive expansion, according to (\[shadow-def\]) and (\[fields-def\]) we would simply have to add $$\begin{aligned}
\label{B3-def}
B_3(t, x) & =
&-4\pi\Big[\Delta^{-1}\big(\nabla\times
j_2\big)+\Delta^{-2}\big(\partial_t^2\nabla\times
j_0\big) \Big](t, x)
\nonumber \\
& = & \int{|z|^{-1}}\nabla\times j_2(t, x)\,dz+\frac
1 2 \int|z|\partial_t^2\nabla\times j_0(t,
x)\,dz \end{aligned}$$ together with the factor $c^{-3}$ to the magnetic field, where $$j_2=\int
p(f^+_2-f^-_2)-\frac{p^2}{2}\,p(f^+_0-f^-_0)\,dp.$$ Therefore the relevant energy ${\cal E}_D$ has not to be changed in comparison to the 1PN order and we would remain with a Hamiltonian system. On the other hand it is known that in the full relativistic system energy is radiated to null infinity. In [@BKRR Theorem 1.4] it is shown that in the limit $c\to \infty$, corresponding to small velocities, the total amount of radiated energy is given by $$\label{lam-for}
\frac{2}{3}c^{-3}|\ddot{D}|^2,$$ where $D$ is the dipole moment of the Newtonian limit of the matter defined by $$\label{dipole-def}
D(t)=\int x\rho_0(t, x)\,dx.$$ This theorem gives a mathematical formulation and a rigorous proof of the Larmor formula in case of Vlasov matter. Hence, we should introduce a radiation reaction force causing this loss of energy. As already suggested in [@KR1; @KR2] we modify the Vlasov equation of the Newtonian distribution by incorporating a small correction into the force term, $$\label{dVPc}
\partial_t f_0^\pm+p\cdot\nabla_x f^\pm_0\pm(E_0
+\frac{2}{3c^3}\dddot{D})\cdot\nabla_p
f^\pm_0 =0.$$ The additional term is the generalization of the radiation reaction force used in particle models, see [@jac formula (16.8)]. In passing we note that for this system the “energy” $$\label{SEne-def}
{\cal E}_{S}= \frac{1}{2} \int\!\!\!\int
p^2(f^+_0+f^-_0)(t, x, p)\,dp\,dx
+\frac{1}{8\pi}\int |E_0(t,
x)|^2\,dx-\frac{2}{3}c^{-3}\dot{D}\cdot\ddot{D}$$ is decreasing, more precisely one obtains $$\label{Ene-dec}
\frac{d}{dt}{\cal E}_S = -\frac{2}{3c^3}
|\ddot{D}(t)|^2,$$ the subscript $S$ referring to the name “Schott”-energy under which this energy can be found in the literature. This decreasing of energy can be attributed to the effect of radiation damping. If we remain with the positiv definite energy of the Vlasov-Poisson system $$\label{energyVP-def}
{\cal E}_{VP}=\frac{1}{2}\int\!\!\!\!\int
p^2(f^+_0+f^-_0)\,dp\,dx +\frac{1}{8\pi}\int
|E_0(t, x)|^2\,dx$$ evaluated along solutions of (\[dVPc\]) the “friction” has a definite sign only in the time average $$\nonumber
\frac{d}{dt}{\cal E}_{VP} = \frac{2}{3} c^{-3}
\big(\ddot{D}\cdot\dddot{D}
-|\ddot{D}|^2\big).$$
### “Unphysical” solutions and the “Reduced Radiating Vlasov-Poisson system” {#cm-sec}
Introducing system (\[dVPc\]) one immediately runs into the problem that an initial datum has to be supplied for $\ddot{D}$ (note that $D(0)$ and $\dot{D}(0)$ are already determined by $f^{\circ\pm}_0$) and there is no obvious way to extract this information from the approximation scheme. This phenomenon is also known in the theory of accelerated, and thus radiating, single charges and leads to the unphysical so-called run-away solutions. In [@KS2] it has been observed that in the particle model this problem has the structure of a singular geometric perturbation problem, and the “physical” dynamics is obtained on a center-like manifold of the full dynamics. In order to adopt this language to the model under consideration here we assume that we are supplied with a (local in time) classical solution $(f^\pm_0, E_0)$ of (\[dVPc\]) and assume that the support of $f^\pm_0(t, \cdot, \cdot)$ remains compact for all $t$ in the interval of existence of the solution. We define the bare mass of the charges by $$M=\int\!\!\!\int(f^+_0+f_0^-)(t, x, p)\,dp\,dx.$$ Mass conservation and charge conservation for both species easily follow from (\[dVPc\]) and integration by parts, $$\partial_t M=0, \qquad \partial_t
\rho_0^\pm+\nabla\cdot j_0^\pm=0,$$ where $j^\pm_0=\int p f^\pm_0\,dp$ and $\rho_0^\pm=\int f^\pm_0\,dp$. We denote the additional degrees of freedom by $y:=\ddot{D}$ and compute with $$\label{eta-def}
\eta:=(2/3)c^{-3} M$$ and exploiting (\[dVPc\]) in combination with integration by parts twice $$\begin{aligned}
y & = D^{[2]}+\eta \dddot{D} \nonumber
\intertext{where $D^{[2]}$ is defined by}
\label{D2-def}
D^{[2]}(t)& =\int\!\!\!\int E_0(t, x)
(f^+_0+f^-_0)(t, x, p)\,dp\,dx. \end{aligned}$$ Thus, we may rewrite (\[dVPc\]) in a form clearly showing the structure of a singular perturbation problem. $$\label{SGPP}
\left.
\begin{array}{lcr}
\dot{f^\pm_0} & = & F^\pm(f^\pm_0, y) \\
\eta\dot{y} & = & G(f^\pm_0, y)
\end{array}\right\}\tag{SGPP${}_\eta$}$$ where $F^\pm$ and $G$ are defined by $$\begin{aligned}
F^\pm(f^\pm_0, y)&=-p\cdot\nabla_x
f^\pm_0\mp(E_0+M^{-1}(y-D^{[2]}))\cdot \nabla_p
f^\pm_0\\
G(f^\pm_0, y) & = y-D^{[2]}.\end{aligned}$$
In contrast to [@KS2] we are dealing with a phase-space of infinite dimension. Thus, the proof of the existence of invariant manifolds is hard. We shall return to that question in a forthcoming paper. For the moment we shall take the existence of a smooth invariant manifold for granted and assume that it is given by means of a smooth function $h_\eta=h_\eta(f^{\circ}_0)$, acting on $C^\infty_0(\R^3\times\R^3)\times
C^\infty_0(\R^3\times\R^3)$ and taking values in $\R^3$. In this subsection $f^\circ_0$ denotes the couple $(f^{\circ, +}_0, f^{\circ,-}_0)$. The same convention will be used for $f_0=(f^+_0,
f^-_0)$ and $F=(F^+, F^-)$. The manifold ${\cal
M_\eta}=\{(f^{\circ}_0, h_\eta(f^{\circ}_0)\}$ is invariant under the flow of (\[SGPP\]) if the solution of (\[SGPP\]) subject to the initial conditions $(f^{\circ}_0,y(0))= (f^{\circ}_0,
h_\eta(f^{\circ}_0))$ satisfies $$y(t)=h_\eta(f_0(t, \cdot, \cdot))$$ for all times the solution exits.
We want to establish a dynamics of Vlasov-Poisson type which is a good approximation of the dynamics on the manifold. For this reason we assume that we can expand $h_\eta$ in $\eta$ about 0, $h_\eta=h_0+\eta h_1+{\cal O}(\eta^2)$. Inserting $\eta=0$ we have $G(f_0, h_0(f_0, 0))=0$ which implies $$\label{h0-rep}
h_0=D^{[2]}.$$ Exploiting this information we find by a formal calculation $$\begin{aligned}
\eta\dot{y} & = G(f_0, h_0(f_0)+\eta
h_1(f_0)+{\cal O}(\eta^2))\\
& = \eta \partial_y G(f_0,
h_0(f_0))h_1(f_0)+{\cal O}(\eta^2)
\intertext{and on the other hand}
\eta\dot{y} & =\eta <h_0^\prime(f_0),
\dot{f_0}>+{\cal O}(\eta^2) \end{aligned}$$ which yields $$h_1(f_0)=\Big[\partial_y G(f_0,
h_0(f_0))\Big]^{-1}<h_0^\prime, F(f_0,
h_0(f_0))>.$$ Here ${}^\prime$ denotes the Frechét derivative and $<\cdot, \cdot>$ is the duality pairing, $$<h_0^\prime(f_0), F(f_0, h_0(f_0))> =
h_0^\prime(f_0)\cdot\int\!\!\!\int
\Big(-p\cdot\nabla_x({ f^+_0}-{ f^-_0})-E_0\cdot
\nabla_p({ f^+_0}+{ f^-_0})\Big)(\cdot,
x, p)\,dp\,dx.$$ In view of the definition of $G$ and (\[h0-rep\]) we have $$y = D^{[2]}+\eta <{D^{[2]}}^\prime,
F(f_0,D^{[2]})>+{\cal O}(\eta^2).$$ After some straightforward computations we find $$<{D^{[2]}}^\prime, F(f_0,D^{[2]})>={D^{[3]}}$$ where $$\label{D3-def}
D^{[3]}(t)=2\int H^+(t, x)j_0^-(t, x)-H^-(t,
x)j^+_0(t, x)\,dx$$ and $$\label{Fpm-def}
H^\pm(t, x):=\oint {|z|^{-3}}\big(-3\bar{z}\otimes\bar{z}+{\rm
id}\big)\rho_0^\pm(t, x+z)\,dz \in \R^{3\times
3}.$$
Note that $$\label{H-def}
H(z)=-3\bar{z}\otimes\bar{z}+{\rm id}$$ is bounded on $\R^3\setminus\{0\}$ homogeneous of degree zero and satisfies $\int_{|z|=1}H(z){\,ds(z)}=0$.
Alternatively ${D^{[3]}}$ may be found by means of the following formal calculations, using (\[dVPc\]). $$\begin{aligned}
\nonumber
\dddot{D}(t) &=& \partial_t^2\int j_0(t,
x)\,dx+{\cal O}(c^{-3})
\\ \nonumber
& = & \dot{D^{[2]}}+{\cal O}(c^{-3})
\\ \nonumber
& = & D^{[3]}(t)+{\cal O}(c^{-3}). \end{aligned}$$ Whereas this second derivation is more simple the first derivation reveals the import connection to the dynamics on the manifold.
We introduce the “reduced radiating Vlasov-Poisson system” $$\label{rrVP}
\left.\begin{array}{c}\partial_t
f^\pm_0+p\cdot\nabla_x f^\pm_0
\pm (E_0+2/3 c^{-3} D^{[3]})\cdot\nabla_p
f^\pm_0=0, \\[1ex]
E_0(t, x)=-\displaystyle\int
|z|^{-2}\bar{z}\,\rho_0(t, x+z)\,dz, \\[2ex]
\rho_0 = \displaystyle \int (f^+-f^-_0)\,dp
\\[2ex]
f^\pm_0(0, x, p)=f^{\circ, \pm}_0(x,
p),\end{array}\right\}\tag{rrVP${}_c$}$$ where $D^{[3]}$ is defined according to (\[D3-def\]) and (\[Fpm-def\]).
Our first proposition addresses the existence and uniqueness of local classical solutions of (\[rrVP\]). Furthermore, it provides us with some estimates useful in the sequel of this paper. In every (even) order $k$ of $c^{-1}$ we shall only consider smooth compactly supported initial data, $$\begin{gathered}
\label{dat-def}
f^{\circ, \pm}_k\in C^\infty_0(\R^3\times \R^3),
\quad f^{\circ, \pm}_k\ge 0
\\ \nonumber
f^{\circ, \pm}_k(x,
p)=0\quad\text{for}\quad|x|\ge
r_0\quad\text{or}\quad
|p|\ge r_0\quad||f^{\circ\pm}_k||_{W^{4,
\infty}}\le S_0 \end{gathered}$$ with some $r_0>0$ and $S_0$ fixed. We call the constants $r_0$ and $S_0$ ’basic’ as all bounds related to $(f^\pm_0, E_0)$ occurring in the sequel will only depend on $r_0$ and $S_0$. The following proposition is proved in [@bau1].
\[f0-th\] If $f^{\circ,\pm}_0$ satisfies (\[dat-def\]) then there exists a constant $0< \tilde{T}\le
\infty$ such that the following holds for $c\ge
1$.
1. \[locex\] There is a unique classical, i.e. $C^1$ solution $(f^\pm_0, E_0)$ of (\[rrVP\]) existing on a time interval $[0, T_c)$ with $\tilde{T}\le T_c\le\infty$.
2. \[sup-th\] For every $T<\tilde{T}$ there is a constant $M_1(T)>0$ such that for all $0\le t\le T$ $$f^\pm_0(t,x, p)=0\quad\text{if}\quad|x|
\ge M_1(T)\quad{or}\quad|p|\ge M_1(T).$$
3. \[est-th\] In fact, $f^\pm_0$ is $C^\infty$ and for every $T<\tilde{T}$ there is constant $M_2(T)$, such that for all $0\le t\le T$ $$|\partial^\alpha f^\pm_0(t, x, p)|
+|\partial_t^\beta E_0(t,
x)|+|\partial_t^\gamma{D^{[3]}}(t)| \le M_2(T)$$ for all $x\in\R^3$, $p\in\R^3$, $|\alpha|\le 4$ and $\beta\le 2$ and $\gamma\le 1$
4. \[dtD2D3-th\] For this solution $(f_0^\pm, E_0)$ we have $$\dot{D}^{[2]}=D^{[3]},$$ where $D^{[2]}$ is defined according to (\[D2-def\]).
Note that the constants $\tilde{T}, M_1(T),
M_2(T)$ appearing in Proposition \[f0-th\] do only depend on $r_0$ and $S_0$, in particular they are independent of $c$.
As the second moment $\int\!\!\!\int p^2(f^+_0+f^-_0)
\,dp\,dx$ cannot be bounded a priori by using energy conservation it seems difficult to prove global existence of classical solutions of (\[rrVP\]). Note that both methods yielding global existence of Vlasov-Poisson type systems essentially rely on such an a priori bound, see [@pfa] or [@schae2] and [@lipe].
We shall use solutions of (\[rrVP\]) instead of solutions of (\[VP\]) in order zero of our post-Newtonian approximation and define $B_1, (f^\pm_2, E_2)$ and $B_3$ according to (\[B1-def\]), (\[LVP\]) and (\[B3-def\]) respectively. It is important to note that $(f_0,
E_0)$ is the solution of (\[rrVP\]). In particular this solution is depending on $c$ and thus all other quantities defined here are depending on $c$. Concerning the solvability of (\[LVP\]) we have the following lemma whose proof is sketched in [@bau1].
\[f2-le\] Let $f^{\circ, \pm}_k$, $k=0, 2$ satisfy (\[dat-def\]). Suppose that $(f^\pm_0, E_0)$ is the solution of (\[rrVP\]). Compute $B_1$ according to (\[B1-def\]). Then (\[LVP\]) has a unique classical solution $(f_2^\pm, E_2)$ existing on $[0, T_c)$ and enjoying the following properties.
1. \[sup-le\] For every $T<\tilde{T}$ there is a constant $M_3(T)>0$, such that for all $0< t\le T$ $$f^\pm_2(t, x, p)=0\quad{if}\quad|x|\ge
M_3(T) \quad\text{or}\quad|p|\ge M_3(T).$$
2. \[est-le\] In fact, $f^\pm_2$ is $C^\infty$ and there is a constant $M_4(T)$, such that for all $0\le
t\le T$ $$|\partial^\alpha f^\pm_2(t, x, p)|\le M_4(T)
\quad\text{for all}\quad x\in\R^3,
p\in\R^3\quad\text{and}\quad|\alpha|\le 2$$
In the following section we want to show that $$\begin{aligned}
\label{rad-def}
\nonumber
f^\pm_R &=& f^\pm_0+c^{-2}f^\pm_2
\\
E^R &=& E_0+c^{-2}E_2+2/3 c^{-3}D^{[3]}
\\ \nonumber
B^R &=& c^{-1}B_1+c^{-3} B_3 \end{aligned}$$ yields a higher order pointwise approximation of (\[RVMC\]) than the Vlasov-Poisson or the Darwin system defined in [@bauku]. We call (\[rad-def\]) the radiation approximation. In the terminology of post-Newtonian approximations it is the 1.5PN approximation. Employing the Vlasov equation and integration by parts it is not difficult to prove the following formulas.
The fields $E^R$ and $B^R$ can be written as
$$\begin{aligned}
\label{radE-rep}
\nonumber
E^R(t, x) & = & -\int|z|^{-2}\bar{z}(\rho_0+
c^{-2}\rho_2)(t, x+z)\,dz
\\ \nonumber
& & +
c^{-2}\frac{1}{2}\int\!\!\!\!\int|z|^{-2}
\bigg\{3(\bar{z}\cdot
p)^2\bar{z}-p^2\bar{z}\bigg\} (f^+_0-f^-_0)
(t, x+z, p)\,dp\,dz
\\ \nonumber
& & -
c^{-2}\int\!\!\!\!\int{|z|^{-1}}\bigg\{\bar{z}\otimes\bar{z}+1
\bigg\}
\nonumber \\
& & \hspace{4em}\Big(E_0(t,
x+z) +c^{-3}2/3 D^{[3]}(t)\Big) (f^+_0
+f^-_0)(t, x+z, p)\,dp\,dz
\nonumber \\
& & +c^{-3}\frac{2}{3}{D^{[3]}}(t)
\end{aligned}$$
and $$\begin{aligned}
\label{radB-rep}
B^R(t, x) & = & c^{-1}\int\!\!\!\!\int
|z|^{-2} \bar{z}\wedge
p(f^+_R-f^-_R)(t, x+z, p)\,dp\,dz
\\ \nonumber
& & -c^{-3}\frac{3}{2}\int\!\!\!\!\int
|z|^{-2}(\bar{z}\cdot p)^2\bar{z}\wedge p
(f^+_0-f^-_0)(t, x+z, p)\,dp\,dz
\\ \nonumber
& & +\frac{c^{-3}}{2}\int\!\!\!\!\int
{|z|^{-1}}\bigg\{(\bar{z}\wedge p)\otimes
\bar{z}+(\bar{z}\cdot p)\bar{z}
\wedge(\cdots)\bigg\}
\\ \nonumber
& & \hspace{7em}\Big(E_0(t,
x+z) +c^{-3}2/3 D^{[3]}(t)\Big) (f^+_0
+f^-_0)(t, x+z, p)\,dp\,dz
\\
& & -c^{-3}\int\bar{z}\wedge \big(H^+(t,
x+z)j^-_0(t, x+z)-H^-(t, x+z)j^+_0(t,
x+z)\big)\,dz.
\nonumber
\end{aligned}$$
Closing this section we want to mention that there is another variant of a damped Vlasov Poisson type system investigated in [@KR1] and [@KR2]. While for that system a global solution theory is at hand, the author did not succeed in comparing approximations based on solutions of that system on the one hand and solutions of the full system on the other hand.
Comparison of the 1.5 PN dynamics with the Vlasov-Maxwell dynamics {#comp-sec}
===================================================================
The Vlasov-Maxwell dynamics with adapted initial values {#comp-ad-sec}
-------------------------------------------------------
For achieving the improved approximation property we match the initial data of (\[RVMC\]) by the data for radiation system. For prescribed initial densities $f^{\circ, \pm}_k$, $k=0,2$ we are able to calculate $(f^\pm_0, E_0)$, $B_1$, $(f^\pm_2, E_2)$ and $B_3$ according to what has been outlined in Section \[PN-sec\]. We then consider the Cauchy problem of (\[RVMC\]) where the initial values are given by $$\label{IC}\tag{IC}
\left.\begin{array}{lcl}
f^\pm(0, x, p) & = & f^{\circ, \pm}_c(x, p)
= f^{\circ, \pm}_0(x, p)+ c^{-2} f^{\circ, \pm}_2
(x, p)+c^{-4}f^{\circ, \pm}_{c, free} , \\[1ex]
E(0, x) & = & E^\circ_c(x):=E_0(0, x)
+c^{-2}E_2(0, x)+\frac{2}{3}c^{-3}D^{[3]}(0)
+c^{-4}E^\circ_{c, free} \\[1ex]
B(0, x) & = & B^\circ_c(x):=c^{-1}B_1(0, x)+
c^{-3}B_3(0, x)+c^{-4}B^\circ_{c, free} .
\end{array}\right\}$$ In contrast to the contributions in the orders 0 to 3, which are fixed by the values of the approximations, $(f^{\circ, \pm}_{c, free}, E^\circ_{c, free},
B^\circ_{c, free})$ are to be chosen freely only subject to the constraints $\nabla\cdot E^\circ_{c, free}=4\pi
\int (f^{\circ, +}_{c, free}-f^{\circ, -}_{c,
free})\,dp$ and $\nabla\cdot B^\circ_{c,
free}=0$. Note that the constraint equations in the lower orders are satisfied by fiat. Furthermore, we shall assume that the following bounds hold uniformly in $c$. $$\begin{gathered}
\nonumber
f^{\circ, \pm}_{c, free}\in C^\infty(\R^3\times\R^3),
\quad E^\circ_{c, free}, B^\circ_{c, free}
\in C_0^\infty(\R^3)
\nonumber \\
f^{\circ, \pm}_{c, free} = 0\quad\text{if}\quad|x|
\ge r_0\quad\text{or }\quad|p|\ge r_0
\label{freedat-def} \\
||f^{\circ, \pm}_{c, free}||_{L^\infty} \le S_0
\nonumber \\
||E^\circ_{c, free}||_{W^{1, \infty}}+
||B^\circ_{c, free}||_{W^{1, \infty}} \le S_0.
\nonumber\end{gathered}$$ Before we formulate our theorem, let us recall that solutions of (\[RVMC\]) with initial data (\[IC\]) exist at least on some time interval $[0, \hat{T})$, which is independent of $c\geq 1$; see [@schaeffer:86 Thm. 1].
\[schaeffer-prop\] Assume that $f^{\circ, \pm}_k, k=0, 2$ satisfies (\[dat-def\]). If $f^{\circ, \pm}_c$, $E^\circ_c$ and $B^\circ_c$ are defined according to (\[IC\]), then there exits $0<\hat{T}\leq\infty$ (independent of $c$) such that for all $c\geq 1$ the system (\[RVMC\]) with initial data (\[IC\]) has a unique $C^1$-solution $(f, E, B)$ on the time interval $[0, \hat{T})$. In addition, for every $0<T<\hat{T}$ there are constants $M_5(T), M_6(T)$ (independent of $c$) such that
$$\begin{gathered}
f^\pm(t, x, p) = 0\quad\mbox{if}\quad|x|\ge
M_5(T)
\quad\text{or}\quad |p|\ge M_5(T),
\label{SchrankeSupport} \\
|f^\pm(t, x, p)|+ |E(t, x)|+|B(t, x)|\le M_6(T),
\label{SchrankeFelder}\end{gathered}$$
for all $x,\,p\in\R^3$, $t\in [0, T]$, and $c\geq
1$.
Actually, in [@schaeffer:86 Thm. 1] $E^\circ_c$ and $B^\circ_c$ do not depend on $c$ but an inspection of the proof shows that the assertions remain valid for initial fields defined by (\[IC\]). After these preparations we can state the first of our main results.
\[Hauptsatz\] Assume that $f^{\circ,\pm}_k, k=0,2$ satisfies (\[dat-def\]). From $f^{\circ, \pm}_k$ calculate $(f_0^\pm, E_0)$, $B_1$, $(f_2^\pm, E_2)$ and $B_3$, choose initial data $(f^{\circ, \pm}_c,
E^\circ_c, B^\circ_c)$ for (\[RVMC\]) according to (\[IC\]) and (\[freedat-def\]). Let $(f,
E, B)$ denote the solution of (\[RVMC\]) with initial data (\[IC\]) and let $(f^\pm_R, E^R,
B^R)$ be defined as in (\[rad-def\]). Then for every $T<\min\{\tilde{T}, \hat{T}\}$ and $R>0$ there are constants $M(T)>0$ and $M(T, R)>0$, such that $$\begin{aligned}
|f^\pm(t, x, p)-f^\pm_R(t, x, p)| & \le &
M(T)c^{-4}\quad\hspace{0.55em} (x\in\R^3),
\nonumber \\
|E(t, x)-E^R(t, x)| & \le & M(T,
R)\,c^{-4}\quad (|x|\le R), \label{diff-esti}
\\
|B(t, x)-B^R(t, x)| & \le & M(T,
R)\,c^{-4}\quad (|x|\le R), \nonumber \end{aligned}$$ for all $p\in\R^3$, $t\in [0, T]$, and $c\geq 1$.
The constants $M(T)$ and $M(T, R)$ are independent of $c\geq 1$, but do depend on the basic constants $r_0, S_0$. Note that if (\[RVMC\]) is compared to the Vlasov-Poisson system (\[VP\]) only, one obtains the estimate $$\begin{aligned}
|f(t, x, p)-f_0(t, x, p)| +|E(t, x)-E_0(t,
x)|+|B(t, x)| & \le M(T)c^{-1},
\intertext{see \cite[Thm.~2B]{schaeffer:86}, and
if compared to the Darwin system the
estimates}
|f(t, x, p)-f^D(t, x, p)|+|B(t, x)-B^D(t, x)|
&\le M(T)c^{-3}\\
|E(t, x)-E^D(t, x)| &\le M(T, R)c^{-3}, \end{aligned}$$ see [@bauku Thm. 1.1]. On first glance it could seem that the time interval $[0, T]\subset [0, \min\{\tilde{T},
\hat{T}\})$, which might be very small, is a strong limitation of the theorem. But if the theorem is formulated in an ${\varepsilon}$-depending fashion using fixed units as indicated in Section \[PN-sec\] and elaborated in [@bauku] the approximation is valid on the time interval $[0, {\varepsilon}^{-3/2}T]$ and thus for long times on that time scale.
The retardet Vlasov-Maxwell dynamics {#retRVMC-sec}
------------------------------------
Following [@cal1] we introduce the retardet relativistic Vlasov-Maxwell system. $$\label{retRVMC}
\left. \begin{array}{c}
\partial_t f^\pm+\hat{p}\cdot\nabla_x f^\pm
\pm(E+c^{-1}\hat{p}\times B)\cdot\nabla_p
f^\pm=0,
\\[1ex]
{\displaystyle}E(t, x) =
-\int\frac{dy}{|x-y|}\big(\nabla\rho+
c^{-2}\partial_t j\big)(t-c^{-1}|x-y|, y)
\\[1ex]
{\displaystyle}B(t, x) =
c^{-1}\int\frac{dy}{|x-y|}\nabla\times
j(t-c^{-1}|x-y|, y)
\\[1ex]
\displaystyle\rho = \int (f^+-f^-)\,dp, \qquad
j = \int\hat{p} (f^+-f^-)\,dp
\end{array}\right\}\tag{${}_{{\rm
ret}}$RVM${}_{c}$}$$ If we assume that $f^\pm$ is a global $C^1$ solution of (\[retRVMC\]) and that also $E$ and $B$ are $C^1$, then, by means of the Vlasov equation, $\rho$ and $j$ satisfy the continuity equation $$\label{cont-eq}
\partial_t\rho+j=0,$$ and therefore the retardet fields are a solution of Maxwell’s equations. Thus, $(f, E, B)$ also solves (\[RVMC\]). Note that it is necessary to know the densities for all time $(-\infty, t]$ in order to compute the fields at time $t$. Hence, there is no sense in the notation of a local solution of this system. As in the case of the Cauchy problem every solution of (\[retRVMC\]) satisfies the identity $$\label{fpmdarst}
f^\pm(t, x, p) = f^{\pm} (0, X^\pm(0;t,x,p),
P^\pm(0;t,x,p)),$$ where $s\mapsto (X^\pm(s; t, x, p), P^\pm(s; t, x,
p))$ solves the characteristic system $$\label{olan}
\dot x = \hat p,\quad\dot p =
\pm(E+c^{-1}\hat{p}\times B),$$ with data $X^\pm(t; t, x, p)=x$ and $P^\pm(t; t,
x, p)=p$. For this reason $$\label{finf-est}
0 \leq f^\pm(t,x,p)
\leq {\|f^{\pm}(0, \cdot, \cdot)\|}_\infty.$$ In order to derive our results on PN-approximations let us assume that we are furnished with a one-parameter familiy of solutions (\[retRVMC\]) satisfying certain plausible a priori bounds and smoothness conditions. We consider solutions of (\[retRVMC\]) passing through a certain density configuration at time $t=0$ $$\label{retfini-def}
f^{\circ, \pm}_c=f^{\circ,\pm}_0+ c^{-2}
f^{\circ,\pm}_2+ c^{-4} f^{\circ, \pm}_{c, free}$$ where $f^{\circ,\pm}_k, k=0, 2$ and $f^{\circ,
\pm}_{c, free}$ satisfy (\[dat-def\]) and (\[freedat-def\]) respectively with some constants $r_0$ and $S_0$.
\[retRVMC-ass\]
- For each $c\ge 1$ there is a solution $f^\pm\in C^4(\R\times\R^3\times\R^3)$ of (\[retRVMC\]), passing through $f^{\circ, \pm}_c$ at time $t=0$, i.e. $$f^\pm(0, x, p)=f^{\circ, \pm}_c(x, p), \quad
x\in\R^3,\,p\in\R^3.$$
- There exists a constant $P_1>0$ such that $f^\pm(t, x, p)=0$ for $|p|\ge P_1$ and $c\ge 1$. In particular, $f^\pm(t, x, p)=0$ for $|x|\ge r_0+P_1|t|$ by (\[olan\]).
- For every $T>0$, $R>0$, and $P>0$ there is a constant $M_7(T, R, P)>0$ such that $$|\partial_t^{\alpha+1} f^\pm(t, x, p)|
+|\partial_t^\alpha\nabla_x f^\pm(t, x, p)|\le
M_7(T, R, P)$$ for $|t|\le T$, $|x|\le R$, $|p|\le P$, and $\alpha=0,...,3$, uniformly in $c\geq 1$.
After these preparations we can state our second main result.
\[Hauptsatz2\]
Assume that $f^{\circ,\pm}_k, k=0,2$ satisfies (\[dat-def\]). From $f^{\circ, \pm}_k$ calculate $(f_0^\pm, E_0)$, $B_1$, $(f_2^\pm, E_2)$, $B_3$ and ${D^{[3]}}$. Define $(f^\pm_R, E^R, B^R)$ according to (\[rad-def\]). Assume that $(f^\pm, E, B)$ is a family of solutions of (\[retRVMC\]) satisfying Assumption \[retRVMC-ass\] with constants $P_1$ and $M_7(T, R, P)$. Take $\Tilde{T}$ from Proposition \[f0-th\]. Then for every $T<\tilde{T}$ and $R>0$ there are constants $M(T)$ and $M(T, R)$ such that $$\begin{aligned}
|f^\pm(t, x, p)-f^\pm_R(t, x, p)| & \le & M(T)
c^{-4}\quad\hspace{0.55em} (x\in\R^3),
\nonumber \\
|E(t, x)-E^R(t, x)| & \le & M(T, R)\,c^{-4}
\quad (|x|\le R), \label{diff-esti2} \\
|B(t, x)-B^R(t, x)| & \le & M(T, R)\,c^{-4}
\quad (|x|\le R), \nonumber\end{aligned}$$ for all $p\in\R^3$, $t\in [0, T]$, and $c\geq
2P_1$.
The constants $M(T)$ and $M(T, R)$ do only depend on $r_0, S_0, P_1$ and $M_7(\cdot , \cdot,
\cdot)$. In particular they are independent of $c\ge P_1$.
Comparison with the particle model
----------------------------------
We shall compare our results for the Vlasov model with the results for the particle model governed by the Abraham system obtained in [@KS1] and [@KS2]. Both systems share the features of Hamiltonian approximations up to 1-PN order and dissipative corrections in the 1.5-PN approximation leading to an increase of the phase-space. The right comparison dynamics is given on a center-like manifold, where the dynamics on this manifold can be approximated by a modified Vlasov-Poisson equation and a second order equation, respectively. In [@KR2 section 3] it is shown that the force terms of the 1.5PN approximation used here do agree with the infinte particle limit of the comparison dynamics used in [@KS2].
In contrast to the PN approximation in this paper in [@KS1] and [@KS2] only the forces are expanded but the main difference is in the treatment of the initial data. For the full particle model the initial data for the fields are supposed to be of “charged soliton” type. One can think of these fields as generated by charges forced to move freely for $-\infty<t\le
0$ with their initial velocity. For the approximation this leads to an initial time slip $t_0$ which the charges need to “forget” their initial data. The initial data of the approximation is fixed by matching the data of the full system at time $t_0$. Therefore the initial data for the approximation are given only implicitly, first one has to compute a solution of the full system over a time span $t_0$. Regarding the Cauchy problem of Vlasov-Maxwell system we do the matching the other way round. For a given initial density one computes the fields of the approximations and imposes their values at $t=0$ as initial data on the fields of the full system. Therefore these initial data are given more explictly. Even more it is possible to calculate them by the values of $f^{\circ, \pm}_0$ and $f^{\circ, \pm}_2$ alone, see (\[radE-rep\]) and (\[radB-rep\]).
Moreover both results in this paper (Theorem \[Hauptsatz\] and Theorem \[Hauptsatz2\]) seem to be stronger than the results obtained for the particle model. In [@KS2] the error bounds of the 1.5-PN approximation in contrast to the error bounds of the 1-PN-approximation are only improved in a certain direction, see [@KS2 formulas (3.21) and (3.32)]. It seems reasonable that a matching of the initial conditions at time $t=0$ according to the treatment of the initial conditions used here might improve those bounds.
Notation {#not-sec}
--------
In the remaining more technical sections of this paper we shall make use of the following notations. $B(0, R)$ denotes the closed ball in $\R^3$ with center at $x=0$ or $p=0$ and radius $R>0$. We write $$g(x, v, t, c)={\cal O}_{cpt}(c^{-m}),$$ if for all $R>0$ and $T>0$ there is a constant $M_R>0$ only depending on the basic constants $r_0, S_0$ and, while dealing with solutions of (\[retRVMC\]), $P_1$ and $M_7(\cdot , \cdot,
\cdot)$ such that $$\label{GOForm}
|g(x, v, t, c)|\leq M_Rc^{-m}$$ for $|x|\le R$, $p\in\R^3$, $t\in [0, T]$, and $c\geq 1$. Similarly, we write $$g(x, v, t, c)={\cal O}(c^{-m}),$$ if for all $T>0$ there is a constant $M>0$ only depending on the basic constants such that (\[GOForm\]) holds for all $x, p\in\R^3$, $t\in
[0, T]$, and $c\geq 1$. In short, ${\cal O}$ is global in $x, p$ and $c$ while ${\cal O}_{cpt}$ is local in $x$ global in $P$ and $c$, and both symbols are local in $t$. In general, generic constants only depending on the basic constants are denoted by $M$. Furthermore, in the following sections we shall use the notation $$f=f^+-f^-,\quad f^\circ_k=f^{\circ,
+}_k-f^{\circ,-}_k,\quad f^{\circ}_c=f^{\circ,
+}_c-f^{\circ, -}_c\text{and}\quad
f_l=f^+_l-f^-_l$$ for $k=0, 2$ and $l=0, 2, R$. (Recall that $f^\circ_0$ and $f_0$ were used in a different way in Subsection \[cm-sec\].)
Proof of Theorem \[Hauptsatz\] {#HS-bew}
==============================
Estimating $E-E^R$ {#diffE-subsec}
------------------
In Section \[repapp-sec\] below we will show that the approximate electric field $E^R$ from (\[rad-def\]) admits the following representation.
$$\label{ER-rep}
E^R=E^R_{ext} +E^R_{int} +E^R_{bound}+{\cal
O}_{cpt}(c^{-4})$$
with $$\begin{aligned}
\label{ERext-rep}
E^R_{ext}(t, x) & = & \int_{|z|\ge ct}
\bigg\{-|z|^{-2}\bar{z}
\big(\rho_0+c^{-2}\rho_2\big)
-c^{-2}|z|^{-1}\partial_t j_0
\\ \nonumber
&& +1/2c^{-2}\bar{z}\partial_t^2\rho_0 +2/3
c^{-3}\partial_t\big[E_0(\rho_0^
++\rho^-_0)\big]\bigg\}(t, x+z)\,dz
\\ \label{ERint-rep}
E^R_{int}(t, x) & = & \int_{|z|\le ct}
\int|z|^{-2} K_1(\bar{z}, c^{-1}p)
f_R(\hat{t}(z), x+z, p)\,dp\,dz
\\ \nonumber
&& +c^{-2}\int_{|z|\le ct} \int|z|^{-1}
K_2(\bar{z}, c^{-1}p) E_0(\hat{t}(z), x+z)
(f^+_R+f^-_R)(\hat{t}(z), x+z, p)\,dp\,dz
\\ \label{ERbound-rep}
E^R_{bound}(t, x) & = & (ct)^{-1} \int_{|z|=ct}
\int K_3(\bar{z}, c^{-1}p)
(f^\circ_0+c^{-2}f^\circ_2)(x+z, p) \,dp\,ds(z)
\\ \nonumber
&& -1/3 tc^{-2}\int_{|z|=ct} \bar{z}
\bar{z}\cdot\partial_t^2 j_0(0, x+z)\,ds(z)
\\ \nonumber
&& +2/3c^{-3} \int_{|z|=ct} \int(\bar{z}\cdot
p)\partial_t j_0(0,x+z, p)\,dp\,ds(z)\end{aligned}$$
where the subscripts ‘ext’, ‘int’ and ‘bound’ refer to the exterior, interior and boundary integration in $z$. The kernels are given by
$$\begin{aligned}
K_1(\bar{z}, \tilde{p}) & = & -\bar{z} +2\bar{z}
(\bar{z}\cdot\tilde{p}) -\tilde{p} -3\bar{z}
(\bar{z}\cdot\tilde{p})^2
+2(\bar{z}\cdot\tilde{p})\tilde{p}
+\bar{z}\tilde{p}^2 +4\bar{z}
(\bar{z}\cdot\tilde{p})^3
\nonumber \\
&& -3\tilde{p}(\bar{z}\cdot\tilde{p})^2
-3\bar{z}(\bar{z}\cdot \tilde{p})\tilde{p}^2
+3/2\tilde{p}\tilde{p}^2
\label{K1-def}\\
K_2(\bar{z}, \tilde{p}) & = &
\bar{z}\otimes\bar{z}-1
-2\bar{z}\otimes\bar{z}(\bar{z}\cdot \tilde{p})
+\bar{z}\cdot\tilde{p} +\tilde{p}\otimes\bar{z}
\label{K2-def} \\
K_3(\bar{z}, \tilde{p}) & = &
\bar{z}(\bar{z}\cdot\tilde{p})
+\tilde{p}(\bar{z}\cdot \tilde{p})
-\bar{z}(\bar{z}\cdot\tilde{p})^2
+\bar{z}(\bar{z}\cdot\tilde{p})^3
-\tilde{p}(\bar{z}\cdot\tilde{p})^2
-1/2\bar{z}(\bar{z}\cdot\tilde{p})\tilde{p}^2.
\nonumber \end{aligned}$$
We also recall that $\bar{z}=z|z|^{-1}$ and $\hat{t}(z)=t-c^{-1}|z|$. On the other hand, according to Section \[repmax-sec\] below, we have
$$\label{E-rep}
E=E_{ext} +E_{int}+E_{bound} +{\cal O}(c^{-4})$$
with $$\begin{aligned}
\label{Eext-rep}
E_{ext}(t, x) & = & \int_{|z|\ge ct} \bigg\{
-|z|^{-2}\bar{z}
\Big[(1+t\partial_t)(\rho_0+c^{-2}\rho_2)
+(1/2t^2\partial_t^2+1/6t^3\partial_t^3)\rho_0
\Big]
\\[1.5ex] \nonumber
& & +1/2c^{-2}\bar{z}
(\partial_t^2+t\partial_t^3)\rho_0
-c^{-2}|z|^{-1}(\partial_t+t\partial_t)j_0
\\[1ex] \nonumber
& & +2/3c^{-3}\partial_t[E_0(\rho^+_0
+\rho^-_0)] \bigg\}(0,
x+z)\,dz
\\
E_{int}(t, x) & = &\int_{|z|\le ct}
\int|z|^{-2}K_1(\bar{z}, c^{-1}p)f(\hat{t}(z),
x+z, p)\,dp\,dz
\\ \nonumber
&& +\int_{|z|\le ct}\int|z|^{-1}
K_2(\bar{z},c^{-1}p)E(\hat{t}(z), x+z)
(f^++f^-)(\hat{t}(z), x+z, p)\,dp\,dz
\label{Eint-rep} \\
E_{bound}(t, x) & = & E^R_{bound}(t, x).
\label{Ebound-rep}\end{aligned}$$
In order to verify (\[diff-esti\]), we fix constants $R>0$ and $0<T<\min\{\tilde{T},
\hat{T}\}$. For $x\in B(0, R)$ and $t\le T$ we start by comparing the exterior fields. We obtain from (\[Eext-rep\]) and (\[ERext-rep\]), due to $|\bar{z}|=1$, and taking into account Theorem \[f0-th\](b)-(c) and Lemma \[f2-le\](a)-(b) $$\begin{aligned}
\label{ext-diff}
\lefteqn{|E^R_{{ ext}}(t, x)-E_{{ext}}(t, x)|}
\nonumber
\\ & \le & \int_{|z|>ct}
\bigg\{|z|^{-2}|\rho_0(t,
x+z)-\big(1+t\partial_t
+1/2\,t^2\partial_t^2+1/6\,t^3\partial_t^3\big){\rho_0}{(0, x+z)}|\nonumber
\\ & & +c^{-2}|z|^{-2}|\rho_2(t,
x+z)-(1+t\partial_t)\rho_2(0, x+z)| \nonumber
\\ & & +c^{-2}|z|^{-1}\int |p|\,|\partial_t
f_0(t, x+z, p)
-(1+t\partial_t)\partial_t f_0(0, x+z, p)|\,dp
\nonumber
\\ & & +1/2\,c^{-2}|\partial_t^2 \rho_0(t, x+z)
-(1+t\partial_t)\partial_t^2 {\rho_0}(0, x+z)|
\nonumber
\\ & &
+2/3\,c^{-3}|\partial_t[{E_0}({\rho_0^+}+{\rho^-_0})]{(t, x+z)}-\partial_t[{E_0}({\rho_0^+}+{\rho^-_0})]{(0, x+z)}\bigg\}\,dz
\nonumber
\\ & \le & M M_2(T) M_1^3(T)\int_{|z|>ct}
|z|^{-2}t^4
{\bf 1}_{B(0,M_1(T))}(x+z)\,dz \nonumber
\\ & & +c^{-2}M M_4(T) M_3^3(T)\int_{|z|>ct}
|z|^{-2}t^2
{\bf 1}_{B(0, M_3(T))}(x+z)\,dz \nonumber
\\ & & + c^{-2} M M_2(T) M^4_1(T)
\int_{|z|>ct}|z|^{-1}t^2
{\bf 1}_{B(0, M_1(T))}(x+z)\,dz \nonumber
\\ & & + c^{-2} M M_2(T) M^3_1(T)
\int_{|z|>ct}t^2
{\bf 1}_{B(0, M_1(T))}(x+z)\,dz \nonumber
\\ & & +c^{-3} M M_2^2(T)
M_1^3(T)\int_{|z|>ct}t
{\bf 1}_{B(0, M_1(T))}(x+z)\,dz \nonumber
\\ & \le & M\int_{|z|>ct} \big(t^4|z|^{-2}+t^2
c^{-2}(|z|^{-2}+{|z|^{-1}}+1)+t c^{-3}\big) {\bf
1}_{B(0, R+M_0)}(z)\,dz
\nonumber
\\ & \le & M_R\,c^{-4};\end{aligned}$$ where $M_0=\max\{M_1(T), M_3(T), M_5(T)\}$ and we used that for instance $$t^4\int_{|z|>ct} |z|^{-2}{\bf 1}_{B(0,
R+M_0)}(x+z)\,dz
\le (ct)^{-4}t^4\int_{|z|\le R+M_0}
|z|^2\,dz\le M_R\,c^{-4}.$$
To bound $|E_{{\rm int}}(t, x)-E^R_{{\rm int}}(t,
x)|$, we first claim that $$\label{ma6}
|E(t, x)-E_0(t, x)|={\cal O}(c^{-2}).$$ which can be proved analogously to [@bauku Theorem 1.1]. (This estimate holds uniformly in $x\in\R^3$.) Next we define $$Q^\pm(t)=\sup\,\{|f^\pm(s, x, p)-f^\pm_R(s, x,
p)|: x\in\R^3, p\in\R^3, s\in [0, t]\}.$$ and $Q(t)=Q^+(t)+Q^-(t)$. Recall that $f^\pm(s, x,
p)=f^\pm_0(s, x, p)=f^\pm_2(s, x, p)=0$ if $|x|\ge M_0$ or $|p|\ge M_0$. Therefore the kernels $K_i$, $i+1,2,3$ are bounded on the domain of the densities by $M(1+M_0^3)$ uniformly in $c\ge 1$. From (\[Eint-rep\]), (\[ERint-rep\]), (\[ma6\]), (\[IC\]), and $0\le\hat{t}(z)\le t$ for $|z|\le ct$ we obtain $$\begin{aligned}
\label{int-diff}
\lefteqn{|E_{{\rm int}}(t, x)-E^R_{{\rm
int}}(t, x)|}
\nonumber \\
& \le & M(1+M_0)^3\int_{|z|\leq
ct}\int\big[\big( |z|^{-2}
+c^{-2}|z|^{-1}|E_0|
\big)\big(|f^+-f^+_R|+|f^--f^-_R|\big)\big)
\nonumber \\
&&+ c^{-2}{|z|^{-1}}|E-E_0|(|f^+|+|f^-|)
\big]{(\hat{t}(z), x+z, p)}{\,dp}{\,dz}\nonumber \\
& \le & M (1+M_0^3)M_0^3(1+M_2(T))\,
Q(t)\int_{|z|\leq ct} (|z|^{-2}+{|z|^{-1}})
\,{\bf 1}_{B(0, M_0)}(x+z)\,dz
\nonumber \\
& & +{c^{-4}}M (1+M_0^3)M_0^3M_6(T)\int_{|z|\leq
ct} {|z|^{-1}}\,{\bf 1}_{B(0, M_0)}(x+z)\,dz
\nonumber \\
& \le & M_R(c^{-4}+Q(t)),\end{aligned}$$ since for instance $$\int_{|z|\leq ct} |z|^{-2}\,{\bf 1}_{B(0,
M_0)}(x+z)\,dz \le\int_{|z|\leq R+M_0}
|z|^{-2}\,dz\le M_R.$$ Recalling that the $E_{ bound}(t, x)=E^R_{
bound}(t, x)$, we can summarize (\[E-rep\]), (\[ER-rep\]), (\[ext-diff\]), and (\[int-diff\]) as
$$\label{abschE}
|E(t, x)-E^R(t, x)|\le M_R(c^{-4}+Q(t)),$$
for $|x|\le R$ and $t\in [0, T]$. Using (\[BR-rep\]) and (\[B-rep\]) below and similar arguments as for the electric fields yield $$\label{abschB}
|B(t, x)-B^R(t, x)|\le M_R(c^{-4}+Q(t)),$$
for $|x|\le R$ and $t\in [0, T]$.
Estimating $f^\pm-f^\pm_R$ {#difff-subsec}
--------------------------
It remains to estimate $q^\pm=f^\pm-f^\pm_R$. Using (\[RVMC\]), (\[rad-def\]), (\[rrVP\]), and (\[LVP\]), it is found that $$\begin{aligned}
\lefteqn{\partial_t q^\pm +\hat{p}\cdot\nabla_x
\,q^\pm \pm(E+c^{-1}\hat{p}\times
B)\cdot\nabla_p\, q^\pm}\hspace{2em}
\\ & = & -\partial_t
f^\pm_R-\hat{p}\cdot\nabla_x f^\pm_R
\mp(E+c^{-1}\hat{p}\times B) \cdot\nabla_p
f^\pm_R
\\ & = & (p -1/2\,c^{-2}p^2\,p -\hat{p})
\cdot\nabla_x f^\pm_0
+c^{-2}(p-\hat{p}) \cdot \nabla_x f^\pm_2
\\ & & \pm((E^R-E)+c^{-1}\hat{p} \times(B^R-B))
\cdot\nabla_p f^\pm_R
\\ & &\pm c^{-1}(p-\hat{p})\times B^R\cdot
\nabla_p f^\pm_0\pm c^{-4} p\times B_3\cdot
\nabla_p f^\pm_0
\\ & & \mp c^{-2}(E^R-E_0)\cdot\nabla_p f^\pm_2
\mp c^{-3}\hat{p}\times B^R\cdot\nabla_p
f^\pm_2. \end{aligned}$$ If $|p|\le M_0$, then also $|\hat{p}|=(1+c^{-2}
p^2)^{-1/2}|p|\le |p|\le M_0$ uniformly in $c$, and hence $$\big|\hat{p}-\big(1-1/2\,c^{-2}p^2\big)p\big|\le
Mc^{-4}.$$ Next we note the straightforward estimates $|B^R(t, x)|\le M c^{-1}$, $|E^R(t, x)-E_0(t,
x)|\le M c^{-2}$ and $|B_3(t, x)|\le M$ for $t\in
[0, T]$. In view of the bounds in Theorem \[f0-th\](c) and Lemma \[f2-le\](c), thus by (\[abschE\]) and (\[abschB\]), $$\begin{aligned}
\label{h-est}
& & |\partial_t q^\pm(t, x, p)
+\hat{p}\cdot\nabla_x q^\pm(t, x, p) +(E(t, x)
+c^{-1}\hat{p}\times B(t, x)) \cdot\nabla_p
\,q^\pm(t, x, p)|
\nonumber \\ & & \qquad\qquad\le M_{M_0}(c^{-4}
+Q(t))=M(c^{-4} +Q(t))\end{aligned}$$ for $|x|\le M_0$, $|p|\le M_0$, and $t\in [0,
T]$. But in $\{(t, x, p): |x|>M_0\} \cup\{(t, x,
p): |p|>M_0\}$ we have $q^\pm=f^\pm-f^\pm_R=0$ by the above definition of $M_0>0$. Accordingly, (\[h-est\]) is satisfied for all $x\in\R^3$, $p\in\R^3$, and $t\in [0, T]$. Now, for any $x\in\R^3$, $p\in\R^3$ and $t\in [0, T]$ we compute using (\[fpmdarst\]) and (\[olan\]) as well as (\[h-est\]) $$\Big|\frac{d}{ds}q^\pm(s, {{ X}^\pm}, {{P}^\pm})\Big|
=|\partial_t q^\pm +\hat{p}\cdot\nabla_x
q^\pm\pm(E +c^{-1}\hat{p}\times
B)\cdot\nabla_p q^\pm||_{(s, {{ X}^\pm}, {{P}^\pm})}\le M c^{-4}+
MQ(s),$$ for $s\in[0, T]$. Here the characteristics are evaluated at $(s; t, x, p)$. Note that $$q^\pm(s, {{ X}^\pm}, {{P}^\pm})|_{s=0}=(f^\pm -f^\pm_R)(0, {{ X}^\pm},
{{P}^\pm})|_{s=0} = c^{-4} (f^{\circ,+}_{c, free}
-f^{\circ,-}_{c, free})({{ X}^\pm}, {{P}^\pm})|_{s=0}.$$ Thus, $$\begin{aligned}
|q^\pm(t, x, p)| & = & |q^\pm(s, ({{ X}^\pm}, {{P}^\pm}) (s;t,
x, p))||_{s=t}
\\
& \le & c^{-4}| f^\pm_{c, free}(0, ({{ X}^\pm}, {{P}^\pm})
(0;t, x, p))|+\int_0^t\Big|\frac{d}{ds}q^\pm(s,
({{ X}^\pm}, {{P}^\pm})(s; t, x, p))\Big|\,ds \\
&\le & M c^{-4}+M\int_0^t Q(s)\,ds.\end{aligned}$$ But by the definition of $Q(t)$ it follows that $$Q(t)\le M c^{-4}+M\int_0^t Q(s)\,ds$$ for $t\in[0, T]$. Then Gronwall’s inequality implies $Q(t)\le M c^{-4}$ for $t\in[0,
T]$. Inserting this into (\[abschE\]) and (\[abschB\]) yields the assertion of Theorem \[Hauptsatz\] and completes the proof. [$\Box$]{}
Proof of Theorem \[Hauptsatz2\] {#main2-th}
===============================
We recall the following representation of the electric field $E$ from [@cal1]. $$\begin{aligned}
\label{retE-rep}
E(t, x) & = & \int\!\!\!\! \int|z|^{-2}
K_T(\bar{z}, \hat{p})\, f(\ast, p) \,dp\,dz
\\ \nonumber
& & + c^{-2}\int\!\!\!\! \int{|z|^{-1}}K_S(\bar{z},
p)(E+c^{-1}\hat{p}\wedge B)(\ast)(f^++f^-)(\ast,
p)\,dp\,dz \end{aligned}$$ where the integrals are to be extended over $\R^3$, $K_T$ and $K_S$ are given by (\[KT-def\]) and (\[KS-def\]), respectively and $$(\ast) = (t-c^{-1}|z|, x+z).$$ Now we fix some constants $T<\tilde{T}$ and $R>0.$ Furthermore, we assume that $c\ge 2P_1$ as well as $|x|\leq R$ and $|t|\leq T$. Define
$$\label{pstern-def}
p^\ast=p^\ast(T)=\max\{P_1, M_1(T), M_3(T)\}$$
and $$\label{rstern-def}
r^\ast=r^\ast(T, R)= \max\{2(r_0+P_1T)+R,
M_1(T), M_3(T)\}$$
Then $|x+z|\ge r^\ast$ implies $$r_0+P_1|t-c^{-1}|z||=r_0+P_1|t-c^{-1}|x+z-x||
\le r_0+P_1T+\frac{1}{2}R+\frac{1}{2}|x+z|\leq
|x+z|.$$ Therefore, $|p|\ge p^\ast $ or $|x+z|\ge r^\ast$ yields $f^\pm(\ast, p)=0$ by Assumption \[retRVMC-ass\](b). This argument shows, that in (\[retE-rep\]) we can replace $\int\!\!\int\,dp\,dz$ by $\int_{|x+z|\le
r^\ast}\int_{|p|\le p^\ast}\,dp\,dz$. In other words, in these integrals we may always assume that both $|z|$ and $|p|$ are bounded by a bound depending on $r_0, P_1, R$ and $T$. Since the $p$-domain is bounded we have
$$\begin{aligned}
K_T(\bar{z}, \hat{p}) & = & K_1(\bar{z},
c^{-1} p)+{\cal O}( c^{-4})
\label{KT-exp} \\
K_S(\bar{z}, p) & = & K_2(\bar{z}, c^{-1}
p)+{\cal O}( c^{-2})
\label{KS-exp}
\end{aligned}$$
where $K_1$ and $K_2$ are defined in (\[K1-def\]) and (\[K2-def\]), respectively. Furthermore, $$\label{pd-exp}
c^{-1} \hat{p} = c^{-1}
p-c^{-3}\frac{p^2}{2}p+{\cal O}(c^{-5}).$$ Firstly, we have using Assumption \[retRVMC-ass\](b),(c) $$\begin{aligned}
\label{retB-bound}
|B(t, x)| & = & c^{-1}\int_{|x+z|\le r^\ast}
{|z|^{-1}}|\nabla\times j(\ast)|\,dz\\ \nonumber
& \leq & c^{-1} M_7(T+R+r^\ast, r^\ast,
p^\ast)\int_{|x+z|\le r^\ast}
|z|^{-1}\,dz\int_{|p|\le p^\ast}\,dp = {\cal
O}_{cpt}(c^{-1}). \end{aligned}$$ In the same waywe are able to bound the electric field $$\begin{aligned}
\label{retE-bound}
|\partial_t^l E(t, x)| & = & \int_{|x+z|\le
r^\ast} |z|^{-1}\Big|\partial^l_t\nabla\rho
+c^{-2} \partial_t^{l+1} j\Big|(\ast) \,dz
\\ \nonumber
& \le & M_7(T+R+r^\ast, r^\ast,
p^\ast)\int_{|x+z|\le r^\ast}
|z|^{-1}\,dz\int_{|p|\le p^\ast}\,dp = {\cal
O}_{cpt}(1)\end{aligned}$$ for $l=0,...,3.$ Hence, using (\[retB-bound\]) $$c^{-2}\int\!\!\int{|z|^{-1}}K_S(\bar{z}, p)
c^{-1} \hat{p}\wedge B(\ast)(f^++f^-)(\ast,
p)\,dp\,dz = {\cal O}_{cpt}( c^{-4}).$$ For this reason and using (\[retE-bound\]) we conclude for the second term in (\[retE-rep\]) $$\begin{aligned}
\lefteqn{ c^{-2}\int\!\!\!\!\int|z|^{-1}
K_S(\bar{z}, p)(E+c^{-1}\hat{p}\wedge
B)(\ast)(f^++f^-)(\ast, p)\,dp\,dz} \\
& = & c^{-2}\int\!\!\!\!\int|z|^{-1} K_S(\bar{z},
p)E\,(\ast)(f^++f^-)(\ast, p)\,dp\,dz +{\cal
O}_{cpt}( c^{-4})
= {\cal O}_{cpt}( c^{-2}).\end{aligned}$$ Secondly we shall expand the retardet time. For every smooth function $\psi$ we have $$\psi(t-c^{-1}|z|) =
\psi(t)-c^{-1}|z|\partial_t\psi(t) +
c^{-2}\frac{|z|^2}{2}\partial_t^2\psi(t)
-c^{-3}\frac{|z|^3}{6}\partial_t^3\psi(t) +
c^{-4}\frac{|z|^4}{24}\partial_t^4\psi(\xi)$$ with some $t-c^{-1}|z|<\xi<t$. Using Assumption \[retRVMC-ass\](b),(c) and (\[retE-bound\]) we have $$\label{retET-exp}
\int_{|x+z|\le r^\ast}|z|^{-2+l} \int_{|p|\le
p^\ast} K_T(\bar{z}, \hat{p})
\partial_t^l f(\xi(t, x, z, p), x+z, p) \,dp\,dz
= {\cal O}_{cpt}(1)$$ for $l=0, \ldots, 4$ as well as $$\label{retES-exp}
\int_{|x+z|\le r^\ast}|z|^{-1+l}\int_{|p|\le
p^\ast}
K_L(\bar{z},
p)\partial_t^l\big[E(f^++f^-)\big](\xi(t, x, z,
p), x+z, p)\,dp\,dz = {\cal O}_{cpt}(1)$$
for $l=0, 1, 2$ and any choice of $t-c^{-1}|z|\le
\xi(t, x, z, p)\le t$ Now from (\[KT-exp\])-(\[retES-exp\]) we can expand the electric field in powers of $c^{-1}$. To order zero we have $$\begin{aligned}
\label{retE0}
E(t, x) & = & -\int_{|x+z|\le
r^\ast}|z|^{-2}\bar{z} \rho(t, x+z)\,dz +
{\cal O}_{cpt}(c^{-1})
\nonumber \\
& = & -\int|z|^{-2}\bar{z} \rho(t, x+z)\,dz +
{\cal O}_{cpt}(c^{-1}).\end{aligned}$$ Here we employed that $f^\pm(t, x+z, p)=0$ if $|x+z|\ge r^\dagger$ with $$\label{rdagger-def}
r^\dagger= r_0+P_1T,$$ see Assumption \[retRVMC-ass\](b) and use $r^\ast>r^\dagger$, see (\[rstern-def\]). In the first order we obtain two terms, the first coming from the expansion of the kernel $K_T$ and the second from the expansion of the retardet time, $$c^{-1}\int\!\!\!\!\int|z|^{-2}(2\bar{z}(\bar{z}\cdot
p)-p)f(t, x+z, p)\,dp\,dz$$ and $$c^{-1}\int|z|^{-1}\bar{z}\partial_t\rho(t, x+z)\,dz.$$ Here and in the sequel we shall treat terms containig time derivatives in the following way. Firstly, we replace $\partial_t f^\pm$ by $-\hat{p}\cdot\nabla_x f^\pm
\mp(E+c^{-1}\hat{p}\wedge B)\cdot \nabla_p f^\pm$ using the Vlasov equation. Secondly we do an integration by parts utilizing $$(E+c^{-1}\hat{p}\wedge B)\cdot \nabla_p f^\pm
=\nabla_p\cdot
\big[(E+c^{-1}\hat{p}\wedge B)f^\pm\big].$$ For the second term we therefore obtain $$-c^{-1}
\int\!\!\!\!\int|z|^{-2}
(2\bar{z}(\bar{z}\cdot\hat{p})-\hat{p})f(t, x+z,
p)\,dp\,dz.$$ Hence, using (\[pd-exp\]), the contribution in first order vanishes and in (\[retE0\]) we can replace ${\cal O}_{cpt}(c^{-1})$ by ${\cal
O}_{cpt}( c^{-2})$, but we have an additional term occuring in the third order, namely $$\label{3o-add}
c^{-3}\int\!\!\!\!\int|z|^{-2}
p^2(\bar{z}(\bar{z}\cdot p)-\frac{1}{2}p)\,f(t,
x+z, p) \,dp\,dz.$$ In the next orders we will simply replace $\hat{p}$ by $p$ according to (\[pd-exp\]) without comment. For the second order we start with the terms with two time derivatives;
$$\begin{aligned}
\label{secotwoder}
\lefteqn{-
c^{-2}\frac{1}{2}\int\bar{z}\partial_t^2
\rho(t, x+z)dz}
\\ \nonumber
& = & - c^{-2}\frac{1}{2}\int\!\!\!\!\int
|z|^{-1}(-\bar{z}(\bar{z}\cdot p) +p)
\partial_t f(t, x+z, p) \,dp\,dz +{\cal
O}(c^{-4}).
\end{aligned}$$
Now we collect and rewrite the terms with one time derivative, including the term coming from (\[secotwoder\]), $$\begin{aligned}
\label{secooneder}
\lefteqn{- c^{-2}\int\!\!\!\! \int|z|^{-1}
\bigg(2\bar{z} (\bar{z}\cdot p)-p+\frac{1}{2}
(-\bar{z}(\bar{z}\cdot p) +p)\bigg)\partial_t
f(t, x+z, p) \,dp\,dz}
\nonumber \\
& = & -\int\!\!\!\! \int|z|^{-2}
\bigg(-\frac{9}{2}\bar{z} (\bar{z}\cdot p)^2
+2 (\bar{z}\cdot p)\,p
+\frac{3}{2}\bar{z}\,p^2\bigg) f(t, x+z, p)
\,dp\,dz
\label{secozeroder}\\
& &
-\int\!\!\!\! \int|z|^{-1}
\bigg(\frac{3}{2}\bar{z}\otimes\bar{z}
-\frac{1}{2}\bigg)
E(t, x+z)(f^+ +f^-)(t, x+z, p) \,dp\,dz + {\cal
O}( c^{-4}).
\nonumber\end{aligned}$$ Together with the remaining terms coming from the expansion of $K_T$ we therefore obtain $$\begin{aligned}
\label{retf2}
E(t, x) & = &
\int\!\!\!\!\int|z|^{-2}\bigg\{-\bar{z}+
c^{-2}\Big(\frac{3}{2}\bar{z}(\bar{z}\cdot
p)^2-\frac{1}{2}\bar{z} p^2\Big)\bigg\}f(t,
x+z, p)\,dp\,dz
\\ \nonumber
& & -
c^{-2}\frac{1}{2}\int\!\!\!\!\int|z|^{-1}
\bigg\{\bar{z}\otimes\bar{z}+1\bigg\}E(t, x+z)
(f^+ +f^-)(t, x+z, p)\,dp\,dz + {\cal
O}_{cpt}(c^{-3}). \end{aligned}$$
At last we turn to the third order and, following the usual route, first treat the term with three time derivatives,
$$\begin{aligned}
\label{thirdotthreeder}
c^{-3}\frac{1}{6} \int z\partial_t^3 \rho(t,
x+z)\,dz & = &
c^{-3}\frac{1}{6} \int\!\!\!\!\int
p\partial_t^2 f(t, x+z, p)\,dp\,dz +{\cal
O}(c^{-5}).
\end{aligned}$$
Regarding terms containing second time derivatives including the term coming from (\[thirdotthreeder\]) we have $$\begin{aligned}
\label{thirdotwoder}
c^{-3} \lefteqn{\int\!\!\!\!
\int\bigg\{\bar{z}(\bar{z}\cdot
p)-\frac{1}{2}p +\frac{1}{6}p\bigg\}
\partial_t^2 f(t, x+z, p)\,dp\,dz}
\nonumber \\ \nonumber &&
=c^{-3}\int\!\!\!\!\int|z|^{-1}
\bigg\{-2\bar{z}(\bar{z}\cdot p)^2
+(\bar{z}\cdot p)p
+\bar{z}\,p^2\bigg\}\partial_t f(t, x+z, p)
\,dp\,dz
\\ & &
+c^{-3}\int\!\!\!\! \int\Bigg
\{\bar{z}\otimes\bar{z} -\frac{1}{3}\bigg\}
\partial_t\Big[E(f^+ +f^-)\Big](t,
x+z, p)\,dp\,dz+{\cal O}( c^{-5})\end{aligned}$$ Now we collect those term which contain exactly one time derivative including those coming from (\[thirdotwoder\]), $$\begin{gathered}
c^{-3}\int\!\!\!\!\int|z|^{-1}\bigg\{-(-3\bar{z}(\bar{z}\cdot
p)^2+2(\bar{z}\cdot
p)\,p+\bar{z}\,p^2)-2\bar{z}(\bar{z}\cdot
p)^2+(\bar{z}\cdot p)\,p+\bar{z}\,p^2
\bigg\}
\nonumber \\
\hspace{-12em}\partial_t f(t, x+z, p)\,dp\,dz
\nonumber \\
+c^{-3}\int\!\!\!\!\int\bigg\{-(\bar{z}\otimes\bar{z}-1)
+\bar{z}\otimes\bar{z}
-\frac{1}{3}\bigg\}
\partial_t\big[E(f^++f^-)\big](t, x+z,
p)\,dp\,dz
\nonumber \\
=c^{-3}\int\!\!\!\!\int|z|^{-2}
\bigg\{-4\bar{z}(\bar{z}\cdot p)^3
+3(\bar{z}\cdot p)^2\,p
+2(\bar{z}\cdot p)\,p^2\bar{z}
-p^2\,p
\bigg\}f(t, x+z, p)\,dp\,dz
\nonumber \\
+c^{-3}\int\!\!\!\!\int\bigg\{2(\bar{z}\cdot
p)\bar{z}\otimes\bar{z} -p\otimes\bar{z}
-\bar{z}\cdot p\bigg\}E(t, x+z)(f^++f^-)(t, x+z,
p)\,dp\,dz
\nonumber \\
+c^{-3}\frac{2}{3}\partial_t\int\!\!\!\!\int
E(t, x+z)(f^++f^-)(t, x+z, p)\,dp\,dz+{\cal
O}(c^{-5}).
\label{thirdooneder} \end{gathered}$$ Note that we do not touch upon the time derivative in the term of the last line. This term is responsible for radiation effects. Collecting all terms without time derivative coming from the third order in the expansion of $K_T$, the first order of the expansion of $K_S$, (\[thirdooneder\]) and (\[3o-add\]) we note that all these terms cancel exactly. Thus, the only remaining contribution in third order is the radiation term $$\partial_t\int\!\!\!\!\int E(t,
x+z)(f^++f^-)(t, x+z, p)\,dp\,dz
= \partial_t\int\!\!\!\!\int E(t, z)(f^++f^-)(t,
z, p)\,dp\,dz=:\tilde{D}(t)$$ Next we try to recast this term making it comparable to ${D^{[3]}}$. Note that using the bounds on $f^\pm$, see Assumption \[retRVMC-ass\](c), and the bounds on the support of $f^\pm$, with some $t-c^{-1}|z|<\xi(t, x, z)<t$ $$\begin{aligned}
\partial_t E(t, x) & = &
-\partial_t\int(\nabla \rho+ c^{-2}\partial_t
j)(\ast)\frac{dz}{|z|} \\
& = & -\int_{|x+z|\le r^\ast}(\partial_t\nabla
\rho+ c^{-2}\partial_t^2
j)(\ast)\frac{dz}{|z|} \\
& = & -\int_{|x+z|\le r^\ast}\partial_t\nabla
\rho(\ast)\frac{dz}{|z|}+{\cal O}_{cpt}(
c^{-2}) \\
& = & -\int_{|x+z|\le r^\ast}\partial_t\nabla
\rho(t, x+z)\frac{dz}{|z|}\\
& & +c^{-1}\int_{|x+z|\le
r^\ast}\partial_t^2\nabla
\rho(\xi(t, x, z),x+z)\,dz + {\cal O}_{cpt}(
c^{-2})\\
& = & -\int|z|^{-2}\bar{z}\partial_t\rho(t,
x+z)\,dz+{\cal O}_{cpt}(c^{-1})
\end{aligned}$$ Furthermore, we recall (\[retE0\]) and emphasize that the continuity equation (\[cont-eq\]) holds for both species seperately. Using these ingredients we compute $$\begin{aligned}
\tilde{D}(t) & = & \partial_t\int_{|z|\le
r^\dagger}\!\!\int E(t,
z)(f^++f^-)(t, z, p)\,dp\,dz \\
& = & \int_{|z|\le
r^\dagger}\int
|y-z|^{-2}(\overline{y-z})\bigg[-\partial_t(\rho^+-\rho^-)(t,
y)(\rho^++\rho^-)(t, z)\\
& & \hspace{2em}-(\rho^+-\rho^-)(t,
y)\partial_t(\rho^++\rho^-)(t, z)
\bigg]\,dy\,dz
+{\cal O}(c^{-1}) \\
& = &
-2\int\!\!\!\!\int|y-z|^{-2}(\overline{y-z})\bigg[\partial_t\rho^+(t,
y)\rho^-(t, z)-\partial_t\rho^-(t, y)\rho^+(t,
z)\bigg]\,dy\,dz +{\cal O}(c^{-1})\\
& = & 2\int\big(\tilde{H}{}^+(t, z) j^-(t,
z)-\tilde{H}{}^-(t,
z) j^+(t, z)\big)\,dz+{\cal O}(c^{-1}),\end{aligned}$$ where $$\nonumber
\tilde{H}{}^\pm(t, x) =
\oint{|z|^{-3}}(-3\bar{z}\otimes\bar{z}+1)\rho^\pm(t,
x+z)\,dz$$ and $\rho^\pm$ and $ j^\pm$ are defined in the obvious way. Summarizing, we have $$\begin{aligned}
\label{retE3}
E(t, x) & = &
\int\!\!\!\!\int|z|^{-2}\bigg\{-\bar{z} +
c^{-2}\Big(\frac{3}{2}\bar{z}(\bar{z}\cdot
p)^2-\frac{1}{2}\bar{z} p^2\Big)\bigg\}f(t,
x+z, p)\,dp\,dz
\\ \nonumber
& & -
c^{-2}\frac{1}{2}\int\!\!\!\!\int|z|^{-1}\bigg\{\bar{z}\otimes\bar{z}+1\bigg\}E
(t, x+z)
(f^++f^-)(t, x+z, p)\,dp\,dz
\\ \nonumber
& & +c^{-3}\frac{4}{3}\int\big(\tilde{H}{}^+(t, z) j^-(t, z)-\tilde{H}{}^-(t,
z) j^+(t, z)\big)\,dz+ {\cal O}( c^{-4}). \end{aligned}$$
Analogous computations lead to
$$\begin{aligned}
\label{retB0}
\lefteqn{B(t, x) = {\cal O}_{cpt}(c^{-1})}
\\ \label{retB2}
& = &
c^{-1}\int\!\!\!\!\int|z|^{-2}\bar{z}\wedge p \,
f(t, x+z, p)\,dp,dz+
{\cal O}_{cpt}(c^{-3})
\\ \nonumber\label{retB3}
& = & c^{-1}\int\!\!\!\!\int|z|^{-2}\bigg\{\bar{z}\wedge p
- c^{-2}\frac{3}{2}\bar{z}\wedge p(\bar{z}\cdot
p)^2 \bigg\} f(t, x+z,
p)\,dp,dz
\\ \nonumber
& &
+\frac{c^{-3}}{2}\int\!\!\!\!\int|z|^{-1}\bigg\{(\bar{z}\cdot
p)\bar{z}\wedge(\cdots)+(\bar{z}\wedge
p)\otimes\bar{z}\bigg\}E(t,
x+z)(f^++f^-)(t, x+z, p)\,dp\,dz
\\ \nonumber
& & +c^{-3}\int\bar{z}\wedge\big(\tilde{H}{}^+(t,
x+z)j^-(t, x+z)-\tilde{H}{}^-(t, x+z) j^+(t,
x+z)\big),dz
\\
& & +{\cal O}_{cpt}( c^{-4}).\end{aligned}$$
Now we have to prove the error estimates step by step, in fact for proving the 1.5PN approximation we need to know that $E=E_0+{\cal O}_{cpt}(
c^{-2})$.
\[new-le\] For all $0<T<\tilde{T}$ and $R>0$ there are constants $M(T)$ and $M(T, R)$ such that $$\begin{aligned}
|f^\pm(t, x, p)-f^\pm_0(t, x, p)| & \le & M(T)
c^{-2}\quad\hspace{0.55em} (x\in\R^3),
\nonumber \\
|E(t, x)-E_0(t, x)| & \le & M(T, R)\,c^{-2}
\quad (|x|\le R), \nonumber \\
|B(t, x)-c^{-1} B_1(t, x)| & \le & M(T, R)\,c^{-2}
\quad (|x|\le R), \nonumber\end{aligned}$$ for all $p\in\R^3$, $t\in [0, T]$, and $c\geq
P_1$.
The proof will not be carried out here because with some rather obvious modifications it closely follows the lines of Section \[HS-bew\] using the representations (\[retE0\]), (\[retB0\]) and $$\label{EradE0-est}
E^R(t, x) = E_0(t, x)+{\cal
O}(c^{-2})\quad\text{and}\quad
B^R(t, x) = c^{-1} B_1(t, x)+ {\cal
O}(c^{-2}),$$ the last two estimates are an easy consequence of (\[rad-def\]) and the bounds of Proposition \[f0-th\] and Lemma \[f2-le\].
If we replace $(f^\pm_0, E_0)$ with a solution of (\[VP\]) the Newton approximation can be obtained in the same way.
As a last step in the proof of Theorem \[Hauptsatz2\] we have to provide an estimate of the differences of the third order terms. Using the Calderon-Zygmund inequality [@ada Thm. 4.31] and (\[H-def\]) we have the estimate
$$\begin{aligned}
\nonumber
\bigg|\bigg|\oint |z|^{-3}H(z)
\rho^\pm_{(0)}(t, \cdot +z)
\,dz\bigg|\bigg|_{L^2}& \leq & C_{CZ}
||\rho^\pm_{(0)}(t, \cdot)||_{L^2}
\\ \nonumber
\end{aligned}$$
with a certain constant $C_{CZ}$. Hence $$\begin{aligned}
\label{diffD-est}
\nonumber
\lefteqn{\bigg|\int \big(\tilde{H}^+
j^--\tilde{H}^-j^+-H^+j^-_0+H^-j^+_0\big)(t,
z)\,dz\bigg|}
\\ \nonumber
& \leq & ||(\tilde{H}{}^+-H^+)(t,
\cdot)||_{L^2}\cdot ||j^-(t,
\cdot)||_{L^2}
+ || H^+(t, \cdot)||_{L^2}\cdot
||(j^--j^-_0)(t,\cdot)||_{L^2}
\\ \nonumber
& & + ||(\tilde{H}{}^--H^-)(t,
\cdot)||_{L^2}\cdot ||j^+(t,
\cdot)||_{L^2}
+ || H^-(t, \cdot)||_{L^2}\cdot
||(j^+-j^+_0)(t,\cdot)||_{L^2}
\\ \nonumber
& \leq & C_{CZ}\bigg(||(\rho^+-\rho_0^+)(t,
\cdot)||_{L^2}\cdot ||j^-(t,
\cdot)||_{L^2}
+ || \rho^+(t, \cdot)||_{L^2}\cdot
||(j^--j^-_0)(t,\cdot)||_{L^2}
\\ \nonumber
& & + ||(\rho^--\rho_0^-)(t,
\cdot)||_{L^2}\cdot ||j^+(t,
\cdot)||_{L^2}
+ || \rho_0^-(t, \cdot)||_{L^2}\cdot
||(j^+-j^+_0)(t,\cdot)||_{L^2}\bigg)
\\ \nonumber
& \leq & C_{CZ} \,Q^0(t)\, M
\,(r^\dagger)^3\,(p^\ast)^7
\\
& \leq & M Q^0(t)\end{aligned}$$ where $$Q^0(t) := \max\big\{ |f^\pm(t, x, p)-f^\pm_0(t,
x, p)|\;:\; x\in\R^3, p\in \R^3, \pm\in\{+,
-\}\big\}$$
Here we used (\[rdagger-def\]) and Assumption \[retRVMC-ass\](a) in combination with (\[retfini-def\]) and (\[finf-est\]).
In view of the formulas (\[radE-rep\]), (\[retE3\]), Lemma \[new-le\], (\[EradE0-est\]), (\[diffD-est\]) and proceeding analogously to Section \[diffE-subsec\] we conclude the estimate $$\nonumber
|E(t, x)-E^R(t, x)|\le M_R( c^{-4}+Q(t)))$$ for all $|t|\le T, |x|\le R$ with a constant $M_R$ independent of $c$ where $$Q(t) = \max\big\{ |f^\pm(t, x, p)-f^\pm_R(t, x, p)|\;:\;
x\in\R^3, p\in \R^3, \pm\in\{+, -\}\big\}$$ Employing the formulas (\[radB-rep\]) and (\[retB3\]) together with an estimate corresponding to (\[diffD-est\]) we conclude $$\nonumber
|B(t, x)-B^R(t, x)|\le M_R( c^{-4}+Q(t)))$$ for all $|t|\le T, |x|\le R$ with a constant $M_R$ independent of $c$. Proceeding analogously to Section \[difff-subsec\] finishes the proof of Theorem \[Hauptsatz2\].[$\Box$]{}
Appendix {#append}
========
Representation of the approximation fields $E^R$ and $B^R$ {#repapp-sec}
----------------------------------------------------------
Here we will derive the representation formula (\[ER-rep\]) for the approximate field $E^R$ from (\[rad-def\]). Since the calculations for the electric and the magnetic field are quite similar, we will only analyse in detail the electric field and simply state the result for its magnetic counterpart. Actually, the representation of the electric field exhibits more difficulties than the representation of the magnetic field, due to the presence of the radiation term.
From (\[rad-def\]) we recall $E^R=E_0+c^{-2}E_2+
(2/3)c^{-3} D^{[3]}$, where
$$\begin{aligned}
E_0(t, x) & = & -\int |z|^{-2}\bar{z}\,\rho_0(t, x+z)\,dz,
\label{ma1} \\
E_2(t, x) & = & \frac{1}{2}\int\bar{z}\,\partial_t^2\rho_0(t, x+z)\,dz
-\int |z|^{-1}\partial_t j_0(t, x+z)\,dz
\nonumber \\
& & -\int |z|^{-2}\bar{z}\,\rho_2(t, x+z)\,dz,\qquad
\label{ma2} \\
D^{[3]}(t) & = & \dot{D}^{[2]}
=\partial_t\int\!\!\!\! \int E_0(t, x)(f^+_0+f^-_0)(t, x, p)\,dp\,dx
\label{ma3}\end{aligned}$$
cf. (\[rrVP\]) and (\[LVP\]). Firstly, we write $$\nonumber
D^{[3]}(t) =\partial_t\int\!\!\!\! \int E_0(t, x+z)(f^+_0+f^-_0)(t, x+z, p)\,dp\,dz,\qquad x\in \R^3.$$ Secondly, we split the domain of integration in $\{|z|>ct\}$ and $\{|z|\leq ct\}$; note that the exterior part exactly gives $E^R_{ext}$ in (\[ER-rep\]). To handle the interior part $\{|z|\leq ct\}$, in the sequel denoted by $\widetilde{E^R_{int}}$, we fix $R>0$ and $0<T<\min\{\tilde{T}, \hat{T}\}$ and put $$r^\ddagger=R+M_0\qquad \text{and}\qquad p^\dagger=M_0,$$ recall that $M_0=\max\{M_1(T), M_3(T), M_5(T)\}$; $r^\ddagger$ and $p^\dagger$ are only depending on our basis constants $r_0$, $S_0$, see (\[dat-def\]), and of course of $R$. For $|x|\le
R$ and $0\le t\le T$, if $|z|\ge r^\ddagger$ or $|p|\ge p^\dagger$, than $f^\pm_0(\tau, x+z,
p)=f^\pm_2(\tau, x+z, p)=0$ for all $0\le\tau\le
T$ by Theorem \[f0-th\](b) and Lemma \[f2-le\](a). This argument shows that we can replace $\int_{|z|\le ct}\int\,dp\,dz$ by $\int_{|z|\le \min\{ct, r^\ddagger\}}\int_{|p|\le
p^\dagger}\,dp\,dz$ in the integrals defining the interior part. In other words, we may always assume that both $|z|$ and $|p|$ are bounded with a bound only depending on the basic constants and $R$, but not on $c$. Next we expand the densities w.r.t. $t$ about the retarded time $\hat{t}(z):=t-c^{-1}|z| $ and obtain e.g.
$$\begin{aligned}
\label{ma1-exp}
\lefteqn{-\int_{|z|\le ct} |z|^{-2}\bar{z}\,\rho_0(t, x+z)\,dz}
\\ \nonumber
&=& -\int_{|z|\le ct}
|z|^{-2}\bar{z}\,\Big((1+c^{-1}|z|\partial_t+1/2c^{-2}|z|^2\partial_t^2+
1/6c^{-3}|z|^3\partial_t^3)
\\ \nonumber
& & \hspace{15em}\rho_0(\hat{t}(z), x+z)+|z|^4{\cal O}(c^{-4})\Big)\,dz
\\ \nonumber
&=& -\int_{|z|\le ct}
|z|^{-2}\bar{z}\,(1+c^{-1}|z|\partial_t+1/2c^{-2}|z|^2\partial_t^2+
1/6c^{-3}\partial_t^3)\,\rho_0(\hat{t}(z), x+z)\,dz
\\ \nonumber
& & +{\cal O}_{cpt}(c^{-4})\end{aligned}$$
where Theorem \[f0-th\](c) was utilized and hence $M_2(T)$ enters the bounds on ${\cal O}(c^{-4})$ and ${\cal O}_{cpt}(c^{-4})$. In the same manner we expand the terms in (\[ma2\]) up to first order and the term from (\[ma3\]) up to zeroth order, employing Theorem \[f0-th\](c) and Lemma \[f2-le\](b), therefore also $M_4(T)$ enters the bounds. $$\begin{aligned}
\label{ma2-exp}
\lefteqn{c^{-2}\int_{|z|\le ct}\big\{1/2\bar{z}\partial_t^2\rho_0-|z|^{-1}\partial_t j_0-|z|^{-2}\bar{z}\rho_2\big\}(t, x+z)\,dz}
\\ \nonumber
& = & c^{-2}\int_{|z|\le ct}\big\{\Big(1+c^{-1}|z|\partial_t\Big)\Big(1/2\bar{z}\partial_t^2\rho_0-|z|^{-1}\partial_t
j_0-|z|^{-2}\bar{z}\rho_2\Big)\big\}(\hat{t}(z), x+z)\,dz
\\ \nonumber
& & +{\cal O}_{cpt}(c^{-4})
\\ \label{ma3-exp}
\lefteqn{2/3 c^{-3}\int_{|z|\le ct} \partial_t\big[E_0(\rho^+_0+\rho^-_0)\big](t, x+z)\,dz}
\\ \nonumber
& = & 2/3 c^{-3}\int_{|z|\le ct} \partial_t\big[E_0(\rho^++\rho^-)\big](\hat{t}(z), x+z)\,dz+{\cal O}_{cpt}(c^{-4}).\end{aligned}$$
Next we sort the terms according to their orders in $c^{-1}$. To zeroth order we have $$\label{0-or}
\widetilde{E^R_{int}}(t, x) = -\int_{|z|\le ct}|z|^{-2}\bar{z}\rho_0(\hat{t}(z), x+z)\,dz+{\cal O}(c^{-1}).$$ In the first order we have the contribution $-c^{-1}\int_{|z|\le ct}|z|^{-1}\bar{z}\partial_t\rho_0(\hat{t}(z), x+p)\,dz$, which gives, using that $\partial_t\rho_0+\nabla\cdot j_0=0$ also holds true for (\[rrVP\]) and further integration by parts,
$$\begin{aligned}
\lefteqn{c^{-1}\int_{|z|\le ct}|z|^{-1}\bar{z}\nabla_x\cdot j_0(\ldots)\,dz}
\nonumber \\
& = & c^{-1}\int_{|z|\le ct}|z|^{-1}(\nabla_z+c^{-1}\bar{z}\partial_t)\cdot\big[j_0(\hat{t}(z), x+z)\big]\,dz
\nonumber \\
& = & c^{-1}(ct)^{-1}\int_{|z|=ct}\bar{z}\bar{z}\cdot j_0(0, x+z)\,ds(z)
-c^{-1}\int_{|z|\le ct}|z|^{-2}(-2\bar{z}\bar{z}\cdot j_0+j_0)(\ldots)\,dz
\nonumber \\
&& +c^{-2}\int_{|z|\le ct}|z|^{-1}\bar{z}\bar{z}\cdot\partial_t j_0(\ldots)\,dz,
\label{r1-or}\end{aligned}$$
where $(\ldots)=(\hat{t}(z), x+z)$. Hence, the first order term of $\widetilde{E^R_{int}}$ can be written as $$\label{1-or}
c^{-1}(ct)^{-1}\int_{|z|=ct}\bar{z}\bar{z}\cdot j_0(0, x+z)\,ds(z)
-c^{-1}\int_{|z|\le ct}|z|^{-2}(-2\bar{z}\bar{z}\cdot j_0+j_0)(\ldots){\,dz}$$
To continue we collect the terms of second order from (\[ma2-exp\]), (\[ma1-exp\]), (\[r1-or\]). $$\begin{gathered}
\nonumber
c ^{-2}\int_{|z|\le ct} \Big(1/2\bar{z}\partial_t^2
\rho_0-|z|^{-1}\partial_t j_0-|z|^{-2}\bar{z}\rho_2-1/2\bar{z}\partial_t^2\rho_0+|z|^{-1}\bar{z}\bar{z}\cdot\partial_t j_0\Big)(\ldots)\,dz\end{gathered}$$ Since the terms containing second derivatives cancel each other we start with the first order time derivatives. Utilizing (\[rrVP\]) and integration by parts we calculate
$$\begin{aligned}
\label{r2-or}
\nonumber
\lefteqn{c^{-2}\int_{|z|\le ct}|z|^{-1}(\bar{z}\bar{z}\cdot\partial_t j_0-\partial_t j_0)(\ldots)\,dz}
\\ \nonumber
& = & c^{-2}\int_{|z|\le ct}|z|^{-1}(\bar{z}\otimes\bar{z}-1)(E_0+c^{-3}2/3{D^{[3]}})(\rho_0^++\rho_0^-)(\ldots)\,dz
\\ \nonumber
& & +c^{-2}\int_{|z|\le
ct}\int|z|^{-2}f_0(\ldots, p)\big(-3\bar{z}(\bar{z}\cdot p)^2+2\bar{z}\cdot p
p+\bar{z}p^2\big)\,dp\,dz
\\ \nonumber
& & -c^{-2}(ct)^{-1}\int_{|z|=ct}\int(\bar{z}\cdot p)(\bar{z}\bar{z}\cdot p-p) f^\circ_0(x+z, p)\,dp\,dz
\\
& & -c^{-3}\int_{|z|\le
ct}\int|z|^{-1}(\bar{z}\cdot
p)(\bar{z}\bar{z}\cdot p-p)\partial_t
f_0(\ldots, p)\,dp\,dz.\end{aligned}$$
Note that since the bounds in Theorem \[f0-th\](b)(c) also imply that $2/3c^{-5}\int_{|z|\le ct}|z|^{-1}(\bar{z}\otimes\bar{z}-1){D^{[3]}}(\rho_0^++\rho_0^-)(\ldots)\,dz={\cal O}_{cpt}(c^{-5})$ this term can be dropped. Using the same argument all terms containing ${D^{[3]}}$ appearing in the sequel are at least ${\cal
O}_{cpt}(c^{-5})$ and hence will be dropped without comment. Thus, the contribution of second order in $\widetilde{E^R_{int}}$ is $$\begin{gathered}
c^{-2}\int_{|z|\le
ct}|z|^{-1}(\bar{z}\otimes\bar{z}-1)E_0(\ldots)(f_0^+
+f_0^-)(\ldots, p) \,dp\,dz
\nonumber \\
+ c^{-2}\int_{|z|\le ct}
|z|^{-2}\bigg\{f_0(\ldots, p)\big(-3\bar{z}(\bar{z}\cdot p)^2+2\bar{z}\cdot p
p+\bar{z}p^2\big)-\bar{z}f_2(\dots, p)\Big\} \,dp\,dz
\nonumber \\
-c^{-2}(ct)^{-1}\int_{|z|=ct}\int(\bar{z}\cdot p)(\bar{z}\bar{z}\cdot p-p){f^\circ_0}(x+z, p)\,dp\,dz.
\label{2-or}\end{gathered}$$
We continue by collecting the terms of third order from (\[ma2-exp\]), (\[ma3-exp\]), (\[ma1-exp\]) and (\[r2-or\]).
$$\begin{gathered}
\nonumber
c^{-3}\int_{|z|\le ct}
\Big\{(-1/6+1/2)z\partial_t^3\rho_0-\partial_t^2j_0-|z|^{-1}\bar{z}\partial_t\rho_2+2/3\partial_t\big[E_0(\rho_0^++\rho_0^-\big]\Big\}(\ldots)\,dz
\\ \label{3-or1}
-c^{-3}\int_{|z|\le ct}\int|z|^{-1}(\bar{z}\cdot p)\big(\Bar{z}\bar{z}\cdot p\big)\partial_t f_0(\ldots)\,dp\,dz \end{gathered}$$
Using (\[rrVP\]) and integration by parts we compute $$\begin{gathered}
\label{3td3-or}
1/3c^{-3}\int_{|z|\le ct} z\partial_t^3\rho_0(\ldots)\,dz
\\ \nonumber
= 1/3c^{-3}\int_{|z|\le ct} \partial_t^2j_0(\ldots)\,dz
-1/3tc^{-2}\int_{|z|=ct} \bar{z}\bar{z}\cdot \partial_t^2 j_0(0, x+z)\,ds(z)+{\cal O}(c^{-4}).\end{gathered}$$ Summing the terms with two time derivatives from (\[3-or1\]) and (\[3td3-or\]) we get $$\begin{gathered}
\label{2td3-or}
-2/3 c^{-3}\int_{|z|\le ct}\partial_t^2 j_0(\ldots)\,dz
= -2/3c^{-3}\int_{|z|\le ct}\partial_t\big[(E_0+(2/3)c^{-3}{D^{[3]}})(\rho_0^++\rho_0^-)\big](\ldots)\,dz
\\ \nonumber
+2/3c^{-3}\int_{|z|=ct}\int\bar{z}\cdot p \,p\,\partial_t f_0(0, x+z, p)\,dp\,ds(z)+{\cal O}(c^{-4}).\end{gathered}$$ Now we collect the first time derivatives from (\[3-or1\]). Starting with the contributions containing $f_0$ we have $$\begin{aligned}
\lefteqn{-c^{-3}\int_{|z|\le ct}\int|z|^{-1}\big(\bar{z}(\bar{z}\cdot p)^2
-(\bar{z}\cdot p) p\big)\partial_t f_0(\ldots, p)\,dp\,dz}
\nonumber \\
& = & -c^{-3}\int_{|z|\le ct}\int|z|^{-1}\big(2(\bar{z}\cdot p)\bar{z}\otimes\bar{z}-\bar{z}\cdot p -p
\otimes\bar{z}\big)
\nonumber \\
& &
\hspace{10em}\Big(E_0(\ldots)+(2/3)c^{-3}{D^{[3]}}(\hat{t}(z))\Big)(f^+_0+f^-_0)(\ldots,
p)\,dp\,dz
\nonumber \\
& & -c^{-3}\int_{|z|\le
ct}\int|z|^{-2}f_0(\ldots, p)\big(-4\bar{z}(\bar{z}\cdot p)^3+3p(\bar{z}\cdot p)^2
+2\bar{z}(\bar{z}\cdot p)p^2-pp^2\big)\,dp\,dz
\nonumber \\
& & +c^{-3}(ct)^{-1}\int_{|z|=ct}\int\big(\bar{z}(\bar{z}\cdot p)^3-p(\bar{z}\cdot p)^2\big){f^\circ_0}(x+z,
p)\,dp\,ds(z) + {\cal O}(c^{-4})
\label{1td3-or1}\end{aligned}$$ and secondly, for the terms containing $f_2$ we compute employing (\[LVP\]) $$\begin{aligned}
\lefteqn{-c^{-3}\int_{|z|\le ct}|z|^{-1}\bar{z}\partial_t\rho_2(\ldots)\,dz}
\nonumber\\
& = & -c^{-3}\int_{|z|\le
ct}\int|z|^{-2}f_2(\ldots, p)\big(-
2\bar{z}(\bar{z}\cdot p)+p\big)\,dp\,dz
\nonumber \\
& & +c^{-3}(ct)^{-1}\int_{|z|=ct} \int
\bar{z}(\bar{z}\cdot p) f^{\circ}_2(x+z, p)
\,dp\,ds(z)
\nonumber \\
& & +1/2c^{-3}\int_{|z|\le
ct}\int|z|^{-2}f_0(\ldots, p)\big(-2\bar{z}(\bar{z}\cdot p)p^2+p^2 p\big)\,dp\,dz
\nonumber \\
& & -1/2 c^{-3}(ct)^{-1}\int_{|z|=ct}\int\bar{z}(\bar{z}\cdot p)p^2{f^\circ_0}(x+z, p)\,dp\,ds(z)
+{\cal O}(c^{-4}).
\label{1td3-or2}\end{aligned}$$ Summarizing (\[3-or1\])-(\[1td3-or2\]) we can identify the contribution in the third order of $\widetilde{E^R_{int}}$. $$\begin{gathered}
\int_{|z|=ct}\int\big\{-1/3 tc^{-2}\bar{z}(\bar{z}\cdot p)\partial_t^2+2/3 c^{-3}(\bar{z}\cdot
p)p\partial_t
\nonumber \\
+c^{-3}(ct)^{-1}\Big(\bar{z}(\bar{z}\cdot p)^3-p(\bar{z}\cdot p)^2-1/2\bar{z}(\bar{z}\cdot
p)p^2\Big)\big\}f_0(0, x+z, p)
\,dp\,ds(z)
\nonumber \\
+c^{-3}(ct)^{-1}\int_{|z|=ct} \int
\bar{z}(\bar{z}\cdot p) f^{\circ}_2(x+z, p)
\,dp\,ds(z)
\nonumber \\
+c^{-3}\int_{|z|\le ct}\int|z|^{-2}f_0(\ldots, p)\Big\{4\bar{z}(\bar{z}\cdot p)^3-3p(\bar{z}\cdot
p)^2-3\bar{z}(\bar{z}\cdot p)p^2+3/2p p^2\Big\}\,dp\,dz
\nonumber \\
-c^{-3}\int_{|z|\le ct}\int|z|^{-1}\big\{2(\bar{z}\cdot p)\bar{z}\otimes\bar{z}-\bar{z}\cdot p -p
\otimes\bar{z}\big\}E_0(\ldots)(f^+_0
+f^-_0)(\ldots, p )\,dp\,dz
\nonumber \\
+c^{-3}\int_{|z|\le ct}\int|z|^{-2}f_2(\ldots, p)\big\{2\bar{z}(\bar{z}\cdot p)-p\big\}\,dp\,dz.
\label{3-or}\end{gathered}$$
Therefore, if we use (\[0-or\]), (\[1-or\]), (\[2-or\]), (\[3-or\]) and add some terms of order ${\cal
O}(c^{-4})$ containing $f^\pm_2$ as e.g. $\int_{|z|\le ct}\int|z|^{-2}\bar{z}(\bar{z}\cdot p)^2 f_2 (\ldots)\,dp\,dz$ it turns out that $E^R$ can be decomposed as it is claimed in (\[ER-rep\]).
Similar calculations for $B^R$ yield $$\label{BR-rep}
B^R = B^R_{ext}+B^R_{int}+B^R_{bound}+{\cal O}_{cpt}(c^{-4})$$ where $$\begin{aligned}
\label{BRext-rep}
B^R_{ext}(t, x) & = & c^{-1}\int_{|z|\ge ct}\Big(|z|^{-2}\bar{z}\times (j_0+ c^{-2} j_2)-1/2 c^{-2}\bar{z}\partial_t^2{ j_0}\big){(t, x+z)}{\,dz}\\ \label{BRint-rep}
B^R_{int}(t, x) & = & \int_{|z|\le ct}\int|z|^{-2} L_1(\bar{z}, c^{-1} p)f_R({\hat{t}(z)}, x+z, p){\,dp}{\,dz}\\\nonumber
& & + c^{-2}\int_{|z|\le ct}\int|z|^{-1} L_2(\bar{z}, c^{-1} p){E_0}(f^+_R+f^-_R)({\hat{t}(z)}, x+z, p){\,dp}{\,dz}\\ \label{BRbound-rep}
B^R_{bound}(t, x) & = &{(ct)^{-1}}\int_{|z|=ct}\int L_3(\bar{z}, c^{-1} p){f^\circ_0}{(x+z, p)}{\,dp}{\,ds(z)}\end{aligned}$$ and the kernels are given by $$\begin{aligned}
\label{L1}
L_1(\bar{z},\tilde{p}) & = & \bar{z}\times\tilde{p}-2\bar{z}\times\tilde{p}(\bar{z}\cdot\tilde{p})3/2\bar{z}\times\tilde{p}\tilde{p}^2+3\bar{z}\times\tilde{p}(\bar{z}\cdot\tilde{p})^2
\\ \label{L2}
L_2(\bar{z},\tilde{p}) & = &\bar{z}\times(\cdots)-\bar{z}\cdot\tilde{p}\bar{z}\times(\cdots)-\bar{z}\cdot\tilde{p}\bar{z}\cdot(\cdots)\in \R^{3\times3}
\\ \label{L3}
L_3(\bar{z}, \tilde{p}) & = & -\bar{z}\times\tilde{p}\bar{z}\cdot\tilde{p}+\bar{z}\times\tilde{p}(\bar{z}\cdot\tilde{p})^2.\end{aligned}$$
Representation of the Maxwell fields $E$ and $B$ {#repmax-sec}
------------------------------------------------
In this section we will verify the representation formula (\[E-rep\]) for the full Maxwell field $E$ by expanding the respective expressions from [@glstr; @schaeffer:86] to higher orders. Once again the computation for the corresponding magnetic field $B$ is very similar and therefore omitted. Let $(f, E, B)$ be a $C^1$-solution of (\[RVMC\]) with initial data $(f^{\circ, \pm}_c, E^\circ_c, B^\circ_c)$ according to (\[IC\]). We recall the following representation from [@schaeffer:86 (A13), (A14), (A3)], $$E = E_D+E_{DT}+E_T+E_S, \label{EFeld}$$ where $$\begin{aligned}
E_D(t, x) & = &
\partial_t\bigg(\frac{t}{4\pi}\int_{|\omega|=1} E^\circ_c(x+ct\omega)\,d\omega\bigg)
+\frac{t}{4\pi}\int_{|\omega|=1}\partial_t E(0, x+ct\omega)\,d\omega, \\
E_{DT}(t, x) & = & (ct)^{-1}\int_{|z|=ct}\int K_{DT}(\bar{z}, \hat{p})
{f^\circ_0}(x+z, p)\,dp\,ds(z), \\
E_T(t, x) & = & \int_{|z|\leq ct} |z|^{-2}\int K_T(\bar{z}, \hat{p})
f(\ldots, p)\,dp\,dz, \\
E_S(t, x) & = &c^{-2}\int_{|z|\leq ct} |z|^{-1}\int K_S(\bar{z}, p)
(E+c^{-1}\hat{p}\times
B)(\ldots)(f^++f^-)(\ldots, p)\,dp\,dz\end{aligned}$$ The kernels are given by
$$\begin{aligned}
\nonumber
K_{DT}(\bar{z}, \hat{p}) & = & -(1+c^{-1}\bar{z}\cdot\hat{p})^{-1}
(\bar{z}-c^{-2}(\bar{z}\cdot\hat{p})\hat{p}),
\\ \label{KT-def}
K_T(\bar{z}, \hat{p}) & = & -(1+c^{-1}\bar{z}\cdot\hat{p})^{-2}(1-c^{-2}\hat{p}^2)
(\bar{z}+c^{-1}\hat{p}),
\\ \nonumber
K_S(\bar{z}, p) & = & -(1+c^{-1}\bar{z}\cdot\hat{p})^{-2}(1+c^{-2} p^2)^{-1/2}
\Big[(1+c^{-1}\bar{z}\cdot\hat{p})+c^{-2}((\bar{z}\cdot\hat{p})\bar{z}-\hat{p})\otimes\hat{p}
\\ & & \hspace{14em} -(\bar{z}+c^{-1}\hat{p})\otimes\bar{z}\Big]\in\R^{3\times 3}
\label{KS-def}\end{aligned}$$
Next we expand these kernels in powers of $c^{-1}$. According to (\[SchrankeSupport\]) the $p$-support of $f^\pm(t, x, \cdot)$ is uniformly bounded in $x\in\R^3$ and $t\in [0, T]$ by $M_5$. Thus, we may suppose that $|p|\le M_5(T)$ in each of the $p$-integrals. Therefore as long as $f^\pm(\ldots, p)\not=0$ we have $$\begin{aligned}
K_T(\bar{z}, \hat{p}) & = & K_1(\bar{z}, c^{-1}p)+{\cal O}(c^{-4}) \\
K_S(\bar{z}, \hat{p}) & = & K_2(\bar{z}, c^{-1}p)+{\cal O}(c^{-2}) \\
K_{DT}(\bar{z}, \hat{p}) & = & -\bar{z}+K_3(\bar{z}, c^{-1}p)+{\cal O}(c^{-4}).\end{aligned}$$ Recall that $f^{\circ, \pm}_c=0$ if $|x|\ge r_0$, see (\[dat-def\]) and (\[IC\]); utilizing this as well as $$\begin{aligned}
\lefteqn{-(ct)^{-1}\int_{|z|=ct}\int_{|p|\le p^\dagger} {\cal O}(c^{-4})
{\bf 1}_{B(0, r_0)}(x+z)\,dp\,ds(z)}
\\
& = & \bigg(ct\int_{|\omega|=1}{\bf 1}_{B(0, r_0)}(x+ct\omega)\,ds(\omega)\bigg){\cal O}(c^{-4})
= {\cal O}(c^{-4})\end{aligned}$$ by [@schaeffer:86 Lemma 1], uniformly in $x\in\R^3$, $t\in [0, T]$, and $c\ge 1$, we arrive at $$\label{EDT-exp}
E_{DT}(t, x) =
(ct)^{-1}\int_{|z|=ct}\int(-\bar{z}+K_3(\bar{z},
c^{-1}p))({f^\circ_0}+c^{-2} f^\circ_2)(x+z, p)\,dp\,ds(z)+{\cal O}(c^{-4}).$$ Concerning $E_T$, we note that $f^\pm(t, x, p)=0$ for $|x|\ge M_5(T)$ and $t\le T$, see (\[SchrankeSupport\]). Since, by distinguishing the cases $|x-y|\ge 1$ and $|x-y|\le 1$, $$\int_{|z|\leq ct} |z|^{-2}\,{\bf 1}_{B(0, M_5(T))}(x+z)\,dz
=\int_{|x-y|\leq ct} |x-y|^{-2}\,{\bf 1}_{B(0, M_5(T))}(y)\,dy={\cal O}(1)$$ uniformly in $x\in\R^3$, $t\in [0, T]$, and $c\ge 1$, similar computations as before show that
$$\begin{aligned}
\label{ET-expa}
E_T(t, x) & = & \int_{|z|\leq ct} |z|^{-2}\int
K_1(\bar{z}, c^{-1}p)f(\ldots, p)\,dp\,dz+{\cal O}(c^{-4}).\end{aligned}$$
In the same manner, elementary calculations using also (\[SchrankeFelder\]) can be carried out to get $$E_S(t, x) = c^{-2}\int_{|z|\leq ct} |z|^{-1}\int K_2(\bar{z}, c^{-1}p)
E(\ldots)(f^+ +f^-)(\ldots, p) \,dp\,dz +{\cal
O}(c^{-4}).
\label{ES-expa}$$ Observe that we have proved $$\label{maxEin-est}
E_T+E_S=E_{int}+{\cal O}(c^{-4}),$$ see (\[Eint-rep\]).
Next we consider the data term $$\label{dat1}
E_D(t, x) =
\partial_t\big(\frac{t}{4\pi}\int_{|\omega|=1} E^\circ_c(x+ct\omega)\,d\omega\big)
+\frac{t}{4\pi}\int_{|\omega|=1}\partial_t E(0,x+ct\omega)\,d\omega = I+II$$ By Maxwell’s equations we have $\partial_t E(0, x)=c\nabla\times B^\circ_c (x)-4\pi j(0, x)$. Recall that $B^\circ_c =c^{-1}B_1+c^{-3}B_3+c^{-4}B_{c, free}$. Using $\Delta=-\nabla\times\nabla\times+\nabla\nabla\cdot$ and $\partial_t\rho_0+\nabla\cdot j_0=0$ as well as $\partial_t\rho_2+\nabla\cdot j_2=0$, the latter an easy consequence of (\[LVP\]), it is easy to check that $\nabla\times B_1=\partial_t E_0+4\pi j_0$ and $\nabla\times B_3=\partial_t E_2+4\pi j_2$. Employing the definition of $E_0$, $E_2$ and (\[formeli\])-(\[formeliv\]) below we compute
$$\begin{aligned}
\label{dte0}
\lefteqn{\frac{t}{4\pi}\int_{|\omega|=1}\partial_t E_0(0, x+ct\omega)\,d\omega =
-\frac{t}{4\pi}\int_{|\omega|=1}\int|z|^{-2}\bar{z}\partial_t\rho_0(0, x+ct\omega+z)\,dz\,d\omega}
\nonumber \\
& = & -\frac{t}{4\pi}\int\partial_t\rho_0(0, y)\int_{|\omega|=1}|y-x-ct\omega|^{-3}(y-x-ct\omega)\,d\omega\,dy
\nonumber \\
& = & -t\int_{|z|\ge ct}|z|^{-2}\bar{z}\partial_t\rho_0(0, x+z)\,dz
\\ \label{dte2}
\lefteqn{\frac{t}{c^24\pi}\int_{|\omega|=1}\partial_t E_2(0, x+ct\omega)\,d\omega}
\\ \nonumber
& = &
\frac{t}{c^24\pi}\int_{|\omega|=1}\int\big\{1/2\bar{z}\partial_t^3\rho_0-|z|^{-1}\partial_t^2 j_0-|z|^{-2}\bar{z}\partial_t\rho_2\big\}(0,
x+ct\omega+z)\,dz\,d\omega
\nonumber \\
& = & \frac{1}{3c^3}\int_{|z|\le ct} z\partial_t^3\rho_0(0, x+z)\,dz+\frac{t}{2c^2}\int_{|z|\ge ct} \bar{z}\partial_t^3\rho_0(0,
x+z)\,dz
\nonumber \\
& &-\frac{t^3}{6}\int_{|z|\ge ct}|z|^{-2}\bar{z}\partial_t^3\rho_0(0, x+z)\,dz
-\frac{1}{c^3}\int_{|z|\le ct}\partial_t^2 j_0(0, x+z)\,dz
\nonumber \\
& & -\frac{t}{c^2}\int_{|z|\ge ct}|z|^{-1}\partial_t^2 j_0(0, x+z)\,dz-
\frac{t}{c^2}\int_{|z|\ge ct}|z|^{-2}\bar{z}\partial_t\rho_2(0, x+z)\,dz.
\nonumber\end{aligned}$$
Using (\[rrVP\]) we also have $$\begin{aligned}
\label{dte2-1}
\lefteqn{\frac{1}{3c^3}\int_{|z|\le ct} z\partial_t^3\rho_0(0, x+z)\,dz-\frac{1}{c^3}\int_{|z|\le ct}\partial_t^2 j_0(0, x+z)\,dz}
\\
& = & -\frac{t}{3c^2}\int_{|z|=ct}\bar{z}\bar{z}\cdot \partial_t^2 j_0(0, x+z)\,dz-\frac{2}{3c^3}\int_{|z|\le ct}\partial_t^2 j_0(0, x+z)\,dz
\nonumber \\
& = & -\frac{t}{3c^2}\int_{|z|=ct}\bar{z}\bar{z}\cdot \partial_t^2 j_0(0, x+z)\,dz+\frac{2}{3c^3}\int_{|z|=ct}\int(\bar{z}\cdot p)\partial_t j_0(0,
x+z)\,dz
\nonumber \\
&&-\frac{2}{3c^3}\int_{|z|\le ct}\partial_t\big[(E_0+(2/3)c^{-3}{D^{[3]}})(\rho^+_0+\rho^-_0)\big](0, x+z)\,dz.
\nonumber \end{aligned}$$ Furthermore, according to (\[dat-def\]) and (\[freedat-def\]) and the definition of $j_0$ and $j_2$ the support of $\tilde{j}:=(j-j_0-c^{-2}j_2)(0, \cdot)$ is bounded by $r_0$ and $$\tilde{j}(x)=\int_{|p|\le r_0}
\Big[(\hat{p}-p+\frac{p^2}{2c^2}){f^\circ_0}+
c^{-2}(\hat{p}-p)f^\circ_2
+c^{-4}\hat{p}f^\circ_{c, free} \Big](x, p)\,dp = {\cal
O}(c^{-4}).$$ Utilizing [@schaeffer:86 Lemma 1] we conclude $$\label{jdat}
\frac{ct}{4\pi}\int_{|\omega|=1}4\pi\tilde{j}(0, x+ct\omega)\,d\omega = {\cal O}(c^{-4}).$$
Now we return to the contributions coming from I in (\[dat1\]). Note that in [@bauku p304/305,(5.21)] the contributions coming from $E_0$ and $E_2$ are already determined to
$$\begin{gathered}
-\int_{|z|>ct} |z|^{-2}\bar{z}\,\rho_0(0, x+z)\,dz
+\frac{1}{2}\,c^{-2}\int_{|z|>ct}\bar{z}\,\partial_t^2\rho_0(0,
x+z)\,dz
\nonumber \\
-\frac{1}{2}\,t^2\int_{|z|>ct}|z|^{-2}\bar{z}
\,\partial_t^2\rho_0(0, x+z)\,dz
-c^{-2}\int_{|z|>ct}|z|^{-1}\partial_t j_0(0, x+z)\,dz
\nonumber \\
+(ct)^{-1}\int_{|z|=ct} \bar{z}\,(\rho_0
+c^{-2}\rho_2)(0, x+z)\,ds(z).
\label{Idat}\end{gathered}$$
Concerning the two remaining terms we have $$\begin{aligned}
\label{D3dat}
\lefteqn{\partial_t\bigg(\frac{t}{4\pi}\int_{|\omega|=1}\frac{2}{3c^3}D^{[3]}(0)\,d\omega\bigg) = \frac{2}{3c^3}D^{[3]}(0)
=\frac{2}{3c^3}\int\partial_t\big[E_0(\rho_0^++\rho_0^-)\big](0, z)\,dz}
\\ \nonumber
& = &
\frac{2}{3c^3}\int_{|z|\ge ct} \partial_t\big[E_0(\rho_0^++\rho_0^-)\big](0, x+z)\,dz
+\frac{2}{3c^3}\int_{|z|\le ct} \partial_t\big[E_0(\rho_0^++\rho_0^-)\big](0, x+z)\,dz,\end{aligned}$$ note that the last term in (\[D3dat\]) cancels the last term in (\[dte2-1\]); and $$\begin{aligned}
\label{efree-dat}
\lefteqn{\partial_t\big(\frac{t}{4\pi}\int_{|\omega|=1}c^{-4}E^\circ_{c, free}(x+ct\omega)\,d\omega\big)}
\\ \nonumber
& = & \frac{c^{-4}}{4\pi}\int_{|\omega|=1}E^\circ_{c, free}(x+ct\omega)\,d\omega+\frac{ct}{c^4}\int_{|\omega|=1}DE_{c,
free}(x+ct\omega)\omega\,d\omega
= {\cal O}(c^{-4})\end{aligned}$$ according to (\[freedat-def\]), [@schaeffer:86 Lemma 1].
Thus, combining (\[EDT-exp\]) and (\[dat1\])-(\[efree-dat\]) the representation formulas (\[Eext-rep\]) and (\[Ebound-rep\]) are proved.
### Representation of $B$ {#B-ref-sec}
We simply state the representation of $B$. $$\label{B-rep}
B = B_{ext}+B_{int}+B_{bound}+{\cal O}(c^{-4})$$ $$\begin{aligned}
\nonumber
B_{ext}(t, x) & = &
c^{-1}\int_{|z|\le ct}\big\{|z|^{-2}\bar{z}\times\big(j^R-t\partial_t{ j_0}-t^2/2{ j_0}\big)
\\ \nonumber
& & -1/2 c^{-2}\bar{z}\times\partial_t^2{ j_0}\big\}(0, x+z){\,dz}\\ \nonumber
B_{int}(t, x) & = & \int_{|z|\le ct}\int|z|^{-2} L_1(\bar{z}, c^{-1} p)f({\hat{t}(z)}, x+z, p){\,dp}{\,dz}\\ \nonumber
& & + c^{-2}\int_{|z|\le ct}\int|z|^{-1} L_2(\bar{z}, c^{-1} p) E (f^++f^-)({\hat{t}(z)}, x+z, p){\,dp}{\,dz}\\ \nonumber
B_{bound}(t, x) & = & B^R_{bound}\end{aligned}$$
Some explicit integrals and a lemma
-----------------------------------
We point out some formulas that have been used in the previous sections. For $z\in\R^3$ and $r>0$ an elementary calculation yields
$$\label{formeli}
\int_{|\omega|=1}|z-r\omega|^{-1}\,d\omega
=\left\{\begin{array}{c@{\quad:\quad}c}
4\pi r^{-1} & r\geq |z| \\ 4\pi |z|^{-1} & r\leq |z|
\end{array}\right. .$$
Differentiation w.r.t. $z$ gives $$\label{formelii}
\int_{|\omega|=1}|z-r\omega|^{-3}(z-r\omega)\,d\omega
=\left\{\begin{array}{c@{\quad:\quad}c}
0 & r>|z| \\ 4\pi |z|^{-2}\bar{z} & r<|z|
\end{array}\right. .$$ Similarly, $$\int_{|\omega|=1}|z-r\omega|\,d\omega
=\left\{\begin{array}{c@{\quad:\quad}c}
4\pi r+\frac{4\pi}{3}z^2 r^{-1} & r\geq |z| \\[1ex]
4\pi |z|+\frac{4\pi}{3}r^2 |z|^{-1} & r\leq |z|
\end{array}\right. ,$$ and thus by differentiation $$\label{formeliv}
\int_{|\omega|=1}|z-r\omega|^{-1}(z-r\omega)\,d\omega
=\left\{\begin{array}{c@{\quad:\quad}c}
\frac{8\pi}{3r}\,z & r>|z| \\[1ex] 4\pi\bar{z}
-\frac{4\pi}{3}r^2|z|^{-2}\bar{z} & r<|z|\end{array}\right. .$$ Finally, for $z\in\R^3\setminus\{0\}$ also $$\label{formelv}
\int |z-v|^{-1}|v|^{-3}v\,dv=2\pi\bar{z}$$ can be computed.
[**Acknowledgments:**]{} The author is indebted to M. Kunze, G. Rein, A. Rendall and H. Spohn for many discussions.
[99]{} . New York: Academic Press, 1975 On the Vlasov-Poisson limit of the Vlasov-Maxwell equation, in [*Patterns and Waves*]{}, Eds. [Nishida T., Mimura M.& Fujii H.]{}, Stud. Math. Appl., Vol. 18, pp. 369-383. Amsterdam: North-Holland Publishing Co., 1986. The reduced Vlasov-Poisson system with radiation damping, in preparation. Post-Newtonian approximation of the Vlasov-Nordström system, to appear in [*Commun. Partial Differ. Equations*]{}. , The Darwin approximation of the relativistic Vlasov-Maxwell system, [*Ann. Henri Poincaré*]{}[ **6**]{}, 283-308 (2005). Multipole radiation in a collisionless gas coupled to electromagnetism or scalar gravitation, preprint [arxiv:math-ph/0508057]{} Post-Newtonian gravitational radiation, in Schmidt B. G (ed.) [*Einstein’s Field Equations and Their Physical Implications*]{}, Springer, Berlin 2000. The non-relativistic limit of the Nordström-Vlasov system, [*Comm. Math. Sci. *]{}[**2**]{}, 19-34 (2004). Global small solutions of the Vlasov-Maxwell system in the absence of incoming radiation, [*Indiana Univ. Math. J. *]{}[**53**]{} No. 5, 1331-1364 (2004). Outgoing radiation from an isolated plasma, [*Ann. Henri Poincaré*]{}[**5**]{} No. 1, 189-201 (2004). Local existence of solutions of the Vlasov-Maxwell equations and convergence to the Vlasov-Poisson equation for infinite light velocity, [*Math. Methods Appl. Sci. *]{}[**8**]{}, 533-558 (1986). An analysis of the Darwin model of approximation to Maxwell’s equations, [*Forum Math. *]{}[**4**]{}, 13-44 (1992). Singularity formation in a collisionless plasma could occur only at high velocities, [*Arch. Rational Mech. Anal. *]{}[**92**]{}, 59-90 (1986). , 3rd edition. New York: Wiley, 1998. The Vlasov-Poisson system with radiation damping, [*Ann. Henri Poincaré*]{}[**2**]{}, 857-886 (2001). Simplified models of electromagnetic and gravitational radiation damping, [*Classical Quantum Gravity*]{}[**18**]{}, 3573-3587 (2001). Slow motion of charges interacting through the Maxwell field, [*Comm. Math. Phys. *]{}[**212**]{}, 437-467 (2000). Post-Coulombian dynamics at order $c^{-3}$, [*J. Nonlinear Sci. *]{}[**11**]{}, 321-396 (2001). The classical limit of the relativistic Vlasov-Maxwell system in two space dimensions, [*Math. Methods Appl. Sci. *]{}[**27**]{}, No. 3, 249-287 (2004). Propagation of moments and regularity for the 3-dimensional Vlasov-Poisson system, [*Invent. Math. *]{}[**105**]{}, 415-430 (1991). Global classical solutions of the Vlasov-Poisson system in three dimensions for general initial data, [*J. Differ. Equations*]{}[**95**]{}, 281-303 (1992). On the definition of post-Newtonian approximation [*Proc. R. Soc. Lond. A*]{}[ **438**]{}, 341-360 (1992). The Newtonian limit for asymptotically flat solutions of the Vlasov-Einstein system, [*Comm. Math. Phys. *]{}[**163**]{}, 89-112 (1994). The classical limit of the relativistic Vlasov-Maxwell system, [*Comm. Math. Phys. *]{}[**104**]{}, 403-421 (1986). Global existence of smooth solutions to the Vlasov-Poisson system in three dimensions, [*Commun. Partial Differ. Equations*]{}[**16**]{}, 1313-1335 (1991). . Cambridge: Cambridge University press, 2004.
[^1]: Supported in parts by DFG priority research program SPP 1095
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'At the faintest radio flux densities ($S_{1.4}<10$mJy), conflicting results have arisen regarding whether there is a flattening of the average spectral index between a low radio frequency (325 or 610MHz), and e.g. 1.4GHz. We present a new catalogue of 843MHz radio sources in the ELAIS-S1 field that contains the sources, their ATLAS counterparts, and the spectral index distributions of the sources as a function of flux density. We do not find any statistically significant evidence for a trend towards flatter spectral indices with decreasing flux density. We then investigate the spectral index distribution with redshift for those sources with reliable redshifts and explore the infrared properties. An initial sample of faint Compact Steep Spectrum sources in ATLAS is also presented, with a brief overview of their properties.'
author:
- |
K. E. Randall$^{1,2}$[^1], A. M. Hopkins$^{3}$, R. P. Norris$^{2}$, P.-C. Zinn$^{4}$, E. Middelberg$^{4}$,M. Y. Mao$^{2,3,5}$, R. G. Sharp$^{6}$\
$^{1}$Sydney Institute for Astronomy, School of Physics, The University of Sydney, NSW 2006, Australia\
$^{2}$CSIRO Astronomy and Space Science (CASS), P.O. Box 76, Epping NSW, 1710, Australia\
$^{3}$Australian Astronomical Observatory, PO Box 296, Epping, NSW, 1710, Australia\
$^{4}$Astronomisches Institut der Ruhr-Universit$\ddot{a}$t Bochum, Universit$\ddot{a}$tsstr. 150, 44801 Bochum, Germany\
$^{5}$School of Mathematics and Physics, University of Tasmania, Private Bag 27, Hobart, Tasmania 7001, Australia\
$^{6}$Research School of Astronomy and Astrophysics, The Australian National University, Cotter Road, Weston, ACT 2611, Australia
date: 'Accepted 2011 December 20. Received 2011 December 16; in original form 2011 October 17'
title: Spectral index properties of milliJansky radio sources
---
\[firstpage\]
catalogues — galaxies: active — galaxies: evolution — radio continuum: galaxies.
Introduction
============
The desire to understand the properties of the faintest radio source populations has led to numerous surveys pushing to ever fainter levels. There are now a large number of deep and wide area radio surveys available, such as the Australia Telescope Large Area Survey [ATLAS; @ray; @Middelberg08], the Cosmic Evolution Survey [COSMOS; @scoville; @smolcic], the ATESP Survey [@prandoni; @prandoni1; @prandoni2], the Phoenix Deep Survey [@phoenix], and many others [@seymour; @owen; @owen1; @ibar; @ibar1]. The bright radio source population ($S_{1.4}>10$mJy) is well studied, and is predominantly composed of active galactic nuclei (AGN) [@condon; @grup; @maglio; @georg; @afonso1]. At fainter flux densities, star-forming galaxies (SFGs) begin to dominate the radio source population, particularly into the $\mu$Jy regime [@windhorst85; @srccnts; @afonso; @seymour]. Understanding these faint radio source populations is essential for understanding galaxy evolution, the role of star formation, AGN and the relationship between the two. The use of multiwavelength data to complement radio surveys is vital in order to distinguish between AGN and star formation processes as the origin of the radio emission down to the faintest flux densities, although there is likely also a composite population at these faint levels [@hill99; @hill01]. The relative proportions of these two populations will have an effect on the average radio spectral index ($\alpha$)[^2] as a function of flux density, and at the faintest flux densities, core-dominated AGN may be more prevalent, and may flatten the average spectral index to $\alpha>-0.7$.
Conflicting evidence has arisen over the nature and properties of sources at frequencies below 1.4GHz, particularly of their spectral index properties, and whether there is a flattening of the average spectral indices for faint ($S_{1.4}<10$mJy) radio sources. @prandoni2 [@prandoni4] found sources with fluxes less than a few milliJansky had an average spectral index which was flatter than that of the brighter radio sources ($\alpha\sim-0.7$). Similarly, @owen found a flattening of the average spectral index between 325MHz and 1.4GHz for 1.4GHz selected radio sources with $S_{1.4}<10$mJy, and angular sizes $>3''$. They found though, that the spectral indices steepen again at the faintest flux density end of the 20cm survey ($\sim\,0.5\,$mJy). In contrast, in the deepest radio field, the Lockman Hole [@ibar], no flattening of the spectral indices between 610MHz and 1.4GHz was seen for flux densities S$_{1.4}>100\mu$Jy.
A flattening of the average spectral indices implies that there is a flat or inverted spectrum population of sources at these milliJansky flux densities, at fluxes fainter than where the steep spectrum star-forming population emerges ($\sim0.5\,$mJy). Investigating the milliJansky and microJansky radio source populations at different frequencies allows us to investigate this suspected flattening of the average spectral indices, and the population of sources causing the flattening, particularly if we can fill in the frequency regime from very low frequencies ($\sim100$MHz) up to 1.4GHz. It is necessary to understand the spectral index properties of these faint radio sources for many reasons, such as the $z$-$\alpha$ relation, used to find the most distant radio galaxies [@debreuck04; @klamer; @ishwara], identifying young radio AGN, such as Gigahertz Peaked Spectrum (GPS) and Compact Steep Spectrum (CSS) sources [@Review; @young; @morganti] or determining the emission mechanism in radio galaxies, whether it is from star formation, or the central AGN [@prandoni2].
Here we present radio observations at 843MHz in the ELAIS-S1 field, for which we have complementary 1.4 and 2.3GHz data from ATLAS that we can use to investigate the possible spectral index flattening at a frequency nearer to the ubiquitous 1.4GHz radio surveys. Section \[sec:elais\] briefly reviews the existing multiwavelength data covering the ELAIS-S1 region and ATLAS, and in Section \[sec:data\] we describe our observations, the data reduction, and the cross-matching process to the ATLAS 1.4 and 2.3GHz catalogue. We present the catalogue in Section \[sec:catalogue\]. Our results and analysis are explored in Section \[sec:results\], and a new faint sample of candidate CSS sources is presented in Section \[sec:props\], discussing the initial selection and properties of these sources. Our results and plans for future work are discussed in Section \[sec:discussion\] and our conclusions are presented in Section \[sec:concl\]. Throughout this analysis, unless otherwise noted, we use the cosmological parameters, $\Omega_{M}=0.27$, $\Omega_{\Lambda}=0.73$ and $H_0=71$kms$^{-1}$Mpc$^{-1}$ [@wmap].
The ELAIS-S1 Region {#sec:elais}
===================
The European Large Area ISO Survey - South 1 Region (ELAIS-S1) has been a target for many multiwavelength surveys and targeted observations over the last decade. The ELAIS-S1 field covers $2^{\circ}\times2^{\circ}$, centred on RA = $00{^h}34{^m}44^{s}.4$ and Dec=$-43^{\circ}28'12''.0$ (J2000.0). Observations covering the ELAIS-S1 field include radio imaging, optical imaging and spectroscopic redshifts, infra-red observations (near, mid and far-infrared), UV, and X-ray observations. Part of the attraction for the multiwavelength surveys in this region is that this field has the lowest Galactic 100$\mu$m cirrus emission in the southern sky [@schlegel], including the absolute minimum.
Multiwavelength data in ELAIS
-----------------------------
### ISO
The field was first observed as part of a deep, wide-angle survey with the Infrared Space Observatory (ISO) at 15 and 90$\mu$m [@oliver]. Follow-up observations were done at 6.7$\mu$m, and the catalogue covering the S1 Region at these three wavelengths is discussed in @rowan. The complete 15$\mu$m catalogue contains all the sources from the ELAIS regions: N1, N2, N3, S1, S2, and was finalized by @vaccari. The three bands of this survey have rms noise levels of 1.0, 0.7 and 70 mJy respectively for the 6.7, 15 and 90$\mu$m bands. Photometric uncertainties for all bands were $\sim10\%$, and the astrometric accuracy is $\sim0.5''$.
### Spitzer/SWIRE {#sec:swire}
The ELAIS-S1 field was observed by the Spitzer Space Telescope as part of the largest Legacy Program, the Spitzer Wide-area Infra-Red Extragalactic survey [SWIRE; @lonsdale; @lonsdale1]. SWIRE aimed to trace the evolution of dusty SFGs, AGN, and evolved stellar populations out to a redshift of $z\sim$3, by imaging large areas of sky in seven different infrared bands. In combination with optical imaging, this would allow SED modeling and exploration of galaxy evolution with environment. The seven imaging bands of SWIRE used two different instruments aboard Spitzer, the Multiband Imaging Photometer [MIPS; @rieke], and the Infrared Array Camera [IRAC; @fazio], covering $\sim7$deg$^2$ over the ELAIS-S1 region. The 24$\mu$m MIPS imaging reaches a 5$\sigma$ sensitivity of 350$\mu$Jy, and the IRAC bands of 3.6, 4.5, 5.8 and 8.0$\mu$m have 5$\sigma$ sensitivities of 3.7, 5.3, 48 and 37.75$\mu$Jy respectively.
### ESIS
The ESO-Spitzer Imaging extragalactic Survey (ESIS) is the optical follow-up to SWIRE, consisting of optical imaging in the B, V, R bands with the Wide Field Imager [WFI; @berta], and I and z band imaging with VIMOS [@berta1]. Currently, the VIMOS observations have been completed, but the WFI observations are not yet fully processed. The BVR observations described in @berta cover 1.5deg$^2$, in the central region of the ELAIS-S1 field. The catalogue of BVR sources is 95% complete to 25$^{m}$ in B and V, and 24.5$^{m}$ in R. 132712 sources are included in this catalogue, with an rms uncertainty of $\sim0.15''$ for the coordinates of the sources. The VIMOS data covers $\sim4$deg$^2$ in the I band and $\sim1$deg$^2$ in the z band, resulting in a completeness of 90% at 23.1$^{m}$ in the I band, and 22.5$^{m}$ in the z band. Over 300,000 sources were catalogued in the I band, and over 50,000 in the z band, with an rms of $\sim0.2''$ in both bands.
### Previous Radio Observations
Prior to the ATLAS observations (detailed below in §\[sec:atlas\]), @grup [G99] observed the ELAIS-S1 field using the Australia Telescope Compact Array (ATCA) in 1997, covering $\sim7$deg$^2$. The observations consisted of a mosaic of 49 different pointings, resulting in an image rms of $\sim80\mu$Jy at 1.4GHz, and a catalogue of 581 radio sources. The lowest flux density sources were catalogued down to a 5$\sigma$ level of 0.2mJy in the centre of the field, and 0.4mJy in the remaining area.
### X-ray Observations
BeppoSAX, an X-ray satellite, first observed 40% of the ELAIS-S1 field in 1999 [@alexander]. These observations reached a sensitivity of $\sim10^{-13}$ergcm$^{-2}$s$^{-1}$ in the 2$-$10keV range, covering $\sim1.7$deg$^2$. The central $\sim0.6$deg$^2$ were then surveyed by XMM-Newton in four deep pointings [@puccetti]. The XMM observations were taken in both soft and hard X-rays, and each pointing had a net exposure time of $\sim60$ks. A total of 478 sources were detected, with 395 in the soft X-ray band (0.5$-$2keV) and 205 in hard X-ray (2-10keV). The flux limits are $\sim5.5\times10^{-16}$ergcm$^{-2}$s$^{-1}$ and $\sim2\times10^{-15}$ergcm$^{-2}$s$^{-1}$ respectively for the soft and hard bands.
### UV Observations
The ultraviolet Galaxy Evolution Explorer [GALEX; @martin] observed the ELAIS field in two of its observing modes; the Deep Imaging and Wide Spectroscopic Survey modes [@burgarella]. The bands observed were centred on 153 and 231nm, and the current data release (GR6[^3]) covers 70% of the sky in the imaging mode, and has 61439 spectroscopic sources. The ELAIS field was observed for $\sim9$hours in the spectroscopic survey mode, and $\sim3$hours in the imaging mode.
In the analysis below we focus on the infrared data as the primary complement to our radio data, although in future work we will also take advantage of the X-ray and UV measurements.
The Australia Telescope Large Area Survey (ATLAS) {#sec:atlas}
-------------------------------------------------
ATLAS is the widest, deep radio survey to date [@ray; @Middelberg08 N06, M08], covering $\sim7$deg$^2$ over two fields, ELAIS-S1, and the Chandra Deep Field South (CDFS). The rms of the 1.4GHz imaging data is currently 30$\mu$Jy, and the aim is to achieve an rms of 10$\mu$Jy across both fields [@banfield]. ATLAS observations have been completed at 1.4 and 2.3 GHz with ATCA from 2006 to 2010, with one data release in 2008, and subsequent data releases to begin in late 2011 [@chales; @banfield]. ATLAS has many scientific goals, primarily focussed on investigating the evolution of galaxies and AGN. Specific goals include distinguishing between AGN and SFGs and determining the contribution of each to a given galaxy’s luminosity, searching for high-$z$ radio galaxies to trace the formation of clusters at high redshift [@minnie], and finding new types of rare sources [N06, M08, @ifrs]. The ATLAS survey regions were chosen because of the large amount of multiwavelength data covering these two fields, including near and far-infrared, and deep optical data, and in some areas, X-ray and UV. We aim to create the most comprehensive multiwavelength survey of faint radio sources to date. ATLAS currently contains $\sim\,2000$ radio sources, and we estimate we will have $\sim\,16000$ radio sources following final analysis of new observations completed in 2010 from ATCA with the Compact Array Broadband Backend [CABB; @cabb]. These new observations consist of 1000 hours of integration time, and cover a bandwidth of 0.5GHz centred on 1.4GHz.
In ELAIS-S1, 1276 radio sources have been catalogued, comprising 1366 radio components, at 1.4GHz [@Middelberg08 M08]. The 2.3GHz image has a lower sensitivity, with an rms noise level of $\sim60\mu$Jy in the central $\sim1$deg$^2$ and $\sim100\mu$Jy over the entire field. We find only 576 radio sources have a well-defined counterpart at 2.3GHz, mainly because the resolution is $\sim3$ times coarser at this frequency than at 1.4GHz as a consequence of the compact configuration of ATCA used for these observations. For the purposes of this paper and catalogue, we use the cross-matches by @zinn [Z11], where we take the 1.4GHz sources matched to the poorer resolution 2.3GHz data, and extract the fluxes from these images for each source. M08 also compared their flux densities and positions to G99 by repeating their source extraction on the G99 1.4GHz radio image. The differences in radio positions were determined to be negligible, but the flux densities of M08, while within $\sim\,3$% of the earlier measurements, were consistently higher than G99.
Observations, data reduction and cross-matching {#sec:data}
===============================================
To obtain low-radio frequency (843MHz) data within ATLAS, we used the Molonglo Observatory Synthesis Telescope [MOST; @mills; @robertson]. The MOST is a 1.6km long cylindrical paraboloid reflector, with its axis aligned in an east-west direction. It operates at a frequency of 843MHz, and has a field of view of $2{\fdg}7cosec(\delta)\times2{\fdg}7$deg$^2$ [@large; @bock]. We have 31 separate 12 hour observations taken with MOST, which were combined into a single image. The CDFS field (centred at RA = $03{^h}32{^m}28^{s}$ and Dec=$-27^{\circ}48'31\farcs0$ (J2000.0)) was not observed, due to strong radio frequency interference (RFI) that increased with lower elevation.
Processing
----------
The data reduction pipeline for MOST data is highly automated and reliable. Unfortunately, the telescope was subject to severe RFI at the time of our observations. The process to remove the RFI required manual identification and excision of the affected data. Once completed, we were able to re-run the data pipeline, and produce final cleaned images. Seven of our 31 observations were subject to RFI that we were unable to remove completely, and these were not included in our final imaging. The remaining 24 images were added together in MIRIAD, first by re-gridding the images to ensure a common astrometry and pixel grid, and then using the task <span style="font-variant:small-caps;">imcomb</span> to combine the images. The final image has an rms sensitivity of $\approx0.6$mJy, a higher value than expected due to the presence of low-level RFI that could not be excised. We catalogue sources down to a $5\sigma$ level of 3mJy. The final image is shown in Figure \[fig:mostimage\] with the approximate borders of the ATLAS 1.4 and 2.3GHz mosaics overlaid.
The MOST image was subject to two kinds of artifacts, grating rings and radial spokes. Due to the nature of the telescope design and observing mode, these artifacts are technically difficult to remove. The artifacts remain in our image, and due care was taken to ensure all objects in our final catalogue were not spurious sources associated with the artifacts. An in-depth discussion of these artifacts is given in @SUMSS2.
Source-finding
--------------
Source finding for the final MOST image was done using the task <span style="font-variant:small-caps;">sfind</span> [@sfind] in MIRIAD. This produces a catalogue of source positions, peak and total flux density and errors, plus the attributes of the Gaussian fits for each source. The final catalogue contains 325 radio sources, after removal of $\approx100$ spurious sources associated with the grating rings and radial spokes. The spurious sources were identified and removed by visual inspection (most lay directly on a strong radial spoke or diffraction ring), and by comparison to the Sydney University Molonglo Sky Survey [SUMSS; @bock; @SUMSS2] images (see §\[sec:fluxdist\] for details). Each MOST ATLAS source was identified in the SUMSS images with a low signal-to-noise source that was below the detection level of the SUMSS catalogue. The MOST image is larger than the ATLAS survey region, so our catalogue encompasses a larger number of sources in total than we subsequently analyse together with the other ATLAS data.
We use the process outlined in @phoenix to calculate the errors associated with our measured flux densities, and use Equation 5 of @phoenix with one small modification:\
$$\sigma_I=I \sqrt{2.5 \frac{\sigma^2}{I^2} + 0.05^2},
\label{eq:error}$$ where $I$ is the total integrated flux density, $\sigma$ is the rms error in the image at the source location, and $\sigma_I$ is the total error on the integrated flux density. We use $0.05^{2}$ as the sum of the squares of the instrumental and pointing errors as given in M08, instead of $0.01^{2}$ as given in @phoenix. We have used the more conservative estimate of these errors as given by M08, due to the calibration accuracy of MOST being comparable to that of ATCA.
Cross-matching to ATLAS {#sec:crossmatch}
-----------------------
The final MOST catalogue was positionally cross-matched to the combined 1.4 and 2.3GHz catalogue produced by Z11, that included the relevant SWIRE and optical data. As mentioned previously in §\[sec:atlas\], only 576 of the 1276 1.4GHz radio sources in ELAIS have a reliable single 2.3GHz counterpart, whilst 460 have no counterpart, and 240 are blended or poorly fitted sources (where the 2.3GHz counterpart encompasses multiple 1.4GHz sources). From our cross-matching, 105 MOST sources were matched to single sources in the Z11 catalogue. Another 30 MOST sources were matched to confused or blended sources from Z11, and 31 MOST sources were found to have multiple isolated Z11 sources matched to a single MOST source.
Resolution Matching {#sec:convolution}
-------------------
The resolution of the MOST image, $62''\times\,43''$, is much coarser than the resolution of the 1.4GHz image ($10''\times7''$), and the 2.3GHz image ($33''\times20''$). To determine accurate spectral indices across the three frequencies, it is necessary to convolve the ATLAS 1.4 and 2.3GHz images to the same size as the MOST beam to ensure the recovered flux for sources at all frequencies includes any emission extended on scales up to those consistent with the MOST beam. Figure \[fig:resol\] shows an example of a complex source from the ATLAS 1.4GHz catalogue in greyscale in the three radio frequencies, to highlight the differences in resolution. In Figure \[fig:resol1\], the same source is shown after convolving the 1.4 and 2.3GHz images to the same resolution as the MOST image. Although convolution removes most of the small scale structure visible in the original ATLAS 1.4GHz image, it ensures we are detecting emission from the same spatial region from each source.
For consistency, we used <span style="font-variant:small-caps;">sfind</span> to detect sources within the convolved 1.4 and 2.3GHz images. This allows us to measure errors on our flux densities that are consistent for the three images, which is important for producing accurate spectral indices and associated uncertainties. Rather than rejecting blended sources entirely, we account for them carefully in our catalogue. These are single or point sources in the 843MHz MOST image (and the 1.4 and 2.3GHz convolved ATLAS images) that encompass multiple ATLAS 1.4GHz sources from the M08 catalogue (Figure \[fig:complex\]). The M08 flux densities for these sources reported in our catalogue are the summed values of the individual M08 1.4GHz flux densities. As the contribution to the total flux density from each individual component within a blended source cannot be determined, we treat them as a single entity for the purpose of the current analysis, even though they may be physically unrelated. We include these sources in our analysis for completeness, but caution that their spectral indices should not be considered as accurate estimates of those for the underlying source components.
### Flux Density Comparisons and Corrections
To ensure our measured 1.4 and 2.3GHz flux densities from the convolved, coarser-resolution, images were robust, we compared these to the flux densities in M08 and Z11. We find that our measurements in the images convolved to a resolution consistent with the MOST image tend to systematically overestimate the 1.4GHz flux densities compared to M08, and underestimate the 2.3GHz flux densities compared to Z11.
The issue at 2.3GHz arises from negative CLEAN bowls around sources in the original image. These artifacts give rise, after convolution, to our observed systematic decrease in flux density. The 2.3GHz beam size ($33''\times20''$) is close to that of our MOST images ($62''\times\,43''$), and the bulk of the sources are unresolved, meaning that there is likely to be little flux missed on extended scales. Consequently, we choose to use the flux densities estimated by Z11 in the original 2.3GHz image for our spectral index estimates.
At 1.4GHz, the difference in resolution from the MOST image is sufficient that we need to use the flux densities from the convolved image. The flux density overestimate compared to M08, which is limited to the fainter sources ($S_{843}<10$mJy), is typically of the order of 10%. This exists for clearly unresolved sources, which should be identical before and after convolution. This issue is associated with side-lobes from the radio sources in each individual telescope pointing, below the level to which the image has been CLEANed, being summed in the convolved image. While this is a small effect, with a minimal impact on our derived spectral indices (quantified below), we can make an empirical correction that minimizes any impact further still. This is implemented through a least-squares fit of our convolved flux densities against those of M08, and scaling our convolved flux densities using this fit to be consistent with those of M08. This accounts for the flux-density dependence of the imaging systematics while retaining contributions from any real extended flux components, and at the same time, minimizing any possible overestimate of the flux density. This correction typically results in a change of only $\alpha_{fit}\sim0.02$.
The impact on our derived two-point spectral index estimates of potential remaining flux density uncertainties at 1.4 and 2.3GHz of $\sim10\%$ is $\sim$0.1. This uncertainty is included in our estimates of spectral index errors below, by being added in quadrature with the flux density errors (given by Equation \[eq:error\]).
### Blended sources
Another effect of the convolution is that most objects classified as radio doubles, triples, or core-jet morphology by M08, appear as one MOST object (single or blended). Only two core-jet morphology sources appear slightly elongated in the direction of the jet in the MOST and convolved 1.4 and 2.3GHz images.
Due to the lower resolution of the MOST image, there are 61 instances where there are ATLAS sources within the beam for a MOST radio source. These 61 MOST sources correspond to a total of 144 ATLAS M08 1.4GHz sources. Of the 61, 30 sources were classified as blended by Z11 due to having multiple 1.4GHz sources within the 2.3GHz beam ($33''\times20''$). The remaining 31 are classified as blended in this paper due to multiple 1.4 or 2.3GHz sources within the MOST beam. The cross-matched 1.4GHz ATLAS source (positionally closest to the MOST radio position) is generally bright while the second, blended, source contributing to the flux density is typically much fainter ($S_{1.4}<1$mJy). As a consequence, the majority of such blended sources are likely to have little contribution to the MOST flux density from the secondary component, and the spectral index estimate for the blended sources can be considered as that for the primary ATLAS counterpart, even though the uncertainties on this estimate will clearly be larger than the formal uncertainties provided in our catalogue. Blended sources are flagged in the catalogue.
For blended sources in our convolved images, which were not classified as blended sources by Z11, the flux densities from separate sources in the ATLAS Z11 catalogue are summed. A source is defined as being part of a blended object if the 1.4GHz radio position lay within an ellipse the size of the Gaussian used to find the total flux density in the convolved image. An example of a blended source is shown in Figure \[fig:complex\] where the cross-matched ATLAS counterpart is in the centre of the MOST source, and the blended object is within the MOST synthesized beam. There is one exception to this, shown in Figure \[fig:complex1\], where the blended object is outside the Gaussian fit, but the radio emission clearly extends from S120 into S107.
![An example of a blended source. The contours are from the convolved 1.4GHz image (at levels of 5, 10 and 20mJy), and the greyscale is the 843MHz image. The white crosses indicate the 1.4GHz radio position of the two sources, where the object in the centre of the image is the MOST cross-matched source, and the blended object indicated by the other cross. \[fig:complex\]](f4.eps)
The MOST ATLAS source Catalogue {#sec:catalogue}
===============================
The total catalogue consists of 325 MOST sources, limited to sources above the 5$\sigma$ cutoff of 3mJy, of which 166 have an ATLAS 1.4 or 2.3GHz counterpart. The remaining 159 sources are outside the ATLAS survey area. In total, there are 310 ATLAS sources in our catalogue, with 205 classified as blended sources. The remaining 105 ATLAS sources correspond to single MOST sources. An extract of the catalogue in shown in Table \[table:catalogue\], and the full version is available online. The MOST radio position is listed first, followed by the flux density and error at 843MHz. For MOST sources with a single ATLAS counterpart, the ATLAS source name (e.g. ATELAIS J002905.22$-$433403.9), and ID (e.g. S100) is listed; for blended sources the given ATLAS IDs correspond to all Z11 sources cross-matched to the MOST source, and the ATLAS source name corresponds to the first ATLAS ID listed. All ATLAS source names and IDs are from M08. The corrected convolved 1.4GHz and Z11 2.3GHz flux density measurements and associated errors are given, along with the three spectral indices (described in §\[sec:specs\] below).
Results and Analysis {#sec:results}
====================
Flux Density Distribution {#sec:fluxdist}
-------------------------
The flux density distributions (Figure \[fig:fluxes\]) do not show any major differences between the single, blended, and non-ATLAS sources. The location of the median of the entire catalogue is shown in Figure \[fig:fluxes\]a. The distributions appear consistent with other faint radio samples, such as the original ATLAS catalogues [@ray; @Middelberg08].
![Distribution of $\alpha_{fit}$ for our catalogue. \[fig:alphahisto\]](f8.eps)
The Sydney University Molonglo Sky Survey [SUMSS; @bock; @SUMSS2] is an 843MHz survey with MOST, that covers the sky south of $\delta<-30^{\circ}$ with $|\textit{b}|>10^{\circ}$. SUMSS has similar resolution and sensitivity to the National Radio Astronomy Observatories (NRAO) Very Large Array (VLA) Sky Survey [NVSS; @NVSS]. SUMSS contains 211063 radio sources, with an rms of $\sim1$mJy. Our observations probe sources fainter than SUMSS by a factor of $\sim2$. Of our catalogue, 178 sources are cross-matched to SUMSS sources. Only one source in SUMSS does not have a counterpart in our observations, as it is located directly on an artifact in our image. Our measured flux densities, positions, and radio differential source counts (see §\[sec:srccnts\] for details) are consistent with those of SUMSS, with only two outliers, likely due to intrinsic source variability. It is known that a few percent of mJy radio sources are variable on the timescale of years [@oort]. Attributing our outliers to variability is consistent with this work, as our data was taken over several years.
Spectral Index Distributions and Properties {#sec:specs}
-------------------------------------------
We have calculated spectral indices for all sources for which flux densities are available at two or three frequencies. $\alpha_{fit}$ is a three-point power-law fit, and $\alpha^{1.4}_{0.843}$ and $\alpha^{2.3}_{1.4}$ are two-point power-law fits for the respective frequencies. Distributions of these spectral indices are given in Figures \[fig:alphaflux\] and \[fig:alphahisto\]. Spectral index $\alpha^{1.4}_{0.843}$ versus spectral index $\alpha^{2.3}_{1.4}$ is also shown in Figure \[fig:specspec\], which indicates that most of our sources are steep-spectrum, with a small fraction of inverted, peaked, and flat spectrum objects, discussed further below. A large proportion of steep-spectrum sources are seen because of the low radio frequency selection, and steep spectrum sources are brighter at lower frequencies. The median spectral index $\alpha_{fit}$ for the single sources is shown on each of the panels in Figure \[fig:alphaflux\]. Only single sources are included in this calculation. The spectral index limits are shown in Figure \[fig:alphaflux\]b,c arising from the limiting flux density of the least-sensitive frequency in the relevant two-point spectral index calculation. Although the three-point spectral index
[|ccccccccccccccc|]{} 0:28:45.438&$-$42:51:39.46&15.12&1.5&...&...&...&...&...&...&...&...&...&...&...\
0:28:54.063&$-$43:12:18.30&6.80&1.8&...&...&...&...&...&...&...&...&...&...&...\
0:28:56.980&$-$42:18:15.37&40.52&3.2&...&...&...&...&...&...&...&...&...&...&...\
0:29:05.054&$-$43:34:07.05&11.34&1.3&ATELAIS J002905.22$-$433403.9&S749&p&9.01&1.16&5.11&0.34&$-$0.81&0.06&$-$0.45&$-$1.14\
0:29:05.968&$-$44:42:00.59&14.18&1.6&...&...&...&...&...&...&...&...&...&...&...\
0:29:09.167&$-$43:44:02.10&11.13&1.3&ATELAIS J002909.26$-$434356.3&S617&p&12.52&1.27&4.63&0.27&$-$0.93&0.09&0.23&$-$2.00\
0:29:13.163&$-$44:52:23.32&5.33&1.4&...&...&...&...&...&...&...&...&...&...&...\
0:29:15.437&$-$43:26:36.78&9.51&1.3&ATELAIS J002915.52$-$432638.3&S868&p&5.45&1.10&3.56&0.25&$-$0.97&0.03&$-$1.10&$-$0.86\
0:29:21.471&$-$42:55:45.26&44.87&2.6&...&...&...&...&...&...&...&...&...&...&...\
0:29:25.667&$-$44:08:25.68&15.37&1.8&ATELAIS J002925.66$-$440822.8&S293, S304&b&9.12&1.13&3.51&0.35&$-$1.49&0.06&$-$1.03&$-$1.92\
0:29:27.845&$-$43:16:15.59&6.61&1.2&ATELAIS J002927.69$-$431614.4&S1014&p&3.85&1.16&2.03&0.18&$-$1.18&0.01&$-$1.06&$-$1.29\
0:29:36.620&$-$42:25:41.20&64.07&8.4&...&...&...&...&...&...&...&...&...&...&...\
0:29:37.206&$-$42:34:18.69&5.05&1.4&...&...&...&...&...&...&...&...&...&...&...\
0:29:38.084&$-$44:23:21.78&42.38&2.7&ATELAIS J002939.19$-$442319.3&S100, S101.1&b&34.1&2.29&18.82&1.88&$-$0.74&0.06&$-$0.43&$-$1.20\
0:29:43.820&$-$42:37:47.77&17.83&1.7&...&...&...&...&...&...&...&...&...&...&...\
0:29:45.687&$-$43:21:50.03&24.18&1.6&ATELAIS J002945.64$-$432149.5&S943&p&16.25&1.41&10&0.52&$-$0.87&0.06&$-$0.78&$-$0.98\
0:29:46.493&$-$43:15:57.34&41.67&2.4&ATELAIS J002946.52$-$431554.5&S1018&p&27.89&1.78&18.49&0.94&$-$0.81&0.05&$-$0.79&$-$0.83\
0:29:47.622&$-$44:16:15.59&6.62&1.1&ATELAIS J002947.37$-$441607.0&S181&p&3.90&1.65&1.75&0.15&$-$1.33&0.01&$-$1.04&$-$1.62\
0:29:50.398&$-$44:05:48.45&7.16&1.1&ATELAIS J002949.89$-$440541.4&S345, S342, S339&b&5.26&1.07&5.13&0.51&$-$0.31&0.03&$-$0.61&$-$0.05\
0:29:51.452&$-$43:45:28.25&8.23&1.3&ATELAIS J002951.48$-$434528.0&S598&p&4.93&1.13&3.23&0.18&$-$0.93&0.02&$-$1.01&$-$0.85\
0:29:53.062&$-$42:50:17.87&5.16&2.0&...&...&...&...&...&...&...&...&...&...&...\
0:30:03.253&$-$42:18:22.97&20.81&8.4&...&...&...&...&...&...&...&...&...&...&...\
0:30:04.749&$-$42:09:58.88&19.10&5.9&...&...&...&...&...&...&...&...&...&...&...\
0:30:10.595&$-$44:09:12.06&6.49&1.1&ATELAIS J003010.82$-$440907.3&S288&p&7.60&0.72&8.96&0.46&0.32&0.04&0.31&0.33\
0:30:17.439&$-$42:24:47.35&436.20&22.4&...&...&...&...&...&...&...&...&...&...&...\
0:30:18.530&$-$45:26:42.05&12.79&2.5&...&...&...&...&...&...&...&...&...&...&...\
0:30:18.793&$-$44:04:35.34&13.70&1.2&ATELAIS J003019.22$-$440438.3&S355&p&9.83&0.78&5.57&0.23&$-$0.89&0.08&$-$0.65&$-$1.14\
0:30:20.948&$-$43:39:44.51&69.27&3.7&ATELAIS J003020.95$-$433942.8&S694&p&51.99&2.55&33.92&1.7&$-$0.66&0.04&$-$0.57&$-$0.86\
0:30:21.663&$-$45:05:09.55&16.15&3.3&...&...&...&...&...&...&...&...&...&...&...\
0:30:22.709&$-$42:37:02.29&14.25&2.4&...&...&...&...&...&...&...&...&...&...&...\
0:30:27.017&$-$42:35:14.14&7.34&3.1&...&...&...&...&...&...&...&...&...&...&...\
0:30:29.096&$-$42:13:50.96&18.19&6.2&...&...&...&...&...&...&...&...&...&...&...\
0:30:34.835&$-$45:29:47.20&15.73&4.1&...&...&...&...&...&...&...&...&...&...&...\
0:30:35.339&$-$44:37:11.21&3.69&1.1&ATELAIS J003035.77$-$443707.2&S18, S19&b&6.53&3.21&2.07&0.21&$-$0.68&0.01&1.13&$-$2.31\
0:30:35.937&$-$43:23:39.75&8.95&3.1&ATELAIS J003035.03$-$432341.6&S926, S923, S930, S930.1&b&5.03&0.62&4.76&0.48&$-$0.32&0.02&$-$1.14&$-$0.11\
0:30:38.957&$-$44:09:56.34&3.95&0.9&ATELAIS J003039.03$-$441000.0&S279&p&3.04&0.62&1.23&0.13&$-$1.29&0.01&$-$0.52&$-$1.82\
0:30:39.015&$-$45:07:07.04&33.29&2.7&...&...&...&...&...&...&...&...&...&...&...\
0:30:39.621&$-$44:42:01.21&28.43&1.9&ATELAIS J003039.68$-$444159.5&S7&p&26.00&3.44&17.02&0.95&$-$0.49&0.05&$-$0.18&$-$0.85\
0:30:40.860&$-$43:23:40.55&11.17&2.4&ATELAIS J003042.10$-$432335.4&S923, S930, S930.1&b&6.02&0.64&5.47&0.55&$-$0.53&0.03&$-$1.22&$-$0.19\
0:30:42.041&$-$43:18:42.39&4.52&1.0&ATELAIS J003041.88$-$431840.7&S987&p&3.26&0.61&2.12&0.13&$-$0.77&0.01&$-$0.64&$-$0.87\
\
<span style="font-variant:small-caps;">Notes.</span>– $\alpha_{fit}$ is the spectral index fitted across the three flux density measurements at 0.843, 1.4 and 2.3GHz. Type (Column 7) refers to whether the source is a single point source (p) at all frequencies, or a blended source (b).\
$\alpha_{\rm fit}$ is shown in the figure, the limit associated with the two-point spectral index calculation provides a clear indication of where we are selection-limited against particularly steep or flat spectrum sources. Distributions of $\alpha^{1.4}_{0.843}$, split into three flux density bins are also shown in Figure \[fig:fluxbins\], analogous to Figure 8 of @SUMSS2. A flattening of the spectral index $\alpha^{1.4}_{0.843}$ is suggested by Figure \[fig:fluxbins\]; however this is primarily due to different numbers of sources in each flux density bin. This is further discussed in §\[sec:specflux\].
Radio spectral classifications
------------------------------
Figure \[fig:specspec\] shows our sample of sources, in a radio colour-colour diagram [@kesteven; @at20gcat], that we now consider in four classes:
1. Steep-spectrum objects, with a steep radio spectrum from 843MHz to 2.3GHz (the lower-left quadrant of Figure \[fig:specspec\]),
2. Peaked sources, where we see the radio spectrum turn over between 843MHz and 2.3GHz (the lower-right quadrant of Figure \[fig:specspec\]),
3. Inverted (or rising) sources, where the radio flux density increases with increasing frequency (the upper-right quadrant of Figure \[fig:specspec\]), and,
4. Upturn sources, that have an upturn in their radio spectrum (the upper-left quadrant of Figure \[fig:specspec\]).
This distribution highlights the fact that our sample is dominated by steep-spectrum sources. The statistics are given in Table \[table:class\], split into single and blended sources, noting that any statistics from the blended sources are not reliable. In comparison, the Australia Telescope 20 GHz Survey [AT20G; @at20gcat], a high-frequency-selected sample (20GHz), produced a colour-colour plot contained 3763 sources, with 14% inverted (rising) spectrum sources, 57% steep-spectrum objects, 21% peaked spectrum sources, and 8% sources with an upturn in their spectra.
As indicated by Table \[table:class\], 8% of our MOST ATLAS sample appear to be possible Gigahertz Peaked Spectrum (GPS) sources (see §\[sec:CSS\] for a description). @Review suggest that $\sim10$% of bright radio sources are GPS sources, whereas @randall found less than 1% of their sample to be GPS sources. Although both these samples are not complete, it is interesting to note that at faint radio fluxes, the proportion of GPS sources appears to be similar to the bright sample of @Review rather than the unbiased bright sample of @randall. This sample of sources will be explored further in a future paper, @randall1.
Interestingly, there is one very steep spectrum single source, S1256, which has $\alpha^{1.4}_{0.843}=-2.92$, that is not detected in the 2.3GHz image. This object is discussed in more detail in §\[sec:s1256\].
![Spectral index $\alpha^{2.3}_{1.4}$ versus $\alpha^{1.4}_{0.843}$ for MOST ATLAS sources. The black symbols in each outer corner of the plot represent the type of source in each quadrant of the plot, e.g. steep, inverted, peaked or upturn. \[fig:specspec\]](f9.eps)
![Distribution of $\alpha^{1.4}_{0.843}$, in three flux density bins, 0-10mJy (a), 10-20mJy (b), and $>20$mJy (c). The solid vertical lines are the median spectral index $\alpha^{1.4}_{0.843}$ for each flux density bin for the single sources only. \[fig:fluxbins\]](f10.eps)
---------------- -------- --------- --------- ------------
Type Total Point Blended Percentage
Number Sources Sources
Steep-Spectrum 143 89 54 86%
Inverted 3 2 1 2%
Peaked 12 8 4 8%
Upturn 6 4 2 4%
---------------- -------- --------- --------- ------------
: Spectral Classifications[]{data-label="table:class"}
<span style="font-variant:small-caps;">Notes.</span>– The percentages of each source are only for the single or point sources.\
Spectral index as a function of flux density {#sec:specflux}
--------------------------------------------
We have investigated the properties of our sample with flux density. Figure \[fig:medians\] shows median spectral index as a function of flux density, both for our sample (Figure \[fig:medians\]a), and for a compilation of samples from the literature along with ours (Figure \[fig:medians\]b). While the uncertainties are large, there is evidence for a mild trend toward a flatter spectral indices with decreasing flux density (Figure \[fig:medians\]a). Figure \[fig:medians\]b shows the comparison of our median spectral indices with @windhorst [@prandoni4; @ibar], and @owen1. Our data is in general consistent with these previous results, and so we cannot rule out the possibility that there is a flattening of the median spectral index with decreasing flux density. We note that the very steep median spectral index in the faintest flux density bin of many surveys is a consequence of the flux density limits preventing the detection of flatter spectral indices close to the survey limits.
-------------------- ------------------------ ----------- ------------ ----------- -----------
Range in $S_{843}$ Median Mean Flux Number 25$^{th}$ 75$^{th}$
or $S_{1.4}$ (mJy) $\alpha^{1.4}_{0.843}$ (mJy) of sources perc. perc.
$S_{1.4}$
0.81-3.7 $-$1.134 2.50 21 $-$1.390 $-$0.676
3.7-6.4 $-$0.717 4.89 21 $-$1.016 $-$0.355
6.4-9.8 $-$0.533 7.68 22 $-$0.876 $-$0.275
13.5-26 $-$0.584 13.65 21 $-$0.769 $-$0.332
26-95.4 $-$0.775 36.21 20 $-$0.945 $-$0.656
$S_{0.843}$
3.3-4.52 $-$0.818 4.04 18 $-$1.273 $-$0.825
4.52-6.7 $-$0.488 5.54 18 $-$1.203 $-$0.733
6.7-11 $-$0.676 8.76 17 $-$1.210 $-$0.817
11-15 $-$0.671 12.68 18 $-$1.045 $-$0.795
15-35 $-$0.687 22.71 18 $-$0.872 $-$0.575
35-204 $-$0.921 62.42 16 $-$0.930 $-$0.741
-------------------- ------------------------ ----------- ------------ ----------- -----------
: Median Spectral Index $\alpha^{1.4}_{0.843}$ Statistics as a function of flux density[]{data-label="table:medians"}
<span style="font-variant:small-caps;">Notes.</span>– Errors on the median spectral indices are calculated from the 25th and 75th percentile. The mean flux densities are calculated as the mean of the flux densities of all the sources in each bin.\
Radio Source Counts {#sec:srccnts}
-------------------
The differential radio source counts for our 843MHz data are presented in Table \[table:mostsc\]. A weighting factor has been applied to the source counts to correct for incompleteness due to the noise level increasing at the edges of the field [see @srccnts]. We do not apply a resolution correction, as the MOST has a large beam, and has a high sensitivity to extended diffuse radio emission [@bock; @SUMSS2]. The differential source counts are shown in Figure \[fig:sourcecounts\], with the SUMSS [@SUMSS2] differential source counts, and a compilation of 1.4GHz differential source counts [@phoenix] as a reference sample. Our data probes the 843MHz source counts a factor of 2 fainter than SUMSS, but we still clearly underestimate the counts in our faintest flux density bin, due to incompleteness.
The reference differential source count sample at 1.4GHz from @phoenix has been shifted to 843MHz assuming an average spectral index value. We have determined that $\alpha=-0.5$ is the appropriate value to use for the average spectral index, using the Kellermann correction, discussed in further detail here.
### The Kellermann Correction {#sec:kell}
@kellermann noted that any observed spectral index distribution is not independent of the observing frequency. The correction accounts for the relationship between observed spectral index distributions at different frequencies. At higher radio frequencies, we tend to see more flat spectrum objects as proportionally more sources are above the flux limit. In contrast, as we move to lower radio frequencies, we tend to see more steep spectrum objects because they are brighter at the lower radio frequencies. The correction stated in the Appendix of @kellermann describes the offset between mean values of spectral indices at different frequencies. We use Equation A5 of @kellermann, where we assume we have two distributions of spectral indices ($P(\alpha)$ and $Q(\alpha)$), at two different frequencies, $\nu_{1}$ and $\nu_{2}$ respectively. Also required is $x$, the slope of the number-flux power-law $N(>S)=k\,S^{x}$ [@longair], where $k$ is a constant scaling factor, and $N$ represents the number of sources above a given flux density, $S$. $$Q(\alpha)=A(\frac{\nu_{2}}{\nu_{1}})^{-\alpha\,x}P(\alpha),
\label{eq:kell}$$ $$\int^{\infty}_{-\infty}Q(\alpha)d\alpha=1.
\label{eq:kell1}$$ $P(\alpha)$ is the known distribution of spectral indices at one frequency and $Q(\alpha)$ is the distribution of spectral indices we wish to infer, given by Equation \[eq:kell\]. Following @kellermann we assume the special case where these two distributions are Gaussian (a reasonable assumption given the shapes of the distributions in Figures \[fig:alphahisto\] and \[fig:fluxbins\]), and have the same dispersion $\sigma$, and A is chosen such that Equation A6 of @kellermann (Equation \[eq:kell1\]) is satisfied. This special case scenario results in the mean value of the inferred spectral index distribution ($Q(\alpha)$) being shifted by a factor of $x\sigma^{2}$ln$(\nu_{1}/\nu_{2})$. For our sample, to shift the distribution of the 1.4GHz source count compilation to 843MHz, we used $\alpha^{1.4}_{0.843}$ to determine the value of the Kellermann correction for only the single sources. The dispersion was found to be $\sigma=0.56$ and $x=-1.60$, resulting in a Kellermann correction of 0.25. This shift in the mean of the spectral index distribution from 843MHz to 1.4GHz is enough to account for the use of the $\alpha\sim-0.5$ to shift the 1.4GHz source counts to the 843MHz source counts, instead of the observed mean spectral index of our sample $\alpha=-0.71$.
-------------------- ------- ------------- ----------------------------- -----------------------
Range in $S_{843}$ *N* *$N_{eff}$* $\langle\,S_{843}\,\rangle$ $(dN/dS)/S^{-2.5}$
(mJy) (mJy) (Jy$^{1.5}$sr$^{-1}$)
3.31-5.81 45 21.58 4.38 4.29
5.81-8.31 45 34.87 6.95 21.90
8.31-10.81 44 41.63 9.48 56.83
10.81-15.81 51 50.02 13.07 76.29
15.81-25.81 48 47.54 20.20 107.61
25.81-45.81 42 42 34.38 179.70
45.81-100 33 33 67.68 283.27
-------------------- ------- ------------- ----------------------------- -----------------------
: Source Counts from the ATLAS ELAIS MOST Observations[]{data-label="table:mostsc"}
\
$N_{eff}$ is the effective number of sources in each bin, after the weighting factor has been applied. The mean flux density given is the geometric mean of the bin limits.
![The differential source counts for our data (circles), SUMSS (diamonds), and a 1.4GHz compilation of source counts (grey squares) shifted to 843MHz assuming a spectral index of $\alpha=-0.5$, from Hopkins et al. 2003 as our reference sample. \[fig:sourcecounts\]](f12.eps)
Faint CSS and GPS candidates sources in ATLAS {#sec:props}
=============================================
Compact Steep Spectrum and Gigahertz Peaked Spectrum Sources {#sec:CSS}
------------------------------------------------------------
Gigahertz Peaked Spectrum (GPS) and Compact Steep Spectrum (CSS) sources [@young; @cfanti; @morganti] are a class of compact, powerful radio sources, suggested to be the beginning of the evolutionary path for large-scale radio sources. The proposed evolutionary path has GPS sources evolving into CSS sources, which then gradually evolve into Fanaroff-Riley Type I and II galaxies [@fr], depending on their initial luminosity. These sources offer an ideal resource to investigate galaxy evolution and formation, as well as AGN feedback, as they are young AGN, but also have star formation occurring due to interactions and mergers [@Review; @labiano; @morganti]. A more detailed discussion of their properties is given in @randall. If bright CSS sources do evolve into FRI/II’s [@fr], then some strong evolution must occur for this to happen [@evolution], as we do not see this high percentage of CSS and GPS sources in the local Universe. It is also possible that the supply mechanism of the energy powering these objects could cut off, leaving only diffuse emission and a steep spectrum core. This would result in faint objects whose radio lobes have ceased to expand or a prematurely dying radio source [@rfanti], explaining the lack of large numbers of these objects nearby.
Bright CSS and GPS sources are common ($\approx30\%$ and $10\%$ respectively) in radio surveys, but few faint samples exist. Previous surveys include @snell00 [@tschager; @kunert07; @fanti and references therein;], which catalogue CSS and GPS sources down to $\sim20$mJy. If these objects are as prevalent at faint flux densities, a whole population of these objects remains mostly unknown [@faintgps; @faint]. Here we present an initial complete sample of faint CSS and GPS sources from ATLAS, to further our understanding of their role in galaxy formation and evolution. Understanding their properties across different wavelength regimes is important, as is studying samples across a wide range of flux densities, as this will help to build a complete picture of their properties and nature. We present here a brief overview of the selection and properties of the faint sample of candidate CSS and GPS sources selected from ATLAS.
A new candidate sample of faint CSS sources {#sec:newfaint}
-------------------------------------------
An initial sample of faint CSS sources has been selected from ATLAS (§\[sec:atlas\]), based upon spectral index information, and angular size. The initial sample is drawn from both ATLAS fields, CDFS and ELAIS, and only includes the 1.4 and 2.3GHz flux density measurements. CDFS was not observed by MOST, and we do not consider the ELAIS MOST observations in this discussion.
Field ELAIS CDFS
----------- ------------------------------------------------------------------------------------------------------ ----------- -----------
Criterion Selection Sources Sources
Number Criterion Remaining Remaining
1 Unresolved in component catalogue 681 271
2 Not part of a radio double/triple/complex object or classified as a sidelobe 638 247
3 Apply spectral index cut of $\alpha^{2.3}_{1.4}<-0.95$ for sample selection 78 8
3a Sources with spectral index between $-0.9>\alpha^{2.3}_{1.4}>-0.95$ selected as supplementary sample 4 1
\
<span style="font-variant:small-caps;">Notes.</span>– Removal of radio doubles/triples/complex sources was done by cross-matching both the component, and source catalogues from CDFS and ELAIS, as the source catalogue lists what components comprise each source.
The selection criteria listed in Table \[table:selection\] were utilized to first select a sample of unresolved, single objects from both the ELAIS and CDFS fields, before applying a spectral index cut. The number of sources remaining in each field after each criterion is also listed. We have chosen a primary spectral index cut of $\alpha^{2.3}_{1.4}<-0.95$ with an additional supplementary set extending to $\alpha^{2.3}_{1.4}<-0.9$, as the errors on the spectral indices are typically $\sim0.2$. These values were chosen to avoid selecting any star forming galaxies (with typical synchrotron spectral indices of $\alpha^{2.3}_{1.4}\sim-0.7$), as at these low flux densities, star forming galaxies are a dominant part of the population [@windhorst85; @windhorst; @phoenix; @seymour]. In the ELAIS catalogue, there are 576 sources with a single 2.3GHz match (blended sources are not included in this analysis) out of 1276, and 399 of 726 in the CDFS. The primary reason for only half the sources having a 2.3GHz counterpart is the lack of equivalently deep 2.3GHz data across the two fields. The 2.3GHz data has lower resolution ($33''\times20''$ and $54''\times21''$ for ELAIS and CDFS respectively), and does not have the same sensitivity as the 1.4GHz data for both ELAIS and CDFS. The rms of the CDFS 2.3GHz data is $\sim3$ times higher than the ELAIS 2.3GHz data. The initial selection criterion of being an unresolved component also removes 63% of the sources from CDFS, but only 47% from ELAIS. The percentage of sources remaining after the initial selection is small, with 86 sources and 5 supplementary sources, from a possible 975 ($\sim10$%).
Of the 86 sources, only 5 are cross-matched to 843MHz sources, due to the relatively high flux limit of the 843MHz data. The relevant information from these sources is added into the sample catalogue, but is not used for any subsequent analysis below. In considering the spectral index selection, described in Table \[table:selection\], for each field, we note that the CDFS 1.4GHz flux densities were taken for the sources as observed, rather than after convolution to match the 2.3GHz resolution. This step has been omitted from the current analysis due to the small numbers of sources involved, and while the spectral index estimates for these few sources may be somewhat less robust than those in the ELAIS field, this does not have a significant impact on any of the results presented here.
In the ELAIS field, we use the spectral indices measured between 2.3GHz and the Z11 1.4GHz flux densities (measured with the same resolution as the 2.3GHz image). If the source does not have a reliable convolved measured flux density, we use the M08 flux (this is only necessary for four sources). To robustly constrain the radio spectra of these sources, we need to explore both higher, and lower frequency data to confirm that our sources are truly steep-spectrum or peaked spectrum objects. We show the basic statistics of our sample in Table \[table:means\], and discuss initial results in §\[sec:discussion\]. The 1.4GHz flux density and spectral index distributions are presented in Figure \[fig:fluxdist\], with the median flux density and spectral index included as the solid black lines. The spectral index distribution of the entire ATLAS catalogue is also shown in Figure \[fig:fluxdist\]b to provide context for the initial CSS sample distribution. Note that the steep spectrum sources not identified as CSS are those that were rejected by criteria 1 or 2 in Table \[table:selection\]. The median of the CSS sample is $\alpha^{2.3}_{1.4}=-1.48$, compared to $\alpha^{2.3}_{1.4}=-0.84$ for the entire ATLAS catalogue.
Spectral Index Spectroscopic Redshift Photometric Redshift R magnitude 1.4GHz flux density 2.3GHz flux density
-------- ---------------- ------------------------ ---------------------- ------------- --------------------- --------------------- -- --
Mean $-1.58$ 0.37 1.21 19.03 1.68 1.08
Median $-1.48$ 0.32 0.97 18.8 0.52 0.36
\
Discussion {#sec:discussion}
==========
Redshift and Luminosity Properties {#sec:lums}
----------------------------------
ATLAS has $\sim$700 spectroscopic [@mao], and $\sim$700 photometric [@rowan] redshifts for radio sources in the CDFS and ELAIS-S1 fields. The sources chosen for the spectroscopic targets were primarily the optically bright sources (80% complete at R magnitude $\sim$20), that are mainly low redshift sources, although some quasars at higher redshift are also selected. For the 105 single sources in the MOST ATLAS catalogue, we found 20 spectroscopic redshifts and 9 photometric redshifts (no analysis is done on the blended sources). In the initial CSS sample of 91, 22 spectroscopic and 17 photometric redshifts were found. In Figure \[fig:lumalpha\] spectral index against redshift and 1.4GHz rest-frame luminosity are shown, for both the MOST ATLAS catalogue and the initial CSS sample. The grey box in Figure \[fig:lumalpha\]b indicates where SFGs would most likely be located, defined by the dividing line between AGN and SF luminosity of L$_{1.4}=\sim10^{24.5}$WHz$^{-1}$, and a spectral index between $-0.9<\alpha^{2.3}_{1.4}<-0.5$, typical for SFGs [@afonso]. This is a conservative upper limit on the star formation rate, assuming that the star formation rate implied from the radio emission is very high. The single inverted spectrum source with a spectroscopic redshift of $z=1.33$ is classified spectroscopically as a quasar, with broad emission lines. It has an unusual radio spectrum that appears to contain an upturn, with $\alpha^{1.4}_{0.843}=-1.03$, and $\alpha^{2.3}_{1.4}=1.18$, one of only 6 MOST ATLAS sources with an upturning radio spectrum.
![(a) Distribution of luminosities for our MOST ATLAS sources and (b) initial CSS sample, at 1.4GHz and 2.3GHz. \[fig:lums\]](f15.eps)
The rest-frame 843MHz, 1.4GHz, and 2.3GHz luminosities were calculated for each source in the MOST ATLAS catalogue with a redshift (spectroscopic or photometric), and the 1.4 and 2.3GHz luminosities are shown in Figure \[fig:lums\]a. The rest-frame 1.4 and 2.3GHz luminosities were also calculated for the CSS sample, and are shown in Figure \[fig:lums\]b. The distribution of luminosities, in conjunction with their spectral index, is consistent with these sources being AGN, as expected. Although it appears in Figures \[fig:lumalpha\]b and \[fig:lums\]b that the CSS sample is on average a lower-luminosity sample, this is due to the flux limit of the 843MHz MOST data, that preferentially selects the higher-luminosity objects. However, as we only find 4 of these MOST ATLAS sources in our CSS candidate sample, we are selecting a low-luminosity sample of young AGN that is ideal for comparison to our bright sample [@randall].
The MOST ATLAS and initial CSS sample sources with a spectroscopic redshift have also been spectroscopically classified as AGN or SFGs [@mao]. Of the 22 MOST ATLAS sources with spectroscopic redshifts, 15 sources were classified as AGN, four were classified as SFGs and the remaining two sources were unclassified. Three of the SFG sources are discussed further in §\[sec:interesting\]. Within the CSS sample, 13 sources were classified as AGN, 5 as SFGs, and four were unclassified. The location of the CSS sample SFGs are shown on Figures \[fig:rfir\] and \[fig:qplot\] by thick black circles surrounding each source.
Radio-Far Infrared Correlation {#sec:rfirc}
------------------------------
The radio-far infrared correlation (RFIRC) is a well known tight correlation between the far and mid-infrared and radio emission from galaxies (primarily SFGs), first noted in the 1970s, and confirmed in later years [e.g. @dickey; @dejong]. @appleton confirmed the correlation holds to a redshift $z=1$, and it has been recently noted that the RFRIC does not evolve above this redshift, but remains relatively constant to a redshift of $z\sim2$ [@frc]. We have utilized the data for both the ELAIS and CDFS field from ATLAS, to further investigate the properties of the sources detected in our MOST observations, and the initial CSS sample (Figure \[fig:rfir\]). The solid black line on Figure \[fig:rfir\] gives the relationship between radio and infrared flux at 24$\mu$m from @appleton, where $q_{24}=log(S_{{24\,\mu}{m}}/S_{20\,cm})=0.84$, and the dashed line is the line used by M08 to classify a source as unambiguously AGN (above the dashed line represented by $q_{{24\,\mu}{m}}<-0.16$) or a SFG, AGN or composite object (below the dashed line). Most of our 843MHz selected sources are AGN by the M08 classification. The MOST ATLAS sources which fall below this AGN line are further discussed in §\[sec:interesting\].
The initial CSS sample is also shown in Figure \[fig:rfir\]. While these are clearly low-luminosity AGN on the basis of their spectral indices (Figure \[fig:lumalpha\]b), they equally clearly fall predominantly in the region occupied by SFGs on the RFIRC. This is an interesting preliminary result which will be investigated further in a follow-up analysis [@randall1]. It may be possible that the radio spectral index is steep, betraying the presence of an AGN, but that the AGN contribution to the 1.4GHz flux density is small enough (or perhaps is balanced by an equivalent contribution from the AGN to the FIR) that the RFIRC is still obeyed. If this is the case, the location of the CSS sources close to the RFIRC may actually be a consequence of the 1.4GHz luminosity being dominated by star formation, and the host galaxy being a composite object. This would be consistent with earlier work suggesting that CSS sources are actively star forming [@Review; @labiano; @morganti; @holtb], and highlight this population as an ideal resource for exploring AGN feedback effects on star formation in galaxies. Only the use of radio discriminants (such as morphology or spectral index) will indicate whether the object hosts an AGN [@roy; @buriedagn].
In Figure \[fig:qplot\] the $q_{24}$ values are shown against redshift, with several different spectral energy distribution (SED) tracks from @elvis94 [@dev99], and @rieke09 overlaid. A suggested dividing line between AGN and SFGs from @seymour is also shown. The CSS sample here are again consistent with having their radio and FIR emission dominated by star formation as are the three nearby MOST ATLAS sources (as labelled in Figures \[fig:rfir\] and \[fig:qplot\]). The SED tracks near these objects include those for Luminous Infrared Galaxies (LIRGs), Ultra-Luminious Infrared Galaxies (ULIRGs), Arp 220 and Mrk 273, two nearby ULIRGs that both host an AGN [@engel; @iwa11]. The three MOST ATLAS sources in the AGN-dominated region at the bottom of the figure, are likely AGN, and this is supported by the SED of the radio-loud quasar [@elvis94] closest to these AGN. Although the SED tracks do not extend to the lowest redshifts of our sample, four of the CSS sample are likely to be LIRGs or ULIRGs, similar to F00183$-$7111, a powerful ULRIG, that has strong radio emission from both an AGN and vigorous star forming activity [@foo], but less luminous in the radio. @foo have suggested that F00183-7111 is in the earliest phase of the formation of a quasar, where a “quasar-mode” AGN [@best06; @croton] is hosted by a star forming ULIRG, where the star formation is fueled by gas from a past major or minor merger event. Whilst the radio jets of F00183$-$7111 are currently confined by the host galaxy, when they break through this dense gas, the jets will begin the transition of this source into a typical radio-loud quasar, quenching the star formation at the same time [@foo]. These CSS sources will be further investigated in a future paper, @randall1.
Unusual Sources {#sec:interesting}
---------------
We have identified several interesting sources in our final MOST ATLAS catalogue, particularly those sources which lie close to the radio-far infrared correlation (and consequently are likely to host, or even be dominated by, star formation), and one possible ultra-steep spectrum object.
### Possible SFGs in ATLAS
Of the four single sources that fall close to the RFIRC, three are disk or spiral-like galaxies based on their luminosity profiles in the 3.6$\mu$m SWIRE images (labelled by name in Figures \[fig:rfir\], and \[fig:qplot\]), which suggests they may be SFGs. The fourth object appears as an elliptical, with no evidence of a disk or interaction, and was not further investigated. Figure \[fig:s707\] shows the SWIRE 3.6$\mu$m images with the 1.4GHz radio contours overlaid, indicating the clear disk or spiral structure of these three galaxies. Of the three blended sources near the RFIRC, there is less evidence for disks or star-formation in the SWIRE data. For the two closest to the RFIRC (S259/S269 and S1235/S1228/S1230/S1243), the radio emission could possibly still be associated with star formation, but it is harder to constrain without clear evidence of a disk or spiral structure in the host galaxy. We briefly discuss S707, S897 and S1160 below.
### S707 (ATELAIS J003828.02-433847.2) {#sec:s707}
S707 is a possible star-forming galaxy at a redshift of $z=0.048$, where the host galaxy has a disk structure, as evidenced by is luminosity profile as seen in Figure \[fig:s707\]a. The radio morphology of this object is a single point source at the centre of the SWIRE galaxy, suggesting the low-luminosity AGN in the centre of the galaxy is the primary source of the radio emission, rather than star formation in the disk. Nuclear star formation may also be possible, and the optical spectrum of this object suggests there is ongoing star formation. The radio source has an overall steep spectrum, $\alpha_{fit}=-1.17$, although it is much flatter between 1.4 and 2.3GHz ($\alpha^{2.3}_{1.4}=-0.45$), than 843MHz and 1.4GHz ($\alpha^{1.4}_{0.843}=-1.66$), as expected from an AGN. S707 has a rest-frame luminosity of L$_{1.5}=3.85\times10^{22}$WHz$^{-1}$, from the observed convolved 1.4GHz flux density, resulting in an upper limit on the star formation rate of $21M_{\sun}$yr$^{-1}$ [assuming all the radio emission is from star formation; @sfr]. It also has $q_{{24\,\mu}{m}}=0.57$, suggesting a small fraction of the radio emission is from star formation, whilst the majority is from the AGN. Optically, this source has a similar morphology to the 3.6$\mu$m galaxy, with magnitudes $B=15.8$, $V=15.2$ and $R=14.7$ from M08.
![$q_{24\mu}$$_{m}$ as a function of redshift for the entire ATLAS catalogue, split into spectroscopic and photometric redshifts. The initial CSS sample is overlaid as the large red circles for those sources with spectroscopic redshifts, and blue squares for those with photometric redshifts. The MOST ATLAS sources with 24$\mu$m detections are shown as the hollow triangles (sources with spectroscopic redshifts) or inverted hollow triangles (sources with photometric redshifts. The SED tracks describe several different models. The heavy dashed line represents an Arp220 type object [@dev99], the thin solid black line is a model for Mrk273 [@dev99], the regular dashed line is a radio-loud QSO [@elvis94], the dotted line is a LIRG model [@rieke09], and the dash-dotted line represents a ULIRG model [@rieke09]. The thick black line shows the dividing line between AGN and SFGs from @seymour. Objects with a thick black circle around the data point are from the CSS sample, and are those spectroscopically classified as SFGs. \[fig:qplot\]](f17.eps)
### S896 (ATELAIS J003429.33-432614.4) {#sec:s896}
S896 appears to be hosted by an edge-on disk galaxy, with some possible ongoing star formation that can be seen in Figure \[fig:s707\]b, at a redshift of $z=0.053$. Although the radio emission appears to be emitted primarily from the central object, there is most likely some contribution to the overall flux density from star formation in the disk, as the optical spectrum suggests there is some star formation ongoing. This object has a rest-frame luminosity of L$_{1.5}=1.92\times10^{22}$WHz$^{-1}$, calculated from the observed convolved 1.4GHz flux density. This implies an upper limit on the star formation rate of $\sim11M_{\sun}$yr$^{-1}$, assuming all the radio emission is from star formation. However, it has an overall steep spectral index of $\alpha_{fit}=-1.21$, which suggests that the source of the radio emission is a low-luminosity AGN. For this source $q_{{24\mu}{m}}=0.87$, consistent with the idea that the radio emission is primarily from the AGN.
### S1160 (ATELAIS J003915.11-430428.5) {#sec:s1160}
S1160 is hosted by a star-forming spiral galaxy, shown in Figure \[fig:s707\]c, at a redshift of $z=0.0135$ [@jones; @jones1]. The radio emission is clearly associated with star formation, as the radio contours trace the star formation in the spiral arms of this galaxy. However, the radio position of the source is offset from the galaxy centre, assigned as the location of the brightest peak in the complex radio structure (M08). The radio luminosity of this source is L$_{1.42}=4.60\times10^{21}$WHz$^{-1}$ which results in an upper limit on the star formation rate of $2.5M_{\sun}$yr$^{-1}$ (if all radio emission is due to star formation). The optical morphology of this galaxy is very similar to the SWIRE morphology, with a clear spiral structure. The optical host galaxy has magnitudes from the SuperCOSMOS Science Archive, although these magnitudes should be treated with caution, as it is a large, relatively nearby galaxy and measuring accurate magnitudes in an automated fashion for such sources is difficult due to their complex structure. The SuperCOSMOS magnitudes are $I=13.4$, $R=12.9$ and B$_{J}=7.9$. The optical spectrum is also indicative of star formation. S1160 is detected in all five of the SWIRE bands, with a $q_{24\mu\,m}=0.25$. This $q_{{24\,\mu}{m}}$ value is still consistent with the source being either a SFG or an AGN, and it most likely there is a non-negligible contribution to the radio emission from star formation in this source. The radio flux of this object is interesting, as it has a similar flux density between 843MHz and 1.4GHz (S$_{0.843}=12.7$mJy compared to S$_{1.4}=11.6$mJy or S$_{M08}=12.9$mJy), but drops rapidly away to S$_{2.3}=5.3$mJy. This gives a spectral index $\alpha^{1.4}_{0.843}=-0.2$ and $\alpha^{2.3}_{1.4}=-1.6$, with an overall spectral index of $\alpha_{fit}=-1.03$. This spectral shape is more indicative of AGN activity, but given the radio morphology and the FIR constraints, the radio emission from this object is likely a combination of star formation and AGN activity.
### S1256 (ATELAIS J003053.26-425215.3): A candidate Ultra-Steep Spectrum source {#sec:s1256}
High redshift radio galaxies are often found by identifying radio sources with steep or ultra steep radio spectra ($\alpha < -1$), via the well-known $z$-$\alpha$ relation [@debreuck04; @klamer; @ishwara and references therein;]. This trend was first noticed in the early 1980s, when it was found that the fraction of radio sources optically identified on optical plates tended to have the steepest radio spectra. S1256 is an isolated point source, with $S_{0.843}=4.38$mJy and $S_{1.4}=0.81$mJy, that is not detected in the 2.3GHz ATLAS image, with a spectral index $\alpha^{1.4}_{0.843}=-3.33$. Following Z11, using the 3$\sigma$ detection limit of 300$\mu$Jy, an upper limit on the spectral index is found to be $\alpha^{2.3}_{1.4}<-2.41$, which suggests that this source is possibly an ultra-steep spectrum source, and therefore at high redshift. Further radio imaging and flux density measurements would be necessary to confirm S1256 as an ultra-steep spectrum source. S1256 has no optical counterpart or confirmed redshift, but is detected in the SWIRE 3.6 and 4.5$\mu$m bands. Its flux densities in these bands are low ($S_{3.6\mu\,m}=17.5\,\mu$Jy and $S_{4.5\mu\,m}=19.9\,\mu$Jy). The infrared morphology is point-like and is shown in Figure \[fig:s1256\]. Another possible cause of this steep spectrum is variability; where the radio source has faded away since the 1.4GHz observations were completed, and before the 2.3GHz observations began. Further observations would also be necessary to confirm whether variability is the cause of the ultra-steep radio spectrum.
Conclusions {#sec:concl}
===========
We have presented a new catalogue of 843MHz radio sources, cross-matched to ATLAS at 1.4 and 2.3GHz, and explored the properties of this catalogue with spectral index as a function of flux density. Our results do not support the hypothesis that there is a significant flattening of the spectral index with decreasing flux density values. However, we cannot rule out the possibility without further, deep radio data at different radio frequencies. Our analysis of the distribution of steep and flat spectrum sources with redshift, luminosity and infrared flux density indicates that most of the sources in our MOST ATLAS catalogue are AGN, inferred from the spectral index and infrared properties. An initial sample of faint CSS sources in ATLAS has also been selected, with their basic properties explored. We will explore the properties of the CSS sample and the MOST ATLAS selected GPS sources in depth in future work, and compare and contrast this faint sample to our bright sample [@randall]. The spectral index properties of ATLAS radio sources across a wide range of frequencies will also be explored in future work, with the aim of distinguishing between the proportions of the populations of low-luminosity, and/or core-dominated AGN and SFGs driving this effect.
Acknowledgements {#acknowledgements .unnumbered}
================
We would also like to thank our referee, Rogier Windhorst, for his extensive and thorough comments on this paper. We thank the staff at Molonglo for completing the observations and Dick Hunstead for performing the initial data reduction process. We also thank Dick and Elaine Sadler for their input; their comments and help are greatly appreciated. The authors would also like to thank Nick Seymour for the nicely formatted SED templates from various places in the literature for the $q_{24}$ plot.\
The Australia Telescope Compact Array is part of the Australia Telescope which is funded by the Commonwealth of Australia for operation as a National Facility managed by CSIRO.\
This research has made use of the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Labratory, Caltech, under contract with NASA.\
[99]{} Afonso J., Georgakakis A., Almeida C., Hopkins A. M., Cram L. E., Mobasher B., Sullivan M., 2005, ApJ, 624, 135 Afonso J., Mobasher B., Koekemoer A., Norris R. P., Cram L., 2006, AJ, 131, 1216 Alexander D. M., et al., 2001, ApJ, 554, 18 Appleton P. N., et al., 2004, ApJS, 154, 147 Banfield J., et al., in prep. Best P. N., Kaiser C. R., Heckman T. M., Kauffmann G., 2006, MNRAS, 368L, 67 Berta S., et al., 2006, A&A, 451, 881 Berta S., et al., 2008, A&A, 488, 533 Bock D. C.-J., Large M. I., Sadler E. M., 1999, ApJ, 117, 1578 Burgarella D., et al., 2005, AJ, 619, L63 Condon J. J., 1984, ApJ, 287, 461 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 Croton D. J., et al., 2006, MNRAS, 365, 11 De Breuck C., Hunstead R. W., Sadler E. M., Rocca-Volmerange B., Klamer I., 2004, MNRAS, 347, 837 de Jong T., Klein U., Wielebinski R., Wunderlich E., 1985, A&A, 147, L6 Devriendt J. E. G., Guiderdoni B., Sadat R., 1999, A&A, 350, 381 Dickey J. M., Saltpeter E. E., 1984, ApJ, 284, 461 Elvis M., et al., 1994, ApJS, 95, 1 Engel H., Davies R. I., Genzel R., Tacconi L. J., Sturm E., Downes D., 2011, ApJ, 729, 58 Fanaroff B. L. & Riley J. M., 1974, MNRAS, 167, 31 Fanti C., 2009a, Astron. Nachr., 330, 120 Fanti R., 2009b, Astron. Nachr., 330, 303 Fanti C., Fanti R., Zanichelli A., Dallacasa D., Stanghellini C., 2011, A&A, 528, A110 Fazio G. G., et al., 2004, ApJS, 154, 10 Georgakakis A., Mobasher B., Cram L., Hopkins A., Lidman C., Rowan-Robinson M., 1999, MNRAS, 306, 708 Gruppioni C., et al., 1999, MNRAS, 305, 297 Hales C. A., et al., in prep. Hill T. L., Heisler C. A., Sutherland R., Hunstead R. W., 1999, AJ, 117, 111 Hill T. L., Heisler C. A., Norris R. P., Reynolds J. E., Hunstead R. W., 2001, AJ, 121, 128 Holt J., 2009, Astron. Nachr., 330, 789 Hopkins A. M., 2004, AJ, 615, 209 Hopkins A. M., Mobasher B., Cram L., Rowan-Robinson M., 1998, MNRAS, 296, 839 Hopkins A. M., Miller C. J., Connolly A. J., Genovese C., Nichol R. C., Wasserman L., 2002, AJ, 123, 1086 Hopkins A. M., Afonso J., Chan B., Cram L. E., Georgakakis, A., Mobasher B., 2003, AJ, 125, 465 Ibar E., Ivison R. J., Biggs A. D., Lal D. V., Best P. N., Green D. A., 2009, MNRAS, 397, 281 Ibar E., Ivison R. J., Best P. N., Coppin K., Pope A., Smail I., Dunlop J. S., 2010, MNRAS, 401L, 53 Ishwara-Chandra C. H., Sirothia S. K., Wadadekar Y., Pal S., Windhorst R, 2010, MNRAS, 405, 436 Iwasawa K., et al., 2011, A&A, 528A, 137 Jones D. H., et al., 2004, MNRAS, 355, 747 Jones D. H., et al., 2009, MNRAS, 399, 683 Kellermann K., 1964, ApJ, 140, 969 Kesteven M. J. L., Bridle A. H., Brandie G. W., 1977, AJ, 82, 541 Klamer I. J., Ekers R. D., Bryant J. J., Hunstead R. W., Sadler E. M., De Breuck C., 2006, MNRAS, 371, 852 Komatsu E., et al., 2011, ApJS, 192, 18 Kunert-Bajraszewska M., Marecki A., 2007, A&A, 469, 437 Labiano A., O’Dea C. P., Barthel P. D., de Vries W. H., Baum S. A., 2008, A&A, 477, 491 Large M. I., Mills B. Y., Little A. G., Crawford D. F., Sutton J. M, 1981, MNRAS, 194, 693 Longair M. S., 1966, MNRAS, 133, 421 Lonsdale C. J., et al., 2003, PASP, 115, 897 Lonsdale, C. J., et al., 2004, ApJS, 154, 54 Magliocchetti M., Maddox S. J., Wall J. V., Benn C. R., Cotter G, 2000, MNRAS, 318, 1047 Mao M. Y., Sharp R., Saikia D. J., Norris R. P., Johnston-Hollitt M., Middelberg E., Lovell J. E. J.., 2010, MNRAS, 406, 2578 Mao M. Y., Huynh M. T., Norris R. P., Dickinson M., Frayer D., Helou G., Monkiewicz J. A., 2011, ApJ, 731, 79 Mao M. Y., et al., in prep. Martin D. C., et al., 2005, AJ, 619, L1 Mauch T., Murphy T., Buttery H. J., Curran J., Hunstead R. W., Piestrzynski B., Robertson J. G., Sadler E. M., 2003, MNRAS, 342, 1117 Middelberg E., et al. (M08), 2008, AJ, 135, 1276 Middelberg E., Norris R. P., Hales C. A., Seymour N., Johnston-Hollitt M., Huynh M. T., Lenc E., Mao M. Y., 2011, A&A, 526A, 8 Mills B. Y., 1981, PASA, 4, 156 Morganti R., Emonts B., Holt J., Tadhunter C., Oosterloo T., Struve C., 2009, Astron. Nachr., 330, 789 Murphy T. et al., 2010, MNRAS, 402, 2403 Norris R. P., et al. (N06), 2006, AJ, 132, 2409 Norris R. P., Middelberg E., Boyle B. J., 2007, ASPC, 380, 229 Norris R. P., Lenc E., Roy A. L., Spoon H., 2011, preprint (arXiv: 1107.3895N) O’Dea C., 1998, PASP, 110, 493 Oliver S., Rowan-Robinson M., Alexander D. M., et al., 2000, MNRAS, 316, 749 Oort M. J. A., Windhorst R. A., 1985, A&A, 145, 405 Owen F. N., Morrison G. E., 2008, AJ, 136, 1889 Owen F. N., Morrison, G. E., Klimek M. D., Greisen E. W., 2009, AJ, 137, 4846 Polatidis, A., & Conway, A., 2003, PASA, 20, 69 Prandoni I., Gregorini L., Parma P., de Ruiter H. R., Vettolani G., Wieringa M. H., Ekers R. D., 2000a, A&AS, 146, 31 Prandoni I., Gregorini L., Parma P., de Ruiter H. R., Vettolani G., Wieringa M. H., Ekers R. D., 2000b, A&AS, 146, 41 Prandoni I., Parma P., Wieringa M. H., de Ruiter H. R., Gregorini L., Mignano A., Vettolani G., Ekers R. D., 2006, A&A, 457, 517 Prandoni I., Bernardi G., Di Vincenzo A., de Bruyn A. G., 2011, submitted to A&A Puccetti S., et al., 2006, A&A, 457, 501 Randall K. E., Hopkins A. M., Norris R. P, Edwards P. G., 2011a, MNRAS, 416, 1135 Randall K. E., et al., in prep. Rieke G. H., Young E. T., Engelbracht C. W., et al., 2004, ApJS, 154, 25 Rieke G. H., Alonso-Herrero A., Weiner B. J., P[é]{}rez-Gonz[á]{}lez P. G., Blaylock M., Donley J. L., Marcillac D., 2009, ApJ, 692, 556 Robertson J. G., 1991, Aust. J. Phys., 44, 729 Rowan-Robinson M., Lari C., Perez-Fournon I., et al., 2004, MNRAS, 351, 1290 Roy A. L., Norris R. P., Kesteven M. J., Troup E. R., Reynolds J. E., 1998, MNRAS, 301, 1019 Schlegel D. J., Finkbeiner D. P., Davis M., 1998, ApJ, 500, 525 Scoville N., et al., 2007, ApJS, 172, 1 Seymour N., et al., 2008, MNRAS, 386, 1695 Smol[č]{}i[ć]{} V., et al., 2008, ApJS, 177, 14 Snellen I. A. G., Schilizzi R. T., de Bruyn A. G., Miley G. K., Rengelink R. B., R[ö]{}ttgering H. J., Bremer M. N., 1998, A&AS, 131, 435 Snellen I. A. G., Schilizzi R. T., Miley G. K., Bremer M. N., R[ö]{}ttgering H. J. A., van Langevelde H. J., 1999, NewAR, 43, 675 Snellen I. A. G., Schilizzi R. T., van Langevelde H. J, 2000, MNRAS, 319, 429 Tschager, W., Schilizzi R. T., R[ö]{}ttgering H. J. A., Snellen I. A. G., Miley G. K., Perley R. A., 2003a, PASA, 20, 75 Tschager, W., Schilizzi R. T., R[ö]{}ttgering H. J. A., Snellen I. A. G., Miley G. K., Perley R. A., 2003b, A&A, 402, 171 Vaccari M., et al., 2005, MNRAS, 358, 397 Wilson W. E., et al., 2011, MNRAS, 416, 832 Windhorst R. A., Miley G. K., Owen F. N., Kron R. G., Koo D. C., 1985, ApJ, 289, 494 Windhorst R. A., Fomalont E. B., Partridge R. B., Lowenthal J. D., 1993, AJ, 405, 498 Zinn P-C., et al., in prep.
Supporting Information {#supporting-information .unnumbered}
======================
An additional table of Supporting Information is available in the online version of this article.\
**Table \[table:catalogue\].** The full version of the catalogue, containing all flux density information, ATLAS cross-IDs, optical magnitudes, SWIRE IDs and other relevant information.\
Please note: Wiley-Blackwell are not responsible for the content or functionality of any supporting material supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: Assuming $S_{\nu}\propto\nu^{\alpha}$, where S is the measured flux density and $\nu$ is the observer’s frame frequency.
[^3]: http://galex.stsci.edu/GR6/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In 1960 Borel proved a “localization” result relating the rational cohomology of a topological space $X$ to the rational cohomology of the fixed points for a torus action on $X$. This result and its generalizations have many applications in Lie theory. In 1934, P. Smith proved a similar localization result relating the mod $p$ cohomology of $X$ to the mod $p$ cohomology of the fixed points for a ${\mathbb{Z}}/p$-action on $X$. In this paper we study ${\mathbb{Z}}/p$-localization for constructible sheaves and functions. We show that ${\mathbb{Z}}/p$-localization on loop groups is related via the geometric Satake correspondence to some special homomorphisms that exist between algebraic groups defined over a field of small characteristic.'
author:
- David Treumann
date: 'July 19, 2011'
title: Smith theory and geometric Hecke algebras
---
Introduction
============
It is often possible to compare the cohomology of a space $X$ to the cohomology of the fixed points for a torus action on $X$, by a technique than can be called “equivariant localization.” The prototype for this family of results goes back to Borel and Leray, e.g.
\[thm:1.1\] Let $X$ be a finite-dimensional space with a ${\mathrm{U}}(1)$-action. If $H^i(X;{\mathbb{Q}}) = 0$ for $i$ odd, then $H^i(X^{{\mathrm{U}}(1)};{\mathbb{Q}}) = 0$ for $i$ odd as well, and $$\sum \dim(H^i(X;{\mathbb{Q}})) = \sum \dim(H^i(X^{{\mathrm{U}}(1)};{\mathbb{Q}}))$$
For instance, if $G$ is a Lie group and $T$ is a maximal torus, the Theorem can be used to compute the rational cohomology of $G/T$ (or at least, its total rank). Descendants of Borel’s result (for instance [@AB] and [@GKM]) are used constantly in the deeper study of these and other spaces attached to a Lie group.
This paper concerns an important antecedent of Borel’s result due to P. Smith:
\[thm:1.2\] Let $p$ be a prime number and let $X$ be a finite-dimensional space with an action of ${\mathbb{Z}}/p$. If $H^i(X;{\mathbb{Z}}/p) = 0$ for $i > 0$, then $H^i(X^{{\mathbb{Z}}/p};{\mathbb{Z}}/p) = 0$ for $i > 0$ as well, and $$H^0(X^{{\mathbb{Z}}/p};{\mathbb{Z}}/p) = H^0(X;{\mathbb{Z}}/p)$$
Theorem \[thm:1.2\] appears weaker than Theorem \[thm:1.1\], but it is an essential ingredient in Borel’s original proof (one may reduce to Theorem \[thm:1.2\] by considering a large cyclic subgroup ${\mathbb{Z}}/p \subset {\mathrm{U}}(1)$). Borel’s result has had an enormous impact on Lie theory. One might hope that Smith’s result would have a comparable impact on characteristic $p$ aspects of Lie theory, but to my knowledge this has not been the case. Theorem \[thm:1.2\] has been influential in algebraic topology and in cohomology of groups (e.g. [@Q]) but I do not know of many applications to representation theory.
The bread and butter of geometric representation theory is sheaf theory. In this paper we develop “Smith theory for sheaves” and give an application to algebraic groups in small characteristic. The main result of the formalism is the following—we will give only a rough version here in the Introduction. Note that by “sheaf” in this paper we usually mean “chain complex of sheaves” or something similar.
\[thm:main\] If $X$ is a space with a ${\mathbb{Z}}/p$-action, and $F$ is a ${\mathbb{Z}}/p$-equivariant sheaf on $X$ defined over a field of characteristic $p$, then we may associate in a functorial way a sheaf ${\mathbf{Psm}}(F)$ on the fixed point set $X^{{\mathbb{Z}}/p}$. This assignment $F \mapsto {\mathbf{Psm}}(F)$, called the *Smith operation*, commutes with all other sheaf operations, including duality, pushforward and pullback, nearby cycles, and microlocal stalks. An equivariant version of ${\mathbf{Psm}}$ carries $G$-equivariant sheaves on $X$ to $Z_G(g)$-equivariant sheaves on $X^g$ whenever $g$ is an element of order $p$ in $G$.
The catch of the Theorem is that, while $F$ might be an ordinary sheaf (or complex of sheaves) of ${\mathbb{F}}_p$-vector spaces, the sheaf ${\mathbf{Psm}}(F)$ is not defined over ${\mathbb{F}}_p$ but over a certain $E_\infty$-ring spectrum ${\mathcal{T}}$, called the “Tate spectrum.” We do not work directly with ${\mathcal{T}}$ in this paper (and in particular the reader does not have to know what an $E_\infty$ ring spectrum is), but it’s category of modules, which is easy to describe: it is the Verdier quotient of the category of bounded complexes of finitely generated ${\mathbb{F}}_p[{\mathbb{Z}}/p]$-modules by the category of bounded complexes of finitely generated free ${\mathbb{F}}_p[{\mathbb{Z}}/p]$-modules.
To explain why the catch arises it is useful to consider what happens at the level of Grothendieck groups, i.e. to develop a “Smith theory for functions.” Let $f$ be a constructible function on $X$—for instance, $f$ might arise from a sheaf $F$ of ${\mathbb{F}}_p$-modules by setting $f(x)$ equal to the Euler characteristic of the stalk of $F$ at $x$. If ${\mathbf{Psm}}$ is to commute with all operations, then it commutes with taking stalks, giving us $$\label{eq:smfx}
{\mathbf{Psm}}(f)(x) = f(x)$$ whenever $x$ is a fixed point. It should furthermore commute with taking global sections which at the level of functions is “integration with respect to Euler measure” (see Definition \[def:eulerint\]), giving us $$\label{eq:intXf}
\int_X f = \int_{X^{{\mathbb{Z}}/p}} f$$ However, Equation \[eq:intXf\] only holds if we reduce both sides mod $p$, in which case it’s a consequence of the fact that the Euler characteristic of a space with a free ${\mathbb{Z}}/p$-action is divisible by $p$.
In other words, to get a good Smith theory at the level of functions we have to reduce the values of those functions mod $p$. Working with the funny coefficients ${\mathcal{T}}$ is analogous to this reduction in the following sense: while the Grothendieck group of the category of ${\mathbb{F}}_p$-vector spaces is ${\mathbb{Z}}$, the Grothendieck group of the category of ${\mathcal{T}}$-modules is ${\mathbb{Z}}/p$. The projection ${\mathbb{Z}}\to {\mathbb{Z}}/p$ is realized by a natural “change of coefficients” functor from ${\mathbb{F}}_p$-vector spaces to ${\mathcal{T}}$-modules.
Hecke algebras and the Satake isomorphism
-----------------------------------------
Let $G$ be a Lie group acting on a space $X$. In reasonable situations (for instance if $X$ and the action are real subanalytic) we may attach a ring to $X$ and $G$, called the Hecke algebra of the action. The additive structure of the algebra is the group of $G$-invariant functions on $X \times X$, and the multiplication is given by $$f_1 \ast f_2 (x,y) = \int f_1(x,z) \cdot f_2(z,y) dz$$ where the right-hand side again denotes the Euler characteristic integral. A similar construction makes sense on a categorical level, yielding monoidal “Hecke categories” that we discuss in Section \[sec:heckecat\].
We give an application of Smith theory by considering the spherical Hecke algebra of Satake, or rather the loop-group analog considered by Beilinson-Drinfeld and others. Let $G$ be a compact group, $LG$ the space of free loops $\alpha:S^1 \to G$, and $\Omega G$ the space of based loops $\beta$ with $\beta(1) = 1$.
As these are infinite-dimensional spaces, we will have to take some care to construct Hecke algebras. In particular we will work with the usual affine Grassmannian model of $\Omega G$, see Section \[sec:ind\] for details.
The spherical Hecke algebra attached to $G$ is a group of $LG$-invariant functions on $\Omega G \times \Omega G$, where $LG$ acts on $\Omega G$ by $$(\alpha \cdot \beta)(t) = \alpha(t) \beta(t) \alpha(1)^{-1}$$
The Satake isomorphism is an identification $${\mathrm{SHA}}_G \cong {\mathrm{Rep}}(G^\vee)$$ where the right hand side denotes the Grothendieck ring of representations of Langlands dual group $G^\vee$ to $G$.
An element $g \in G$ acts by conjugation on $LG$ and $\Omega G$, and the fixed points are $L(Z_G(g))$ and $\Omega(Z_G(g))$—the free and based loop spaces of the centralizer of $g$. In case $g$ has prime order $p$, Smith theory for functions provides a map $${\mathbf{Psm}}:{\mathrm{SHA}}_G \otimes_{\mathbb{Z}}{\mathbb{F}}_p \to {\mathrm{SHA}}_{Z_G(g)} \otimes_{\mathbb{Z}}{\mathbb{F}}_p$$ This map simply restricts an ${\mathbb{F}}_p$-valued function on $\Omega G \times \Omega G$ to the fixed points $\Omega Z_G(g) \times \Omega Z_G(g)$. Since this operation commutes with the Euler integral, it is a homomorphism of algebras.
It is natural to search for a representation-theoretic meaning of the corresponding homomorphism $${\mathbf{Psm}}:{\mathrm{Rep}}(G^\vee) \otimes_{{\mathbb{Z}}} {\mathbb{F}}_p \to {\mathrm{Rep}}(Z_G(g)^\vee) \otimes_{{\mathbb{Z}}} {\mathbb{F}}_p$$ The group $Z_G(g)^\vee$ is called an “endoscopic group” for $G^\vee$. It often happens that $Z_G(g)$ is a Levi subgroup of $G^\vee$, in which case $Z_G(g)^\vee$ is a Levi subgroup of $G^\vee$ and the map ${\mathbf{Psm}}$ is given by restriction of representations. However the endoscopic group is not in general a subgroup—for instance we may realize $Z = {\mathrm{Sp}}(2n) \times {\mathrm{Sp}}(2m)$ as the centralizer of an element of order 2 in $G = {\mathrm{Sp}}(2n+2m)$, but there is no way to include the subgroup ${\mathrm{SO}}(2n+1) \times {\mathrm{SO}}(2m+1)$ into ${\mathrm{SO}}(2n+2m+1)$.
We make two observations:
1. If we regard $G^\vee$ as an algebraic group over the field $K$, the representation ring ${\mathrm{Rep}}(G^\vee)$ is not sensitive to $K$.
2. There *is* an inclusion of ${\mathrm{SO}}(2n+1)\times {\mathrm{SO}}(2m+1)$ into ${\mathrm{SO}}(2n+2m+1)$ so long as $K$ has characteristic 2. See Section \[subsubsec:Cn\].
More generally we prove the following Theorem:
\[thm:1.4\] Let $G$ be a compact simply connected simple Lie group, and let $g \in G$ be an element of order $p$ whose centralizer $Z_G(g)$ is semisimple. Then the endoscopic group $Z_G(g)^\vee$ injects into $G^\vee$ when these groups are regarded as algebraic groups of characteristic $p$. The restriction homomorphism ${\mathrm{Rep}}(G^\vee) \to {\mathrm{Rep}}(Z_G(g)^\vee)$ and the Smith homomorphism ${\mathrm{SHA}}_G \otimes_{\mathbb{Z}}{\mathbb{F}}_p \to {\mathrm{SHA}}_{Z_G(g)} \otimes_{\mathbb{Z}}{\mathbb{F}}_p$ are compatible with each other under Satake.
In particular, there is a natural lift of the Smith operator ${\mathrm{SHA}}_G \otimes_{{\mathbb{Z}}} {\mathbb{F}}_p \to {\mathrm{SHA}}_{Z_G(g)} \otimes_{\mathbb{Z}}{\mathbb{F}}_p$ to ${\mathrm{SHA}}_G \to {\mathrm{SHA}}_{Z_G(g)}$. The Smith homomorphism does nothing more than restrict an ${\mathbb{F}}_p$-valued function on $\Omega G \times \Omega G$ to the subset $\Omega Z_G(g) \times \Omega Z_G(g)$, but the lift to ${\mathbb{Z}}$-valued functions is necessarily more intricate. I do not know how to interpret this lift geometrically.
It’s natural to place this result in the context of results of Borel-de Siebenthal and Kac [@bds; @kac]. Let $G$ be a connected complex reductive group. Then
1. If $H \subset G$ is a connected subgroup of the same rank of $G$, then $H = Z_G(Z(H))$, i.e. $H$ is the centralizer of its center.
2. If $H \subset G$ is furthermore maximal among subgroups of full rank, then $H$ is the centralizer of an element of prime order.
3. Up to conjugacy, the elements of $G$ of order $k$ whose centralizer is semisimple are in one-to-one correspondence with simple roots whose coefficient in the maximal root of $G$ is equal to $k$.
Results (1) and (2) do not hold in characterstics 2 and 3. In those characteristics maximal semisimple subgroups have been classified by Liebeck and Seitz [@ting]. According to this classification, every such subgroup arises in the manner of Theorem \[thm:1.4\]. The converse is almost true—with the exception of a single conjugacy class of order 2 in ${\mathrm{F}}_4$, whenever $Z_G(g) \subset G$ is maximal, the subgroup $Z_G(g)^\vee \subset G^\vee$ is also maximal.
Hecke categories and the geometric Satake correspondence of Mirkovi[ć]{}-Vilonen {#sec:heckecat}
--------------------------------------------------------------------------------
In the end, our proof of Theorem \[thm:1.4\] is a case-by-case analysis. But let us speculate about an alternative “Tannakian” proof. Let $R$ be a commutative ring and let ${\mathrm{Sat}}(G,R)$ denote the “$R$-linear Satake category”—this is the triangulated category of suitably constructible $LG$-equivariant sheaves of $R$-modules on $\Omega G \times \Omega G$. ${\mathrm{Sat}}(G,R)$ is a categorification of the spherical Hecke algebra, with a monoidal structure that lifts the algebra structure on ${\mathrm{SHA}}_G$. There is a second “fusion” product on ${\mathrm{Sat}}(G,R)$ as well, exhibiting it as a symmetric monoidal category. A theorem of Mirkovi[ć]{} and Vilonen [@MV] (following Lusztig, Beilinson-Drinfeld, and Ginzburg in case $R = {\mathbb{C}}$) identifies the abelian subcategory ${\mathcal{P}}(G,R) \subset {\mathrm{Sat}}(G,R)$ of perverse sheaves with the tensor category of representations of the split $R$-form of $G^\vee$.
We have called ${\mathrm{Sat}}(G,R)$ a “symmetric monoidal category,” but that is somewhat misleading. We are regarding ${\mathrm{Sat}}(G,R)$ as a triangulated category in the sense of Verdier, but the more natural object is a certain stable $\infty$-category whose homotopy category is ${\mathrm{Sat}}(G,R)$. In the $\infty$-categorical world, there is a hierarchy of commutativity constraints $E_2,E_3,\ldots,E_\infty$ on monoidal structures, and the one usually considered on ${\mathrm{Sat}}(G,R)$ is only $E_3$, not $E_\infty$, at the stable $\infty$-level. The difference between $E_3$ and $E_\infty$ monoidal structures vanishes when we restrict attention to the subcategory ${\mathcal{P}}$.
There is a good version of the monoidal category ${\mathrm{Sat}}(G,R)$ when $R = {\mathcal{T}}$ as well. The categorical version of our Smith theory, Theorem \[thm:main\], gives us a functor $${\mathbf{Psm}}:{\mathrm{Sat}}(G,{\mathbb{F}}_p) \to {\mathrm{Sat}}(Z_G(g),{\mathcal{T}})$$ As ${\mathbf{Psm}}$ commutes with all operations, including those used to define the convolution and fusion products on ${\mathrm{Sat}}(G,{\mathbb{F}}_p)$ and ${\mathrm{Sat}}(Z_G(g),{\mathcal{T}})$, this can be shown to be a symmetric (or better, per the Remark, an $E_3$) monoidal functor.
There is no perverse $t$-structure on ${\mathrm{Sat}}(Z_G(g),{\mathcal{T}})$. In fact, there can be no $t$-structure on ${\mathrm{Sat}}(Z_G(g),{\mathcal{T}})$ at all, for in the category of ${\mathcal{T}}$-modules the identity functor is isomorphic to the shift-by-two functor. However, we can “extend coefficients” from ${\mathbb{F}}_p$ to ${\mathcal{T}}$, which gives us a functor $$\otimes_{{\mathbb{F}}_p} {\mathcal{T}}: {\mathcal{P}}(Z_G(g);{\mathbb{F}}_p) \to {\mathrm{Sat}}(Z_G(g);{\mathcal{T}})$$ As ${\mathcal{P}}(Z_G(g);{\mathbb{F}}_p)$ is equivalent to the category of representations of the ${\mathbb{F}}_p$-form of $Z_G(g)^\vee$. A consequence of Theorem \[thm:1.4\] is that we may fill in the dotted arrow in the diagram $$\xymatrix{
& {\mathcal{P}}(Z_G(g);{\mathbb{F}}_p) \ar[d]^{\otimes_{{\mathbb{F}}_p} {\mathcal{T}}} \\
{\mathcal{P}}(G;{\mathbb{F}}_p) \ar[r]_{{\mathbf{Psm}}\quad} \ar@{-->}[ur] & {\mathrm{Sat}}(Z_G(g);{\mathcal{T}})
}$$ Under the equivalence of Mirkovi[ć]{} and Vilonen, the dotted arrow corresponds to the restriction functor of the inclusion $Z_G(g)^\vee \to G^\vee$ that exists over ${\mathbb{F}}_p$. With a better understanding of how ${\mathbf{Psm}}$ interacts with the theory of perverse sheaves, one might be able to give a Smith-theoretic construction of the dotted arrow, and presumably standard Tannakian considerations could then be used to deduce Theorem \[thm:1.4\]. We make a conjecture in Section \[sec:conjecture\] along these lines.
Notation and conventions
------------------------
To get a good theory of constructible functions and sheaves, we will work with real subanalytic sets. We refer to [@KS] for the basic theory of subanalytic geometry. We also often work with complex algebraic varieties, which we regard in their usual, locally compact and Hausdorff topology. We will let $\chi_c$ denote the compactly supported Euler characteristic of a subanalytic or complex algebraic set, which is always well-defined.
The symbol $p$ always denotes a prime number. We let ${\mathbf{k}}$ be a commutative ring (usually ${\mathbb{Z}}$ or ${\mathbb{Z}}/p$), and $K$ an algebraically closed field (usually $\overline{{\mathbb{F}}_p}$).
If $G$ is a group, we write $Z_G(g)$ for the centralizer of an element $g \in G$. The symbol $\varpi$ will usually denote a group of order $p$, and we sometimes write $Z_G(\varpi)$ instead of $Z_G(g)$ if $g$ has order $p$.
Constructible functions and Hecke algebras
==========================================
Let $X$ be a real subanalytic set. A function $f:X \to {\mathbf{k}}$ is called constructible if it is constant along the strata of a real subanalytic stratification of $X$, and zero away from a finite union of strata. Write ${\mathrm{Fun}}_{{\mathbb{R}-c}}(X,{\mathbf{k}})$ for the module of ${\mathbf{k}}$-valued functions on $X$ that are constructible with respect to a real subanalytic stratification of $X$. If $X$ has a complex algebraic structure write ${\mathrm{Fun}}_{{\mathbb{C}-c}}(X,{\mathbf{k}}) \subset {\mathrm{Fun}}_{{\mathbb{R}-c}}(X,{\mathbf{k}})$ for the ${\mathbf{k}}$-valued functions that are constructible with respect to a complex algebraic stratification of $X$. When it is clear from context whether we are working in the real or complex setting, and which ring ${\mathbf{k}}$ we are considering, we will often write simply ${\mathrm{Fun}}(X)$.
\[def:eulerint\] Let $\int:{\mathrm{Fun}}_{{\mathbb{R}-c}}(X,{\mathbf{k}}) \to {\mathbf{k}}$ denote the operator that takes a function $f$ to $$\sum_{i \in {\mathbf{k}}} \chi_c(f^{-1}(i)) \cdot i$$ Here $\chi_c$ denotes compactly-supported Euler characteristic.
Note that if $f$ is in ${\mathrm{Fun}}_{{\mathbb{C}-c}}(X,{\mathbf{k}})$, then we have the alternative formula $$\int f = \sum_{i \in {\mathbf{k}}} \chi(f^{-1}(i)) \cdot i$$ because of the relation $H^i_c(X) \cong H^{2n-i}(X)$ when $X$ is a smooth affine variety.
If $u:X \to Y$ is a subanalytic (resp. complex algebraic) map, we define operations $u^*:{\mathrm{Fun}}(Y) \to {\mathrm{Fun}}(X)$ and $u_!:{\mathrm{Fun}}(X) \to {\mathrm{Fun}}(Y)$ by the formulas $$\begin{array}{rcl}
u^*(f)(x) & = & f(u(x)) \\
u_!(f)(x) & = & \int_{u^{-1}(x)} f\vert_{u^{-1}(x)}
\end{array}$$
We have, essentially by definition, the fundamental “base-change” relation:
Suppose that the square $$\xymatrix{
X \ar[r]^v \ar[d]_{u'} & Y \ar[d]^u \\
Z \ar[r]_{v'} & W
}$$ is Cartesian. Then the operators $(v')^* \circ u_!:{\mathrm{Fun}}(Y) \to {\mathrm{Fun}}(Z)$ and $u'_! \circ v^*:{\mathrm{Fun}}(Y) \to {\mathrm{Fun}}(Z)$ are the same.
\[rem:cor1\] We can reformulate this relation in the language of categories. Let ${\mathcal{C}\mathrm{or}}_{\mathbb{R}}$ (resp. ${\mathcal{C}\mathrm{or}}_{\mathbb{C}}$) denote the category whose objects are real subanalytic spaces and whose morphisms are “correspondences” $u = (\overleftarrow{u},\overrightarrow{u})$, i.e. diagrams of the form $$X \stackrel{\overleftarrow{u}}{\leftarrow} U \stackrel{\overrightarrow{u}}{\rightarrow} Y$$ The composite of $u:X \to Y$ with $v:Y \to Z$ is given by the diagram $$X \stackrel{\overleftarrow{u}\circ \mathrm{proj}}{\longleftarrow} U \times_{\overrightarrow{u},\overleftarrow{v}} V \stackrel{\overrightarrow{v} \circ \mathrm{proj}}{\longrightarrow} Z$$
A correspondence $u = (\overleftarrow{u},\overrightarrow{u})$ determines an operation $\overrightarrow{u}_! \overleftarrow{u}^*:{\mathrm{Fun}}(X) \to{\mathrm{Fun}}(Y)$. The base-change relation is equivalent to the statement that $X \mapsto {\mathrm{Fun}}(X)$, $u \mapsto \overrightarrow{u}_! \overleftarrow{u}^*$ is a functor from ${\mathcal{C}\mathrm{or}}$ to the category of ${\mathbf{k}}$-modules.
Hecke algebras {#sec:heckealgebras}
--------------
Let $G$ be a Lie group (resp. complex algebraic group) and suppose that it acts subanalytically (resp. algebraically) on $X$. We define ${\mathrm{Fun}}_G(X)$ to be the ${\mathbf{k}}$-submodule of ${\mathrm{Fun}}(X)$ consisting of constructible functions that are constant along $G$-orbits. If $u:X \to Y$ is a $G$-equivariant map then the operations $u^*$ and $u_!$ carry the $G$-invariant submodules into each other.
We may use these operations to define a natural ring structure on ${\mathrm{Fun}}_G(X \times X)$, where $G$ acts by the diagonal action. Specifically, given $f_1$ and $f_2$ in ${\mathrm{Fun}}_G(X \times X)$, we define $$(f_1 \ast f_2)(x,y) = \int_z f_1(x,z) f_2(z,y)$$ It is clear that this new function is $G$-invariant.
The *Hecke algebra* associated to a $G$-space is the algebra ${\mathrm{Fun}}_G(X \times X)$ just described.
\[ex:doublecosets\] Let $X$ be a homogeneous space for $G$, say $X = G/H$. Let $H \times H$ act on $G$ by $(h_1,h_2) \cdot g = h_1 g h_2^{-1}$. The map ${\mathrm{Fun}}_G(X \times X) \to {\mathrm{Fun}}_{H \times H}(G)$ given by sending the function $f(g_1H,g_2H)$ to the function $f(1H,gH)$ is an isomorphism. This identifies ${\mathrm{Fun}}_G(X \times X)$ with the group of functions on $G$ that are constant on $H-H$ cosets. In particular, suppose $G$ is finite and $H$ is the trivial subgroup. Then ${\mathrm{Fun}}_G(G \times G;{\mathbf{k}})$ is naturally equivalent to the group ring of $G$ (though note that there are two such natural equivalences, which are exchanged by the involution $g \mapsto g^{-1}$ of the group ring).
The Smith operator
------------------
Suppose now that ${\mathbf{k}}$ has characteristic $p > 0$ and let $\varpi$ be the group ${\mathbb{Z}}/p$. Let $X$ be a space with a $\varpi$ action, and let $X^\varpi$ denote the fixed points. We define ${\mathrm{Fun}}_\varpi(X) \subset {\mathrm{Fun}}(X)$ to be the submodule of maps that are constant on $\varpi$-orbits. The Smith operator is the map $${\mathbf{Psm}}:{\mathrm{Fun}}_\varpi(X) \to {\mathrm{Fun}}(X^\varpi):f \mapsto f\vert_{X^\varpi}$$
It is worth emphasizing that on functions ${\mathbf{Psm}}$ does nothing more than restrict $f:X \to {\mathbb{F}}_p$ to the set of $\varpi$-fixed points.
Suppose $u:X \to Y$ is a $\varpi$-equivariant map, and by abuse of notation write $u$ also for the map $X^\varpi \to Y^\varpi$. We clearly have $u^*({\mathbf{Psm}}(f)) = {\mathbf{Psm}}(u^*(f))$. More interestingly, since ${\mathbf{k}}$ has characteristic $p$, we also have $u_!({\mathbf{Psm}}(f)) = {\mathbf{Psm}}(u_!(f))$.
\[prop:Smith\] Let $X$ and $Y$ be $\varpi = {\mathbb{Z}}/p$-spaces and let $u:X \to Y$ be a $\varpi$-equivariant map. Let $X^\varpi$ and $Y^\varpi$ denote the fixed points. Suppose that ${\mathbf{k}}$ has characteristic $p$. Then the square $$\xymatrix{
{\mathrm{Fun}}_\varpi(X,{\mathbf{k}}) \ar[r]^{{\mathbf{Psm}}}
\ar[d]_{u_!}
& {\mathrm{Fun}}(X^\varpi,{\mathbf{k}})
\ar[d]^{u_!} \\
{\mathrm{Fun}}_\varpi(Y,{\mathbf{k}}) \ar[r]_{{\mathbf{Psm}}} & {\mathrm{Fun}}(Y^\varpi,{\mathbf{k}})
}$$ is commutative
By the base-change relation, we may assume $Y$ is a point, i.e. we only have to show that $$\int_{X^\varpi} {\mathbf{Psm}}(f) = \int_X f$$ when ${\mathbf{k}}$ has characteristic $p$. By definition it suffices to show that $$\chi_c(f^{-1}(t)) - \chi_c(f^{-1}(t)^\varpi) = 0 \text{ mod }p$$ for every $t$. This follows from the fact that each $f^{-1}(t) - f^{-1}(t)^\varpi$ is triangulable compatible with the free $\varpi$-action.
\[rem:cor2\] Let ${\mathcal{C}\mathrm{or}}_\varpi$ denote the category of $\varpi$-spaces and $\varpi$-equivariant correspondences, defined in the evident way. The compatibility of the Smith operator with the operators $u_!$ and $u^*$ is equivalent to the statement that ${\mathbf{Psm}}$ defines a natural transformation between the functors of Remark \[rem:cor1\] $$\xymatrix{
{\mathcal{C}\mathrm{or}}_\varpi \ar[rr]^{X \mapsto X^\varpi} \ar[dr]_{{\mathrm{Fun}}_\varpi} & & {\mathcal{C}\mathrm{or}}\ar[dl]^{{\mathrm{Fun}}} \\
& {\mathbf{k}}\text{-mod} &
}$$
\[rem:pgroup\] We may generalize the Smith operator and Proposition \[prop:Smith\] to the case of finite $p$-groups. Let $\varrho$ be a $p$-group with $p^n$ elements and let ${\mathbf{Psm}}_{\varrho}$ denote the restriction map ${\mathrm{Fun}}_{\varrho}(X;{\mathbf{k}}) \to {\mathrm{Fun}}(X^\varrho;k)$. We may always find a normal subgroup $\varrho' \subset \varrho$ of order $p^{n-1}$, in which case ${\mathbf{Psm}}_\varrho$ factors as $${\mathrm{Fun}}_\varrho(X;{\mathbf{k}}) \stackrel{{\mathbf{Psm}}_{\varrho'}}{\longrightarrow} {\mathrm{Fun}}_{\varrho/\varrho'}(X^{\varrho'};{\mathbf{k}}) \stackrel{{\mathbf{Psm}}_{\varrho/\varrho'}}{\longrightarrow} {\mathrm{Fun}}(X^\varrho;{\mathbf{k}})$$. By induction on $n$, we conclude that $$\int {\mathbf{Psm}}_{\varrho} f = \int f$$ when $f \in {\mathrm{Fun}}_{\varrho}(X;{\mathbf{k}})$.
The Borel operator: localization for torus actions {#sec:u1}
--------------------------------------------------
We can cast some torus-localization results in similar terms, and use the Smith operator to deduce them. Let $T$ be a group isomorphic to ${\mathrm{U}}(1)^{\times n}$ if we are working in the real subanalytic setting or $({\mathbb{C}}^*)^{\times n}$ if we are working in the complex algebraic setting. Suppose that $T$ acts subanalytically or complex algebraically on $X$. We have a restriction map $${\mathrm{Fun}}_T(X;{\mathbb{Z}}) \to {\mathrm{Fun}}(X^T;{\mathbb{Z}}): f \mapsto f\vert_{X^T}$$ We refer to this map as the *Borel operator* ${\mathbf{Bor}}$.
\[rem:BorSmcompat\] We have a basic compatibility between the Borel and Smith operators. For each $p$, we may find a $p$-group $\varrho$ contained in $T$ with $X^\varrho = X^T$. In that case it is clear that reducing the Borel map for $T$ mod $p$ yields the Smith map for $\varrho$.
Suppose $T$ acts on both $X$ and $Y$ and that $f:X \to Y$ is $T$-equivariant. The square $$\xymatrix{
{\mathrm{Fun}}_T(X,{\mathbb{Z}}) \ar[r]^{\mathbf{Bor}}
\ar[d]_{u_!}
& {\mathrm{Fun}}(X^T,{\mathbb{Z}})
\ar[d]^{u_!} \\
{\mathrm{Fun}}_T(Y,{\mathbb{Z}}) \ar[r]_{\mathbf{Bor}} & {\mathrm{Fun}}(Y^T,{\mathbb{Z}})
}$$ commutes.
Let us give the proof in the subanalytic setting, the complex algebraic version is similar. We may reduce to the case $T = {\mathrm{U}}(1)$ by induction on the dimension of $T$. For each prime $p$ we have an inclusion ${\mathbb{Z}}/p \subset {\mathrm{U}}(1)$, and since the action is subanalytic we have $X^{{\mathrm{U}}(1)} = X^{{\mathbb{Z}}/p}$ for $p$ sufficiently large. The Proposition now follows from Proposition \[prop:Smith\] and Remark \[rem:BorSmcompat\], and the observation that a map between free abelian groups is determined by its reduction mod infinitely many primes.
The Smith operator for $G$-spaces
---------------------------------
Let $G$ be a Lie group (or a complex algebraic group) and let $X$ be a $G$-space. If we fix a subgroup of $G$ of the form $\varpi = {\mathbb{Z}}/p$, then ${\mathrm{Fun}}_G(X) \subset {\mathrm{Fun}}_\varpi(X)$. Moreover, if $f \in {\mathrm{Fun}}_G(X)$ then ${\mathbf{Psm}}(f):X^\varpi \to {\mathbf{k}}$ is constant along $Z_G(\varpi)$-orbits, where $Z_G(\varpi)$ is the centralizer of $\varpi$ in $G$. We define the equivariant Smith operator to be the function $${\mathbf{Psm}}:{\mathrm{Fun}}_G(X) \to {\mathrm{Fun}}_{Z(\varpi)}(X^\varpi)$$
\[rem:normalizer\] In fact ${\mathbf{Psm}}(f)$ is constant along orbits for the slightly larger group $N_G(\varpi) \supset Z_G(\varpi)$, the normalizer of $\varpi$ in $G$. If $\varpi$ is a larger $p$-group or torus, then $N_G(\varpi)$ can be quite a bit larger than $Z_G(\varpi)$ and this extra invariance is useful. However at the level of sheaves this larger group turns out to play a different role than $Z_G(\varpi)$, so we prefer to regard ${\mathbf{Psm}}$ as taking values in ${\mathrm{Fun}}_{Z_G}(X)$.
Combining the results of this section gives us the following:
\[thm:smithhecke\] Let $G$ act on $X$ and let $\varpi$ be a subgroup of $G$ of the from ${\mathbb{Z}}/p$. The Smith operator $${\mathbf{Psm}}:{\mathrm{Fun}}_G(X \times X,{\mathbf{k}}) \to {\mathrm{Fun}}_{Z_G(\varpi)}(X^\varpi \times X^\varpi,{\mathbf{k}})$$ is a ${\mathbf{k}}$-algebra homomorphism.
Ind-varieties {#sec:ind}
-------------
We will explain how to extend some of the formalism of constructible functions, Hecke algebras, and the Smith operator to the setting of ind-complex varieties. In this paper (in section \[sec:SHA\]) we only treat the case of $G({\mathcal{K}})$ acting on the affine Grassmannian, but we make a few general remarks here.
For us, an ind-variety is a topological space $X$ together with an exhaustive filtration $X \supset \cdots X_i \supset X_{i-1} \supset \cdots \supset X_0$ by closed subspaces, each equipped with the structure of a complex algebraic variety and each inclusion $X_j \hookrightarrow X_{j+1}$ being algebraic. See [@Nadler Section 2.2] for an exposition of ind-varieties relevant for our applications.
Suppose $X = \bigcup X_i$ is an ind-variety. Denote by ${\mathrm{Fun}}^{{\mathrm{fd}}}(X;{\mathbf{k}})$ the ${\mathbf{k}}$-module of functions on $X$ that are supported on one of the $X_i$, and that are constructible there. We call ${\mathrm{Fun}}^{{\mathrm{fd}}}(X;{\mathbf{k}})$ the ${\mathbf{k}}$-module of constructible functions with finite-dimensional support. On ind-varieties of the form $X \times X$ it is useful to consider a larger class of functions. Let us say a function $f:X \times X \to {\mathbf{k}}$ has “property H” if for every finite-dimensional subvariety $Z \subset X$, the functions $f \vert_{Z \times X}$ and $f\vert_{X \times Z}$ have finite-dimensional support. We ${\mathrm{Fun}}^{{\mathrm{H}}}$ denote the group of functions satisfying property $H$. It is clear that the operation $(f,g) \mapsto f \ast g$ where $$f \ast g(x,y) = \int_z f(x,z) g(z,y)$$ is well-defined whenever both $f$ and $g$ have property H.
In case $X$ is discrete, we may identify ${\mathrm{Fun}}^{{\mathrm{H}}}(X \times X;{\mathbf{k}})$ with the ${\mathbf{k}}$-module of infinite square matrices each row and each column of which have only finitely many nonzero entries.
If $G$ is an infinite discrete group then we may identify ${\mathrm{Fun}}_G^{\mathrm{H}}(G \times G)$ with the group ring of $G$ as in Example \[ex:doublecosets\].
It is clear that whenever $Y$ is an ind-subvariety of $X$, a function $f:X \times X \to {\mathbf{k}}$ with property H restricts to a function on $Y \times Y$ with property $H$. In particular if $\varpi$ acts on $X$ then we have a natural restriction map $${\mathbf{Psm}}:{\mathrm{Fun}}_\varpi^{{\mathrm{H}}}(X \times X;{\mathbf{k}}) \to {\mathrm{Fun}}^{{\mathrm{H}}}(X^\varpi \times X^\varpi;{\mathbf{k}})$$ which is an algebra homomorphism if ${\mathbf{k}}$ has characteristic $p$ and which we call the “Smith operator” as usual.
It is natural to consider actions of ind-pro algebraic groups ${\mathbf{G}}$ on ind-varieties $X$. In that case there is a subalgebra ${\mathrm{Fun}}_{{\mathbf{G}}}^{{\mathrm{H}}}(X\times X;{\mathbf{k}}) \subset {\mathrm{Fun}}^{{\mathrm{H}}}(X \times X;{\mathbf{k}})$ of H-functions that are constant on the ${\mathbf{G}}$-orbits of $X \times X$. If we have an inclusion $\varpi \subset {\mathbf{G}}$ then we have a ind pro algebraic subgroup $Z_{\mathbf{G}}(\varpi) \subset {\mathbf{G}}$ that centralizes $\varpi$. In that case the Smith operator defines an algebra homomorphism
$${\mathrm{Fun}}^{{\mathrm{H}}}_{{\mathbf{G}}}(X \times X;{\mathbf{k}}) \to {\mathrm{Fun}}^{{\mathrm{H}}}_{Z_{{\mathbf{G}}}(\varpi)}(X^\varpi \times X^\varpi;{\mathbf{k}})$$
Smith and the natural operators on constructible functions
----------------------------------------------------------
Let us recall some other important operations on functions, and show that the Smith operators is compatible with all of them.
### Duality
Let $M$ be a real analytic manifold. Let us call an open subset $U$ of $M$ *good* if there exists a subanalytic triangulation of $M$ for which $U$ is the star of a vertex. We moreover require that the triangulation is fine enough that the boundary of $U$ is a topological sphere. In that case we define constructible functions $i_U$ and $j_U$ by $$\begin{array}{c}
i_U(x) := \bigg\{
\begin{array}{ll}
1 & \text{if $x$ is in the closure of $U$}\\
0 & \text{otherwise}
\end{array} \\
j_U(x) := \bigg\{
\begin{array}{ll}
(-1)^{\dim(U)} & \text{if $x$ is in $U$}\\
0 & \text{if $x$ is outside of or on the boundary of $U$}
\end{array}
\end{array}$$ We refer to $i_U$ and $j_U$ as the *standard* and *costandard* functions on $M$ associated to $U$ respectively. Standard results on the existence of subanalytic triangulations imply that the functions $i_U$ (resp. $j_U$) generate ${\mathrm{Fun}}(M;{\mathbf{k}})$ as a ${\mathbf{k}}$-module.
The following variations on this fact are useful:
1. Suppose $\{{\mathcal{U}}_i\}_{i \in I}$ is an open cover of $M$. Then the collection of functions $j_U$ where $U$ is a good open subset of $M$ entirely contained in one of the ${\mathcal{U}}_i$ span ${\mathrm{Fun}}(M;{\mathbf{k}})$.
2. Suppose $X \subset M$ is a subanalytic set. Then the functions of the form $j_U\vert_X$ span ${\mathrm{Fun}}(X;{\mathbf{k}})$.
From this we can deduce the following:
\[thm:dualityfun\] There is a unique system of operators ${\mathbb{D}}_X:{\mathrm{Fun}}(X) \to {\mathrm{Fun}}(X)$ satisfying the following conditions:
1. When $M$ is a real analytic manifold and $U \subset M$ is a subanalytic open subset with smooth boundary, then ${\mathbb{D}}_X(i_U) = {\mathbb{D}}_X(j_U)$.
2. When $f:X \to Y$ is a closed immersion, then ${\mathbb{D}}_Y \circ f_! = f_! \circ {\mathbb{D}}_X$.
The operators ${\mathbb{D}}_X$ satisfy the following additional properties:
1. When $X \stackrel{\overleftarrow{u}}{\leftarrow} Z \stackrel{\overrightarrow{u}}{\rightarrow} Y$ is a subanalytic correspondence with $\overleftarrow{u}$ étale and $\overrightarrow{u}$ proper, the operator $\overrightarrow{u}_! \overleftarrow{u}^*:{\mathrm{Fun}}(X) \to {\mathrm{Fun}}(Y)$ intertwines ${\mathbb{D}}_X$ and ${\mathbb{D}}_Y$.
2. ${\mathbb{D}}_X \circ {\mathbb{D}}_X = \mathrm{id}$
See [@Schapira Theorem 2.5]
Let ${\mathbb{D}}_X$ be the operators of Theorem \[thm:dualityfun\]. Let $u:X \to Y$ be a subanalytic map between subanalytic varieties. We introduce the following notation:
1. We refer to ${\mathbb{D}}_X$ as the duality operator of $X$
2. We let $u_*:{\mathrm{Fun}}(X) \to {\mathrm{Fun}}(Y)$ denote the unique operator with ${\mathbb{D}}_Y \circ u_* = u_! \circ {\mathbb{D}}_X$.
3. We let $u^!:{\mathrm{Fun}}(Y) \to {\mathrm{Fun}}(X)$ denote the unique operator with ${\mathbb{D}}_Y \circ u^! = u^* \circ {\mathbb{D}}_Y$.
Thus, if $u$ is proper then $u_* = u_!$, and if $u$ is étale then $u^! = u^*$.
If $\varpi$ acts subanalytically on $X$, then applying property (2) to the translation maps in $\varpi$ we see that ${\mathbb{D}}_X$ must carry $\varpi$-invariant functions to $\varpi$-invariant functions. Similarly $u_*$ and $u^!$ carry $\varpi$-invariant functions to $\varpi$-invariant functions whenever $u$ is $\varpi$-equivariant.
\[prop:Smithduality\] Let $X$ and $Y$ be subanalytic varieties equipped with $\varpi$-actions, and let $u$ be a $\varpi$-equivariant map. Suppose ${\mathbf{k}}$ has characteristic $p$.
1. We have ${\mathbf{Psm}}\circ {\mathbb{D}}_X = {\mathbb{D}}_X \circ {\mathbf{Psm}}$
2. We have $u_* \circ {\mathbf{Psm}}= {\mathbf{Psm}}\circ u_*$
3. We have $u^! \circ {\mathbf{Psm}}= {\mathbf{Psm}}\circ u^!$
Pick a subanalytic embedding $X \hookrightarrow V$ into a real vector space $V$, and suppose that the embedding is equivariant for a linear $\varpi$-action on $V$. It suffices to verify that ${\mathbb{D}}_{V^\varpi} \circ {\mathbf{Psm}}(f) = {\mathbf{Psm}}\circ {\mathbb{D}}_{V}(f)$ for all $f \in {\mathrm{Fun}}_\varpi(V;{\mathbf{k}})$, and since the costandard functions $j_U$ where $U$ is $\varpi$-invariant good open subset generate we may assume $f$ is of this form. Since the $\varpi$-action is linear, $U^\varpi$ is a good open subset of $V^\varpi$, so $j_U\vert_{V^\varpi} = j_{U^\varpi}$. Then ${\mathbb{D}}_{V^\varpi} \circ {\mathbf{Psm}}(j_U) = i_{U^\varpi} = {\mathbf{Psm}}(i_U) = {\mathbf{Psm}}\circ {\mathbb{D}}_V(j_U)$. This proves property (1).
Properties (2) and (3) follow from property (1), Proposition \[prop:Smith\], and the definitions.
### Specialization {#sec:specialization}
Let $X$ be a subanalytic variety and let $u:X \to {\mathbb{R}}$ be a subanalytic map. Let $i$ denote the inclusion $u^{-1}(0) \hookrightarrow X$ and $j$ the inclusion $u^{-1}\{ t \mid t > 0\} \hookrightarrow X$. The *upper specialization* operator is the homomorphism $\psi_u^+:{\mathrm{Fun}}(X;{\mathbf{k}}) \to {\mathrm{Fun}}(u^{-1}(0);{\mathbf{k}})$ given by $$\psi_u^+(f) = i^* j_* j^* f$$
If $\varpi$ acts on $X$ and $u$ is constant on $\varpi$-orbits, $\varpi$ also acts on $u^{-1}(0)$ and the specialization map carries ${\mathrm{Fun}}_\varpi(X;{\mathbf{k}})$ to ${\mathrm{Fun}}_\varpi(u^{-1}(0);\varpi)$.
\[prop:speccompat\] Let $\varpi$ act on $X$ and let ${\mathbf{k}}$ have characteristic $p$. Suppose $u:X \to {\mathbb{R}}$ is constant on $\varpi$-orbits. We have a commutative square $$\xymatrix{
{\mathrm{Fun}}_\varpi(X;{\mathbf{k}}) \ar[r]^{{\mathbf{Psm}}} \ar[d]_{\psi^+_u} & {\mathrm{Fun}}(X^\varpi,{\mathbf{k}}) \ar[d]^{\psi_{u\vert_{X^\varpi}}^+} \\
{\mathrm{Fun}}_\varpi(u^{-1}(0);{\mathbf{k}}) \ar[r]_{{\mathbf{Psm}}} & {\mathrm{Fun}}(u^{-1}(0)^\varpi,{\mathbf{k}})
}$$
### Fourier-Sato transform {#sec:FT}
Let $V$ be a real vector space. We say a constructible function $f:V \to {\mathbf{k}}$ is *conic* if $f(t \cdot v) = f(v)$ for all real numbers $t > 0$. Let ${\mathrm{Fun}}_{{\mathbb{R}}_{>0}}(V;{\mathbf{k}}) \subset {\mathrm{Fun}}(V;{\mathbf{k}})$ denote the ${\mathbf{k}}$-module of conical constructible functions. The *Fourier-Sato transform* ${\mathcal{FT}}:{\mathrm{Fun}}_{{\mathbb{R}}_{>0}}(V;{\mathbf{k}}) \to {\mathrm{Fun}}_{{\mathbb{R}}_{>0}}(V^*;{\mathbf{k}})$ given by $${\mathcal{FT}}(f)(\xi) = \int_{\{v \in V \mid \xi(v) < 1\}} f$$
We have an evident relative version of this notion: if $V$ is a vector bundle over a space $X$, then ${\mathcal{FT}}$ carries dilation-invariant functions on the total space of $V$ to dilation-invariant functions on the total space of $V^*$.
Suppose now that $\varpi$ acts linearly on $V$. Note that the restriction map $(V^*)^\varpi \to (V^\varpi)^*$ is an isomorphism; we compute the inverse by sending a functional $\xi:V^\varpi \to {\mathbb{R}}$ to the $\varpi$-invariant functional $\tilde{\xi}$ given by $$\tilde{\xi}(v) = \frac{1}{p} \xi(\sum_{g \in \varpi} gx)$$ Write ${\mathrm{Fun}}_{{\mathbb{R}}_{>0},\varpi}(V;{\mathbf{k}})$ for the ${\mathbf{k}}$-module of constructible functions that are conical and constant on $\varpi$-orbits. The Smith operator defines a map $${\mathbf{Psm}}:{\mathrm{Fun}}_{{\mathbb{R}}_{>0},\varpi}(V;{\mathbf{k}}) \to {\mathrm{Fun}}_{{\mathbb{R}}_{>0}}(V^\varpi;{\mathbf{k}})$$
\[prop:FTcompat\] Let $V$ be a real vector space with a linear $\varpi$-action, and endow the dual vector space $V^*$ with the contragredient $\varpi$-action. The Smith operator is compatible with the Fourier-Sato transform and the identification $(V^\varpi)^* = (V^*)^\varpi$. That is, the following diagram commutes $$\xymatrix{
{\mathrm{Fun}}_{{\mathbb{R}}_{>0},\varpi}(V,{\mathbf{k}}) \ar[rr]^{{\mathcal{FT}}} \ar[d]_{{\mathbf{Psm}}} & & {\mathrm{Fun}}_{{\mathbb{R}}_{>0},\varpi}(V^*;{\mathbf{k}}) \ar[d]^{{\mathbf{Psm}}} \\
{\mathrm{Fun}}_{{\mathbb{R}}_{>0}}(V^\varpi;{\mathbf{k}}) \ar[r]^{{\mathcal{FT}}} & {\mathrm{Fun}}_{{\mathbb{R}}_{>0}}((V^\varpi)^*;{\mathbf{k}}) \ar[r]^{\xi \mapsto \tilde{\xi}} & {\mathrm{Fun}}_{{\mathbb{R}}_{>0}}((V^*)^\varpi;{\mathbf{k}})
}$$
As the averaging map ${\mathrm{Fun}}_{{\mathbb{R}}_{>0}}((V^\varpi)^*;{\mathbf{k}}) \to {\mathrm{Fun}}_{{\mathbb{R}}_{>0}}((V^*)^\varpi;{\mathbf{k}})$ is inverse to the restriction map ${\mathrm{Fun}}_{{\mathbb{R}}_{>0}}((V^*)^\varpi;{\mathbf{k}}) \to {\mathrm{Fun}}_{{\mathbb{R}}_{>0}}((V^\varpi)^*;{\mathbf{k}})$, the Proposition is equivalent to the assertion that the diagram $$\xymatrix{
{\mathrm{Fun}}_{{\mathbb{R}}_{>0},\varpi}(V,{\mathbf{k}}) \ar[rr]^{{\mathcal{FT}}} \ar[d]_{{\mathbf{Psm}}} & & {\mathrm{Fun}}_{{\mathbb{R}}_{>0},\varpi}(V^*;{\mathbf{k}}) \ar[d]^{{\mathbf{Psm}}} \\
{\mathrm{Fun}}_{{\mathbb{R}}_{>0}}(V^\varpi;{\mathbf{k}}) \ar[r]^{{\mathcal{FT}}} & {\mathrm{Fun}}_{{\mathbb{R}}_{>0}}((V^\varpi)^*;{\mathbf{k}}) & {\mathrm{Fun}}_{{\mathbb{R}}_{>0}}((V^*)^\varpi;{\mathbf{k}}) \ar[l]
}$$ commutes. Let $f:V \to {\mathbf{k}}$ be a conical, $\varpi$-invariant constructible function on $V$. The associated map $(V^\varpi)^* \to {\mathbf{k}}$ given by traveling through the upper right corner of the square is given by $$\xi \mapsto \int_{\{v \in V \mid \xi(v) < 1\}} f$$ while going through the bottom left corner is given by $$\xi \mapsto \int_{v \in V^\varpi \mid \xi(v) < 1\}} f\vert_{V^{\varpi}}$$ That these agree follows from the Proposition \[prop:Smith\].
### Specialization and microlocalization
Let $X$ be a manifold and $M \subset X$ a closed submanifold. Let $T_M X$ denote the normal bundle and $T_M^* X$ the conormal bundle to $M$ in $X$. Using the operators of \[sec:specialization\] and \[sec:FT\] one defines operators $$\begin{array}{c}
\nu_M:{\mathrm{Fun}}(X;{\mathbf{k}}) \to {\mathrm{Fun}}(T_M X;{\mathbf{k}}) \\
\mu_M:{\mathrm{Fun}}(X;{\mathbf{k}}) \to {\mathrm{Fun}}(T_M^* X;{\mathbf{k}})
\end{array}$$ called “specialization along $M$” and “microlocalization along $M$” respectively. The definition involves a new manifold $\widetilde{X}_M$ (see [@KS Chapter IV] for a construction) called the “deformation to the normal bundle” of $M$ in $X$. $\widetilde{X}_M$ is equipped with an action of ${\mathbb{R}}_{>0}$ and a map $\pi:\widetilde{X}_M \to {\mathbb{R}}$ with the following properties:
1. $\pi$ is ${\mathbb{R}}_{>0}$-equivariant
2. $\pi^{-1}(t)$ is naturally identified with $X$ for $t \neq 0$
3. $\pi^{-1}(0)$ is naturally identified with the normal bundle $T_M X$.
For each constructible function $f:X \to {\mathbf{k}}$ we may find a ${\mathbb{R}}_{>0}$-invariant constructible function $f_1:\widetilde{X}_M \to {\mathbf{k}}$ whose restriction to each nonzero fiber agrees with $f$ under the identification (2). Then we set $\nu_M(f) = \psi_\pi^+(f_1)$ and (as $\nu_M(f)$ is ${\mathbb{R}}_{>0}$-invariant) $\mu_M(f) = {\mathcal{FT}}(\nu_M(f))$.
The construction $(X,M) \mapsto \widetilde{X}_M$ is functorial, in particular if $\varpi$ acts on $X$ and $M$ is stable for this action then $\varpi$ acts on $\widetilde{X}_M$. Thus we have $\varpi$-equivariant versions of the operators $\nu_M$ and $\mu_M$. $$\begin{array}{c}
\nu_M:{\mathrm{Fun}}_\varpi(X;{\mathbf{k}}) \to {\mathrm{Fun}}_{{\mathbb{R}}_{>0},\varpi}(T_M X;{\mathbf{k}})\\
\mu_M:{\mathrm{Fun}}_\varpi(X;{\mathbf{k}}) \to {\mathrm{Fun}}_{{\mathbb{R}}_{>0},\varpi}(T^*_M X;{\mathbf{k}})
\end{array}$$
We can identify $(T_M X)^\varpi \cong T_{M^\varpi} X^\varpi$ and $(T_M X)^{*\varpi} \cong T_{M^\varpi}^* X^\varpi$ By Proposition \[prop:speccompat\] and \[prop:FTcompat\] we have
Suppose $\varpi$ acts on a manifold $X$ and $M$ is a $\varpi$-invariant submanifold. Then we have commutative squares $$\xymatrix{
{\mathrm{Fun}}_\varpi(X;{\mathbf{k}}) \ar[r]^{\nu_M} \ar[d]_{{\mathbf{Psm}}} & {\mathrm{Fun}}_\varpi(T_M X;{\mathbf{k}}) \ar[d]^{{\mathbf{Psm}}} &
{\mathrm{Fun}}_\varpi(X;{\mathbf{k}}) \ar[r]^{\mu_M} \ar[d]_{{\mathbf{Psm}}}&
{\mathrm{Fun}}_\varpi(T^*_M X;{\mathbf{k}}) \ar[d]^{{\mathbf{Psm}}} \\
{\mathrm{Fun}}(X^\varpi;k) \ar[r]_{\nu_{M^\varpi}} \ar[r] & {\mathrm{Fun}}(T_{M^\varpi} X^\varpi; {\mathbf{k}}) & {\mathrm{Fun}}_\varpi(X^\varpi;{\mathbf{k}}) \ar[r]_{\mu_{M^\varpi}} & {\mathrm{Fun}}(T^*_{M^\varpi} X^\varpi;{\mathbf{k}})
}$$
### Singular support {#sec:singsupp}
Let $M$ be a manifold and let $f:M \to {\mathbf{k}}$ be a constructible function. We define a subset ${\mathit{SS}}(f) \subset T^*M$ to be the closure of the set of $(x,\xi) \in T^*M$ with the following property: in every sufficiently small neighborhood $U$ of $x$ and $\epsilon >0$, if $\psi:U \to {\mathbb{R}}$ is a smooth function with $d\psi_x = \xi$, then $$\int_{\{u \in U \mid \psi(u) \leq \psi(x) + \epsilon\}} f \neq \int_{\{u \in U \mid \psi(u) \leq \psi(x) - \epsilon\}} f$$
If $\varpi$ acts on $M$ and $f$ is $\varpi$-invariant, then ${\mathit{SS}}(f)$ is also $\varpi$-invariant. We may identify ${\mathit{SS}}(f)^\varpi$ with a subset of $T^* (M^\varpi)$, and we have the following:
Suppose ${\mathbf{k}}$ has characteristic $p$. Then we have ${\mathit{SS}}({\mathbf{Psm}}(f)) = {\mathit{SS}}(f)^\varpi$.
The equality is local on $M$, so it suffices to consider the case where $M$ is an open subset of a vector space $V$. In that case we identify ${\mathit{SS}}(f)^\varpi$ with a subset of $T^*(V^\varpi)$ by sending $(x,\xi) \in {\mathit{SS}}(f)^\varpi \subset V \times V^*$ to $(x,\xi\vert_{V^\varpi})$.
A pair $(x,\xi) \in V^\varpi \times (V^\varpi)^*$ belongs to ${\mathit{SS}}({\mathbf{Psm}}(f))$ if and only if in a neighborhood $U$ of $x$ we may find a smooth $\psi:U \to {\mathbb{R}}$ such that $d\psi_x = f$ and $$\int_{\{u \in U \mid \psi(u) \leq \psi(x) + \epsilon\}} {\mathbf{Psm}}(f) \neq \int_{\{u \in U \mid \psi(u) \leq \psi(x) - \epsilon\}} {\mathbf{Psm}}(f)$$ We may always extend $U$ to a $\varpi$-stable open subset $U_1 \subset V$ with $U_1^\varpi = U$, and we may define $\psi_1:U_1 \to {\mathbb{R}}$ by $$\psi_1(u) = \psi(\frac{1}{p}\sum_{g \in \varpi} gu)$$ Then $\psi_1 \vert_U = \psi$ and $\{u \in U \mid \psi(u) \leq t\} = \{u \in U_1 \mid \psi_1(u) \leq t\}^\varpi$, so that by Proposition \[prop:Smith\] and the equation above we have $$\int_{\{u \in U_1 \mid \psi_1(u) \leq \psi_1(x) + \epsilon\}} f \neq \int_{\{u \in U_1 \mid \psi_1(u) \leq \psi_1(x) - \epsilon\}} f$$ As $(d\psi_1)_x = d\psi_x = \xi$, this is equivalent to $(x,\xi)$ belonging to ${\mathit{SS}}(f)$.
Smith theory for the spherical Hecke algebra {#sec:SHA}
============================================
In this section we work with complex algebraic varieties and ind-varieties. Let ${\mathcal{K}}$ denote the field of Laurent series ${\mathbb{C}}((t))$ and ${\mathcal{O}}$ denote the ring of Taylor series ${\mathbb{C}}[[t]]$. If $G$ is a complex reductive algebraic group then we have an ind-group $G({\mathcal{K}})$, a subgroup $G({\mathcal{O}})$, and a coset space $G({\mathcal{K}})/G({\mathcal{O}})$ which is an ind-variety in a natural way.
Let ${\mathbf{k}}$ be a commutative ring and let $G$ be a complex connected reductive algebraic group. The *spherical Hecke algebra* is the ${\mathbf{k}}$-module of constructible functions ${\mathrm{Fun}}^{{\mathrm{H}}}_{G({\mathcal{K}})}(G({\mathcal{K}})/G({\mathcal{O}}) \times G({\mathcal{K}})/G({\mathcal{O}}))$ endowed with the convolution product defined in Sections \[sec:heckealgebras\] and \[sec:ind\]. We will denote the spherical Hecke algebra attached to $G$ and ${\mathbf{k}}$ by ${\mathrm{SHA}}_{G,{\mathbf{k}}}$.
Note that if $G$ is not connected then $G({\mathcal{K}})/G({\mathcal{O}})$ is naturally identified with $G^\circ({\mathcal{K}})/G^{\circ}({\mathcal{O}})$, where $G^\circ$ is the neutral component of $G$. In that case $\pi_0(G) = \pi_0(G({\mathcal{K}}))$ acts on ${\mathrm{SHA}}_{G^{\circ},{\mathbf{k}}}$ and the invariant subring is ${\mathrm{SHA}}_{G,{\mathbf{k}}}$.
Suppose $G$ is connected. The Satake isomorphism identifies ${\mathrm{SHA}}_{G,{\mathbf{k}}}$ with the representation ring ${\mathrm{Rep}}(G^\vee)$ of the Langlands dual group to $G^\vee$. If $\varpi$ is a subgroup of order $p$ in $G$ and ${\mathbf{k}}$ has characteristic $p$ then it turns out the Smith homomorphism maps ${\mathrm{SHA}}_{G,{\mathbf{k}}}$ to ${\mathrm{SHA}}_{Z_G(\varpi);{\mathbf{k}}}$. If $G$ is simply connected then $Z_G(\varpi)$ is connected, and we can ask whether the corresponding homomorphism $${\mathrm{Rep}}(G^\vee) \otimes_{\mathbb{Z}}{\mathbf{k}}\to {\mathrm{Rep}}(Z_G(\varpi)^\vee) \otimes_{{\mathbb{Z}}} {\mathbf{k}}$$ has an interpretation. Before stating the main result of this section we make a couple of more observations:
- We can regard $G^\vee$ and $Z_G(\varpi)^\vee$ as algebraic groups defined over any algebraically closed field (or split algebraic groups defined over any field at all) without changing the structure of ${\mathrm{Rep}}(G^\vee)$ and ${\mathrm{Rep}}(Z_G(\varpi)^\vee)$.
- As $\varpi$ is commutative, $Z_G(\varpi)$ contains a maximal torus of $G$ and therefore has the same rank as $G$. It follows that $Z_G(\varpi)^\vee$ has the same rank as $G^\vee$.
\[thm:centralizer\] Let $G$ be a simply connected simple complex algebraic group. Let $\varpi \subset G$ be a subgroup of order $p$ and suppose that the centralizer $Z_G(\varpi)$ of $\varpi$ in $G$ is semsimple. Let $G^\vee$ and $Z_G(\varpi)^\vee$ be the Langlands dual groups over an algebraically closed field $K$
1. If $K$ has characteristic $p$, then there is an inclusion of $Z_G(\varpi)^\vee$ into $G^\vee$
2. The canonical bijections $\mathrm{coroots}(Z_G(\varpi)^\vee)) \cong \mathrm{roots}(Z_G(\varpi))$ and $\mathrm{coroots}(G^\vee) \cong \mathrm{roots}(G)$ commute with the inclusions induced by $Z_G(\varpi) \hookrightarrow G$ and $Z_G(\varpi)^\vee \hookrightarrow G^\vee$.
3. The square $$\xymatrix{
{\mathrm{Rep}}(G^\vee) \otimes_{\mathbb{Z}}{\mathbf{k}}\ar[r]^{\mathrm{res}\quad} \ar[d]_{\cong} & {\mathrm{Rep}}(Z_G(\varpi)^\vee) \otimes_{{\mathbb{Z}}} {\mathbf{k}}\ar[d]^{\cong} \\
{\mathrm{SHA}}_{G,{\mathbf{k}}} \ar[r]_{{\mathbf{Psm}}} & {\mathrm{SHA}}_{Z_G(\varpi),{\mathbf{k}}}
}$$ commutes.
\[rem:autom\] Assertion (2) is automatic from assertion (1): if there in an inclusion of $Z_G(\varpi)^\vee$ into $G^\vee$ then it must induce an inclusion of roots that preserves lengths and angles, and there is only one possibility. Moreover according to Lemma \[lem:37\] below, assertion (3) is implied by the following consequence of assertion (2): the induced map on Weyl groups $W_{Z_G(\varpi)^\vee} \to W_{G^\vee}$ is the same as the map $W_{Z_G(\varpi)} \hookrightarrow W_G$ induced by the inclusion of $Z_G(\varpi)$ into $G$, and the natural isomorphisms $W_{G^\vee} \cong W_G$, $W_{Z_G(\varpi)^\vee} \cong W_{Z_G(\varpi)}$. To prove the Theorem we will therefore only have to check (1).
In Section \[sec:31\] we prove the Satake isomorphism with the aid of the Smith and Borel operators. In Section \[sec:33\], we verify Theorem \[thm:centralizer\] by a case-by-case analysis.
The Satake isomorphism via Smith theory {#sec:31}
---------------------------------------
A subgroup $H$ of $G$ acts on $G({\mathcal{K}})/G({\mathcal{O}})$ by translation, and on $G({\mathcal{K}})$ by conjugation. We have the following basic fixed-point calculation:
Let $H$ be a reductive subgroup of $G$, and let $Z$ be the centralizer of $H$ in $G$.
- $G({\mathcal{K}})^H = Z({\mathcal{K}})$
- $(G({\mathcal{K}})/G({\mathcal{O}}))^H = Z({\mathcal{K}})/Z({\mathcal{O}})$
We will focus on the following special cases:
- If $H$ is a finite $p$-group and ${\mathbf{k}}$ has characteristic $p$, then we may consider the Smith map $${\mathbf{Psm}}:{\mathrm{SHA}}_{G,{\mathbf{k}}} \to {\mathrm{SHA}}_{Z_G(H),{\mathbf{k}}}$$ of Remark \[rem:pgroup\].
- If $H$ is a connected torus we may consider the Borel map $${\mathbf{Bor}}:{\mathrm{SHA}}_{G,{\mathbb{Z}}} \to {\mathrm{SHA}}_{Z_G(H),{\mathbf{k}}}$$ of Section \[sec:u1\].
Let us consider the Borel map first, in the case when $H = T$ is a maximal torus. With it we may deduce a version of the classical Satake isomorphism.
Let $G$ be a complex reductive algebraic group, let $T$ be a maximal torus, and let $W$ be the Weyl group. If we identify ${\mathrm{SHA}}_{T,{\mathbb{Z}}}$ with the group ring ${\mathbb{Z}}[X_*(T)]$ by the method of Example \[ex:doublecosets\], then the Borel operator for $T$ $${\mathbf{Bor}}:{\mathrm{SHA}}_{G,{\mathbb{Z}}} \to {\mathrm{SHA}}_{T,{\mathbf{k}}} \cong {\mathbb{Z}}[X_*(T)]$$ is an isomorphism onto the ring of $W$-invariants ${\mathbb{Z}}[X_*(T)]^W$.
The $G({\mathcal{O}})$-orbits on $G({\mathcal{K}})/G({\mathcal{O}})$ are in one-to-one correspondence with the $W$-orbits on $X_*(T)$ by the map $O \mapsto O \cap X_*(T)$. Equivalently, the $G({\mathcal{K}})$-orbits on $G({\mathcal{K}})/G({\mathcal{O}}) \times G({\mathcal{K}})/G({\mathcal{O}})$ are in one-to-one correspondence with the $X_*(T) \rtimes W$-orbits on $X_*(T) \times X_*(T)$. It follows that restricting a function on $f \in {\mathrm{SHA}}_{G,{\mathbb{Z}}}$ to a function on $X_*(T) \times X_*(T)$ is an isomorphism onto $W$-invariants.
### Note on representation rings {#sec:norr}
Let $K$ be an algebraically closed field and let $G$ be a reductive algebraic group defined over $K$. Let $T \subset G$ be a maximal torus for $G$ and let $W$ be the Weyl group of $G$. Let ${\mathrm{Rep}}(G)$ denote the representation ring of $G$, i.e. the Grothendieck ring of the $K$-linear tensor category of finite-dimensional algebraic representations of $G$. Recall that
1. The representation ring of $T$ itself is naturally identified with the Laurent polynomial ring ${\mathbb{Z}}[X^*(T)]$, where $X^*(T)$ denotes the character lattice of $T$. The action of $W$ on $T$ induces an action of $W$ on ${\mathbb{Z}}[X^*(T)]$.
2. Restricting to the maximal torus yields an injective homomorphism ${\mathrm{Rep}}(G) \to {\mathrm{Rep}}(T)$, which is an isomorphism onto the ring of $W$-invariants ${\mathrm{Rep}}(T)^W$.
We have the following trivial consequence:
\[lem:37\] Let $K$ be an algebraically closed field, let $G$ be an reductive algebraic $K$-group and let $H \subset G$ be a reductive algebraic $K$-subgroup of the same rank as $G$. Let $W_G$ and $W_H$ denote the Weyl groups of $G$ and $H$ respectively. There is a commutative square $$\xymatrix{
{\mathrm{Rep}}(G) \ar[r]^{\mathrm{res}^{G}_{H}} \ar[d]_{\cong} & {\mathrm{Rep}}(H) \ar[d]^{\cong} \\
{\mathbb{Z}}[X^*(T)]^{W_G} \ar[r]_{\mathrm{inclusion}} & {\mathbb{Z}}[X^*(T)]^{W_H}
}$$
Proof of Theorem \[thm:centralizer\] {#sec:33}
------------------------------------
By Remark \[rem:autom\], to prove Theorem \[thm:centralizer\] we only have to verify part (1). That is, we have to show that given a subgroup $\varpi \subset G$ of order $p$, the group $Z_G(\varpi)^\vee$ injects into $G^\vee$.
Recall some features of Kac’s classification of semsimple elements whose centralizer is semisimple:
1. Besides the identity element, there is one for each node in the Dynkin diagram associated to $G$.
2. The Dynkin diagram of the centralizer is obtained by deleting this vertex from the extended Dynkin diagram of $G$.
3. If $\{\alpha_i\}_{i \in I}$ are the simple roots of $G$ and $$\alpha_{\mathrm{top}} = \sum_{i \in I} c_i \alpha_i$$ is the maximal root, then if $c_i = 1$ the semisimple element corresponding to $\alpha_i$ is central, and otherwise it is of order $c_i$. In the noncentral case the element itself is the image of a $c_i$th root of unity under the fundamental coweight $\beta_i:{\mathbb{G}_\mathrm{m}}\to T$ corresponding to $\alpha_i$.
4. The center of $Z_G(\varpi)$ is a split extension of $\varpi$ by the center of $G$.
### Type $\mathrm{A}_n$
The only semisimple elements with semisimple centralizers are in the center of ${\mathrm{SL}}(n)$, in particular their centralizer is all of ${\mathrm{SL}}(n)$. Thus there is nothing to prove.
### Type $\mathrm{B}_n$
In types $\mathrm{B}_n$ and $\mathrm{C}_n$, the existence of the subgroups has to do with the peculiar nature of quadratic forms in characteristic 2. The essential fact for $\mathrm{B}_n$ is the following: if $q$ is a nondegenerate quadratic form on an even-dimensional vector space over a field of characteristic 2, then the associated bilinear form is alternating (i.e. we have $B(v,v) =0$). This gives us an inclusion $\iota_a:{\mathrm{SO}}(2a) \hookrightarrow {\mathrm{Sp}}(2a)$.
Let us use this fact to construct the interesting subgroups of $G^\vee$ in characteristic 2. The group $G$ is ${\mathrm{Spin}}(2n+1)$, which covers ${\mathrm{SO}}(2n+1)$. The Langlands dual is ${\mathrm{Sp}}(2n)/\mu_2$. The noncentral semisimple elements with semisimple centralizers are lifts of $$\left(
\begin{array}{cc}
-1_{2a} & 0 \\
0 & 1_{2(n-a)+1}
\end{array}
\right)$$ where $2 \leq a \leq 2n$. The centralizer is a double cover of ${\mathrm{SO}}(2a) \times {\mathrm{SO}}(2(n-a)+1)$, so that the Langlands dual is a quotient of ${\mathrm{SO}}(2a) \times {\mathrm{Sp}}(2(n-a))$ by a diagonal central $\mu_2$. The inclusion of this into ${\mathrm{Sp}}(2n)/\mu_2$ is covered by the inclusion $${\mathrm{SO}}(2a) \times {\mathrm{Sp}}(2(n-a)) \stackrel{\iota_a \times \mathrm{id}}{\longrightarrow} {\mathrm{Sp}}(2a) \times {\mathrm{Sp}}(2(n-a)) \to {\mathrm{Sp}}(2n)$$
### Type $\mathrm{C}_n$ {#subsubsec:Cn}
The group is ${\mathrm{Sp}}(2n)$, and the interesting centralizers are all of the form ${\mathrm{Sp}}(2a) \times {\mathrm{Sp}}(2b)$ where $a+b = n$. These can be described as the centralizers of an element of order 2, that acts as $-1$ on the $2a$-dimensional symplectic subspaces and $+1$ on the $2b$-dimensional symplectic subspace. We may describe the Langlands dual inclusion ${\mathrm{SO}}(2a + 1) \times {\mathrm{SO}}(2b+1)$ into ${\mathrm{SO}}(2a+2b+1)$ by noting the following feature of quadratic spaces of odd dimension in characteristic 2.
If $K$ has characteristic 2 and we endow $K^{2a+1}$ and $K^{2b+1}$ with the quadratic forms $$\begin{array}{c}
x_1x_2 + \cdots + x_{2a-1} x_{2a} + x_{2a+1}^2 \\
y_1y_2 + \cdots + y_{2b-1} y_{2b} + y_{2b+1}^2
\end{array}$$ then the line $\ell \subset K^{2a+1} \oplus K^{2b+1}$ given by setting $x_{2a+1} = y_{2b+1}$ and all other coordinates zero has the following properties with respect to the direct sum quadratic form $q$ on $K^{2a+1} \oplus K^{2b+1}$: it is perpendicular to everything, and $q$ vanishes identically on it. The quotient by this line is equivalent the standard quadratic form on $K^{2a+2b+1}$. In this way we get a homomorphism $\mathrm{O}(2a+1) \times \mathrm{O}(2b+1) \hookrightarrow \mathrm{O}(2a+2b+1)$ and ${\mathrm{SO}}(2a+1) \times {\mathrm{SO}}(2b+1) \hookrightarrow {\mathrm{SO}}(2a+2b+1)$.
### Type $\mathrm{D}_n$
The group is ${\mathrm{Spin}}(2n)$ and the Langlands dual is ${\mathrm{SO}}(2n)/\mu_2$. The interesting elements all have order 2: $$\left(
\begin{array}{cc}
-1_{2a} & 0 \\
0 & 1_{2(n-a)}
\end{array}
\right)$$ The centralizers are double covers of ${\mathrm{SO}}(2a) \times {\mathrm{SO}}(2(n-a))$, whose Langlands duals are $({\mathrm{SO}}(2a) \times {\mathrm{SO}}(2(n-a)))/\mu_2$. The inclusion into ${\mathrm{SO}}(2n)/\mu_2$ is covered by the usual inclusion of ${\mathrm{SO}}(2a) \times {\mathrm{SO}}(2(n-a))$ into ${\mathrm{SO}}(2n)$, which exists in every characteristic.
### Exceptional types {#sec:exc}
$$\begin{array}{c}
\begin{array}{ccc}
{\mathrm{G}}_2: \, {\mathbf 2} \equiv \!\!\!\!\! {>}\!\!\!\!\! \equiv {\mathbf 3} & \qquad & {\mathrm{F}}_4: \, {\mathbf 2} - {\mathbf 3} = \!\!\!\! {>}\!\!\!\! = 4 - {\mathbf 2}
\end{array}
\\
\\
\\
\begin{array}{ccc}
\begin{array}{lcccccc}
& & & \mathbf{2} & & &\\
{\mathrm{E}}_6: &1 & \mathbf{2} & \mathbf{3} & \mathbf{2} & 1
\end{array}
&
\begin{array}{lccccccc}
& & & \mathbf{2} & & & & \\
{\mathrm{E}}_7: &\mathbf{2} & \mathbf{3} & 4 & \mathbf{3} & \mathbf{2} & 1
\end{array}
&
\begin{array}{lcccccccc}
& & & \mathbf{3} & & & & &\\
{\mathrm{E}}_8: & \mathbf{2} & 4 & 6 & \mathbf{5} & 4 & \mathbf{3} & \mathbf{2}
\end{array}
\end{array}
\end{array}$$ The diagram displays the highest roots in the exceptional root systems, with the prime coefficients in boldface. We have to investigate subgroups of these groups in characteristics 2, 3, and 5. Maximal subgroups of the exceptional groups have been classified by Liebeck and Seitz [@ting], and it can be seen by consulting Table 10.3 of their manuscript that, with a single exception, each endoscopic group of $G^\vee$ associated to an element of order $p$ in $G$ does appear as a maximal subgroup of $G^\vee/K$ when $K$ has characteristic $p$.
The exception is the element $x \in {\mathrm{F}}_4$ of order 2 corresponding to the left-most node of the displayed diagram. As the affine Dynkin diagrm of ${\mathrm{F}}_4$ is $$\circ - \bullet - \bullet = \!\!\!\! {>}\!\!\!\! = \bullet - \bullet$$ the centralizer must have Dynkin diagram $$\circ \qquad \quad \bullet = \!\!\!\! {>}\!\!\!\! = \bullet - \bullet$$ i.e. it should be a quotient of ${\mathrm{SL}}(2) \times {\mathrm{Sp}}(6)$, where the highest root of ${\mathrm{SL}}(2)$ corresponds to the highest long root of ${\mathrm{F}}_4$ under the inclusion. In fact it must be ${\mathrm{SL}}(2) \times {\mathrm{Sp}}(6)$ modulo the diagonal copy of ${\mathbb{Z}}/2 \cong Z({\mathrm{SL}}(2)) \cong Z({\mathrm{Sp}}(6))$. To see this, one can reason as follows: the image of ${\mathrm{Sp}}(6)$ in ${\mathrm{F}}_4$ centralizes the image of ${\mathrm{SL}}(2)$, and in particular it must contain the center of ${\mathrm{SL}}(2) \to {\mathrm{F}}_4$ in its center. This map is injective by the next Lemma, so the image of the map from ${\mathrm{Sp}}(6)$ does not map $Z({\mathrm{Sp}}(6))$ to $1$, so it is itself injective.
\[lem:f41\] If $\alpha$ is a root (resp. coroot) of ${\mathrm{F}}_4$, then $\alpha$ is primitive in the weight lattice, i.e. $\alpha/k$ is not an integral weight for any $k \in {\mathbb{Z}}$. Because of this, the coroot homomorphism ${\mathrm{SL}}(2) \to {\mathrm{F}}_4$ is injective in any characteristic.
As both the center and fundamental group of ${\mathrm{F}}_4$ are trivial, it suffices to check the first assertion for the roots, and by symmetry it suffices to check it for the simple roots. In the weight basis, the simple roots of ${\mathrm{F}}_4$ are the columns of its Cartan matrix $$\left(
\begin{array}{rrrr}
2 & -1 & 0 & 0 \\
-1 & 2 & -1 & 0 \\
0 & -2 & 2 & -1 \\
0 & 0 & -1 & 2
\end{array}
\right)$$ which are evidently all primitive.
Then $Z_{{\mathrm{F}}_4}(x)^\vee$ must be ${\mathrm{SL}}(2) \times {\mathrm{Spin}}(7)$ mod its diagonal copy of ${\mathbb{Z}}/2$. The coroots of $Z_{{\mathrm{F}}_4}(x)^\vee$ correspond to the coroots of ${\mathrm{F}}_4^\vee \cong {\mathrm{F}}_4$ in the same way that the roots of $Z_{{\mathrm{F}}_4}(x)$ correspond to the roots of ${\mathrm{F}}_4$. If we are to have a map ${\mathrm{SL}}(2) \times {\mathrm{Spin}}(7) \to {\mathrm{F}}_4^\vee$, the ${\mathrm{SL}}(2)$ factor must map the highest root of ${\mathrm{SL}}(2)$ to the highest *short* root of ${\mathrm{F}}_4^\vee$. Let us denote this root by $\gamma$—its coefficients are given by $$1 - 2 = \!\!\!\! {>}\!\!\!\! = 3 - 2$$ The previous Lemma shows that the coroot map $\gamma^\vee:{\mathrm{SL}}(2) \to {\mathrm{F}}_4$ is injective, even in characteristic 2.
For each root $\alpha$ of ${\mathrm{F}}_4$, write $U_\alpha:{\mathbb{G}_\mathrm{a}}\hookrightarrow {\mathrm{F}}_4$ for the corresponding root subgroup. Let $\mathfrak{u}_\alpha \subset \mathfrak{f}_4$ denote the root space in the Lie algebra of ${\mathrm{F}}_4$.
\[lem:f42\] If $\gamma$ denotes the highest short root of the root system ${\mathrm{F}}_4$, then the centralizers of $U_\gamma$ and of $\mathfrak{u}_\gamma$ in ${\mathrm{F}}_4$ coincide in characteristic 2.
First, one notes that in characteristic 2 the unipotent radical $U$ of the Borel subgroup $B$ of ${\mathrm{F}}_4$ is contained in both centralizers. For $\mathfrak{u}_\gamma$, this follows from [@spalt Table 1]. For $U_\gamma$ one can consult the commutation relations in [@sh Section 2]. Let $T'$ denote the kernel of $\gamma$ regarded as a weight $T \to {\mathbb{G}_\mathrm{m}}$. $T'$ and $U$ together generate the centralizer of $U_\gamma$ (resp. of $\mathfrak{u}_\gamma$) in $B$.
Let $W$ denote the Weyl group of ${\mathrm{F}}_4$, and for each $w \in W$ fix an element $n_w \in N(T)$ mapping to $W$, such that $n_w U_\alpha(a) n_w^{-1} = U_{w \alpha}(a)$. Using the Bruhat decomposition we can write a general element of ${\mathrm{F}}_4$ as $g = t u_1 n_w u_2$ where $t \in T$ and $u_1,u_2 \in U$. As both $u_1$ and $u_2$ are automatically in the centralizer of $U_\gamma$ (resp. $\mathfrak{u}_\gamma$), we see that $g$ is in the centralizer if and only if $t \in T'$ and $n_w$ commutes with $U_\gamma$ (resp. with $\mathfrak{u}_\gamma$). This holds if and only if $w$ stabilizes $\gamma$, which holds if and only if $w$ is generated by the simple reflections associated to the left three nodes. In particular $g$ belongs to the centralizer of $U_\gamma$ if and only if it belongs to the centralizer of $\mathfrak{u}_\gamma$.
The centralizer of ${\mathrm{SL}}(2) \to {\mathrm{F}}_4$ is the reductive part of the centralizer of $U_\gamma$, which by the Lemma coincides with the reductive part of the centralizer of $\mathfrak{u}_\gamma$. By [@spalt Table 1], this is a simple algebraic group of type $\mathrm{B}_3$, i.e. it is either ${\mathrm{Spin}}(7)$ or ${\mathrm{SO}}(7)$. By Lemma \[lem:f41\], the center of this centralizer must contain the center of ${\mathrm{SL}}(2)$, so we see that it is ${\mathrm{Spin}}(7)$. This produces a map ${\mathrm{SL}}(2) \times {\mathrm{Spin}}(7) \to {\mathrm{F}}_4$ whose kernel is the diagonal copy of $\mu_2 = Z({\mathrm{SL}}(2)) = Z({\mathrm{Spin}}(7))$, as required.
By inspecting the table in [@ting], one sees that in these exceptional types there are only three endoscopic groups that are not subgroups in all characteristics: ${\mathrm{PGL}}(3)$ for ${\mathrm{G}}_2$, correspoding to the node labeled “3” in the ${\mathrm{G}}_2$ diagram. ${\mathrm{Sp}}(8)$ for ${\mathrm{F}}_4$ corresponding to the right-most node in the ${\mathrm{F}}_4$ diagram, and the ${\mathrm{SL}}(2) \times {\mathrm{Spin}}(7) /\mu_2$ in ${\mathrm{F}}_4$ that we have just discussed, corresponding to the left-mode node in the ${\mathrm{F}}_4$ diagram.
Smith theory for sheaves
========================
Let $X$ be a real subanalytic or complex algebraic variety. Let $K$ be a commutative ring. We let $D^b_{{\mathbb{R}}-c}(X;K)$ (resp. $D^b_{{\mathbb{C}}-c}(X;K)$) denote the triangulated category of bounded cohomologically ${\mathbb{R}}$-constructible (resp. ${\mathbb{C}}$-constructible) sheaves of $K$-modules on $X$. We will usually abuse notation and write $D^b_c(X;K)$ for one of these categories, and it should be clear from context whether we are in the subanalytic or complex algebraic settings. If $G$ is a Lie group (resp. complex algebraic group) acting subanalytically (resp. algebraically) on $X$, write $D^b_G(X;K)$ for the $G$-equivariant version of this category.
The Tate coefficient category
-----------------------------
Let $\varpi = {\mathbb{Z}}/p$ and let $K$ be a field of characteristic $p$. Let $K[\varpi]$ be the group ring of $\varpi$, and let $D^b(K[\varpi])$ be the bounded derived category of finitely-generated $K[\varpi]$-modules. We have a thick subcategory ${\mathrm{Perf}}(K[\varpi])$ spanned by bounded complexes of free $K[\varpi]$-modules.
The *Tate category* is the Verdier quotient category $D^b(K[\varpi])/{\mathrm{Perf}}(K[\varpi])$. Write ${\mathrm{Perf}}({\mathcal{T}})$ for the Tate category.
The Grothendieck group of the Tate category is ${\mathbb{Z}}/p$, generated by the class of the trivial $K[\varpi]$-module $K$.
The algebra $K[{\mathbb{Z}}/p]$ is local: it has only one simple module $K$. It follows that $K$ generates the Grothendieck group of $D^b(K[\varpi])$ as well as the Grothendieck group of any localization of $D^b(K[\varpi])$. To show that the Grothendieck group of ${\mathrm{Perf}}({\mathcal{T}})$ is ${\mathbb{Z}}/p$ then, it suffices to exhibit a ${\mathbb{Z}}/p$-valued invariant $\chi$ of objects of $D^b(K[\varpi])$ with the following properties:
1. $\chi$ is additive for exact triangles
2. $\chi(M^\bullet) = 0$ when $M^\bullet$ belongs to ${\mathrm{Perf}}(K[\varpi])$
3. $\chi(K) = 1$
It is easy to check that the invariant $$\chi(M^\bullet) = \sum_{i \in {\mathbb{Z}}} (-1)^i \dim_K(M^i) \text{ mod }p$$ has the required properties.
\[prop:periodic\] The shift-by-2 functor ${\mathrm{Perf}}({\mathcal{T}}) \to {\mathrm{Perf}}({\mathcal{T}}): M \mapsto M[2]$ is naturally isomorphic to the identity functor. If $p = 2$, then the shift-by-1 functor is naturally isomorphic to the identity functor.
Let $g$ be a generator of $\varpi$. For any $K[\varpi]$-module $M$, we have the exact sequence $$0 \to M \to M \otimes_K K[\varpi] \stackrel{1 - g}{\to} M \otimes_K K[\varpi] \to M \to 0$$ where $\varpi$ acts diagonally on the middle two terms. The associated short exact sequence of cochain complexes $$\xymatrix{
0 \ar[r] & M \ar[r] \ar[d] & M \otimes_K K[\varpi] \ar[r] \ar[d] & 0 \ar[r] \ar[d] & 0 \\
0 \ar[r] & 0 \ar[r] & M \otimes_K K[\varpi] \ar[r] & M \ar[r] & 0
}$$ induces a map $M \to M[2]$ whose cone is a 2-term complex of free modules. The proposition follows.
When $p = 2$ we may use the shorter exact sequence $$0 \to M \to M \otimes_K K[\varpi] \to M \to 0$$ to deduce the Proposition.
\[rem:tatespec\] The proposition shows that ${\mathrm{Perf}}({\mathcal{T}})$ is not the derived category of any abelian category, and indeed can carry no $t$-structure at all. However it can be shown that ${\mathrm{Perf}}({\mathcal{T}})$ is equivalent to the homotopy category of a certain category of module spectra over an $E_\infty$-ring spectrum ${\mathcal{T}}$. Basically, ${\mathcal{T}}$ is the natural ring spectrum whose homotopy groups are the Tate cohomology groups of ${\mathbb{Z}}/p$ with coefficients in $K$. If $p$ is odd, then the homotopy groups of ${\mathcal{T}}$ are $$\begin{array}{c}
\pi_{2i}({\mathcal{T}}) = K \text{ with generator }x^i\\
\pi_{2i+1}({\mathcal{T}}) = K \text{ with generator } x^i y
\end{array}$$ with the evident ring structure. If $p=2$, then we have $\pi_\bullet({\mathcal{T}}) = K[y,y^{-1}]$ with $y \in \pi_1$. The fact that the natural class in $\pi_2$ is invertible accounts for Proposition \[prop:periodic\].
### Tensor structure {#sec:tensorcoef}
Related to Remark \[rem:tatespec\], it is possible to endow ${\mathrm{Perf}}({\mathcal{T}})$ with a symmetric monoidal structure. If $M^\bullet$ and $N^\bullet$ are two bounded complexes of finitely generated $K[\varpi]$-modules, the tensor product $M^\bullet \otimes_K N^\bullet$ is another bounded complex equipped with the diagonal $K[\varpi]$-module structure, endowing $D^b(K[\varpi])$ with a symmetric monoidal structure.
If $M^\bullet \in D^b(K[\varpi])$ and $N^\bullet \in {\mathrm{Perf}}(K[\varpi])$, then $M^\bullet \otimes_K N^\bullet \in {\mathrm{Perf}}(K[\varpi])$. Thus, $\otimes_K$ descends to a symmetric monoidal structure on ${\mathrm{Perf}}({\mathcal{T}})$, which we denote by $\otimes_{{\mathcal{T}}}$.
\[rem:notKpilinear\] Since $K[\varpi]$ is a commutative ring, we can define the $K[\varpi]$-linear tensor product $\stackrel{\mathbf{L}}{\otimes}_{K[\varpi]}$. Note that this is not the one we are considering when we define $\otimes_{{\mathcal{T}}}$. Indeed, $D^b(K[\varpi])$ is not even closed under $\stackrel{\mathbf{L}}{\otimes}_{K[\varpi]}$, e.g. $K \stackrel{\mathbf{L}}{\otimes}_{K[\varpi]} K$ is unbounded below.
### Duality {#sec:dualitycoef}
If $M^\bullet$ is a bounded complex of finitely generated $K[\varpi]$-modules, we let $(M^\bullet)^*$ denote the complex of dual $K$-vector spaces equipped with the contragredient $\varpi$-action. As in Remark \[rem:notKpilinear\], we note that this functor differs from the duality functor $M^\bullet \mapsto \mathbf{R}Hom_{K[\varpi]}(M^\bullet,K[\varpi])$.
If $M^\bullet$ is a bounded complex of free modules then so is $(M^\bullet)^*$. The functor $M \mapsto M^*$ therefore descends to a duality functor on ${\mathrm{Perf}}({\mathcal{T}})$, which we denote by ${\mathbb{D}}$.
Tate coefficients and the Smith operation
-----------------------------------------
If $Y$ is a real subanalytic variety and $\varpi$ acts trivially on $Y$, then we make the identification $$D^b_\varpi(Y;K) \cong D^b(Y;K[\varpi])$$ Let us denote by ${\mathrm{Perf}}(Y;K[\varpi]) \subset D^b(Y;K[\varpi])$ the full subcategory spanned by sheaves of $K[\varpi]$-modules all of whose stalks are perfect. We will denote the Verdier quotient of $D^b(Y;K[\varpi])$ by ${\mathrm{Perf}}(Y;K[\varpi])$ by ${\mathrm{Perf}}(Y;{\mathcal{T}})$.
\[lem:link\] Let $X$ be a finite-dimensional space on which $\varpi$ acts freely. Then the global sections functor $\Gamma:D^b_\varpi(X;K) \to D^b_\varpi(\mathit{pt};K) = D^b(K[\varpi])$ takes values in ${\mathrm{Perf}}(K[\varpi])$.
It suffices to show that $\Gamma(F)$ is a perfect complex of $K[\varpi]$-modules when $F$ is the constant sheaf on a $\varpi$-invariant closed subset $Y$, as these sheaves generate $D^b_\varpi(X;K)$. Pick a $\varpi$-invariant triangulation of $Y$. Then $\Gamma(F)$ is quasi-isomorphic to the simplicial cochain complex of this simplicial with coefficients in $K$, together with its natural $\varpi$-action. As $\varpi$ acts freely on $Y$ it acts freely on the set of $i$-simplices in $Y$, and therefore this cochain complex is perfect.
\[thm:link\] Let $X$ be a $\varpi$-space and let $i$ denote the inclusion of $X^\varpi$ into $X$. The cone on the natural map $i^! \to i^*$ belongs to ${\mathrm{Perf}}(X^\varpi;K[\varpi])$.
Let $x$ be a $\varpi$-fixed point of $X$, and let $U$ be a regular neighborhood of $x$. As $\varpi$ is finite we may assume $U$ is $\varpi$-invariant. Let $L = U - (U \cap X^\varpi)$. A standard computation identifies the stalk of $C$ at $x$ with the cohomology of $L$ with coefficients in $F\vert_L$. By Lemma \[lem:link\] this is perfect. It follows that $C$ is perfect.
The sheaf-theoretic *Smith operation* is the composite functor $$D^b_\varpi(X;K) \stackrel{i^*}{\to} D^b_\varpi(X^\varpi,K) \cong D^b(X^\varpi,K[\varpi]) \to {\mathrm{Perf}}(X^\varpi;{\mathcal{T}})$$ We denote the functor by ${\mathbf{Psm}}$.
The previous theorem shows that we could define this operation with $i^!$ in place of $i^*$.
\[rem:hyperbolic\] If $X$ is a complex algebraic variety carrying an action of ${\mathbb{C}}^*$, then in between $X$ and $X^{{\mathbb{C}}^*}$ we have the attracting set $X^+$. The hyperbolic localization functor is defined to be the composition of shriek and star restriction functors $$(X^{{\mathbb{C}}^*} \hookrightarrow X^+)^! \circ (X^+ \hookrightarrow X^{{\mathbb{C}}^*})^*$$ Smith localization is analogous to hyperbolic localization in the following sense: instead of combining the two restriction functors in a clever way, we simply erase the distinction between them.
Six operations with Tate coefficients
-------------------------------------
Suppose that $Y$ is a variety equipped with the trivial $\varpi$-action.
### Duality and tensor product
Under the identification $D^b_\varpi(Y;K) \cong D^b(Y;K[\varpi])$, the $\varpi$-equivariant Verdier duality operation is a sheaf version of the operation considered in section \[sec:dualitycoef\]. Since an object of $D^b(Y;K[\varpi])$ belongs to the subcategory ${\mathrm{Perf}}(Y;K[\varpi])$ if and only if each stalk belongs to ${\mathrm{Perf}}(K[\varpi])$, the duality operation preserves ${\mathrm{Perf}}$ and descends to an operation on ${\mathrm{Perf}}(Y;{\mathcal{T}})$. Similarly the tensor product considered in section \[sec:tensorcoef\] gives a symmetric monoidal structure on $D^b(Y;K[\varpi])$ that descends to a symmetric monoidal structure on ${\mathrm{Perf}}(Y;{\mathcal{T}})$.
### Pushforward and pullback {#subsubsec:pandp}
Let $Y'$ be a second variety equipped with the trivial $\varpi$-action, and let $u:Y \to Y'$ be a morphism. If $F'$ is a sheaf of $K[\varpi]$-modules on $Y'$, then the stalk of $u^* F'$ at $y$ is isomorphic to the stalk of $F'$ at $u(y)$. It follows that $u^*$ carries perfect sheaves of $K[\varpi]$-modules to perfect sheaves of $K[\varpi]$-modules, and descends to an operation $u^*:{\mathrm{Perf}}(Y';{\mathcal{T}}) \to {\mathrm{Perf}}(Y;{\mathcal{T}})$. Similary:
If $F$ is a sheaf of perfect $K[\varpi]$-modules on $Y$ then $u_! F$ is a sheaf of perfect $K[\varpi]$-modules on $Y'$.
By proper base-change we may assume $Y'$ is a point. By induction on the length of $F$ we may furthermore assume assume that $F$ is a sheaf of free $K[\varpi]$-modules concentrated in a single degree, i.e. $F = F_1 \otimes_K K[\varpi]$ where $F_1$ is a sheaf of $K$-modules concentrated in a single degree. Then $u_!(F) = u_!(F_1) \otimes_K K[\varpi]$ is perfect since $u_!(F_1)$ vanishes in degrees $\geq \dim(Y)$.
It follows that $u_!$ induces a functor ${\mathrm{Perf}}(Y;{\mathcal{T}}) \to {\mathrm{Perf}}(Y';{\mathcal{T}})$. Since Verdier duality preserves perfect sheaves we also have well-defined functors $u_*$ and $u^!$.
Compatibility of Smith with six operations
------------------------------------------
Two classical applications of Smith theory are the following:
1. In [@Q], Quillen extends Smith’s original result (Theorem \[thm:1.2\]), and shows that the cohomology of a finite-dimensional space with mod $p$ coefficients is closely related to the cohomology of the ${\mathbb{Z}}/p$-fixed points with mod $p$ coefficients.
2. In [@Bre] and [@CK], it is shown that the fixed points of a ${\mathbb{Z}}/p$-action on a space that satisfies Poincaré duality mod $p$ again satisfies Poincaré duality mod $p$.
These results are consequences of the following general principle: the Smith operation commutes with all other operations. A generalization of (1) states that ${\mathbf{Psm}}$ is compatible with pushforwards, and a generalization of (2) states that ${\mathbf{Psm}}$ is compatible with Verdier duality. A somewhat more trivial result is that ${\mathbf{Psm}}$ commutes with pullback; let us prove this result first.
\[thm:Smithpull412\] Let $X$ and $Y$ be real subanalytic varieties with an action of $\varpi$. Let $f:X \to Y$ be a $\varpi$-equivariant morphism between them. The square $$\xymatrix{
D^b_\varpi(Y;K) \ar[r]^{f^*} \ar[d]_{{\mathbf{Psm}}} & D^b_\varpi(X;K) \ar[d]^{{\mathbf{Psm}}} \\
{\mathrm{Perf}}(Y^\varpi;{\mathcal{T}}) \ar[r]_{f^*} & {\mathrm{Perf}}(X^\varpi;{\mathcal{T}}) }$$ commutes up to a natural isomorphism.
Let $i_X:X^\varpi \to X$ and $i_Y:Y^\varpi \to Y$ denote the inclusion maps. As $f \circ i_X = i_Y \circ (f\vert_{X^\varpi})$ we have a natural isomorphism $i_X^* \circ f^* F \cong (f\vert_{X^\varpi})^*\circ i_Y^* F$ in $D^b_\varpi(X^\varpi;K)$, which induces an isomorphism between ${\mathbf{Psm}}\circ f^* F$ and $(f\vert_{X^\varpi})^* \circ {\mathbf{Psm}}(F)$ in ${\mathrm{Perf}}(X^\varpi;{\mathcal{T}})$.
\[thm:duality\] Let $X$ be a real subanalytic variety with an action of $\varpi$. The square $$\xymatrix{
D^b_\varpi(X;K) \ar[r]^{{\mathbb{D}}_X} \ar[d]_{{\mathbf{Psm}}} & D^b_\varpi(X;K) \ar[d]^{{\mathbf{Psm}}} \\
{\mathrm{Perf}}(X^\varpi;{\mathcal{T}}) \ar[r]_{{\mathbb{D}}_{X^\varpi}} & {\mathrm{Perf}}(X^\varpi;{\mathcal{T}})
}$$ commutes up to a natural isomorphism.
Let $i:X^\varpi \to X$ denote the inclusion. We construct a natural transformation ${\mathbb{D}}i^* F \to i^* {\mathbb{D}}F$ by composing the natural isomorphism ${\mathbb{D}}i^* F \cong i^! {\mathbb{D}}F$ with the natural transformation $i^! {\mathbb{D}}F \to i^* {\mathbb{D}}F$. To show that the induced map ${\mathbf{Psm}}{\mathbb{D}}F \to {\mathbb{D}}{\mathbf{Psm}}(F)$ is an isomorphism it suffices to show that the cone on ${\mathbb{D}}i^* F \to i^* {\mathbb{D}}F$ is perfect—this follows from Theorem \[thm:link\].
\[thm:Smithpush\] Let $X$ and $Y$ be real subanalytic varieties with an action of $\varpi$. Let $f:X \to Y$ be a $\varpi$-equivariant morphism between them. The squares $$\xymatrix{
D^b_\varpi(X;K) \ar[r]^{f_!} \ar[d]_{{\mathbf{Psm}}} & D^b_\varpi(Y;K) \ar[d]^{{\mathbf{Psm}}} & & D^b_\varpi(X;K) \ar[r]^{f_*} \ar[d]_{{\mathbf{Psm}}} & D^b_\varpi(Y;K) \ar[d]^{{\mathbf{Psm}}} \\
{\mathrm{Perf}}(X^\varpi;{\mathcal{T}}) \ar[r]_{(f\vert_{X^\varpi})_!} & {\mathrm{Perf}}(Y^\varpi;{\mathcal{T}}) & & {\mathrm{Perf}}(X^\varpi;{\mathcal{T}}) \ar[r]_{(f\vert_{X^\varpi})_*}& {\mathrm{Perf}}(Y^\varpi;{\mathcal{T}})
}$$ commute up to natural isomorphisms.
As ${\mathbb{D}}\circ f_! \circ {\mathbb{D}}= f_*$, it is enough to consider the left-hand square. Let $i_X:X^\varpi \to X$ and $i_Y:Y^\varpi \to Y$ be the inclusion maps. We have an adjunction morphism $$i_Y^* f_! \to (f\vert_{X^\varpi})_! i_X^*$$ in the category of functors from $D^b_\varpi(X;K) \to D^b_\varpi(Y^\varpi;K)$, which induces a morphism $n:{\mathbf{Psm}}\circ f_! \to f_! \circ {\mathbf{Psm}}$.
To show that $n$ is an isomorphism it suffices to show that the cone on $i_Y^* f_! F \to (f\vert_{X^\varpi})_! i_X^*F$ is a perfect. Since this may be checked on stalks, we may as well assume that $Y$ is a single point. We may furthermore reduce to the case where $F$ is a constant sheaf on a closed subanalytic $\varpi$-invariant subset, as these sheaves generate $D^b_\varpi(X;K)$. Now we only have to check that the cone on the map $$\Gamma_c(K_X) \to \Gamma_c(K_{X^\varpi})$$ has a perfect cone. This may be verified as in the proof of Lemma \[lem:link\].
\[thm:Smithnearby\] Let $X$ be a complex algebraic variety with an action of $\varpi$, and let $f:X \to {\mathbb{C}}$ be a $\varpi$-invariant map, and let $\psi_f$ denote the nearby cycles functor $$\psi_f:D^b_\varpi(X;K) \to D^b_\varpi(f^{-1}(0);K)$$ The square $$\xymatrix{
D^b_\varpi(X;K) \ar[r]^{\psi_f\quad} \ar[d]_{{\mathbf{Psm}}} & D^b_\varpi(f^{-1}(0);K) \ar[d]^{{\mathbf{Psm}}} \\
{\mathrm{Perf}}(X^\varpi;{\mathcal{T}}) \ar[r]_{\psi_f\quad} & {\mathrm{Perf}}(f^{-1}(0)^\varpi;{\mathcal{T}})
}$$ commutes up to natural isomorphism.
We have $\psi_f = i^*j_* j^*$ where $j$ is the inclusion of $f^{-1}({\mathbb{R}}_{>0})$ into $X$ and $i$ is the inclusion of $f^{-1}(0)$ into $X$. The Theorem the follows from Theorems \[thm:Smithpull412\] and \[thm:Smithpush\].
Smith theory for equivariant sheaves
------------------------------------
Let $X$ be a subanalytic space and $G$ a Lie group acting subanalytically on $X$. The bar construction $[X/G]$ is the simplicial space whose $n$th term is $G^{\times n} \times X$, and whose face and degeneracy maps are given by multiplication and insertion. A $G$-equivariant sheaf on $X$ is a complex of simplicial sheaves on $[X/G]$ whose cohomology simplicial sheaves are “effective” or “Cartesian.” We refer to [@BL] for more details; informally we are given a complex of sheaves $F^k$ on each $G^{\times k} \times X$, together with a quasi-isomorphism $\phi^* F^\ell \to F^k$ for each structure map $\phi:G^{\times k} \times X \to G^{\times \ell} \times X$, and these quasi-isomorphisms are required to be compatible with each other in a suitable sense. We call an equivariant sheaf bounded if each $F^k$ is finitely many nonzero cohomology sheaves, or equivalently if $F^0$ has finitely many nonzero cohomology sheaves.
Let us denote the triangulated category of bounded $G$-equivariant sheaves on $X$ by $D^b([X/G];K)$. If we have a subgroup $\varpi \subset G$, then $\varpi$ acts on each term $G^{\times k} \times X$ of $[X/G]$ in the following way: $$h \cdot (g_1,\ldots,g_k,x) = (hg_1h^{-1},\ldots,hg_k h^{-1},hx)$$ We let $D^b_\varpi([X/G];K)$ denote the category of Cartesian sheaves on the bisimplicial space $\varpi^j \times G^k \times X$. Roughly speaking, to give an object of $D^b_\varpi([X/G];K)$ one gives a $\varpi$-equivariant sheaf $F^k$ on each $G^{\times k} \times X$, together with a $\varpi$-equivariant quasi-isomorphism for each structure map $G^{\times k} \times X \to G^{\times \ell} \times X$.
Applying ${\mathbf{Psm}}$ term-by-term gives us a functor $$D^b_\varpi([X/G];K) \to D^b([X^\varpi/Z_G(\varpi)];{\mathcal{T}})$$ We denote this functor by ${\mathbf{Psm}}'$.
A standard argument shows that the category of Cartesian bisimplicial sheaves on a bisimplicial space is equivalent to the category of Cartesian simplicial sheaves on the diagonal. In the case of $\varpi^{\times j} \times G^{\times k} \times X$ we may describe this diagonal concretely:
1. The simplicial space $\varpi^{\times n} \times G^{\times n} \times X$ is naturally isomorphic to the bar construction $[X/(G \rtimes \varpi)]$, where the semidirect product $G \rtimes \varpi$ is constructed using the conjugation action of $\varpi$ on $G$, and $G \rtimes \varpi$ acts on $X$ via $((g,h),x) \mapsto ghx$.
2. The map $G \rtimes \varpi \to G \times \varpi:(g,h) \mapsto (gh,h)$ is a group isomorphism.
It follows from (2) that we have a simplicial map $f:[X/(G \rtimes \varpi)] \to [X/G]$. We define the Smith operator for equivariant sheaves as the composite
$$D^b([X/G];K) \stackrel{f^*}{\to} D^b([X/G \rtimes \varpi];K) \cong D^b_\varpi([X/G];K) \stackrel{{\mathbf{Psm}}'}{\longrightarrow} D^b([X^\varpi/Z_G(\varpi)];{\mathcal{T}})$$
Conjecture on perverse sheaves {#sec:conjecture}
------------------------------
By Proposition \[prop:periodic\], the category ${\mathrm{Perf}}(Y;{\mathcal{T}})$ can carry no t-structure. Nevertheless, I believe that the Smith operator $${\mathbf{Psm}}:D^b_\varpi(X;K) \to {\mathrm{Perf}}(X^\varpi;{\mathcal{T}})$$ interacts well with the perverse $t$-structure on $D^b_\varpi(X;K)$ when $X$ is a complex algebraic variety. Before stating the conjecture, note that there is a natural functor $$D^b(X^\varpi;K) \to {\mathrm{Perf}}(X^\varpi;{\mathcal{T}})$$ that carries a sheaf $F$ first to $F \otimes_K K[\varpi]$ in $D^b_\varpi(X^\varpi;{\mathcal{T}})$ and then to its image in ${\mathrm{Perf}}(X^\varpi;{\mathcal{T}})$. Let us denote this functor by $\otimes_K {\mathcal{T}}$, as suggested by Remark \[rem:tatespec\].
Let $X$ be a complex algebraic variety equipped with a $\varpi$-action, and let $P$ be a $\varpi$-equivariant perverse sheaf of $K$-vector spaces on $X$. Then there exist perverse sheaves of $K$-vector spaces $P_1, \ldots, P_n$ on $X^\varpi$ and integers $a_1,\ldots,a_n$ such that $${\mathbf{Psm}}(P) \cong (P_1[a_1] \oplus \cdots \oplus P_n[a_n]) \otimes_K {\mathcal{T}}$$ in ${\mathrm{Perf}}(X^\varpi;{\mathcal{T}})$
I furthermore believe that the $P_i$ are at least some of the time functorially associated to $P$, but I do not know how to formulate this precisely. Note that by Proposition \[prop:periodic\], the $a_i$ are irrelevant for $p = 2$ and only relevant mod $2$ for $p > 2$.
Let me discuss some evidence for this conjecture
1. Let $U$ be a smooth affine open subset of $X$ and let $i:U \to X$ denote the inclusion map—then $K_U[\dim(U)]$ is a perverse sheaf on $U$ and a theorem of Artin shows that $i_! K_U[\dim(U)]$ and $i_* K_U[\dim(U)]$ are perverse on $X$. If $U$ is stable for the $\varpi$-action then these are $\varpi$-equivariant, and the space of fixed points $U^\varpi$ is again a smooth affine open subset of $X^\varpi$, and $i_! K_{U^\varpi}$ and $i_*K_{U^\varpi}$ are also shifts of perverse sheaves. The conjecture then holds for this class of perverse sheaves by Theorem \[thm:Smithpush\].
2. A similar argument shows that the Conjecture holds for perverse sheaves of the form $\psi_f K_{f^{-1}(1)}$, where $f:Y \to {\mathbb{C}}$ is a family of varieties whose general fiber is smooth.
3. Microlocal considerations can be used to justify the Conjecture. Nadler and Zaslow [@NZ] construct a dictionary between Lagrangian submanifolds of $T^*M$ and constructible sheaves on $M$, lifting the less functorial, many-to-one dictionary between constructible sheaves and conic Lagrangian subsets considered by Kashiwara and Schapira [@KS]. When $M$ is a complex manifold it is natural to ask which Lagrangians in $T^*M$ correspond to perverse sheaves on $M$. The answer, up to shift, is those Lagrangians which are also complex submanifolds (possibly immersed) in the natural complex structure on $T^*M$—this is an unpublished result of Nadler’s.
Now one can reason as follows: we may replace the perverse sheaf $P$ by a complex Lagrangian $L \subset T^* M$. We may hope to express the $\varpi$-equivariance of $P$ by saying that $L$ is stable for the $\varpi$-action on $T^*M$, and that the Smith operator should carry $L$ to $L^\varpi$ (Section \[sec:singsupp\] gives some evidence for this idea). If $L$ is a complex submanifold of $T^* M$ then $L^\varpi$ will be a complex submanifold of $T^* (M^\varpi)$, which is consistent with the Conjecture.
4. We argued in Remark \[rem:hyperbolic\] that the Smith operator is analogous to hyperbolic localization for $T$-actions. It is proved in [@Braden] that hyperbolic localization interacts well with perverse sheaves.
### Example
Let $X$ be the one-point compactification of the total space of the line bundle ${\mathcal{O}}(n)$ on ${\mathbb{P}}^1$—denote the cone point by $\ast$. If $n$ is positive, this is an algebraic variety. A basic example of a perverse sheaf on $X$ is the intersection homology sheaf $IC_X = j_{!*}K[2]$ of Deligne-Goresky-MacPherson. At a smooth point of $X$, the stalk of $IC_X$ is $K[2]$, while at the point $\ast$ the stalk cohomology is “half the cohomology of the link,” i.e. it is $H^0(L;K)$ in degree $-2$ and $H^1(L;K)$ in degree $-1$ and zero in other degrees, where $L$ denotes the 3-dimensional link of $\ast$ in $X$.
In fact $L$ is the lens space $(S^3 - \{0,0\})/\mu_n$. If $p$ does not divide $n$ then $H^1(L;K) = 0$, so that $IC_X$ is just the shifted constant sheaf. Let us instead suppose that $p$ does divide $n$, so that $H^1(L,K) = K$.
Let ${\mathbb{Z}}/p$ act on ${\mathbb{C}}^2$ by $(x,y) \mapsto (x,\eta \cdot y)$, where $\eta$ is a $p$th root of unity. This induces an action of ${\mathbb{Z}}/p$ on $X = \ast \cup ({\mathbb{C}}^2 - \{0\})/\mu_n \cup {\mathbb{P}}^1$ whose fixed point set is the one-point compactification of the two lines ${\mathcal{O}}(n)_0 \coprod {\mathcal{O}}(n)_\infty$—topologically, this is a wedge of two ${\mathbb{P}}^1$s, call them $Y_1$ and $Y_2$.
On each $Y_i$, there is up to isomorphism a unique indecomposable perverse sheaf $P_i$ that is constant on $Y_i - \ast$, that is isomorphic to its Verdier dual, and that admits a surjection (in the perverse $t$-structure) onto the skyscraper sheaf at $\ast$. (It can be described as the projective cover as well as the injective hull of this skyscraper sheaf, and also as the tilting extension of the constant perverse sheaf on $Y_i - \ast$).
Since $IC_X$ is the constant sheaf $K[2]$ along $Y_i - \ast$ for $i=1,2$, we have a map $$IC_X \vert_{Y_1 \cup Y_2} \to f_{1*} K[2] \oplus f_{2*} K[2]$$ where $f_i$ denotes the inclusion of $Y_i - \ast$ into $Y_i$. The cone on this map is a skyscraper sheaf supported on $\ast$ placed in degree $-1$, denote it by $\delta[1]$.
Note that $f_1$ and $f_2$ are affine, so that $f_{i*} K[1]$ is perverse. Thus ${\mathbf{Psm}}(IC_X)$ is isomorphic to $P[1] \otimes_K {\mathcal{T}}$, where $P$ is the kernel of the surjective map of perverse sheaves $f_{1*} K[1] \oplus f_{2*} K[1] \to \delta[1]$
*Acknowledgments:* I thank Florian Herzig, Gopal Prasad, and Ting Xue for help with algebraic groups. In particular, most of the material in Section \[sec:exc\] I learned from Ting. While developing these ideas I benefited from discussions with Paul Goerss, David Nadler, and Zhiwei Yun.
[99]{} M. Atiyah and R. Bott, “The moment map and equivariant cohomology,” Topology [**23**]{} (1984) 1–28 J. Bernstein and V. Lunts, “Equivariant sheaves and functors,” Springer (1994). A. Borel, “Seminar on transformation groups” Ann. of Math. studies [**46**]{}, (1960). A. Borel and J. de Siebenthal, “Les sous-groupes fermŽs de rang maximum des groupes de Lie clos.” Comment. Math. Helv. [**23**]{} (1949) 200–221. T. Braden, “Hyperbolic localization of intersection cohomology” Transform. Groups [**8**]{} (2003) 209Ð216 G. Bredon, “Fixed point sets of actions on Poincaré duality spaces,” Topology [**12**]{} (1973) 159–175.
R. Bott, “The space of loops on a Lie group,” Michigan Math. J. [**5**]{} (1958) 35–61. T. Chang, T. Skjelbred, “Group actions on Poincaré duality spaces,” Bull. Amer. Math. Soc. [**78**]{} (1972) 1024–1026.
M. Goresky, R. Kottwitz, and R. MacPherson, “Equivariant cohomology, Koszul duality, and the localization theorem,” Inv. Math. [**131**]{} (1998), 25–83 V. Kac, “Automorphisms of finite order of semisimple Lie algebras,” Funkts. Anal. Prilozh., [**3**]{} (1969) 252–254.
M. Kashiwara and P. Schapira, “Sheaves on manifolds.” Springer (1990) M. Liebeck and G. Seitz, “The maximal subgroups of positive dimension in exceptional algebraic groups” Mem. Am. Math. Soc. [**169**]{} (2004)
I. Mirkovi[ć]{} and K. Vilonen, “Geometric Langlands duality and representations of algebraic groups over commutative rings,” Ann. of Math. [**166**]{} (2007), 95–143. D. Nadler, “Perverse sheaves on real loop Grassmannians,” Invent. Math. [**159**]{} (2005) 1–73 D. Nadler, “Microlocal branes are constructible sheaves,” Selecta Math. [**15**]{} (2009), no. 4, 563–619 D. Nadler and E. Zaslow, “Constructible sheaves and the Fukaya category,” J. Amer. Math. Soc. [**22**]{} (2009), 233–286. D. Quillen, “The spectrum of an equivariant cohomology ring” Ann. of Math. [**94**]{} (1971) P. Schapira, “Operations on constructible functions” J. pure and applied algebra (1991) K. Shinoda, “The conjugacy classes of type $({\mathrm{F}}_4)$ over finite fields of characteristic 2.” J. Fac. Sci. Univ. Tokyo, [**21**]{} (1974)
P. Smith, “A theorem on fixed points for periodic transformations,” Ann. of Math. (1934) “Nilpotent classes in Lie algebras of type ${\mathrm{F}}_4$ over fields of characteristic 2.” J. Fac. Sci. Univ. Tokyo, [**30**]{} (1984), 517–524.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the preservation of the properties of automaticity and prefix-automaticity in Rees matrix semigroups over semigroupoids and small categories. Some of our results are new or improve upon existing results in the single-object case of Rees matrix semigroups over semigroups.'
title: |
Automatic [R]{}ees matrix semigroups\
over categories
---
Mark Kambites\
Fachbereich Mathematik / Informatik, Universität Kassel\
34109 Kassel, Germany\
`[email protected]`\
Introduction {#sec:intro}
============
In recent years, one of the most productive areas of combinatorial group theory has been the theory of *automatic groups*. The expression of a finitely generated group using an *automatic structure* provides a basis for efficiently performing computations in the group which may be hard or impossible given only a presentation. Groups which admit automatic structures also display a number of remarkable algebraic and geometric properties. [@Epstein92]
The usual language-theoretic definition of automaticity in groups lends itself naturally to application in wider contexts. It is observed in [@Epstein92] that the notion of automaticity extend naturally from groups to *groupoids*. More recently, Hudson [@Hudson96] has introduced a definition of automaticity in semigroups and monoids and a theory of automatic semigroups has begun to emerge [@Campbell99; @Campbell01; @Campbell00a; @Descalco01; @Duncan99; @Hoffmann00a; @Hoffmann01; @Hoffmann02; @Hoffmann00b; @Hoffmann01a]. In [@KambitesAutoCat] we developed a common framework for these generalisations, in the form of a theory of automaticity for *small categories* and *semigroupoids*. We explored what could be learnt about automatic small categories and semigroupoids, by applying the existing theory of automatic semigroups. Semigroupoids and small categories play a vital role in the structural theory of semigroups, so it also seems natural to ask, conversely, whether automatic categories can tell us anything about automatic semigroups.
A key recurring theme in structural semigroup theory is that of the *Rees matrix construction* (see, for example, [@Howie95]). A number of interesting results have been proved concerning the relationship between automaticity properties and Rees matrix constructions over groups [@Campbell00a] and more generally over semigroups [@Descalco01; @Silva00b]. At the same time, Lawson [@Lawson00] has applied to the study of abundant semigroups a form of Rees matrix construction over semigroupoids. In [@Kambites05] we explored combinatorial aspects of this construction showing that, under certain assumptions, combinatorial properties such as finite generation and finite presentability are preserved. In this paper, we consider the relationship between automaticity and these constructions. Some of our results are new, or improve upon existing results, even when specialised to the single-object case of Rees matrix semigroups over semigroups.
In addition to this introduction, this paper comprises four sections. In Section \[sec:background\], we briefly recall the notions of generators and relations for partial algebras which we studied in [@Kambites05], and the definitions of regular path languages and of automatic and prefix-automaticity introduced in [@KambitesAutoCat]. We also recall some key results from [@KambitesAutoCat] which will be applied in this paper. In Section \[sec:technical\], we prove some technical results concerning automaticity in small categories and semigroupoids, which will be needed in the sections that follow.
Section \[sec:autorees\] considers the relationship between automaticity and Rees matrix constructions with zero over semigroupoids. We show that a finitely generated Rees matrix semigroup over an automatic semigroupoid is always automatic, and provide some sufficient conditions for the underlying semigroupoid of an automatic Rees matrix semigroup to be automatic. In Section \[sec:pautorees\] we prove some related results for prefix-automaticity. Section \[sec:reeswithoutzero\] extends these results to the case of Rees matrix constructions without zero. Finally, Section \[sec:remarks\] contains some remarks and open questions.
Background {#sec:background}
==========
In this section, we briefly recall a number of definitions and results from [@Kambites05] and [@KambitesAutoCat]. For a more detailed introduction, the reader should consult those papers.
By a *(directed) graph* $X$ we mean a collection $X^0$ of *vertices* together with a collection $X^1$ of *edges* and two functions $\alpha, \omega : X^1 \to X^0$ which determine respectively the *source* and *target* of each edge.
A *path* $\pi$ in $X$ is a finite sequence $e_1 e_2 \dots e_n$ of (not necessarily distinct) edges in $X$ such that $e_i \omega = e_{i+1} \alpha$ for $1 \leq i < n$. We define $\pi \alpha = e_1 \alpha$, $\pi \omega = e_n \omega$, and call these the *source* and *target* respectively of the path $\pi$. The *length* of the path $\pi$ is the number $n$ of edges; we denote it by $|\pi|$. There is also a distinct *empty path* of length $0$ at each vertex, with source and target that vertex. We identify each path of length $1$ with its single edge, and each vertex with the empty path at that vertex. Thus, $X^0$ and $X^1$ are the sets of paths in $X$ of length $0$ and of length $1$ respectively. We extend this notation by writing $X^n$ for the set of paths of length $n$ in $X$. We also write $X^{\geq n}$ for the set of paths in $X^*$ of length $n$ or more; $X^{> n}$, $X^{\leq n}$ and $X^{< n}$ are defined analogously in the obvious way, including empty paths where appropriate.
Let $X$ and $Y$ be graphs. The *direct product* $X \times Y$ of $X$ and $Y$ is the (finite) graph with vertex set $X^0 \times Y^0$, and edge set $X^1 \times Y^1$ $(f,g) \alpha = (f \alpha, g \alpha)$ and $(f,g) \omega = (f \omega, g \omega)$ for all $f \in X^1, g \in Y^1$. A *graph morphism* $\sigma : X \to Y$ consists of functions $\sigma^0 : X^0 \to Y^0$ and $\sigma^1 : X^1 \to Y^1$ such that $x \sigma^1 \alpha = x \omega \alpha^0$ and $x \sigma^1 \omega = x \omega \sigma^0$ for all edges $x \in X^1$. If, in addition, $X^0 = Y^0$ and $\sigma^0 : X^0 \to X^0$ is the identity function, then we call $\sigma$ an *edge-morphism*, and, for notational convenience, identify $\sigma$ with $\sigma^1$.
A *semigroupoid* $S$ is a small graph $S$, together with a partial multiplication on the edges of $S$ such that, for any two edges $e, f \in S^1$
- the product $ef$ is defined if and only if $e \omega = f \alpha$;
- if $ef$ is defined, then $(ef)\alpha = e \alpha$ and $(ef)\omega = f \omega$; and
- if $ef$ and $fg$ are defined then $e(fg) = (ef)g$.
The vertices and edges of a semigroupoid are called respectively *objects* and *arrows*. Where no ambiguity arises, we abuse the notation slightly by writing $S$ to mean the set $S^1$ of arrows in $S$.
If $s_0 \in S^0$ is such that there exists an $s \in S^1$ with $s \alpha = s \omega = s_0$, we say that the *local semigroup* of $S$ at the object $s_0$ is the semigroup with elements $\lbrace s \in S \mid s \alpha = s \omega = s_0 \rbrace$ and multiplication defined by restricting that in $S$. A semigroupoid is *strongly connected* if for every pair $s_0, t_0 \in S^0$ of objects, there is an arrow (or equivalently, a non-empty path) in $S$ with source $s_0$ and target $t_0$. An object $x \in S^0$ of a semigroupoid $S$ is *isolated* if is not the source or target of any arrow; if a semigroupoid has no isolated objects then it is *isolation-free*.
The *free semigroupoid $X^+$* on a small graph $X$ is the semigroupoid whose objects are the vertices of $X$, and whose arrows are the *non-empty* paths in $X$, with $\alpha$, $\omega$ and the partial multiplication defined in the obvious way. The *free category $X^*$* is the category obtained by adjoining a local identity (which can be thought of as the empty path) at each object of $X^+$. A *path language* or just a *language* over $X$ is a (possibly empty) collection of (possibly empty) paths in $X$, that is, a subset of the free category $X^*$.
Let $x = x_1 \dots x_n$ be a path in a free semigroupoid $X^+$ where each $x_i \in X^1$. A *prefix* of $x$ is a (non-empty) path of the form $x_1 \dots x_j$ for some $1 \leq j \leq n$. A *suffix* of $x$ is a (non-empty) path of the form $x_j \dots x_n$ for some $1 \leq j \leq n$. A *factor* of $x$ is a (non-empty) path of the form $x_j \dots x_k$ for some $1 \leq j \leq k \leq n$. An *internal factor* of $x$ is a (non-empty) path of the form $x_j
\dots x_k$ for some $1 < j \leq k < n$, that is, a factor of $x_2 \dots
x_{n-1}$.
Given a path language $L \subseteq X^+$, we write ${\operatorname{Pref}}(L)$ \[respectively ${\operatorname{Suff}}(L)$, ${\operatorname{Fact}}(L)$, ${\operatorname{Int}}(L)$\] for the set of **non-empty** prefixes \[respectively suffixes, factors, internal factors\] of paths in $L$. For $n \in {\mathbb{N}}$ we write ${\operatorname{Pref}}_n(L)$ to denote ${\operatorname{Pref}}(L) \cap X^n$, and similarly for ${\operatorname{Suff}}_n(L)$, ${\operatorname{Fact}}_n(L)$ and ${\operatorname{Int}}_n(L)$; note that all four are empty when $n = 0$.
A *(path) automaton* $M$ over a small graph $X$ is a small graph $M$, together with a graph morphism $\sigma : M \to X$, a set of distinguished *start vertices* of $M$ and a set of distinguished *terminal vertices* of $M$. We think of the automaton as the graph $M$ with each edge $e \in M^1$ labelled by $e \sigma^1 \in X^1$, and each vertex $v \in M^0$ labelled by $v \sigma^0 \in X^0$. The vertices and edges are called *states* and *transitions* respectively. The unique *language accepted* or *language recognised* by $M$ is the set of paths in $X$ which label paths from a start vertex to a terminal vertex in $M$.
The automaton is called a *complete, deterministic* automaton if (i) there is exactly one start vertex in the pre-image of each object in $X$ and (ii) for every state $m \in M^0$ and every edge $e \in X^1$ with $e \alpha = m \sigma^0$ there is a unique edge $f \in M^1$ with $f \sigma^1 = e$ and $f \alpha = m$. The automaton is called *finite* if the graphs $M$ and $X$ are finite; we shall be concerned exclusively with finite automata.
A path language which is accepted by some finite path automaton is called *regular*. We recall from [@KambitesAutoCat Section 3] that a language $L \subseteq X^+ \subseteq (X^1)^+$ is regular in this sense if and only if it is regular in the usual sense as a language over the alphabet $X^1$. We recall also that the set of regular path languages over $X$ contains $X^+$, $X^*$ and all finite path languages, and is closed under concatenation, finite intersection, finite union, complement, set difference, generation of subcategories and subsemigroupoids, and prefix-closure. We will use these properties without further comment.
Given a graph $X$, we denote by $X^\$$ the graph $X$ with an extra edge $\$_s$ adjoined for every vertex $s \in X^0$, with source and target $s$. We define a function $\delta_X : (X^+ \times X^+) \to (X^\$ \times X^\$)^+$ by $$(a, b) \delta_X =
\begin{cases}
(a_1, b_1) \dots (a_n, b_n) & \text{ if } m = n \\
(a_1, b_1) \dots (a_n, b_n) (a_{n+1}, \$_{b_n \omega}) \
\dots (a_m, \$_{b_n \omega}) & \text{ if } m > n \\
(a_1, b_1) \dots (a_m, b_m) (\$_{a_m \omega}, b_{n+1}) \
\dots (\$_{a_m \omega}, b_n) & \text{ if } n > m \\
\end{cases}$$ where $a = a_1 \dots a_m$ and $b = b_1 \dots b_n$. We observe that the function $\delta_X$ is injective, and in particular that it distributes over intersection, that is, that $$(R_1 \cap R_2) \delta_X = R_1 \delta_X \cap R_2 \delta_X$$ for all binary relations $R_1, R_2 \subseteq X^+ \times X^+$.
A *(binary) synchronous path automaton* over $X$ is a finite path automaton over the graph $X^\$ \times X^\$ $, with the property that the language accepted is contained within the image $(X^+ \times X^+) \delta_X$ of $\delta_X$. For convenience, we shall say that a synchronous path automaton accepts a pair $(a,b) \in X^+ \times X^+$ if it accepts $(a,b) \delta_X$. A binary relation $R \subseteq X^+ \times X^+$ is called *synchronously regular* if there exists a synchronous path automaton accepting exactly the language $R \delta_X$, that is, if $R \delta_X$ is regular.
[@KambitesAutoCat Lemma 3.7]\[lem:syncrat\_prop\] Let $X$ be a finite graph. Then:
- If $R \subseteq X^+ \times X^+$ is synchronously regular, then the binary relation $$R^{-1} = \lbrace (v,u) \mid (u, v) \in R \rbrace$$ is synchronously regular.
- If $R \subseteq X^+ \times X^+$ is synchronously regular, then the projections $R \pi_1$ and $R \pi_2$ of $R$ onto its first and second coordinates are regular.
- If $K, L \subseteq X^+$ are regular languages of non-empty paths, then the binary relation $K \times L$ is synchronously regular.
- Synchronously regular binary relations over $X$ are closed under finite union, finite intersection and relational composition.
- If $K \subseteq X^+$ is regular then the *diagonal binary relation* $$\lbrace (w, w) \mid w \in K \rbrace$$ is synchronously regular.
Let $X$ and $Y$ be finite graphs, $A \subseteq X^*$ and $\phi : A \to Y^*$ be a function. We say that $\phi$ is *strongly regularity preserving* if for every regular language $L \subseteq X^+$, we have that $(L \cap A) \phi \subseteq Y^+$ is a regular language.
Let $X$ and $Y$ be finite graphs, and for $i = 1,2$ suppose we have subsets $A_i \subseteq X^+$ and functions $\phi_i : A_i \to Y^+$. We say that $\phi_1$ and $\phi_2$ are *strongly mutually synchronous regularity preserving* if for every synchronously regular relation $R \subseteq X^+ \times X^+$, the relation $$\lbrace (u \phi_1, v \phi_2) \mid (u, v) \in R \cap (A_1 \times A_2) \
\rbrace \subseteq Y^+ \times Y^+$$ is synchronously regular.
A *sliding window inverse* for a function $\phi : A \to Y^+$ is a quadruple $(n, f, g, h)$ consisting of a positive integer $n$ and three functions $$f : {\operatorname{Pref}}_n(A \phi \cap Y^{> n}) \to X^*,$$ $$g : {\operatorname{Int}}_n(A \phi) \to X^* \text{ and}$$ $$h : {\operatorname{Suff}}_n(A \phi \cap Y^{> n}) \to X^*,$$ with the property that for any $w \in A$ and $y_1 \dots y_k \in Y^+$ with $y_1, \dots, y_k \in Y^1$ and $k > n$ such that $w \phi = y_1 \dots y_k$, we have $$\begin{aligned}
w &= (y_1 \dots y_n) f \ (y_2 \dots y_{n+1}) g \ (y_3 \dots y_{n+2}) g \
\dots \\
& \hspace{18ex} \dots \ (y_{k-n} \dots y_{k-1}) g \ (y_{k-n+1} \dots y_k) h.\end{aligned}$$ If $f$, $g$ and $h$ are functions with domains containing those given above, we shall say that $(n, f, g, h)$ is a sliding window inverse for $\phi$ if the restrictions of $f$, $g$ and $h$ to the appropriate domains have the given properties.
[[@KambitesAutoCat Lemma 3.8]]{} \[lem:regpreserve\] Let $X$ and $Y$ be finite graphs, $A \subseteq X^*$ and $\phi : A \to Y^*$ be a function. If $A \phi \subseteq Y^*$ is regular, and $\phi$ has a sliding window inverse, then $\phi$ is strongly regularity preserving.
Let $\psi : {\mathbb{N}}\to {\mathbb{N}}$ be a function. We say that sliding window inverses $(n_1, f_1, g_1, h_1)$ for $\phi_1$ and $(n_2, f_2, g_2, h_2)$ for $\phi_2$ are *synchronised by $\psi$* if
- $n_1 = n_2$;
- for every $i = 1, 2$ and every $w \in {\operatorname{Pref}}_{n_1}(A_i \phi_i \cap Y^{> n_1})$, we have $|w f_i| = 0 \psi$;
- for every $i=1, 2$ and every $xyz \in A_i \phi_i$ with $x \in Y^m$, $m \geq 1$, $y \in Y^{n_1}$ and $z \in Y^+$ we have $|y g_i| = m \psi$; and
- for every $i = 1, 2$ and every $xy \in A_i \phi_i$ with $x \in Y^m$, $m \geq 1$ and $y \in Y^{n_1}$ we have $|y h_i| \leq m \psi$.
We say that two sliding window inverses are *synchronised* if they are synchronised by some function $\psi : {\mathbb{N}}\to {\mathbb{N}}$.
[[@KambitesAutoCat Lemma 3.9]]{} \[lem:syncratpreserve\] Let $X$ and $Y$ be finite graphs, $A_1, A_2 \subseteq X^+$ and $\phi_1 : A_1 \to Y^+$ and $\phi_2 : A_2 \to Y^+$ be injective functions. If $A_1 \phi_1$ and $A_2 \phi_2$ are regular and $\phi_1$ and $\phi_2$ have synchronised sliding window inverses, then $\phi_1$ and $\phi_2$ are strongly mutually synchronous regularity preserving.
A *choice of representatives* for a semigroupoid $S$ is a triple $(X, K, \rho)$, of a graph $X$ with $X^0 = S^0$, a (surjective) semigroupoid morphism $\rho : X^+ \to S$ of the free semigroupoid $X^+$ onto $S$, and a language $K \subseteq X^+$ such that $K \rho = S$. The choice of representatives is called *finitely generated* if $X$ has finitely many edges. Clearly, a semigroupoid has a finitely generated choice of representatives if and only if it is finitely generated. The choice of representatives is called a *cross-section* if the restriction of $\rho$ to $K$ is bijective, that is, if $K$ contains a unique representative for every arrow in $S$. The choice of representatives is called *prefix-closed* if $K$ is closed under the taking of non-empty prefixes.
An *automatic structure* for a semigroupoid $S$ is a finitely generated choice of representatives $(X, K, \rho)$ with the property that for every edge or empty path $a \in X^0 \cup X^1$, the binary relation $$K_a = \lbrace (u, v) \in K \times K \mid u \omega = a
\alpha, (ua) \rho = v \rho \rbrace$$ is synchronously regular. Equivalently [@KambitesAutoCat Proposition 4.1], $(X, K, \rho)$ is an automatic structure exactly if $K_a$ is synchronously regular for every $a \in X^1$, and the union $$K_= = \lbrace (u,v) \in K \times K \mid u \rho = v \rho \rbrace = \bigcup_{a \in X^0} K_a$$ is synchronously regular.
A *prefix-automatic structure* for a semigroupoid $S$ is an automatic structure $(X, K, \rho)$ with the additional property that the binary relation $$K_=' = \lbrace (u, v) \in K \times {\operatorname{Pref}}(K) \mid u \rho = v \rho \rbrace$$ is synchronously regular. A semigroupoid admits a prefix-automatic structure if and only if it admits an automatic structure which is prefix-closed [@KambitesAutoCat Corollary 4.6].
Now let $S$ be a semigroupoid and $0$ be a new symbol not in $S^1$. The *consolidation* of $S$ is the semigroup with set of elements $T = S^1 \cup \lbrace 0 \rbrace$, and multiplication defined by $$s t = \begin{cases}
\text{the $S$-product } st & \text{ if } \
s, t \in S \text{ and } s \omega = t \alpha \\
0 & \text{ otherwise} \\
\end{cases}$$ for all $s, t \in S$. We shall need the following key results from [@KambitesAutoCat].
[[@KambitesAutoCat Theorem 4.3]]{}\[thm:cons\_iff\_sgpoid\] Let $T$ be the consolidation of a semigroupoid $S$. Then $S$ is automatic \[prefix-automatic\] if and only if $T$ is automatic \[respectively, prefix-automatic\].
[[@KambitesAutoCat Theorem 5.6]]{}\[thm:auto\_cofinite\_subsgpoid\] Let $S$ be a semigroupoid and $U$ a cofinite subsemigroupoid of $S$. Then $S$ is automatic \[prefix-automatic\] if and only if $U$ is automatic \[respectively, prefix-automatic\].
Some Technical Results {#sec:technical}
======================
We shall need the following technical results in the remaining sections.
\[prop:syncrat\_changelastletter\] Let $X$ be a finite graph and $R \subseteq X^+ \times X^+$ be synchronously regular. Let $m \geq 0$ and let $f : X^m \to X^*$ be such that $x f \alpha = x \alpha$ for all $x \in X^m$. Then the binary relation $$R' = \lbrace (u, v (xf)) \mid (u, vx) \in R, u, v \in X^*, \
x \in X^m \rbrace$$ is synchronously regular.
First, notice that $R'$ is the composition of $R$ with the relation $$T = \lbrace \left( vx, v(xf) \right) \mid v \in X^* \rbrace.$$ By Lemma \[lem:syncrat\_prop\], it will suffice to show that $T$ is synchronously rational. But clearly $T \delta_X$ is the concatenation of $D \delta_X$ with $E \delta_X$ where $D$ is the diagonal relation on $X$ and $E$ is the finite relation $\lbrace (x, xf) \mid x \in X^m \rbrace$. Both of these are regular, and regular path languages are closed under concatenation, so it follows that $T \delta_X$ is regular, and so $T$ is synchronously regular as required.
\[prop:no\_duplicate\_letters\] Let $S$ be a semigroupoid with an automatic cross-section \[prefix-closed automatic structure, prefix-automatic cross-section\], and let $T$ be a finite subset of $S$. Then $S$ has an automatic cross-section \[prefix-closed automatic structure, prefix-automatic cross-section\] $(Y, L, \sigma)$ such that the restriction of $\sigma$ to $Y$ is injective and has image containing $T$.
Let $(X, K, \rho)$ be an automatic cross-section \[prefix-closed automatic structure, prefix-automatic cross-section\] for $S$. Choose a subgraph $Y$ of $X$ such that $Y^1 \rho = X^1 \rho$ and the restriction of $\rho$ to $Y^1$ is injective. For each element $t \in T$ which does not already have a representative in $X$, adjoin a new edge $y_t$ to $Y$ with $y_t \alpha = t \alpha$ and $y_t \omega = t \omega$, to obtain a new finite graph $Z$.
For each $x \in X^1$, let $x \sigma$ be the unique element $y \in Z$ such that $y \rho = x \rho$. Extend $\sigma$ to a semigroupoid morphism $\sigma' : X^+ \to Z^+$. Let $\rho' : Z^+ \to S$ be defined by $$x \rho' = \begin{cases} x \rho &\text{ if } x \in X \\
t &\text{ if } x = y_t.
\end{cases}$$ It is a routine exercise to verify that $(Z, K \sigma', \rho')$ is an automatic cross-section \[prefix-closed automatic structure, prefix-automatic cross-section\] for $S$ and that the restriction of $\rho'$ to $Z$ is injective and has image containing $T$.
\[lem:limitedwork1\] Let $(X, K, \sigma)$ be a regular choice of representatives for a semigroupoid $S$, and let $L$ be a cofinite subset of $K$. If $K_= \cap (L \times L)$ is synchronously regular and $K_c \cap (L \times K)$ is synchronously regular for every $c \in X^1$, then $(X, K, \sigma)$ is an automatic structure for $S$.
First, we claim that for any $w \in K$, the language $$w \sigma \sigma^{-1} \cap K = \lbrace x \in K \mid x \sigma = w \sigma \rbrace$$ is regular. If $L$ contains no representatives for $w \sigma$, then $w \sigma \sigma^{-1} \cap K \subseteq K \setminus L$ must be finite and hence regular. Otherwise, let $u$ be a representative in $L$ for $w \sigma$. Now we have $$w \sigma \sigma^{-1} \cap K = (w \sigma \sigma^{-1} \cap (K \setminus L)) \cup ((K_= \cap (L \times L)) \cap (\lbrace u \rbrace \times L)) \pi_2$$ where $w \sigma \sigma^{-1} \cap K \setminus L \subseteq K \setminus L$ is finite, $K_= \cap (L \times L)$ is synchronously regular by assumption, and $L$ is regular. It follows by Lemma \[lem:syncrat\_prop\] that $w \sigma \sigma^{-1} \cap K$ is regular. Now we see that $$\begin{aligned}
K_= = (K_= \cap (L \times L)) \cup \
&\left( \bigcup_{w \in K \setminus L} \lbrace w \rbrace \times (w \sigma \sigma^{-1} \cap K) \right) \cup \\
&\left( \bigcup_{w \in K \setminus L} (w \sigma \sigma^{-1} \cap K) \times \lbrace w \rbrace \right)\end{aligned}$$ is synchronously regular by Lemma \[lem:syncrat\_prop\](iii) and (iv). Similarly, for any $c \in C^1$ we see that $$K_c = (K_c \cap (L \times K)) \cup \
\bigcup_{w \in K \setminus L} \lbrace w \rbrace \times ((wc) \sigma \sigma^{-1} \cap K)$$ is synchronously regular by Lemma \[lem:syncrat\_prop\](iii) and (iv). Thus, $(X, K, \sigma)$ is an automatic structure for $S$, as required.
\[lem:limitedwork2\] Let $(X, L, \sigma)$ be an automatic structure for a semigroupoid $S$, and suppose $L \subseteq K \subseteq X^+$ with $K \setminus L$ finite. Then $(X, K, \sigma)$ is an automatic structure for $S$.
First, note that $K_= \cap (L \times L) = L_=$ is synchronously regular and $K_c \cap (L \times L) = L_c$ is synchronously regular for every $c \in X^1$. We can deduce as in the proof of Lemma \[lem:limitedwork1\] that $w \sigma \sigma^{-1} \cap K$ is regular for every $w \in K$, and that $K_=$ is synchronously regular.
Now let $c \in X^1$. For any $w \in K$, choose $u \in L$ such that $w \sigma = u \sigma$, and consider the set $$\begin{aligned}
&\lbrace x \in K \mid x \omega = c \alpha, (xc) \sigma = w \sigma \rbrace \\
= &\lbrace x \in K \setminus L \mid x \omega = c \alpha, (xc) \sigma = w \sigma \rbrace
\cup \lbrace x \in L \mid x \omega = c \alpha, (x c) \sigma = w \sigma \rbrace \\
= &\lbrace x \in K \setminus L \mid x \omega = c \alpha, (x c) \sigma = w \sigma \rbrace
\cup ((K_c \cap (L \times L)) \cap (L \times \lbrace u \rbrace)) \pi_1.\end{aligned}$$ Certainly the set $\lbrace x \in K \setminus L \mid x \omega = c \alpha, (x c) \sigma = w \sigma \rbrace$ is finite, and we know that $K_c \cap (L \times L)$ is synchronously regular and $L$ is regular. It follows by Lemma \[lem:syncrat\_prop\](ii) (iii) and (iv) that $\lbrace x \in K \mid x \omega = c \alpha, (x c) \sigma = w \sigma \rbrace$ is regular.
It follows that $$\begin{aligned}
K_c = (K_c \cap (L \times L)) \ \cup \
&\bigcup_{w \in K \setminus L} \lbrace w \rbrace \times ((wc) \sigma \sigma^{-1} \cap K) \ \cup \\
&\bigcup_{w \in K \setminus L} \lbrace x \in K \mid x \omega = c \alpha, (x c) \sigma = w \sigma \rbrace \times \lbrace w \rbrace\end{aligned}$$ is synchronously regular by Lemma \[lem:syncrat\_prop\](iii) and (iv). Thus, $(X, K, \sigma)$ is an automatic structure for $S$, as required.
Rees Matrices and Automaticity {#sec:autorees}
==============================
Rees matrix constructions over semigroupoids were introduced by Lawson [@Lawson00], who used them to construct a class of locally adequate abundant semigroups. They represent an alternative formulation of certain special cases of the *blocked Rees matrix semigroup* constructions introduced by Fountain [@Fountain82] and subsequently employed by Armstrong [@Armstrong86]. In this section, we extend some results of [@Descalco01] concerning Rees matrix constructions over semigroups to cover similar cosntructions over semigroupoids and small categories. In some cases, our new results are stronger than previously known results, even when specialised to the case of semigroups.
Let $S$ be a non-empty, isolation-free semigroupoid, and $I$ and $\Lambda$ be indexing sets. Let $F : I \to S^1 \alpha \subseteq S^0$ and $G : \Lambda \to S^1 \omega \subseteq S^0$ be surjective functions. Let $0$ be a new symbol not in $S$, and let $P$ be a $\Lambda \times I$ matrix with entries drawn from $S \cup \lbrace 0 \rbrace$, with the property that $P_{\lambda i} \omega = iF$ and $P_{\lambda i} \alpha = \lambda G$ for all $i \in I, \lambda \in \Lambda$ such that $P_{\lambda i} \neq 0$. The *Rees matrix semigroup with zero $M = M^0(S; F(I), G(\Lambda); P)$* is the semigroup with set of elements $$M = \lbrace 0 \rbrace \cup \lbrace (i, x, \lambda) \in I \times S \times \Lambda \mid i F = x \alpha, x \omega = \lambda G \rbrace$$ and multiplication given by $$(i, x, \lambda)(j, y, \mu) = \begin{cases}
(i, x P_{\lambda j} y, \mu) & \text{ if } P_{\lambda j} \neq 0 \\
0 & \text{ otherwise}
\end{cases}$$ for all $(i, x, \lambda), (j, y, \mu) \in M \setminus \lbrace 0 \rbrace$, and $0 m = m 0 = 0$ for all $m \in M$.
Note that in the expression “$M^0(S; F(I), G(\Lambda); P)$”, the use of the notation “$F(I)$” and “$G(\Lambda)$” is purely symbolic. It is intended to remind the reader of the relationship between $F$ and $I$ and between $G$ and $\Lambda$.
We call $P$ the *sandwich matrix* of the construction. If $P$ contains no zero entries, then $M \setminus \lbrace 0 \rbrace$ is a subsemigroup of $M$, which we call the *Rees matrix semigroup (without zero) $M(S; F(I), G(\Lambda); P)$*.
We remark briefly upon the choice of codomains for the indexing functions $F$ and $G$. This restriction does not limit the range of semigroups which appear as Rees matrix semigroups. Indeed, if $I$, $\Lambda$, $F$, $G$, $S$ and $P$, do not satisfy this requirement but otherwise satisfy the requirements for the Rees matrix construction, then one can instead perform the construction with subsets of $I$ and $\Lambda$, a subsemigroupoid of $S$ and the corresponding restrictions of $F$ and $G$ and submatrix of $P$, to obtain the same Rees matrix semigroup as would be obtained by using the construction on the more general semigroupoid. Similarly, given a semigroupoid $S$ with isolated objects, one can remove those objects to obtain an isolation-free semigroupoid before using a Rees matrix construction. The purpose of the restrictions is to allow the following straightforward proposition.
\[prop:all\_sgpoid\_used\] Let $$M = M^0(S; F(I), G(\Lambda); P)$$ or $$M = M(S; F(I), G(\Lambda); P)$$ be a Rees matrix semigroup with or without zero over a semigroupoid $S$. Then
- for every $i \in I$ there exist $s \in S^1$ and $\lambda \in \Lambda$ with $(i, s, \lambda) \in M$;
- for every $s \in S^1$ there exist $i \in I$ and $\lambda \in \Lambda$ with $(i, s, \lambda) \in M$;
- for every $\lambda \in \Lambda$ there exist $i \in I$ and $s \in S^1$ with $(i, s, \lambda) \in M$; and
- for every $v \in S^0$ there exists $i \in I$ with $i F = v$ or $\lambda \in \Lambda$ with $\lambda G = v$.
The following theorem is an amalgamation of results which can be found in [@Kambites03]; the same results appear, in slightly less generality, in [@Kambites05]. For a precise definition of what it means for a semigroupoid to be finitely presentable, the reader is directed to [@Kambites05].
\[thm:fgp\_rees\] Let $$M = M(S; F(I), G(\Lambda); P)$$ or $$M = M^0(S; F(I), G(\Lambda); P)$$ be a Rees matrix semigroup (with or without zero) over an isolation-free semigroupoid $S$. Then $M$ is finitely generated \[finitely presentable\] if and only if
- the indexing sets $I$ and $\Lambda$ are finite;
- the semigroupoid $S$ is finitely generated \[respectively, finitely presentable\]; and
- $S P' S$ is a cofinite subsemigroupoid of $S$, where $P'$ is the set of non-zero entries in the sandwich matrix $P$.
In order for a semigroup or semigroupoid to be automatic it is, of course, necessary that it be finitely generated. Theorem \[thm:fgp\_rees\] gives necessary and sufficient conditions for a Rees matrix semigroup over a finitely generated semigroupoid to be finitely generated. We begin by showing that the same conditions suffice to ensure that a Rees matrix semigroup over an automatic small category is automatic. We shall then extend this result to cover Rees matrix semigroups over general automatic semigroupoids.
\[thm:auto\_cat\_imp\_rees\] Let $M = M^0(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup with zero over a small category $S$. Suppose $S$ is automatic (as a semigroupoid) and $M$ is finitely generated. Then $M$ is automatic.
Let $P'$ be the set of non-zero entries in the sandwich matrix $P$, and let $U = S P' S$. Since $M$ is finitely generated and $S$ is a small category, we deduce, by Theorem \[thm:fgp\_rees\], that $I$ and $\Lambda$ are finite and that $U$ is a cofinite subsemigroupoid of $S$. It follows by Theorem \[thm:auto\_cofinite\_subsgpoid\] that $U$ is automatic, by [@KambitesAutoCat Corollary 4.6] that $U$ has an automatic cross-section, and then by Proposition \[prop:no\_duplicate\_letters\] that $U$ has an automatic cross-section $(X, K, \sigma)$ with the property that the restriction of $\sigma$ to $X$ is injective.
By the definition of $U$, for each $y \in X$ we can choose $s_y, t_y \in S$, $i_y \in I$ and $\lambda_y \in \Lambda$ such that $y \sigma = s_y P_{\lambda_y i_y} t_y$. Now let $$\label{eqn:hdefine}
H = \lbrace t_y, s_y \mid y \in X \rbrace \cup \lbrace 1_s \mid s \in S^0 \rbrace \subseteq S.$$ We define new alphabets $$A = \lbrace a_{i g h \lambda} \mid i \in I, g, h \in H, \lambda \in \Lambda, i F = g \alpha, g \omega = h \alpha, h \omega = \lambda G \rbrace$$ and $$B = \lbrace b_{i s \lambda} \mid i \in I, s \in S \setminus U, \lambda \in \Lambda, i F = s \alpha, s \omega = \lambda G \rbrace.$$ Clearly, $A$ and $B$ are finite. Let $z$ be a new symbol not in $A$ or $B$, which will represent the zero element $0 \in M$. We define a morphism of semigroups $$\rho : (A \cup B \cup \lbrace z \rbrace)^+ \to M$$ by $(a_{i g h \lambda}) \rho = (i, gh, \lambda)$, $(b_{i s \lambda}) \rho = (i, s, \lambda)$ and $z \rho = 0$.
For every $i \in I$ and $\lambda \in \Lambda$, let $X_{i \lambda}$ be the set $$X_{i \lambda} = \lbrace x \in X^+ \mid i F = x \alpha, x \omega = \lambda G \rbrace$$ of words in $X^+$ starting at $i F$ and finishing at $\lambda G$, and define a function $\phi_{i \lambda} : X_{i \lambda} \to A^+$ by $$(w_1 w_2 \dots w_n) \phi_{i \lambda} = a_{i,1_{i F},s_{w_1},\lambda_{w_1}} a_{i_{w_1},t_{w_1},s_{w_2},\lambda_{w_2}} \dots a_{i_{w_n}, t_{w_n}, 1_{\lambda G}, \lambda}$$ for all $w_1, w_2, \dots, w_n \in X$ with $w_1 \dots w_n \in X_{i \lambda}$. It follows from the fact that the restriction of $\sigma$ to $X$ is injective, that each function $\phi_{i \lambda}$ is injective. Indeed, if $(w_1 \dots w_n) \phi_{i \lambda} = (w'_1 \dots w'_{n'}) \phi_{i \lambda}$ then from the definition of $\phi_{i \lambda}$ we have $$\begin{aligned}
&a_{i,1_{i F},s_{w_1},\lambda_{w_1}} a_{i_{w_1},t_{w_1},s_{w_2},\lambda_{w_2}} \dots a_{i_{w_n}, t_{w_n}, 1_{\lambda G}, \lambda}
= (w_1 \dots w_n) \phi_{i \lambda} \\
= &(w'_1 \dots w'_{n'}) \phi_{i \lambda}
= a_{i,1_{i F},s_{w'_1},\lambda_{w'_1}} a_{i_{w'_1},t_{w'_1},s_{w'_2},\lambda_{w'_2}} \dots a_{i_{w'_{n'}}, t_{w'_{n'}}, 1_{\lambda G}, \lambda}\end{aligned}$$ so clearly $n = n'$ and for $1 \leq k \leq n$ we have $s_{w_k} = s_{w'_k}$, $i_{w_k} = i_{w'_k}$, $\lambda_{w_k} = \lambda_{w'_k}$ and $t_{w_k} = t_{w'_k}$. Now each $$w_k \sigma = s_{w_k} P_{\lambda_{w_k} i_{w_k}} t_{w_k} = s_{w'_k} P_{\lambda_{w'_k} i_{w'_k}} t_{w'_k} = w'_k \sigma.$$ But $\sigma$ is by assumption injective on $X$, so it follows that each $w_k = w'_k$.
Note also that $$\label{eqn:autoreesa}
(i, w \sigma, \lambda) = w \phi_{i \lambda} \rho$$ for all paths $w \in X_{i \lambda}$.
Our aim is to show that each function of the form $\phi_{i \lambda}$ is strongly regularity preserving, and that each pair of such functions is strongly mutually synchronous regularity preserving. We shall do so by showing that each such function has a regular image and that the functions have sliding window inverses which are pairwise synchronised, and then invoking Lemmas \[lem:regpreserve\] and \[lem:syncratpreserve\].
We claim first that the image $X_{i \lambda} \phi_{i \lambda}$ of each $\phi_{i \lambda}$ is regular. We say that a two-letter word $a_{i_1, g_1, h_1, \lambda_1} a_{i_2, g_2, h_2, \lambda_2} \in A^2$ is *compatible* if there exists $y \in X$ such that $s_y = h_1$, $\lambda_y = \lambda_1$, $i_y = i_2$ and $t_y = g_2$, and *incompatible* otherwise. Then clearly, $X_{i \lambda} \phi_{i \lambda}$ is the set of words in $A^{\geq 2}$ which begin with a letter of the form $a_{i,1_{i F}, h, \mu}$ for some $h \in H$ and $\mu \in \lambda$, end with a letter of the form $a_{j, g, 1_{\lambda G}, \lambda}$ for some $j \in I$ and $g \in H$, and contain no incompatible factors of length $2$. Thus, $X_{i \lambda} \phi_{i \lambda} = (A_1 A^* A_2) \setminus (A^* A_3 A^*)$ where $$A_1 = \lbrace a_{i, 1_{i F}, h, \mu} \in A \mid h \in H, \mu \in \Lambda, i F = h \alpha, h \omega = \mu G \rbrace,$$ $$A_2 = \lbrace a_{j, g, 1_{\lambda G}, \lambda} \in A \mid j \in I, g \in H, j F = g \alpha, g \omega = \lambda G \rbrace, \text{ and}$$ $$A_3 = \lbrace ab \in A^2 \mid ab \text{ is not compatible} \rbrace$$ are all finite. It follows that $X_{i \lambda} \phi_{i \lambda}$ is regular, as required.
Next, we claim that the $\phi_{i \lambda}$ have pairwise synchronised sliding window inverses. Fix $i \in I$ and $\lambda \in \Lambda$. We shall define a function $$f : \lbrace ab \in A^2 \mid ab \text{ is compatible} \rbrace \to X.$$ By our observations above, the domain of $f$ will then include all two-letter factors of words in $X_{i \lambda} \phi_{i \lambda}$.
Suppose $a_{i_1 g_1 h_1 \lambda_1} a_{i_2 g_2 h_2 \lambda_2}$ is compatible. Then by definition, there is some $y \in X$ with $s_y = h_1$, $\lambda_y = \lambda_1$, $i_y = i_2$ and $t_y = g_2$. Furthermore, we have $y \sigma = h_1 P_{\lambda_1 i_2} g_2$, and $\sigma$ is injective on $X$, so $y$ is the unique letter in $X$ with this property. Thus, we can well-define $f$ by the rule $$(a_{i_1 g_1 h_1 \lambda_1} a_{i_2 g_2 h_2 \lambda}) f = y.$$
We claim that $(2, f, f, f)$ is a sliding window inverse for $\phi_{i \lambda}$. To show this, suppose $w = w_1 w_2 \dots w_n \in X_{i \lambda}$ with $w_1, \dots, w_n \in X$. Then by the definition of $\phi_{i \lambda}$ we have $$(w_1 w_2 \dots w_n) \phi_{i \lambda} = a_{i,1_{i F},s_{w_1},\lambda_{w_1}} a_{i_{w_1},t_{w_1},s_{w_2},\lambda_{w_2}} \dots a_{i_{w_n}, t_{w_n}, 1_{\lambda G}, \lambda}.$$ Let $$y_1 = a_{i,1_{i F},s_{w_1},\lambda_{w_1}},
\hspace{1em} y_2 = a_{i_{w_1},t_{w_1},s_{w_2},\lambda_{w_2}},
\hspace{1em} \dots,
\hspace{1em} y_{n+1} = a_{i_{w_n}, t_{w_n}, 1_{\lambda G}, \lambda}$$ so that $$(w_1 w_2 \dots w_n) \phi_{i \lambda} = y_1 y_2 \dots y_{n+1}.$$ Now from the definition of $f$, we see that $(y_i y_{i+1}) f = w_i$ for each $1 \leq i \leq n$, so we have $$(y_1 y_2) f (y_2 y_3) f (y_3 y_4) f \dots (y_{n-1} y_{n}) f (y_{n} y_{n+1}) f = w_1 \dots w_n$$ as required. It follows by Lemma \[lem:regpreserve\] that each $\phi_{i \lambda}$ is strongly regularity preserving.
Furthermore, if we define $\psi : {\mathbb{N}}\to {\mathbb{N}}$ to be the constant function given by $n \psi = 1$, then $m \psi = 1 = |w f|$ for any $w \in {\operatorname{Fact}}_2(X_{i \lambda} \phi_{i \lambda})$ which occurs starting in position $m$ (numbered from zero). Since the definition of $\psi$ is independent of the choice of $i$ and $\lambda$, it follows that $\psi$ synchronises the sliding window inverses constructed above for any pair of functions of the form $\phi_{i \lambda}$. We have already observed that the $\phi_{i \lambda}$ are injective and have regular images, so it follows by Lemma \[lem:syncratpreserve\] that any pair of such functions are strongly mutually synchronous regularity preserving.
We now define a language $L \subseteq (A \cup B \cup \lbrace z \rbrace)^+$ by $$L = B \cup \lbrace z \rbrace \cup \bigcup_{i \in I, \lambda \in \Lambda} (K \cap X_{i \lambda}) \phi_{i \lambda}$$ We claim that $(A \cup B \cup \lbrace z \rbrace, L, \rho)$ is an automatic structure (indeed, an automatic cross-section) for $M$.
We show first that $L$ contains a unique representative for every element of $M$. Certainly $z \in L$ is the unique word representing the zero element $0 \in M$. Now consider a non-zero element $(i, s, \lambda) \in M$. If $s \in S \setminus U$ then $b_{i s \lambda} \in B$ is clearly the unique word in $L$ representing $(i, s, \lambda)$, as required. Otherwise, we have $s \in U$. Now $(X, K, \sigma)$ is a choice of representatives for $U$, so we must have $s = w \sigma$ for some path $w \in K$. Furthermore, $w \alpha = s \alpha = i F$ and $w \omega = s \omega = \lambda G$, so that $w \in X_{i \lambda}$. Now by we have $$(i, s, \lambda) = (i, w \sigma, \lambda) = w \phi_{i \lambda} \rho$$ where $w \phi_{i \lambda} \in (K \cap X_{i \lambda}) \phi_{i \lambda} \subseteq L$. Furthermore, if $i' \in I$ and $\lambda' \in \Lambda$ and $v \in K \cap X_{i' \lambda'}$ are such that $v \phi_{i' \lambda'} = (i, s, \lambda)$, then by we must have $(i', v \sigma, j') = (i, w \sigma, j)$ from which it follows that $i = i'$, $j = j'$ and, since $K$ contains a unique representative for every element in $U$, $v = w$. Hence $w \phi_{i \lambda}$ is the unique representative for $(i, s, \lambda)$ in $L$. Thus, $(A \cup B \cup \lbrace z \rbrace, L, \rho)$ is a cross-section for $M$.
Next, we show that $L$ is a regular language. Clearly, $B$ and $\lbrace z \rbrace$ are finite and hence regular. Furthermore, $K$ is regular and each $\phi_{i \lambda}$ is strongly regularity preserving, so each $(K \cap X_{i \lambda}) \phi_{i \lambda}$ is regular. It follows that $L$ is a finite union of regular languages, and hence is regular.
Since $(A \cup B \cup \lbrace z \rbrace, L, \rho)$ is a cross-section for $M$, it follows by Lemma \[lem:syncrat\_prop\](v) that the binary relation $$L_= = \lbrace (u, v) \in L \times L \mid u \rho = v \rho \rbrace = \lbrace (u, u) \mid u \in L \rbrace$$ is synchronously regular.
It remains to show that $L_a$ is synchronously regular for all $a \in A \cup B \cup \lbrace z \rbrace$. First, since $z$ is the unique representative in $L$ for the zero element $0 \in M$, we have $L_z = \lbrace (w, z) \mid w \in L \rbrace = L \times \lbrace z \rbrace$ which is synchronously regular by Lemma \[lem:syncrat\_prop\](iii).
Now suppose $a \in A \cup B$, with $a \rho = (i_a, s_a, \lambda_a)$. By Lemma \[lem:limitedwork1\], it will suffice to show that $$L_a \cap ((A^+ \cap L) \times L) = L_a \cap (A^+ \times L)$$ is synchronously regular. We write $L_a \cap (A^+ \times L)$ as a union $$\begin{aligned}
L_a \cap (A^+ \times L) = &(L_a \cap (A^+ \times B)) \\
&\cup (L_a \cap (A^+ \times \lbrace z \rbrace)) \\
&\cup (L_a \cap (A^+ \times A^+)),\end{aligned}$$ and show that each component in this union is synchronously regular.
Suppose $(u, v) \in L_a \cap (A^+ \times L)$, so that $(u \rho) (a \rho) = v \rho$. Clearly $v \notin B$, since letters in $B$ represent indecomposable elements of $M$, so $L_a \cap (A^+ \times B)$ is empty and hence synchronously regular.
If $v = z$, then we have $(u \rho) (a \rho) = v \rho = z \rho = 0$. Since $u \in A^+ \cap L$ cannot represent the zero element, this can happen only if $P_{u \rho \pi_3, a \rho \pi_1} = 0$. Thus, $$\begin{aligned}
L_a \cap (A^+ \times \lbrace z \rbrace) &= \lbrace (u, z) \mid u \in A^+, P_{u \rho \pi_3, a \rho \pi_1} = 0 \rbrace \\
&= \left(\bigcup_{b \in A, P_{(b \rho \pi_3), a \rho \pi_1} = 0} (A^* b \cap L) \right) \times \lbrace z \rbrace\end{aligned}$$ is a product of finite unions of regular languages, and by Lemma \[lem:syncrat\_prop\](iii) is synchronously regular.
It remains only to show that $L_a \cap (A^+ \times A^+)$ is synchronously regular. Certainly we have $$L_a \cap (A^+ \times A^+) = \bigcup_{i \in I, \lambda \in \Lambda} L_{a, i, \lambda}$$ where each $$L_{a,i,\lambda} = \lbrace (u, v) \in L_a \mid u, v \in A^+, u \rho \pi_1 = v \rho \pi_1 = i, u \rho \pi_3 = \lambda \rbrace.$$ If $P_{\lambda, (a \rho \pi_1)} = 0$ then $L_{a, i, \lambda}$ is empty and hence synchronously regular. Otherwise, let $w$ be a path in $K$ such that $w \sigma = P_{\lambda, (a \rho \pi_1)} (a \rho \pi_2)$. It follows from [@KambitesAutoCat Proposition 4.2] the language $K_w$ is regular. Now using we have $$\begin{aligned}
L_{a, i, \lambda} &= \lbrace (u, v) \in (K \cap X_{i \lambda}) \phi_{i \lambda} \times (K \cap X_{i,a \rho \pi_3}) \phi_{i, a \rho \pi_3} \mid (u \rho) (a \rho) = v \rho \rbrace \\
&= \lbrace (u, v) \in (K \cap X_{i \lambda}) \phi_{i \lambda} \times (K \cap X_{i,a \rho \pi_3}) \phi_{i, a \rho \pi_3} \\
&\hskip 17em \mid (u \rho \pi_2) P_{\lambda, a \rho \pi_1} (a \rho \pi_2) = v \rho \pi_2 \rbrace \\
&= \lbrace (x \phi_{i \lambda}, y \phi_{i, a \rho \pi_3}) \mid (x, y) \in K_w \cap (X_{i \lambda} \times X_{i, a \rho \pi_3}) \rbrace.\end{aligned}$$
But $\phi_{i \lambda}$ and $\phi_{i, a \rho \pi_3}$ are strongly mutually synchronous regularity preserving, so it follows that $L_{a, i, \lambda}$ is synchronously regular. Hence, $L_a$ is a finite union of synchronously regular binary relations, and by Lemma \[lem:syncrat\_prop\](iv) is synchronously regular as required.
This result extends easily from small categories to semigroupoids.
\[thm:auto\_sgpoid\_imp\_rees\] Let $M = M^0(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup with zero over a semigroupoid $S$. Suppose $S$ is automatic and $M$ is finitely generated. Then $M$ is automatic.
Let $P'$ be the set of non-zero entries in the sandwich matrix $P$ and let $U = S P' S$. Since $M$ is finitely generated, Theorem \[thm:fgp\_rees\] tells us that $I$, $\Lambda$ and $S \setminus U$ are all finite. Let $\overline{S}$ be the category obtained from $S$ by adjoining a new identity arrow $1_e$ at every object $e \in S^0$ (even if there is already an identity arrow at $e$). Since $I$ and $\Lambda$ are finite, the images $I F$ and $\Lambda G$ of $F$ and $G$ are finite. But since $S$ is isolation-free, it follows from Proposition \[prop:all\_sgpoid\_used\] that $S^0$ is the union of $I F$ and $G \Lambda$, so $S$ has only finitely many objects. It follows that $S$ is a cofinite subsemigroupoid of $\overline{S}$ so, by Theorem \[thm:auto\_cofinite\_subsgpoid\], $\overline{S}$ is automatic.
Define $\overline{I} = I \cup \lbrace i_s \mid s \in S^0 \setminus I F \rbrace$ and extend $F$ to a function $\overline{F} : \overline{I} \to S^0$ by defining $i_s \overline{F} = s$ for all $s \in S^0 \setminus I F$. Similarly, define $\overline{\Lambda} = \Lambda \cup \lbrace \lambda_s \mid s \in S^0 \setminus
\Lambda G \rbrace$ and extend $G$ to a function $\overline{G} : \overline{\Lambda} \to S^0$ by defining $\lambda_s \overline{G} = s$ for all $s \in S^0 \setminus \Lambda G$. Extend $P$ to a $\overline{\Lambda} \times \overline{I}$ matrix $\overline{P}$, by defining all new entries to be zero.
Now let $\overline{M} = M^0(\overline{S}; \overline{F}(\overline{I}),
\overline{G}(\overline{\Lambda}); \overline{P})$. We know that $\overline{I}$ and $\overline{\Lambda}$ are finite and that $\overline{S}$ is automatic. Furthermore, the set of non-zero entries in $\overline{P}$ is exactly the set $P'$ of non-zero entries in $P$. Now if we let $\overline{U} = \overline{S} \ P' \ \overline{S}$, then we have $U \subseteq \overline{U}$, from which it follows that $\overline{S} \setminus \overline{U}$ is finite. It follows by Theorem \[thm:auto\_cat\_imp\_rees\] that $\overline{M}$ is automatic.
Now clearly every non-zero element of $M$ is also contained in $\overline{M}$. Moreover, the multiplication in $M$ is clearly the same as that in $\overline{M}$, so we conclude that $M$ is a subsemigroup of $\overline{M}$. Furthermore, the only elements of $\overline{M}$ not in $M$ are those of the form $(i, 1_e, \lambda)$ for $i \in \overline{I}$, $\lambda \in \overline{\Lambda}$ and $e \in S^0$ with $$i \overline{F} = 1_e \alpha = e = 1_e \omega = \lambda \overline{G}.$$ Since $S$ has finitely many objects and $\overline{I}$ and $\overline{\Lambda}$ are finite it follows that $M$ is a cofinite subsemigroup of $\overline{M}$. By Theorem \[thm:auto\_cofinite\_subsgpoid\] (or [@Hoffmann02 Theorem 1.1]), it follows that $M$ is automatic.
Combining Theorem \[thm:auto\_sgpoid\_imp\_rees\] with Theorem \[thm:fgp\_rees\] we obtain a more explicit sufficient condition, without reference to $M$ being finitely generated.
\[thm:auto\_rees\_condition\] Let $M = M^0(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup with zero over a semigroupoid $S$. If
- the indexing sets $I$ and $\Lambda$ are finite;
- the semigroupoid $S$ is automatic; and
- $S \setminus SP'S$ is finite, where $P'$ is the set of non-zero entries in the sandwich matrix $P$
then $M$ is automatic.
Next, we give a sufficient condition for the underlying semigroupoid of an automatic Rees matrix semigroup to be automatic.
Let $M = M^0(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup with zero over a semigroupoid $S$. Let $T$ be a subset of $S$. We say that $T$ is *strongly right-ideal-generated by a row cross-section of $P$* if there exists a subset $\Lambda' \subseteq \Lambda$ such that the restriction of $G$ to $\Lambda'$ is bijective, and every arrow in $T$ can be written in the form $P_{\lambda i} s$ for some $\lambda \in \Lambda'$, $i \in I$ and $s \in S^1$. We say that $T$ is *(weakly) right-ideal-generated by a row cross-section of $P$* if there exists $\Lambda'$ as above, such that every arrow in $T$ can be written either in the form $P_{\lambda i} s$ or in the form $P_{\lambda i}$ (or both) for some $\lambda \in \Lambda'$, $i \in I$ and (where appropriate) $s \in S^1$.
The next result says that the set of non-identity elements of $S$ being strongly right-ideal-generated by a row cross-section of $P$ is a sufficient condition for automaticity in $M$ to imply automaticity in $S$. We shall subsequently strengthen the result by weakening the hypothesis, showing that it suffices for the set of non-identity elements of $S$ to be *weakly* right-ideal-generated by a row cross-section of $P$.
\[thm:auto\_rees\_mayimp\_sgpoid1\] Let $M = M^0(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup with zero over a semigroupoid $S$. Suppose $M$ is automatic, and the set of non-identity elements of $S$ is strongly right-ideal-generated by a row cross-section of $P$. Then $S$ is automatic.
It follows from [@KambitesAutoCat Corollary 4.6] (or [@Campbell01 Corollary 5.6]) that $M$ has an automatic cross-section and then from Proposition \[prop:no\_duplicate\_letters\] that $M$ has an automatic cross-section $(A, L, \rho)$ with the property that the restriction of $\rho$ to $A$ is injective. Let $\Lambda' \subseteq \Lambda$ be such that the restriction of $G$ to $\Lambda'$ is bijective, and every non-identity arrow in $S$ can be written in the form $P_{\lambda i} s$ for some $\lambda \in \Lambda'$, $i \in I$ and $s \in S^1$.
Choose some subset $I' \subseteq I$ such that $F$ restricts to a bijection on $I'$. Let $F' : S^1 \alpha \to I'$ and $G' : S^1 \omega \to \Lambda'$ be the inverses of the restrictions of $F$ and $G$ to $I'$ and $\Lambda'$ respectively. We define sets $$C = \lbrace c_s \mid s \in S, (i, s, \lambda) \in A \rho \text{ for some } i \in I, \lambda \in \Lambda \rbrace$$ and $$D = \lbrace d_{\lambda i} \mid \lambda \in \Lambda, i \in I, P_{\lambda i} \neq 0 \rbrace.$$ Clearly, $C$ and $D$ are finite. Let $X$ be a graph with vertex set $S^0$, and edge set $C \cup D$ where $c_s \alpha = s \alpha$, $c_s \omega = s \omega$, $d_{\lambda i} \alpha = P_{\lambda i} \alpha$ and $d_{\lambda i} \omega = P_{\lambda i} \omega$. Then there is a natural morphism $\sigma : X^+ \to S$, given by $c_s \sigma = s$ and $d_{\lambda i} \sigma = P_{\lambda i}$. Note that $w \sigma \alpha = w \alpha$ and $w \sigma \omega = w \omega$ for every $w \in X^1$ and hence for every $w \in X^+$.
Let $V$ be the language $$V = (I' \times S \times \Lambda') \rho^{-1}$$ of words in $A^+$ which represent non-zero elements of the form $(i, s, \lambda) \in M$ with $i \in I'$ and $\lambda \in \Lambda'$. Define a function $\phi : V \to X^+$ by $$\label{eqn:phidefine}
(a_1 a_2 \dots a_n) \phi = c_{s_1} d_{\lambda_1 i_2} c_{s_2} d_{\lambda_2 i_3} \dots d_{\lambda_{n-1} i_n} c_{s_n}$$ where each $a_k \rho = (i_k, s_k, \lambda_k)$. That $\phi$ is a well-defined function into $X^+$ follows from the fact that words in $V$ represent non-zero elements of $M$. Furthermore, if $$(a_1 a_2 \dots a_n) \phi = c_{s_1} d_{\lambda_1 i_2} c_{s_2} d_{\lambda_2 i_3} \dots d_{\lambda_{m-1} i_m} c_{s_m} = (b_1 b_2 \dots b_{n'}) \phi$$ then clearly $n = m = n'$ and for $1 \leq k \leq n$ we have $$a_k \rho = (i_k, s_k, \lambda_k) = b_k \rho$$ where $i_1 = s_1 \alpha F'$ and $\lambda_m = s_m \omega G'$. But $\rho$ is by assumption injective when restricted to the alphabet $A$, so it follows that $\phi$ is injective.
Notice also that for any $w \in V$ we have $$\label{eqn:autoreesb}
w \rho = (w \phi \sigma \alpha F', w \phi \sigma, w \phi \sigma \omega G').$$
Our aim is to show that the function $\phi$ has a regular image and a self-synchronised sliding window inverse. We claim first that the image $V \phi$ of the function $\phi$ is regular. We say that an ordered triple $(d_{\lambda i}, c_s, d_{\mu j}) \in D \times C \times D$ is a *valid internal triple* if there exists a letter $a \in A$ with $a \rho = (i, s, \mu)$. We say that $(c_s, d_{\lambda i}) \in C \times D$ is a *valid start pair* if there exists $a \in A$ with $a \rho = (i', s, \lambda)$ for some $i' \in I'$. We say that $(d_{\lambda i}, c_s) \in D \times C$ is a *valid end pair* if there exists $a \in A$ with $a \rho = (i, s, \lambda')$ for some $\lambda' \in \Lambda'$. Now it is easily verified that $$V \phi \cap X^{\geq 5} = Q_1 (CD)^* C Q_2 \setminus (C \cup D)^* Q_3' (C \cup D)^*$$ where $$Q_1 = \lbrace cd \in CD \mid (c, d) \text{ is a valid start pair} \rbrace,$$ $$Q_2 = \lbrace dc \in DC \mid (d, c) \text{ is a valid end pair} \rbrace, \text{ and }$$ $$Q_3' = \lbrace cde \in DCD \mid (c, d, e) \text{ is \textit{not} a valid internal triple} \rbrace$$ are all finite. It follows that $V \phi \cap X^{\geq 5}$ is regular, and hence that $V \phi$ is regular, as required.
Next, we claim that $\phi$ has a self-synchronised sliding window inverse. Consider a three-letter prefix $v_1 v_2 v_3$ of a word $$(a_1 a_2 \dots a_n) \phi \in V \phi.$$ We shall define $$(v_1 v_2 v_3) f = a_1.$$ To show that $f$ is well-defined, suppose $(b_1 b_2 \dots b_m) \phi$ also has the prefix $v_1 v_2 v_3$. Then by the definition of $\phi$ we have $$(a_1 a_2 \dots a_n) \phi = c_{s_1} d_{\lambda_1 i_2} c_{s_2} d_{\lambda_2 i_3} \dots d_{\lambda_{n-1} i_n} c_{s_n}$$ where each $a_k \rho = (i_k, s_k, \lambda_k)$, and similarly $$(b_1 b_2 \dots b_m) \phi = c_{t_1} d_{\mu_1 j_2} c_{t_2} d_{\mu_2 j_3} \dots \dots \dots d_{\mu_{m-1} j_m} c_{t_m}$$ where each $b_k \rho = (j_k, t_k, \mu_k)$. But now we have $c_{s_1} = v_1 = c_{t_1}$ and $d_{\lambda_1 i_2} = v_2 = d_{\mu_1 j_2}$. But by the definitions of $C$ and $D$, it follows that $s_1 = t_1$ and $\lambda_1 = \mu_1$. Furthermore, we must have $i_1 F = s_1 \alpha = t_1 \alpha = j_1 F$, but $i_1, j_1 \in I'$ and $F$ is injective on $I'$, so we must have $i_1 = j_1$. Thus, we obtain $$a_1 \rho = (i_1, s_1, \lambda_1) = (j_1, t_1, \mu_1) = b_1 \rho.$$ But $\rho$ is, by assumption, injective on the alphabet $A$, so we must have $a_1 = b_1$, as required to show that $f$ is well-defined.
Similarly, given a word $$(a_1 a_2 \dots a_n) \phi = c_{s_1} d_{\lambda_1 i_2} c_{s_2} d_{\lambda_2 i_3} \dots d_{\lambda_{n-1} i_n} c_{s_n}$$ we define $$(d_{\lambda_{p-1} i_p} c_{s_p} d_{\lambda_p i_{p+1}}) g = a_p \text{ and } \
(c_{s_p} d_{\lambda_p i_{p+1}} c_{s_{p+1}}) g = \epsilon$$ where $\epsilon$ denotes the empty word in $A^*$. A similar argument to that for $f$ shows that $g$ is well-defined.
Finally, given a three-letter suffix $v_1 v_2 v_3$ of a word $(a_1 a_2 \dots a_n) \phi \in V \phi,$ we shall define $(v_1 v_2 v_3) h = a_n.$ Once again, a similar argument to that for $f$ shows that $h$ is well-defined.
Now suppose $w = a_1 \dots a_n \in V$ and $w \phi = y_1 \dots y_m$. Then by the definition of $\phi$, we have $$y_1 \dots y_m = (a_1 a_2 \dots a_n) \phi = c_{s_1} d_{\lambda_1 i_2} c_{s_2} d_{\lambda_2 i_3} \dots d_{\lambda_{n-1} i_n} c_{s_n}$$ where each $a_k \rho = (i_k, s_k, \lambda_k)$. But now by the definitions of $f$, $g$ and $h$, it follows that $$\begin{aligned}
(y_1 y_2 y_3) f (y_2 y_3 y_4) g (y_3 y_4 y_5) g \dots (y_{m-3} &y_{m-2} y_{m-1}) g (y_{m-2} y_{m-1} y_m) h \\
&= a_1 a_2 \epsilon a_3 \epsilon \dots \epsilon a_{n-1} a_n = w\end{aligned}$$ We have shown that $(3, f, g, h)$ is a sliding window inverse for $\phi$, and by Lemma \[lem:regpreserve\], it follows that $\phi$ is strongly regularity preserving.
Furthermore, if we define a function $\psi : {\mathbb{N}}\to {\mathbb{N}}$ by $$n \psi = \begin{cases}
1 &\text{ if } n = 0 \text{ or } n \text{ is odd} \\
0 &\text{otherwise}
\end{cases}$$ then since factors of paths in $V \phi$ begin with a letter from $C$ exactly if they begin in an odd position, $\psi$ synchronises $(3, f, g, h)$ with itself. We have already observed that $\phi$ is injective and has a regular image. It follows by Lemma \[lem:syncratpreserve\] that $\phi$ is strongly synchronous regularity preserving.
We now define $K = (L \cap V) \phi \subseteq X^+$, and claim that $(A, K, \sigma)$ is an automatic structure (indeed, an automatic cross-section) for $S$.
First, we show that $\sigma$ maps $K$ bijectively onto $S$. To this end, let $s \in S$. Then there is an element $m = (s \alpha F', s, s \omega G') \in M$, so there exists a word $w \in L$ representing $m$. Indeed, since $s \alpha F' \in I'$ and $s \omega G' \in \Lambda'$, we have $w \in V$, and so $w \phi \in K$. But by , $w \phi \sigma = w \rho \pi_2 = s$, so $w \phi$ is a representative in $K$ for $s$.
Furthermore, if $v \phi \in (L \cap V) \phi = K$ also represents $s$, then we must have $v \rho = (i', s, \lambda')$ where $i' \in I'$ and $\lambda' \in \Lambda'$. But for $(i', s, \lambda') \in M$, we must have $i' F = s \alpha = i F$ and $\lambda' G = s \omega = \lambda G$. Since $F$ and $G$ are bijective when restricted to $I'$ and $\Lambda'$, we must have $i = i'$ and $\lambda = \lambda'$, and so $v \rho = w \rho$. But $v, w \in L$ and $(A, L, \rho)$ is a cross-section for $M$, so we deduce that $v = w$, and hence $v \phi = w \phi$. Thus, $(X, K, \sigma)$ is a cross-section for $S$.
We know that $L$ is a regular language, and that $\phi$ is strongly regularity preserving, so it is immediate that $K = (L \cap V) \phi$ is a regular language. Since $(X, K, \sigma)$ is a cross-section for $S$, it follows that the binary relation $$K_= = \lbrace (u, v) \in K \times K \mid u \sigma = v \sigma \rbrace = \lbrace (u, u) \mid u \in K \rbrace$$ is synchronously regular by Lemma \[lem:syncrat\_prop\](v).
Now let $b$ be an edge in $X$. We must show that $$K_b = \lbrace (u, v) \in K \times K \mid u \omega = b \alpha, (u b) \sigma = v \sigma \rbrace$$ is synchronously regular.
If $b \sigma$ is an identity arrow, then it follows easily from the fact that $K_=$ is synchronously regular that $K_b$ is synchronously regular, and we are done.
Otherwise, $b \sigma$ is not an identity arrow in $S$. Now by assumption, we can write $b \sigma = P_{\lambda i} c$ for some $\lambda \in \Lambda'$, $i \in I$ and $c \in S^1$. Furthermore, we must have $b \alpha = b \sigma \alpha = (P_{\lambda i} c) \alpha = \lambda G$, so that $\lambda = b \alpha G'$. Let $\mu = b \omega G'$. Now $M$ has an element $(i, c, \mu)$. It follows from [@KambitesAutoCat Proposition 4.2] that the language $$L_{(i,c,\mu)} = \lbrace (u, v) \in L \times L \mid (u \rho) (i, c, \mu) = v \rho \rbrace$$ is synchronously regular. We claim that $$K_b = \lbrace (u \phi, v \phi) \mid (u, v) \in L_{(i, c, \mu)} \cap (V \times V), u \rho \pi_3 = \lambda \rbrace.$$ It will then follow that $$K_b = \lbrace (u \phi, v \phi) \mid (u, v) \in L_{(i, c, \mu)} \cap (A^* D \times A^+) \cap (V \times V) \rbrace$$ where $D = \lbrace a \in A \mid a \rho \pi_3 = \lambda \rbrace$ is finite, so that $A^* D$ is regular by Lemma \[lem:syncrat\_prop\]. By Lemma \[lem:syncrat\_prop\](iii) and (iv) it will follow that $L_{(i, c, \mu)} \cap (A^* D \times A^+)$ is synchronously regular. Finally, since $\phi$ is strongly synchronous regularity preserving, we shall deduce that $K_b$ is synchronously regular, as required.
To prove the claim, first suppose that $(a_1, a_2) \in K_b$. Then certainly $a_1, a_2 \in K = (L \cap V) \phi$, so $a_1 = u_1 \phi$ and $a_2 = u_2 \phi$ for some $u_1, u_2 \in L \cap V$. Now by we have $$u_k \rho = (u_k \phi \sigma \alpha F', u_k \phi \sigma, u_k \phi \sigma \omega G')$$ for $k = 1$ and $k = 2$. Now we have $$u_1 \rho \pi_3 = u_1 \phi \sigma \omega G' = a_1 \sigma \omega G' = a_1 \omega G' = b \alpha G' = \lambda.$$ Similarly, we obtain $$u_1 \rho \pi_1 = u_1 \phi \sigma \alpha F' = a_1 \sigma \alpha F' = a_2 \sigma \alpha F' = u_2 \phi \sigma \alpha F' = u_2 \rho \pi_1$$ and $$u_2 \rho \pi_3 = u_2 \phi \sigma \omega G' = a_2 \sigma \omega G' = b \sigma \omega G' = \mu = (i, c, \mu) \pi_3.$$ Finally, we have $$u_2 \rho \pi_2 = u_2 \phi \sigma = a_2 \sigma = (a_1 \sigma) (b \sigma) = (a \sigma) P_{\lambda i} c = ((u_1 \rho) (i, c, \mu)) \pi_2.$$ We have shown that $(u_1 \rho) (i, c, \mu)$ and $u_2 \rho$ are equal in all three components and certainly $u_1, u_2 \in L \cap V$, from which it follows that $(u_1, u_2) \in L_{(i, c, \mu)}$ as required.
Conversely, suppose $(u_1, u_2) \in L_{(i, c, \mu)} \cap (V \times V)$, and that $u_1 \rho \pi_3 = \lambda$. Then $u_2 \rho = (u_1 \rho) (i, c, \mu)$, and using and equating second components, we obtain $$u_2 \phi \sigma = u_2 \rho \pi_2
= (u_1 \rho \pi_2) P_{u_1 \rho \pi_3, i} c
= (u_1 \phi \sigma) P_{\lambda i} c
= (u_1 \phi \sigma) (b \sigma).$$ But now by the definition of $K_b$, it follows that $(u_1 \phi, u_2 \phi) \in K_b$. It follows, as discussed above, that $K_b$ is synchronously regular, as required to complete the proof.
This result extends to cover the case where $S$ is only weakly right-ideal-generated by a row cross-section of $P$.
\[thm:auto\_rees\_mayimp\_sgpoid2\] Let $M = M^0(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup with zero over a semigroupoid $S$. Suppose $M$ is automatic, and the set of non-identity elements of $S$ is weakly right-ideal-generated by a row cross-section of $P$. Then $S$ is automatic.
Since $M$ is automatic it is certainly finitely generated, so Theorem \[thm:fgp\_rees\] tells us that $S$ is finitely generated and that the indexing sets $I$ and $\Lambda$ are finite. Let $\overline{S}$ be the category obtained from $S$ by adjoining a new identity arrow at every object $e \in S^0$ which does not already have one. Define $\overline{I}$, $\overline{\Lambda}$, $\overline{P}$, $\overline{M}$ and $\overline{U}$ just as in the proof of Theorem \[thm:auto\_sgpoid\_imp\_rees\], again noting that $\overline{I}$ and $\overline{\Lambda}$ are finite. Still reasoning as in the proof of Theorem \[thm:auto\_sgpoid\_imp\_rees\], we deduce that $M$ is a cofinite subsemigroupoid of $\overline{M}$. It follows by Theorem \[thm:auto\_cofinite\_subsgpoid\] (or [@Hoffmann02 Theorem 1.1]) that $\overline{M}$ is automatic. We also deduce that $\overline{S} \setminus \overline{U}$ is finite.
Next, we wish to show that the set of non-identity elements of $\overline{S}$ is strongly right-ideal-generated by a row cross-section of $\overline{P}$. Since the non-identity elements of $S$ are weakly right-ideal-generated by a row cross-section of $P$, we can choose some set $\Lambda' \subseteq \Lambda$ such that the restriction of $G$ to $\Lambda'$ is bijective, and every non-identity element of $S$ can be written in the form $P_{\lambda i} c$ or $P_{\lambda i}$ for some $\lambda \in \Lambda'$, $i \in I$ and (where appropriate) $c \in S$. Let $$\overline{\Lambda}' = \Lambda' \cup (\overline{\Lambda} \setminus \Lambda).$$
Now suppose $s$ is a non-identity element of $\overline{S}$. Notice that, because we have adjoined an identity only where $S$ did not already have one, every identity in $S$ remains an identity in $\overline{S}$. Thus, we can assume that $s$ is a non-identity element of $S$. Thus, there exist $\lambda \in \Lambda'$ and $i \in I$ such that either $s = P_{\lambda i} c$ for some $c \in S \subseteq \overline{S}$, or $s = P_{\lambda i} = P_{\lambda i} e$ where $e \in \overline{S}$ is the identity at $P_{\lambda i} \omega = i F$. Thus, the set of non-identity elements of $\overline{S}$ is strongly right-ideal-generated by a row cross-section of $P$.
It now follows by Theorem \[thm:auto\_rees\_mayimp\_sgpoid1\] that $\overline{S}$ is automatic. But as in the proof of Theorem \[thm:auto\_sgpoid\_imp\_rees\], we deduce that $S$ is a cofinite subsemigroupoid of $\overline{S}$. Now by Theorem \[thm:auto\_cofinite\_subsgpoid\], it follows that $S$ is automatic, as required.
Rees Matrices and Prefix-Automaticity {#sec:pautorees}
=====================================
In this section, we turn our attention to prefix-automaticity. The following result, which provides a sufficient condition for a Rees matrix semigroup over a semigroupoid to be prefix-automatic, generalises [@Silva00b Theorem 7.2] in the case of prefix-automaticity. As with Theorems \[thm:auto\_rees\_mayimp\_sgpoid1\] and \[thm:auto\_rees\_mayimp\_sgpoid2\], we shall prove this result first with the hypothesis that the semigroupoid $S$ is a small category and that the non-identity elements of $S$ are strongly right-ideal-generated by a row cross-section of the sandwich matrix $P$. We shall then extend the result to the more general case in which $S$ is a semigroupoid with non-identity elements weakly right-ideal-generated by a row cross-section of $P$.
\[thm:pauto\_cat\_mayimp\_rees\] Let $M = M^0(S; F(I), G(\Lambda); P)$ be a finitely generated Rees matrix semigroup with zero over a small category $S$. If $S$ is prefix-automatic, and the non-identity elements of $S$ are strongly right-ideal-generated by a row cross-section of $P$, then $M$ is prefix-automatic.
Let $\Lambda' \subseteq \Lambda$ be such that the restriction of $G$ to $\Lambda'$ is bijective, and every non-identity arrow in $S$ can be written in the form $P_{\lambda i} s$ for some $\lambda \in \Lambda'$, $i \in I$ and $s \in S$.
We define $U$ as in the proof of Theorem \[thm:auto\_cat\_imp\_rees\], once again deducing from the fact that $M$ is finitely generated, that $I$, $\Lambda$ and $S \setminus U$ are all finite and that $U$ is a subsemigroupoid of $S$. It follows by Theorem \[thm:auto\_cofinite\_subsgpoid\] that $U$ is prefix-automatic, and by [@KambitesAutoCat Proposition 4.4] that $U$ has a prefix-closed automatic structure. Note that since $S$ is finitely generated, it has finitely many objects and hence finitely many local identities. It follows by Proposition \[prop:no\_duplicate\_letters\] that $U$ has a prefix-closed automatic structure $(X, K, \sigma)$ with the property that the restriction of $\rho$ to $X$ is injective. (Note that, since we require our automatic structure to be prefix-closed, we cannot insist that it should also be a cross-section.)
We define $A$, $B$, $z$, $L$, $H$, $\rho$ and each $X_{i \lambda}$ and $\phi_{i \lambda}$ exactly as in the proof of Theorem \[thm:auto\_cat\_imp\_rees\], and claim now that $(A \cup B \cup \lbrace z \rbrace, L, \rho)$ is a prefix-automatic structure for $S$.
By exactly the same arguments as in the proof of Theorem \[thm:auto\_cat\_imp\_rees\], we deduce that $(A \cup B \cup \lbrace z \rbrace, L, \rho)$ is a regular choice of representatives for $M$, and that $L_a$ is synchronously regular for all letters $a \in A \cup B \cup \lbrace z \rbrace$. However, since $(X, K, \sigma)$ is not here assumed to be a cross-section for $S$, we cannot deduce that $(A \cup B \cup \lbrace z \rbrace, L, \rho)$ is a cross-section for $M$, and we must work a little harder to show that $L_=$ is regular.
By Lemma \[lem:limitedwork1\], it will suffice to show that $L_= \cap (A^+ \times A^+)$ is synchronously regular. Now certainly $$L_= \cap (A^+ \times A^+) = \bigcup_{i \in I, \lambda \in \Lambda} L_{i, \lambda}$$ where $$L_{i,\lambda} = \lbrace (u, v) \in L_= \cap (A^+ \times A^+) \mid u \rho \pi_1 = v \rho \pi_1 = i, u \rho \pi_3 = v \rho \pi_3 = \lambda \rbrace.$$ But $$\begin{aligned}
L_{i, \lambda} &= \lbrace (u, v) \in (K \cap X_{i \lambda}) \phi_{i \lambda} \times (K \cap X_{i \lambda}) \phi_{i \lambda} \mid u \rho = v \rho \rbrace \\
&= \lbrace (x \phi_{i \lambda}, y \phi_{i \lambda}) \mid (x, y) \in K_= \cap (X_{i \lambda} \times X_{i \lambda}) \rbrace.\end{aligned}$$ From our argument in the proof of Theorem \[thm:auto\_cat\_imp\_rees\], we know that $\phi_{i \lambda}$ is strongly synchronous regularity preserving, and we know that $K_=$ is synchronously regular, so it follows that each $L_{i,\lambda}$ is synchronously regular. Hence, $L_= \cap (A^+ \times A^+)$ is a finite union of synchronously regular binary relations, and by Lemma \[lem:syncrat\_prop\](iv) is synchronously regular as required.
Thus, we conclude that $(A \cup B \cup \lbrace z \rbrace, L, \rho)$ is an automatic structure for $M$. It remains only to show that the language $$L'_= = \lbrace (u,v) \mid u \in L, v \in {\operatorname{Pref}}(L), u \rho = v \rho \rbrace$$ is synchronously regular. First, observe that we can write $$\begin{aligned}
L'_= = &(L'_= \cap (B \times {\operatorname{Pref}}(L))) \cup (L'_= \cap (L \times B)) \cup (L'_= \cap (\lbrace z \rbrace \times {\operatorname{Pref}}(L)) \\
&\cup (L'_= \cap (L \times \lbrace z \rbrace)) \cup (L'_= \cap (A^+ \times A^+)).\end{aligned}$$
Clearly, $z$ and letters in $B$ are unique representatives in ${\operatorname{Pref}}(L)$ for the respective elements they represent. Thus, we have $$L'_= \cap (B \times {\operatorname{Pref}}(L))) = L'_= \cap (L \times B) = \lbrace (b, b) \mid b \in B \rbrace$$ and $$L'_= \cap (\lbrace z \rbrace \times {\operatorname{Pref}}(L)) = L'_= \cap (L \times \lbrace z \rbrace) = \lbrace (z,z) \rbrace$$ so that the first four components of the union are synchronously regular.
It remains to show that $L'_= \cap (A^+ \times A^+)$ is synchronously regular. Define $$\overline{L} = L \cup \lbrace a_{i 1_s 1_s \lambda} \mid i \in I, s \in S^0, \lambda \in \Lambda, i F = s = \lambda G \rbrace.$$ Clearly, since $I$ and $\Lambda$ are finite and $S$ has finitely many objects, $\overline{L} \setminus L$ is finite. It follows by Lemma \[lem:limitedwork2\] that $\lbrace A \cup B \cup \lbrace z \rbrace, \overline{L}, \rho)$ is an automatic structure for $M$.
Given any letter $a = a_{i s t \lambda} \in A$, we define $\overline{a} = a_{i s 1_{s \omega} \lambda'}$, where $\lambda' = s \omega G'$ is the unique element in $\Lambda'$ satisfying $\lambda' G = s \omega$.
Now consider a non-empty (not necessarily proper) prefix $a_1 \dots a_k$ of a word $a_1 \dots a_n$ in $L \cap A^+$. We claim that $a_1 \dots a_{k-1} \overline{a_k} \in \overline{L}$. If $k=1$ then $\overline{a_k}$ is of the form $a_{i 1_s 1_s \lambda}$, and so is by definition in $\overline{L}$. Otherwise, we observe that from the definitions of $L$ and $\phi_{i \lambda}$ we must have $$a_1 \dots a_n = (y_1 \dots y_{n-1}) \phi_{i \lambda}$$ for some $i \in I$, $\lambda \in \Lambda$ and some word $y_1 \dots y_{n-1} \in K \cap X_{i \lambda}$. But $K$ is prefix-closed, so also $y_1 \dots y_{k-1} \in K$. Furthermore, if we let $\mu \in \overline{\Lambda}$ with $\mu G = y_{k-1} \omega$ then we have $y_1 \dots y_{k-1} \in K \cap X_{i \mu}$ and $$(y_1 \dots y_{k-1}) \phi_{i \mu} = a_1 \dots a_{k-1} \overline{a_k}.$$ Thus, $a_1 \dots a_{k-1} \overline{a_k} \in L \subseteq \overline{L}$ as claimed.
Recall from page the definitions of the subset $H \subseteq S$ and the alphabet $A$. Let $s \in H$ and $\lambda \in \Lambda$ be such that $s \omega = \lambda F$. We consider separately the cases in which $s$ is and is not an identity element in $S$.
First, suppose that $s$ is not an identity element. Then by assumption, we can choose $\lambda' \in \Lambda'$, $i_s \in I$ and $t_s \in S$ such that $s = P_{\lambda' i_s} t_s$. Then certainly $i_s F = t_s \alpha$ and $t_s \omega = s \omega = \lambda F$, so there exists an element $(i_s, t_s, \lambda) \in M$.
Since we have shown that $(A \cup B \cup \lbrace z \rbrace, \overline{L}, \rho)$ is an automatic structure for $M$, we deduce using [@KambitesAutoCat Proposition 4.2] that the language $$\overline{L}_{(i_s, t_s, \lambda)} = \lbrace (u, v) \in \overline{L} \times \overline{L} \mid (u \rho) (i_s, t_s, \lambda) = v \rho \rbrace$$ is synchronously regular. We define a new language $$\begin{aligned}
N_{s, \lambda} = \lbrace (u, a_1 \dots a_k) \mid &u \in L, a_1 \dots a_k \in {\operatorname{Pref}}(L), a_k
= a_{i, s', s, \lambda}, \\
&(a_1 \dots a_{k-1} \overline{a_k}, u) \in \overline{L}_{(i_s, t_s, \lambda)} \rbrace\end{aligned}$$ It follows easily from Proposition \[prop:syncrat\_changelastletter\] and Lemma \[lem:syncrat\_prop\] that $N_{s,\lambda}$ is synchronously regular.
Now for any $u \in L$ and $a_1 \dots a_k \in {\operatorname{Pref}}(L)$ with $a_k$ of the form $a_{i,s',s,\lambda}$ we have $$\overline{a_k} \rho \pi_3 = s' \omega G' = s \alpha G' = P_{\lambda' i_s} \alpha G' = \lambda'$$ so that $$\begin{aligned}
(u, a_1 \dots a_k) \in N_{s,\lambda} &\iff (a_1 \dots a_{k-1} \overline{a_k}, u) \in \overline{L}_{(i_s, t_s, \lambda)} \\
&\iff (a_1 \dots a_{k-1} \overline{a_k}) \rho (i_s, t_s, \lambda) = u \rho \\
&\iff (a_1 \dots a_{k-1}) \rho (i, s' P_{\lambda' i_s} t_s, \lambda) = u \rho \\
&\iff (a_1 \dots a_{k-1}) \rho (i, s' s, \lambda) = u \rho \\
&\iff (a_1 \dots a_k) \rho = u \rho \\
&\iff (u, a_1 \dots a_k) \in L_='.\end{aligned}$$
Next, we consider the case in which $s$ is a local identity in $S$. In this case, we define $$\begin{aligned}
N_{s, \lambda} = \lbrace (&b_1 \dots b_n, a_1 \dots a_k) \mid
a_1 \dots a_k \in {\operatorname{Pref}}(L), b_1 \dots b_n \in L, \\
&a_k = a_{i, s', s, \lambda},
b_n = a_{j, t, 1_{t \alpha}, \lambda},
(a_1 \dots a_{k-1} \overline{a_k}, b_1 \dots b_{n-1} \overline{b_n})
\in \overline{L}_= \rbrace.\end{aligned}$$
Since $(A \cup B \cup \lbrace z \rbrace, \overline{L}, \rho)$ is an automatic structure for $M$, we deduce that $\overline{L}_=$ is regular, and then by Proposition \[prop:syncrat\_changelastletter\] and Lemma \[lem:syncrat\_prop\](i) that $N_{s,\lambda}$ is synchronously regular.
Now for any $a = a_1 \dots a_k \in {\operatorname{Pref}}(L)$ and $b = b_1 \dots b_n \in L$ with $a_k$ of the form $a_{i,s',s,\lambda}$ and $b_n$ of the form $a_{j, t, 1_{t \alpha}, \lambda}$ we have $$\begin{aligned}
(b, a) \in N_{s,\lambda} &\iff
(a_1 \dots a_{k-1} \overline{a_k}, b_1 \dots b_{n-1} \overline{b_n}) \in \overline{L}_= \\
&\iff (a_1 \dots a_{k-1} \overline{a_k}) \rho = (b_1 \dots b_{n-1} \overline{b_n}) \rho \\
&\iff (a_1 \dots a_{k-1}) \rho (i, s', \lambda G G') = (b_1 \dots b_{n-1}) \rho (j, t, \lambda G G') \\
&\iff (a_1 \dots a_{k-1}) \rho (i, s's, \lambda) = (b_1 \dots b_{n-1}) \rho (j, t 1_{t \alpha}, \lambda) \\
&\iff (a_1 \dots a_k) \rho = (b_1 \dots b_n) \rho \\
&\iff (b_1 \dots b_n, a_1 \dots a_k) \in L_='.\end{aligned}$$
It follows from the two cases considered that $$L'_= = \bigcup_{s \in H, \lambda \in \Lambda} N_{s,\lambda}$$ so that $L'_=$ is a union of finitely many synchronously regular languages, and hence by Lemma \[lem:syncrat\_prop\](iv) is synchronously regular.
We now extend this result as described above.
\[thm:pauto\_sgpoid\_mayimp\_rees\] Let $M = M^0(S; F(I), G(\Lambda); P)$ be a finitely generated Rees matrix semigroup with zero over a semigroupoid $S$. If $S$ is prefix-automatic, and the set of non-identity elements in $S$ is weakly right-ideal-generated by a row cross-section of $P$, then $M$ is prefix-automatic.
We combine the methods used to prove Theorems \[thm:auto\_sgpoid\_imp\_rees\] and \[thm:auto\_rees\_mayimp\_sgpoid2\].
As in the proof of Theorem \[thm:auto\_sgpoid\_imp\_rees\], we deduce that the indexing sets $I$ and $\Lambda$ are finite. We define $\overline{S}$, $\overline{I}$, $\overline{\Lambda}$, $\overline{P}$, $\overline{M}$ and $\overline{U}$ as in the proof of Theorem \[thm:auto\_rees\_mayimp\_sgpoid2\], noting that $\overline{S}$ has an adjoined identity only where there was not already an identity in $S$.
As in Theorem \[thm:auto\_rees\_mayimp\_sgpoid2\], we deduce that $S$ is a cofinite subsemigroupoid of $\overline{S}$, that $M$ is a cofinite subsemigroup of $\overline{M}$, and that the non-identity elements of $\overline{S}$ are strongly right-ideal-generated by a row cross-section of $\overline{P}$.
It follows by Theorem \[thm:auto\_cofinite\_subsgpoid\] that $\overline{S}$ is prefix-automatic, by Theorem \[thm:pauto\_cat\_mayimp\_rees\] that $\overline{M}$ is prefix-automatic and then by Theorem \[thm:auto\_cofinite\_subsgpoid\] that $M$ is automatic, as required.
We now show that prefix-automaticity in a Rees matrix semigroup is a sufficient condition for prefix-automaticity in the underlying semigroupoid. The following result generalises [@Descalco01 Theorem 4.2].
\[thm:pauto\_rees\_imp\_sgpoid\] Let $M = M^0(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup with zero over a semigroupoid $S$. If $M$ is prefix-automatic then $S$ is prefix-automatic.
Suppose $M$ is prefix-automatic. Then by [@KambitesAutoCat Proposition 4.4] $M$ has a prefix-closed automatic structure, and by Proposition \[prop:no\_duplicate\_letters\], $M$ has a prefix-closed automatic structure $(A, L, \rho)$ with the property that the restriction of $\rho$ to $A$ is injective. (Once again, we note that, because we require an automatic structure which is prefix-closed, we cannot ask also that it be a cross-section.)
Choose some subsets $I' \subseteq I$ and $\Lambda' \subseteq \Lambda$ such that $F$ and $G$ restrict to bijections on $I'$ and $\Lambda'$ respectively. Define $F'$, $G'$, $C$, $D$, $X$, $\sigma$, $V$, $\phi$ and $K$ exactly as in the proof of Theorem \[thm:auto\_rees\_mayimp\_sgpoid1\]. Reasoning as before, we deduce that $\phi$ is strongly regularity preserving and strongly synchronous regularity preserving, and that for any $w \in V$ we have $$\label{eqn:autoreesb_repeat}
w \rho = (w \phi \sigma \alpha F', w \phi \sigma, w \phi \sigma \omega G').$$ We deduce also that $(X, K, \sigma)$ is a regular choice of representatives for $S$, although we can no longer conclude that it is a cross-section. We claim that $(X, K, \sigma)$ is a prefix-automatic structure for $S$.
Because of the limited role played by the alphabet in the definition of an automatic structure, we can assume without loss of generality that for every letter $a \in A$, there is a letter in $A$ representing the element $$(a \rho \pi_1, a \rho \pi_2, a \rho \pi_3 G G') \in M.$$ For each $a \in A$, let $\overline{a} \in A$ be such a letter.
Let $c \in X^1 \cap X^0$ be an edge or a path of length 0 in $X$. Consider the binary relation $$K_c' = \lbrace (x, y) \in {\operatorname{Pref}}(K) \times K \mid x \omega = c \alpha, (xc) \sigma = y \sigma \rbrace.$$ Our aim is to show that $K_c'$ is synchronously regular. We claim that $$\label{eqn:kcclaim}
K_c' = \left( \bigcup_{\lambda \in \Lambda, i \in I, b \in C \cup CD} K_{\lambda,i,b} \right) \cup \left( \bigcup_{x \in {\operatorname{Pref}}(K) \cap X^{\leq 4}} K_c' \cap ( \lbrace x \rbrace \times K) \right)$$ where each $$\begin{aligned}
K_{\lambda,i,b} = \lbrace & (((w_1 \dots w_{n-1} \overline{w_n}) \phi) d_{\lambda i} b, v \phi) \mid w_1, w_2, \dots, w_n \in A, v \in L \cap V \\
&w = w_1 \dots w_n \in L, w \rho \pi_3 = \lambda, (w_1 \dots w_n, v) \in L_{(i,(bc) \sigma, c \omega G')} \rbrace.\end{aligned}$$ To prove the claim, suppose first that $(x,y) \in K_c'$. Clearly, if $|x| \leq 4$ then, since $(x,y) \in K_c' \cap ( \lbrace x \rbrace \times K)$, we see that $(x,y)$ is contained in the right-hand-side of and we are done.
Now suppose $|x| \geq 5$. Certainly $y \in K = (L \cap V) \phi$, so we can write $y = v \phi$ for some $v \in L \cap V$. Also $x \in {\operatorname{Pref}}(K) = {\operatorname{Pref}}((L \cap V) \phi)$, so certainly we can choose $w \in L \cap V$ and $z \in X^*$ with $xz = w \phi$. From the definition of $\phi$, we have $xz \in C(DC)^*$. Since $|x| \geq 5$, it follows that we can write $x = a d_{\lambda i} b$ for some $a \in CDC(DC)^*$, $\lambda \in \Lambda$, $i \in I$ and either $b \in C$ or $b \in CD$.
Suppose $w = w_1 \dots w_k$. Then it is easily verified from the definition of $\phi$ (see page ) that $a = (w_1 \dots w_{n-1} \overline{w_n}) \phi$ where $n = (|a|+1) / 2$. Let $u = w_1 \dots w_n$, noticing that we have $u \rho \pi_3 = \lambda$. Now we have $$\begin{aligned}
(u \rho) (i, (bc) \sigma, c \omega G')
&= (u \rho \pi_1, u \rho \pi_2, \lambda) (i, (bc) \sigma, c \omega G') \\
&= (u \rho \pi_1, a \sigma, \lambda) (i, (bc) \sigma, c \omega G') \\
&= (u \rho \pi_1, (a \sigma) P_{\lambda i} (bc) \sigma, c \omega G') \\
&= (u \rho \pi_1, (a d_{\lambda i} b c) \sigma, c \omega G') \\
&= (v \rho \pi_1, (x c) \sigma, v \rho \pi_3) \\
&= (v \rho \pi_1, y \sigma, v \rho \pi_3) \\
&= v \rho\end{aligned}$$ so that $(u, v) \in L_{(i, (bc) \sigma, c \omega G')}$ and hence $\left( ((w_1 \dots w_{n-1} \overline{w_n}) \phi \right) d_{\lambda i} b, v \phi) \in K_{\lambda,i,b}$.
Conversely, suppose $w = w_1 \dots w_n \in L$ with $$\left( (w_1 \dots w_{n-1} \overline{w_n}) \phi \right) d_{\lambda i} b, v \phi) \in K_{\lambda,i,b}$$ for some $\lambda \in \Lambda, i \in I$ and $b \in C \cup CD$. Let $a = (w_1 \dots w_{n-1} \overline{w_n}) \phi.$ Now $$\begin{aligned}
(v \rho \pi_1, (((w_1 \dots w_{n-1} \overline{w_n}) \phi) d_{\lambda i} b c) \sigma, v \rho \pi_3)
&= (w \rho \pi_1, (a d_{\lambda i} b c) \sigma, c \omega G') \\
&= (w \rho \pi_1, (a \sigma) P_{\lambda i} (bc) \sigma, c \omega G') \\
&= (w \rho \pi_1, a \sigma, \lambda) (i, (bc) \sigma, c \omega G') \\
&= (w \rho \pi_1, w \rho \pi_2, \lambda) (i, (bc) \sigma, c \omega G') \\
&= (w \rho) (i, (bc \sigma), c \omega G') \\
&= v \rho\end{aligned}$$ so in particular we have $$(((w_1 \dots w_{n-1} \overline{w_n}) \phi) d_{\lambda i} b \sigma) (c \sigma) = v \phi \sigma$$ so that $(((w_1 \dots w_{n-1} \overline{w_n}) \phi) d_{\lambda i} b \sigma, v \phi \sigma) \in K'_c$ as required.
Also, it is clear that if $x \in {\operatorname{Pref}}(K) \cap X^{\leq 4}$ and $y \in K$ are such that $(x,y) \in K_c' \cap ( \lbrace x \rbrace \times K)$ then we must have $(x,y) \in K_c'$. Thus, we have justified our claim that holds.
Now using [@KambitesAutoCat Proposition 4.2] we see that each binary relation of the form $L_{(i, (bc) \sigma, c \omega G')}$ is synchronously regular. Furthermore, the functions given by $$A \to A, a \mapsto \overline{a}, \text{ and}$$ $$X \to X^{|b| + 2}, x \mapsto x d_{\lambda i} b$$ both satisfy the conditions of Proposition \[prop:syncrat\_changelastletter\]. Since $\phi$ is also synchronous regularity preserving, it follows that each $K_{\lambda,i,b}$ is synchronously regular.
Now let $x \in {\operatorname{Pref}}(K) \cap X^{\leq 4}$. Choose a word $u$ in $L$ representing the element $(x \alpha F', (x \sigma) (c \sigma), c \omega G')$. Now for any word $v \phi \in (L \cap V) \phi = K$, we have $(x c) \sigma = v \phi \sigma$ if and only if $u \rho = v \rho$, which in turn is true exactly if $(u, v) \in L_=$. Thus, we have $$K_c' \cap ( \lbrace x \rbrace \times K) = \lbrace x \rbrace \times ((\lbrace u \rbrace \times L) \cap L_=) \pi_2 \phi.$$ We know that $L$ is regular and that $L_=$ is synchronously regular, so it follows by Lemma \[lem:syncrat\_prop\](ii), (iii) and (iv) and the fact that $\phi$ is strongly synchronous regularity preserving that $K_c' \cap (\lbrace x \rbrace \times K)$ is synchronously regular.
We have shown that each $K_c'$ is a finite union of synchronously regular binary relations, and it follows by Lemma \[lem:syncrat\_prop\](iv) that each $K_c'$ is synchronously regular. Now for every $c \in X^1 \cup X^0$ we see that $K_c = K_c' \cap (K \times K)$ is synchronously regular by Lemma \[lem:syncrat\_prop\](iv), so that $(X, L, \sigma)$ is an automatic structure for $S$. Moreover, $$K_=' = \bigcup_{c \in X^0} (K_c')^{-1}$$ is also synchronously regular by Lemma \[lem:syncrat\_prop\](iv), so that $(X, K, \sigma)$ is a prefix-automatic structure for $S$, as required.
Rees Matrix Semigroups Without Zero {#sec:reeswithoutzero}
===================================
Theorem \[thm:auto\_cofinite\_subsgpoid\] ensures that our results about Rees matrix semigroups with zero adapt easily to the case of Rees matrix semigroups without zero.
\[thm:auto\_sgpoid\_imp\_reeswithoutzero\] Let $M = M(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup (without zero) over a semigroupoid $S$. Suppose $S$ is automatic and $M$ is finitely generated. Then $M$ is automatic.
Consider the Rees matrix semigroup with zero $$M' = M^0(S; F(I), G(\Lambda); P).$$ Then $M = M' \setminus \lbrace 0 \rbrace$ is a cofinite subsemigroup of $M'$. So by [@Ruskuc98 Theorem 1.1], $M'$ is finitely generated. Now by Theorem \[thm:auto\_sgpoid\_imp\_rees\], $M'$ is automatic, and by Theorem \[thm:auto\_cofinite\_subsgpoid\] (or [@Hoffmann02 Theorem 1.1]) it follows that $M$ is automatic.
\[thm:auto\_reeswithoutzero\_mayimp\_sgpoid\] Let $M = M(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup (without zero) over a semigroupoid $S$. Suppose the set of non-identity elements of $S$ is weakly right-ideal-generated by a row cross-section of $P$. If $M$ is automatic then $S$ is automatic.
Consider the Rees matrix semigroup with zero $$M' = M^0(S; F(I), G(\Lambda); P).$$ Then $M = M' \setminus \lbrace 0 \rbrace$ is a cofinite subsemigroupoid of $M'$. Now if $M$ is automatic then by Theorem \[thm:auto\_cofinite\_subsgpoid\], $M'$ is automatic, and by Theorem \[thm:auto\_rees\_mayimp\_sgpoid2\], $S$ is automatic.
\[thm:pauto\_sgpoid\_mayimp\_reeswithoutzero\] Let $M = M(S; F(I), G(\Lambda); P)$ be a finitely generated Rees matrix semigroup (without zero) over a semigroupoid $S$. If $S$ is prefix-automatic, and the set of non-identity elements of $S$ is weakly right-ideal-generated by a row cross-section of $P$, then $M$ is prefix-automatic.
Consider the Rees matrix semigroup with zero $$M' = M^0(S; F(I), G(\Lambda); P).$$ Then $M = M' \setminus \lbrace 0 \rbrace$ is a cofinite subsemigroup of $M'$. So by [@Ruskuc98 Theorem 1.1], $M'$ is finitely generated. Now by Theorem \[thm:pauto\_sgpoid\_mayimp\_rees\], $M'$ is prefix-automatic, and by Theorem \[thm:auto\_cofinite\_subsgpoid\] it follows that $M$ is prefix-automatic.
\[thm:pauto\_reeswithoutzero\_imp\_sgpoid\] Let $M = M(S; F(I), G(\Lambda); P)$ be a Rees matrix semigroup (without zero) over a semigroupoid $S$. If $M$ is prefix-automatic then $S$ is prefix-automatic.
Consider the Rees matrix semigroup with zero $$M' = M^0(S; F(I), G(\Lambda); P).$$ Then $M = M' \setminus \lbrace 0 \rbrace$ is a cofinite subsemigroupoid of $M'$. Now if $M$ is prefix-automatic then by Theorem \[thm:auto\_cofinite\_subsgpoid\], $M'$ is prefix-automatic. By Theorem \[thm:pauto\_rees\_imp\_sgpoid\], it follows that $S$ is prefix-automatic.
Closing Remarks {#sec:remarks}
===============
The curious relationship between the results of Section \[sec:pautorees\] and those of Section \[sec:autorees\] seems to demand comment. In the case of automaticity, showing that automaticity in a Rees matrix semigroup is a *sufficient* condition for automaticity in the underlying semigroupoid requires a right-ideal-generation condition, while the converse implication does not. In the case of prefix-automaticity, the situation is entirely reversed – the right-ideal-generation condition is required only to show that prefix-automaticity in a Rees matrix semigroup is a *necessary* condition for prefix-automaticity in the underlying semigroupoid.
In [@KambitesAutoCat] we remarked that it is an open question whether every automatic semigroup is prefix-automatic. Theorem \[thm:cons\_iff\_sgpoid\] implies that this question is no harder in the ostensibly more general semigroupoid context, that is, that every automatic semigroupoid is prefix-automatic, exactly if every automatic semigroup is prefix-automatic. However, the following questions do naturally arise.
\[qn:existautorees\_over\_nonauto\_sgp\] Does there exist an automatic Rees matrix semigroup over a non-automatic semigroup?
\[qn:existautorees\_over\_nonauto\_sgpoid\] Does there exist an automatic Rees matrix semigroup over a non-automatic semigroupoid?
Clearly, if every automatic semigroup (and hence every automatic semigroupoid) is prefix-automatic, then Theorem \[thm:pauto\_rees\_imp\_sgpoid\] gives a negative answer to both of these questions. However, if there *are* automatic semigroups (and hence semigroupoids) which are not prefix-automatic, then the answers to Questions \[qn:existautorees\_over\_nonauto\_sgp\] and \[qn:existautorees\_over\_nonauto\_sgpoid\] could be both positive, both negative, or negative and positive respectively.
Acknowledgements {#acknowledgements .unnumbered}
================
This paper was written while the author was at Carleton University, supported by the Leverhulme Trust. The research documented was conducted while the author was a research student at the University of York, funded by an EPSRC Doctoral Studentship. The author would like to thank John Fountain for all his advice and guidance, as well as Vicky Gould, Nik Ruškuc and the anonymous referees for many helpful comments. He would also like to thank Kirsty for all her support and encouragement.
[10]{}
S. Armstrong. Structure of concordant semigroups. , 118:205–260, 1986.
C. M. Campbell, E. F. Robertson, N. Ruškuc, and R. M. Thomas. Direct products of automatic semigroups. , 69:19–24, 2000.
C. M. Campbell, E. F. Robertson, N. Ruškuc, and R. M. Thomas. Automatic semigroups. , 250:365–391, 2001.
C. M. Campbell, E. F. Robertson, N. Ruškuc, and R. M. Thomas. Automatic completely-simple semigroups. , 95:201–215, 2002.
L. Descalço and N. Ruškuc. On automatic [R]{}ees matrix semigroups. , 30:1207–1226, 2002.
A. J. Duncan, E. F. Robertson, and N. Ruškuc. Automatic monoids and change of generators. , 127:403–409, 1999.
D. B. A. Epstein et al. . Jones and Bartlett, 1992.
J. B. Fountain. Abundant semigroups. , 44:103–129, 1982.
M. Hoffmann. . PhD thesis, Dept. of Mathematics and Computer Science, University of Leicester, 2000.
M. Hoffmann, D. Kuske, F. Otto, and R. M. Thomas. Some relatives of automatic and hyperbolic groups. In G. M. S. Gomes, J.-E. Pin, and P. V. Silva, editors, [ *Semigroups, Algorithms, Automata and Languages*]{}, 2003.
M. Hoffmann, N. Ruškuc, and R. M. Thomas. Automatic semigroups with subsemigroups of finite [R]{}ees index. , 12:463–476, 2002.
M. Hoffmann and R. M. Thomas. Automaticity and commutative semigroups. , 44:167–176, 2002.
M. Hoffmann and R. M. Thomas. Notions of automaticity in semigroups. , 66:337–361, 2003.
J. M. Howie. . Clarendon Press, 1995.
J. F. P. Hudson. Regular rewrite systems and automatic structures. In [*Semigroups, Automata and Languages*]{}, pages 145–152. World Scientific, 1996.
M. E. Kambites. . PhD thesis, Dept. of Mathematics, University of York, 2003.
M. E. Kambites. Automatic semigroups and categories. preprint, 2005.
M. E. Kambites. Presentations for semigroups and semigroupoids. , 15:291–308, 2005.
M. V. Lawson. ees matrix semigroups over semigroupoids and the structure of a class of abundant semigroups. , 66:517–540, 2000.
N. Ruškuc. On large subsemigroups and finiteness conditions of semigroups. , 76:383–405, 1998.
P. V. Silva and B. Steinberg. Extensions and submonoids of automatic monoids. , 289:727–754, 2002.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Victor Pambuccian
title: The sum of irreducible fractions with consecutive denominators is never an integer in a very weak arithmetic
---
amssymb.sty
Most problem solvers have encountered at some stage of their lives the problem asking for a proof that the sum
$$1+\frac{1}{2}+\ldots + \frac{1}{n}$$ for $n\geq 2$ is never an integer. The proof one usually finds offered for this fact is based on Chebyshev’s theorem (Bertrand’s postulate). If one asks for a proof that, more generally, the sum
$$\frac{1}{n}+\frac{1}{n+1}+\ldots + \frac{1}{n+k}$$ with $k\geq 1$ can never be an integer, then the proof based on Chebyshev’s theorem needs to be amended. One first notes that, if $k<n$, then the above sum must be less than 1, and thus cannot be an integer, and if $k\geq n$, then one applies the same proof based on Chebyshev’s theorem (this fact seems to have been overlooked in [@obl], where the author wants to use Chebyshev’s Theorem, but finds it necesary to make the proof dependent on another deep result, the Sylvester-Schur Theorem, as well). However, Kürschák [@kuer] (see also [@ps]) found a much simpler proof, which relies on the very simple observation that among any number ($\geq 2$) of consecutive positive integers there is precisely one, which is divisible by the highest power of 2 from among all the given numbers. Aside from its didactical use, one may wonder whether Kürschák’s proof is not in a very formal way much simpler, i. e. whether it does not require simpler methods of proof in the sense of formal logic.
When formalized, arithmetic is usually presented as Peano Arithmetic, which contains an induction axiom schema, stating, loosely speaking, that any set that can be defined by an elementary formula in the language of arithmetic (i.e. in terms of some undefined operation and predicate symbols, such as $+$, $\cdot$, $1$, $0$, $<$), which contains 1, and which contains $n+1$ whenever it contains $n$, is the set of all numbers. Several weak arithmetics have been studied, in which the types of elementary formulas allowed in the definitions of the sets used in induction are restricted by certain syntactic constraints (see [@daq]), and one might think that Kürschák’s proof would make it in a weaker formal arithmetic than the one dependent on Chebyshev’s theorem. It turns out that, in fact, no amount of induction is needed at all!
To see this, let’s first generalize the problem further, along the lines of the generalization in [@obl1], so that there can be no proof based on Chebyshev’s Theorem.
The sum
$$\label{k1}
\frac{m_0}{n}+\frac{m_1}{n+1}+\ldots + \frac{m_k}{n+k}$$
with $(m_i,n+i)=1$, $m_i<n+i$, and $k\geq 1$ is never an integer.
(Kürschák [@kuer]). Let $a = \max \{ \alpha :
2^{\alpha} | (n+i) \mbox{ for some
$0\leq i \leq k$ } \}$. Then $2^a$ divides exactly one of the numbers $n$, $n+1$, …, $n+k$. Let $l= \mbox{lcm } (n, n+1, \ldots, n+k)$. Suppose the sum in (\[k1\]) is an integer $b$. Multiplying both (\[k1\]) and $b$ by $l$, we obtain on the one hand an odd number, and on the other an even number, which have to be equal.
Moreover, to make it a theorem of arithmetic, we will do away with the fractions appearing in it, and state it, for all positive $k\in
\mathbb{N}$, as $\varphi_k$, the following statement (where we denote by $\overline{u}$ the term $((\ldots ((1+1)+1)+\ldots)+1)$, in which there are $u$ many $1$’s; the terms $\overline{u}$ will be referred to as [*numerals*]{}) $$\begin{aligned}
& & (\forall n)(\forall m_0)\ldots (\forall m_k)(\forall p)\,
\bigvee_{i=0}^k ((\forall a)(\forall b)\, m_ia \neq (n+\overline{i})b +1)\vee \bigvee_{i=0}^k n+\overline{i}<m_i\\
& & \vee \sum_{i=0}^k (m_i\prod_{0\leq j\leq k, j\neq i}
(n+\overline{j})) \neq p \prod_{j=0}^k (n+\overline{j}).\end{aligned}$$
The arithmetic we will show it holds in is PA$^-$, which is expressed in a language containing as undefined operation and predicate symbols only $+$, $\cdot$, $1$, $0$, and $<$, and whose axioms A1-A15 were presented in [@kay pp. 16-18]. We will repeat them here for the reader’s convenience, and we will omit the universal quantifiers for all universal axioms.
$(x+y)+z = x+(y+z)$
$x+y = y+x$
$(x\cdot y)\cdot z = x\cdot (y\cdot z)$
$x\cdot y = y\cdot x$
$x\cdot (y+z) = x\cdot y+x\cdot z$
$x+0=x \wedge x\cdot 0 = 0$
$x\cdot 1 = x$
$(x<y \wedge y<z) \rightarrow x<z$
$\neg x<x$
$x<y \vee x=y \vee y<x$
$x<y \rightarrow x+z < y+z$
$(0<z \wedge x<y) \rightarrow x\cdot z < y\cdot z$
$(\forall x)(\forall y)(\exists z)\, x<y \rightarrow x+z =y$
$0<1 \wedge (x>0 \rightarrow (x>1 \vee x=1))$
$x>0 \vee x=0$
What is missing from $PA^-$, and makes it so weak (indeed, the positive cone of every discretely ordered ring is a model of $PA^-$), is the absence of any form of induction.
The proof that $\varphi_k$ holds in $PA^-$ will be carried out in an arbitrary model $\mathfrak{M}$ of $PA^-$. The idea of proof will be to show that all variables that appear in $\varphi_k$ must be numerals. An essential ingredient of the proof is the following fact, which holds in $PA^-$ (see [@kay Lemma 2.7, p. 22]), for all positive $k\in\mathbb{N}$ $$x<\overline{k} \rightarrow x=0\vee x=1\vee \ldots \vee x=\overline{k-1},$$ and which allows us to deduce that any element that is bounded from above by a numeral must be a numeral.
Suppose that, for some positive $k\in \mathbb{N}$, $\varphi_k$ does not hold in $\mathfrak{M}$. Then, for all $i=0, \ldots, k$, there are $m_i, p, a_i$ and $b_i$ with $m_ia_i=(n+i)b_i + 1$ and such that $$\label{1}
\sum_{i=0}^k (m_i\prod_{0\leq j\leq k, j\neq i} (n+\overline{j})) = p
\prod_{j=0}^k
(n+\overline{j}).$$
This can be rewritten, by leaving only the first term of the sum on the left-hand side, and sending all others to the right-hand side with changed sign, as $m_0(n+\overline{1})\ldots (n+\overline{k})=
nq$, where by $q$ we have denoted $p\prod_{j=1}^k (n+\overline{j}) -
(\sum_{i=1}^k m_i\prod_{0\leq j\leq k, j\neq i} (n+\overline{j}))$. The product $(n+\overline{1})\ldots (n+\overline{k})$ can also be written as a polynomial in $n$, whose free term is $\overline{k!}$, i. e. as $nr + \overline{k!}$, thus $m_0(nr + \overline{k!})=nq$. Given that there are $a_0$ and $b_0$ such that $m_0a_0=nb_0
+1$, if we multiply both sides of the equality $m_0(nr + \overline{k}!)=nq$ by $a_0$ we obtain $(nb_0+1)\overline{k!} = n(a_0q-a_0m_0r)$, thus $\overline{k!} = n(a_0q-a_0m_0r-b_0\overline{k!})$. We know that $\mathfrak{M}$ must contain a copy of $\mathbb{N}$, and it may contain other elements as well, called [*nonstandard*]{} numbers. Could $n$ be in $\mathfrak{M}$ but not of the form $\overline{m}$ for some $m\in \mathbb{N}$? If it were such an element of $\mathfrak{M}$, then it would be greater than all $\overline{m}$ with $m\in \mathbb{N}$, and thus so would $n(a_0q-a_0m_0r-b_0\overline{k!})$, unless $a_0q-a_0m_0r-b_0\overline{k!}=0$, which cannot be the case, as $\overline{k!}$ is not zero. However, $n(a_0q-a_0m_0r-b_0\overline{k!})$ cannot be greater than all $\overline{m}$ with $m\in \mathbb{N}$, for it is equal to such a number, namely to $\overline{k!}$. Thus $n$ must be an $\overline{m}$ for some $m\in \mathbb{N}$. This means that in (\[1\]) all variables are numerals, i. e. $\overline{i}$’s for some $i\in \mathbb{N}$. However, we know, from Kürschák’s proof, that such an equation cannot exist, so, for all $k\in \mathbb{N}$, $\varphi_k$ holds in PA$^-$.
Another generalization of the original problem, proved by T. Nagell in [@nag], states that the sum $$\frac{1}{m}+\frac{1}{m+n}+\frac{1}{m+2n}+\cdots+\frac{1}{m+kn}$$ is never an integer if $n, m, k$ are positive integers. The proof is rather involved and uses both a Kürschák-style argument and Chebyshev’s theorem. This statement turns out to be, with $\overline{k}$ instead of $k$, valid in $PA^{-}$ as well. To see this, let, for all positive $k\in \mathbb{N}$, $\nu_k$ stand for
$$\label{nagell}
(\forall m)(\forall n)(\forall p)\, m>0 \wedge n>0 \rightarrow
\sum_{i=0}^{k}\prod_{0\leq j\leq k, j\neq i} (m+\overline{j}n) \neq
p\prod_{0\leq j\leq k} (m+\overline{j}n),$$
and let $\mathfrak{M}$ be again a model of $PA^{-}$. Notice that, if $m> \overline{k}$, then, then $\neg\nu_k$ cannot hold for any $n$ and $p$. To see this, suppose that, for some $n$ and $p$, we have equality in (\[nagell\]). Given that $(m+n)\ldots
(m+\overline{k}n)$ is the largest of all the summands on the left-hand side, and there are $\overline{k}$ summands, the sum on the left-hand side is $\leq \overline{k}(m+n)\ldots
(m+\overline{k}n)$, and thus $< m(m+n)\ldots (m+\overline{k}n)$, thus equality cannot hold in (\[nagell\]). Thus $m\leq
\overline{k}$, and thus (see [@kay p. 20]) $m$ must be standard, i. e. it must be $\overline{u}$ for some $0<u\leq k$. It remains to be shown that $n$ must be standard as well. To see this, suppose again that, for some $n>0$ and $p$, we have equality in (\[nagell\]). Notice that, since $m(m+2n)\ldots (m+\overline{k}n)$ is the largest product among all $\prod_{0\leq j\leq k, j\neq i}
(m+\overline{j}n)$, for $i=1,2, \ldots k$, the sum on the left hand side of our equality is $\leq (m+n)\ldots (m+\overline{k}n) +
\overline{k}m(m+2n)\ldots (m+\overline{k}n)$ (with equality if and only if $k=1$). Thus $pm(m+n)\ldots (m+\overline{k}n) \leq (m+n +
\overline{k}m)(m+2n)\ldots (m+\overline{k}n)$, which implies $pm(m+n)\leq m+n + \overline{k}m$. If $n$ were nonstandard, then this inequality were possible only if $pm=1$, i. e. if $p=m=1$, which is not possible, for in that case the first summand on the left hand side of (\[nagell\]) is equal to the right hand side, thus the left hand side must be larger than the right hand side, so equality could not have taken place in (\[nagell\]). Now that $m,n,k$ have all been shown to be standard, Nagell’s proof implies the truth of our statement, which thus holds in $PA^{-}$.
By Gödel’s completeness theorem, there must exist syntactic proofs that $\varphi_k$ and $\nu_k$ hold in PA$^-$, i. e. formal derivation of $\varphi_k$ and $\nu_k$ from the axioms of PA$^-$. Such a formal proof for $\varphi_k$ cannot use the idea behind Kürschák’s proof, for it is not even true in PA$^-$ that among $n$, $n+1$, $\ldots$ $n+k$, there is a multiple of 2 (see [@kay p.18]). Thus there must exist even simpler proofs for both $\varphi_k$ and $\nu_k$ and they are worth finding. Such proofs would reveal the real reasons why these results hold, and the reason must be of an algebraic nature, for there is no traditional number theory to be found in $PA^-$.
[17]{}
P. D’Aquino, Weak fragments of Peano Arithmetic, in: *The Notre Dame Lectures*, (ed. P. Cholak), Lecture Notes in Logic, 18, pp. 149–185, A. K. Peters, 2005.
R. Kaye, *Models of Peano Arithmetic*, Oxford University Press, Oxford, 1991.
J. Kürschák, On the harmonic series (Hungarian), Mat. Phys. Lapok 27 (1918), 299–300.
T. Nagell, Eine Eigenschaft gewisser Summen, Norsk Mat. Forenings Skrifter 13 (1923), 10–15.
R. Oblath, Some number-theoretical theorems (Hungarian), Math. és phys. lapok 27 (1918), 91–94.
R. Oblath, Über einen arithmetischen Satz von Kürschák, Comment. Math. Helv. 8 (1935), 186–187.
G. Pólya, G. Szegö, *Aufgaben und Lehrsätze der Analysis*, Springer-Verlag, Berlin, 1971, vol. II, VIII. Abschnitt, Aufgabe 251, S. 159.
*Department of Integrative Studies, Arizona State University - West Campus, Phoenix, AZ 85069-7100, U.S.A\
[email protected]*
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'P.H. Chavanis'
date: To be included later
title: 'Nonlinear mean field Fokker-Planck equations. Application to the chemotaxis of biological populations '
---
Introduction {#sec_introduction}
============
The study of Fokker-Planck equations (Fokker (1914) [@fokker], Planck (1917) [@planck]) is of considerable interest in physics since the pioneering work of Einstein (1905) [@einstein] on the Brownian motion. The simplest Fokker-Planck equation is the Smoluchowski (1915) [@smoluchowski] equation. This is a drift-diffusion equation describing the diffusion of particles in physical space in the presence of an external potential (e.g. the sedimentation of colloids in a gravitational field). A more general Fokker-Planck equation is the Kramers (1940) [@kramers] equation, previously introduced by Klein (1921) [@klein], which takes into account inertial effects and describes the diffusion of particles in phase space when they experience a friction force. The Smoluchowski equation is recovered from the Kramers equation in a strong friction limit where inertial effects are negligible. In this sense, the Smoluchowski equation describes an overdamped evolution. These Fokker-Planck equations [@risken] are consistent with usual thermodynamics in the canonical ensemble. They monotonically decrease the Boltzmann free energy ($H$-theorem) and relax towards the equilibrium Boltzmann distribution.
Recently, several researchers have tried to extend the usual concepts of thermodynamics and kinetic theory in order to describe complex systems that are characterized by non-Boltzmannian distributions. In that respect, some generalized forms of entropic functionals [^1] have been introduced. One of the most popular “generalized entropy” is the Tsallis (1988) [@tsallis] entropy, but other entropies have been presented by Abe (1997) [@abe], Borges & Roditi (1998) [@br], Kaniadakis (2001) [@k1], Naudts (2004) [@naudts], and Kaniadakis [*et al.*]{} (2005) [@k2]. It was later realized that these entropic functionals are special cases of the one-parameter family of entropies introduced earlier by Harvda & Charvat (1967) [@hc] or of the two-parameters family of entropies introduced by Mittal (1975) [@mittal] and Sharma & Taneja (1975) [@st]. Other famous forms of entropies have been presented by Reyni (1970) [@reyni] and Sharma & Mittal (1975) [@sm]. We refer to Kaniadakis & Lissia [@kl] for a very interesting discussion of these historical aspects, starting from the early works of Euler in 1779.
Following these developments, some researchers have tried to develop out-of-equilibrium theories associated to a generalized thermodynamical framework. In particular, it has been first shown by Plastino & Plastino (1995) [@pp] that the Tsallis $q$-distributions are the steady states of a nonlinear Fokker-Planck equation taking into account anomalous diffusion. This type of equations had been previously considered by mathematicians to describe porous media [@spohn]. The seminal work of Plastino & Plastino [@pp] has been further developed by Tsallis & Bukman (1996) [@bukman], Stariolo (1997) [@stariolo], Borland (1998) [@borland] and Nobre [*et al.*]{} (2004) [@nobre] among others. On the other hand, Kaniadakis & Quarati (1994) [@kq] have introduced nonlinear Fokker-Planck equations whose steady states are the Fermi-Dirac [^2] and Bose-Einstein statistics. These kinetic equations take into account an exclusion (fermions) or inclusion (bosons) principle leading to quantum-like statistics at equilibrium. The case of intermediate statistics, interpolating between fermions and bosons, has also been considered in [@kq]. Recently, the bosonic Kramers equation has been studied in [@bose] and was shown to reproduce the phenomenology of the Bose-Einstein condensation in the canonical ensemble.
The above-mentioned nonlinear Fokker-Planck (NFP) equations are associated with [*special*]{} forms of entropic functionals (Tsallis, Fermi-Dirac, Bose-Einstein). More recently, Martinez [*et al.*]{} (1998) [@martinez], Kaniadakis (2001) [@k1], Frank (2002) [@frank1] and Chavanis (2003) [@gen] have studied generalized forms of NFP equations associated with an almost [*arbitrary*]{} entropic functional. They can be viewed as generalized Kramers and Smoluchowski equations where the coefficients of diffusion, friction and drift explicitly depend on the local density of particles. Physically, this can take into account microscopic constraints (exclusion volume effects, steric hindrance, non-extensive effects...) that modify the dynamics of the particles at small scales and lead to non-standard equilibrium distributions [^3]. Martinez [*et al.*]{} [@martinez] determined the NFP equation in order to recover, as a steady state, the equilibrium state produced by minimizing a generalized form of free energy at fixed mass. Kaniadakis [@k1] obtained the NFP equation from the Master equation (see also Curado & Nobre [@cn]) by allowing the transition probabilities to depend on the concentration of particles in the initial and arrival states. Frank [@frank1] derived the NFP equation from a generalized free energy functional by using the linear thermodynamics of Onsager. Chavanis [@gen] obtained the NFP equation by using a form of Maximum Entropy Production Principle (MEPP). This corresponds to a variational version of the linear thermodynamics of Onsager. We refer to the book of Frank [@frank] for a first survey on nonlinear Fokker-Planck equations.
Another topic of active research in statistical physics concerns the dynamics and thermodynamics of systems with long-range interactions [@houches]. Several Hamiltonian systems with long-range interactions have been studied in the microcanonical ensemble such as self-gravitating systems, two-dimensional vortices, the HMF model, the free-electron laser etc... For such systems the mean field approximation provides a very good description of the system and becomes exact in a proper thermodynamic limit $N\rightarrow
+\infty$. The microcanonical ensemble is the correct description of [*isolated*]{} systems evolving at fixed energy. On the other hand, some authors have introduced a canonical version of these models so as to treat systems that are [*dissipative*]{}. This leads to the notion of Brownian systems with long-range interactions. Their dynamics is described by mean field Fokker-Planck equations where the temperature is fixed (instead of the energy). These mean field Fokker-Planck equations were introduced early by Kuramoto (1984) [@kuramoto] to describe the synchronization of globally coupled nonlinear oscillators and more recently by Marzel & Aslangul (2001) [@ma], Chavanis (2006) [@hb] and Frank [@frank] in a more general context. Some specific studies have been made for self-gravitating Brownian particles [@crs; @sc] and for the BMF model [@cvb] which is the canonical version of the HMF model [@antoni].
In view of the importance of these two topics: generalized thermodynamics and long-range interactions, we have introduced in [@gen] a class of nonlinear mean field Fokker-Planck equations (see Eq. (81) of [@gen]) that incorporate [*both*]{} a generalized free energy functional and a long-range potential of interaction. As an illustration, we studied with C. Sire a model of self-gravitating Langevin particles [@lang] (see also [@shiino]) that combines self-gravity (long-range interactions) and anomalous diffusion (generalized thermodynamics) related to the Tsallis entropy and to the polytropic equation of state. In other words, this model couples the NFP equation introduced by Plastino & Plastino [@pp] to the gravitational Poisson equation. Unfortunately, this model of self-gravitating Langevin particles has no clear application in astrophysics (because self-gravitating systems are generally not overdamped and not dissipative) and was introduced essentially as an interesting dynamical model with rich mathematical properties. However, it was noted in [@lang] that this model could have applications in unexpected area, like in the chemotaxis of bacterial populations...
The name chemotaxis refers to the motion of organisms induced by chemical signals [@murray]. In some cases, the biological organisms (bacteria, amoebae, endothelial cells, ants...) secrete a substance (pheromone, smell, food, ...) that has an attractive effect on the organisms themselves. Therefore, in addition to their diffusive motion, they move preferentially along the gradient of concentration of the chemical they secrete (chemotactic flux). When attraction prevails over diffusion, the chemotaxis can trigger a self-accelerating process until a point at which aggregation takes place. This is the case for the slime mold [*Dictyostelium discoideum*]{} and for the bacteria [*Escherichia coli*]{}. This is referred to as chemotactic collapse. A model of slime mold aggregation has been introduced by Patlak (1953) [@patlak] and Keller & Segel (1971) [@ks] in the form of two coupled differential equations. The first equation is a drift-diffusion equation describing the evolution of the concentration of bacteria and the second equation is a diffusion equation with terms of source and degradation describing the evolution of the concentration of the chemical. In the simplest model, the diffusion coefficient $D$ and the mobility $\chi$ of the bacteria are constant. This forms the standard Keller-Segel model. However, the original Keller-Segel model allows these coefficients to depend on the concentration of the bacteria and of the chemical. The case where these coefficients depend on the concentration of the chemical $c({\bf r},t)$, but not on the concentration of bacteria $\rho({\bf r},t)$, has been considered by Othmer & Stevens [@os]. This leads to ordinary mean field Fokker-Planck equations (with respect to $\rho({\bf r},t)$) with space and time dependent coefficients. On the other hand, if we assume that the diffusion coefficient and the mobility of the bacteria depend on their concentration $\rho({\bf r},t)$, but not on the concentration $c({\bf r},t)$ of the secreted chemical, the original Keller-Segel model takes the form of a generalized mean field Fokker-Planck equation. [*Therefore, the Keller-Segel model represents a fundamental example of nonlinear mean field Fokker-Planck equation with physical applications in biology, thereby justifying a notion of generalized thermodynamics.*]{} The analogy between the Keller-Segel model and NFP equations was first pointed out in [@gen] and further developed in subsequent papers (see, e.g., [@csbio]). This analogy makes possible to interprete results of chemotaxis in terms of nonlinear Fokker-Planck equations and generalized thermodynamics, which has not been considered so far by applied mathematicians working on this domain [@horstmann]. It thus allows to make a bridge between two different communities. In this analogy, the model of self-gravitating Langevin particles introduced by Chavanis & Sire [@lang] also provides a generalized Keller-Segel model of chemotaxis taking into account anomalous diffusion (the application of this model to chemotaxis has been emphasized in [@csmass]). More generally, we can use the numerous results accumulated in the context of generalized thermodynamics to propose new forms of generalized Keller-Segel models with potential applications in biology.
The aim of this paper is to develop a simple and rich formalism that allows to deal with nonlinear mean field Fokker-Planck equations. We shall illustrate this formalism on several examples and show the inter-connections between different topics. The paper is organized as follows. In Sec. \[sec\_phys\], we consider NFP equations in physical space. This corresponds to overdamped models where inertial effects are neglected. We review and complete the basic properties of these equations in relation with an effective generalized thermodynamical (E.G.T.) formalism. In Sec. \[sec\_gle\], we show that they can be obtained from generalized Langevin equations. In Sec. \[sec\_h\], we show that they admit an $H$-theorem (in the canonical ensemble) for a generalized free energy. We stress that the Legendre structure of the free energy and the Einstein relation are preserved in this generalized thermodynamical framework. In Sec. \[sec\_sta\], we determine the steady states of these nonlinear mean field Fokker-Planck equations and show that they are solutions of an integrodifferential equation. In Sec. \[sec\_min\] (and in Appendix \[sec\_conn\]), we show that a steady state of a nonlinear mean field Fokker-Planck equation is linearly dynamically stable if and only if (iff) it is a (local) minimum of the free energy at fixed mass. In Sec. \[sec\_gs\], we show that a NFP equation in physical space with a constant mobility and a density-dependent diffusion coefficient can be written in the form of a generalized Smoluchowski equation incorporating a barotropic equation of state. In Secs. \[sec\_phen\] and \[sec\_kin\], we show the correspondance between the phenomenological derivations of the NFP equations given by Frank [@frank1] and Chavanis [@gen] and the kinetic derivation given by Kaniadakis [@k1]. In Sec. \[sec\_ex\], we present several explicit examples of NFP equations and mention their potential applications to the problem of chemotaxis (see Sec. \[sec\_ks\]). In particular, we introduce a new model of chemotaxis that incorporates both effects of anomalous diffusion and exclusion principle (volume filling). The corresponding generalized entropy is expressed in the form of integrals that can be explicited in particular cases. In Sec. \[sec\_phase\], we consider NFP equations in phase space taking into account inertial effects. In Secs. \[sec\_gk\]-\[sec\_eos\], we list their main properties. In Sec. \[sec\_strong\], we consider the strong friction limit and derive the generalized Smoluchowski equation from the generalized Kramers equation. We use a method of moments that is simpler than the Chapman-Enskog expansion presented in [@lemou]. In Sec. \[sec\_ee\], we consider explicit examples corresponding to the Boltzmann, Tsallis and Fermi-Dirac entropies. The Appendices contain important results that complete the basic properties of the NFP equations discussed in the text. In Appendix \[sec\_iso\], we show that a generalized isotropic BGK operator has properties similar to those possessed by a nonlinear Kramers operator. In Appendix \[sec\_conn\], we establish a simple relation showing the equivalence between linear dynamical stability (exponential damping of the perturbation) and generalized thermodynamical stability (minimum of free energy at fixed mass). In Appendix \[sec\_bif\], we study the stability of a spatially homogeneous solution of the nonlinear mean field Fokker-Planck equation and evidence a critical point. In Appendices \[sec\_passage\] and \[sec\_equi\], we show that a distribution function $f$ in phase space is a minimum of the free energy $F[f]$ at fixed mass iff the corresponding distribution $\rho$ in physical space is a minimum of the corresponding free energy $F[\rho]$ at fixed mass. This implies that a distribution function $f$ in phase space is a stable steady state of the generalized Kramers equation iff the corresponding distribution $\rho$ in physical space is a stable steady state of the corresponding generalized Smoluchowski equation. In Appendix \[sec\_eddington\], we extend to $d$ dimensions the Eddington formula that allows to obtain the distribution function $f=f(\epsilon)$ in phase space from the knowledge of the barotropic equation of state $p=p(\rho)$ in physical space. In Appendix \[sec\_ht\], we derive the $H$-theorems associated with the NFP equations and in Appendix \[sec\_pol\] we derive the polytropic equation of state associated with the Tsallis statistics.
Nonlinear mean field Fokker-Planck equations in physical space {#sec_phys}
==============================================================
We first describe nonlinear mean field Fokker-Planck equations in physical space where the inertia of the particles is neglected. They can be viewed as models describing an overdamped dynamics.
Generalized Langevin equations {#sec_gle}
------------------------------
We consider a system of $N$ particles whose individual dynamics is described by the stochastic Ito-Langevin equations $$\begin{aligned}
\label{gle1} \frac{d{\bf
r}_i}{dt}=-\chi(\rho_i)\nabla\Phi_i+\sqrt{2D(\rho_i)}{\bf R}_i(t),\end{aligned}$$ where ${\bf R}_i(t)$ is a white noise satisfying $\langle {\bf
R}_i(t)\rangle={\bf 0}$ and $\langle
R_{i,\alpha}(t)R_{j,\beta}(t')\rangle
=\delta_{ij}\delta_{\alpha,\beta}\delta(t-t')$ where $i=1,...,N$ label the particles and $\alpha=1,...,d$ label the coordinates of space. We have noted $\rho_i=\rho({\bf r}_{i}(t),t)$ and $\Phi_i=\Phi({\bf r}_{i}(t),t)$. In ordinary models, the mobility $\chi$ and the diffusion coefficient $D$ are constant. In that case, the statistical equilibrium state is the Boltzmann distribution $\rho\sim e^{-\Phi/T}$ where the temperature $T=1/\beta$ is given by the Einstein relation $T=D/\chi$. In the present study, we shall consider more general situations and allow the mobility $\chi(\rho)$ and the diffusion coefficient $D(\rho)$ to depend on the local concentration of particles $\rho({\bf
r},t)=\langle \sum_{i=1}^{N}\delta({\bf r}-{\bf
r}_{i}(t))\rangle$. This is an heuristic approach to take into account microscopic constraints that affect the dynamics of particles at small scales and lead to non-Boltzmannian equilibrium distributions. Indeed, it is not surprising that the mobility or the diffusive properties of a particle depend on its environment. For example, in a dense medium, its motion can be hampered by the presence of the other particles so that its mobility is reduced.
On the other hand, in ordinary models, the particles move in a fixed external potential $\Phi_{ext}({\bf r})$. In the present study, we want to be more general and take into account the possibility that the potential $\Phi({\bf r},t)$ is created self-consistently by the particles themselves. In this paper, we shall neglect statistical correlations and use a mean field description (for more general models taking into account statistical correlations see, e.g., [@hb; @new]). Therefore, we assume that the potential is given by a relation of the form $$\begin{aligned}
\label{gle2} \Phi({\bf r},t)=\int \rho({\bf r}',t)u(|{\bf r}-{\bf
r}'|)d{\bf r}',\end{aligned}$$ where $u(|{\bf r}-{\bf r}'|)$ is a binary potential of interaction and $\rho({\bf r},t)$ is the smooth distribution of particles. In general, the mean field approximation gives a very good description of systems with weak long-range binary interactions and it becomes exact in a proper thermodynamic limit $N\rightarrow +\infty$ [@hb]. In Eq. (\[gle2\]), the potential is expressed as a convolution product: $\Phi=u* \rho$. Of course, the potential can be due to the combined effect of the self-interaction and an external field, in which case $\Phi=\Phi_{ext}+u*\rho$. We shall also consider the case where it is determined by an equation of the form $$\begin{aligned}
\label{gle3}
\epsilon\frac{\partial \Phi}{\partial t}=\Delta \Phi-k^{2}\Phi-\lambda\rho,\end{aligned}$$ where $\epsilon$ and $\lambda$ are positive constants. For $\epsilon=0$, Eq. (\[gle3\]) becomes the screened Poisson equation $$\begin{aligned}
\label{gle4}
\Delta \Phi-k^{2}\Phi=\lambda\rho. \end{aligned}$$ Therefore, we can identify $k^{-1}$ as the screening length. If we assume furthermore that $k=0$, we get the Poisson equation $$\begin{aligned}
\label{gle5}
\Delta \Phi=\lambda\rho. \end{aligned}$$ These last two equations can be put in the form of Eq. (\[gle2\]). Note also that, in the stationary state, Eq. (\[gle3\]) reduces to Eq. (\[gle4\]).
Drift-diffusion equations {#sec_dde}
-------------------------
For the stochastic process (\[gle1\]), the evolution of the smooth density of particles $\rho({\bf r},t)$ is governed by the nonlinear mean field Fokker-Planck equation [@gen; @hb]: $$\begin{aligned}
\label{dde1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left\lbrack \nabla(D(\rho)\rho)+\chi(\rho)\rho\nabla\Phi\right\rbrack,\end{aligned}$$ coupled to Eq. (\[gle2\]) or (\[gle3\]). Let us introduce the notations $$\begin{aligned}
\label{dde2} Dh(\rho)=\frac{d}{d\rho}(\rho D(\rho)), \qquad \chi
g(\rho)=\rho\chi(\rho),\end{aligned}$$ where $D$ and $\chi$ are positive constants and $h(\rho)$ and $g(\rho)$ are positive functions. These notations are chosen such that the usual stochastic equations with constant diffusion $D(\rho)=D$ and constant mobility $\chi(\rho)=\chi$ are recovered for $h(\rho)=1$ and $g(\rho)=\rho$. With these notations, the nonlinear Fokker-Planck equation (\[dde1\]) can be rewritten $$\begin{aligned}
\label{dde3} \frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D
h(\rho)\nabla\rho+\chi
g(\rho)\nabla\Phi\right ).\end{aligned}$$ It can be put in the conservative form $$\begin{aligned}
\label{dde4}
\frac{\partial\rho}{\partial t}=-\nabla\cdot {\bf J},\end{aligned}$$ where $$\begin{aligned}
\label{dde5}
{\bf J}=-\left (Dh(\rho)\nabla\rho+\chi g(\rho)\nabla\Phi\right ),\end{aligned}$$ is a diffusion current. This structure guarantees the conservation of mass $M=\int \rho d{\bf r}$ provided that the normal component of the current at the boundary vanishes.
Relation to some known models {#sec_known}
-----------------------------
Known models can be recovered as particular cases of Eq. (\[dde1\]). When $\Phi_{ext}$ is an external potential and when $D(\rho)=D$ and $\chi(\rho)=\chi$ are constant, we recover the Smoluchowski equation $\partial_{t}\rho=\nabla\cdot
(D\nabla\rho+\chi\rho\nabla\Phi_{ext})$ describing the Brownian motion of colloidal suspensions in a fixed gravitational field [@smoluchowski]. When $\rho D(\rho)=K\rho^{\gamma}$ and $\Phi_{ext}=0$, we recover the porous medium equation $\partial_{t}\rho=K\Delta
\rho^{\gamma}$ [@spohn], and when $\Phi_{ext}\neq 0$, we recover the nonlinear Fokker-Planck equation $\partial_{t}\rho=\nabla\cdot
(K\nabla\rho^{\gamma}+\chi\rho\nabla\Phi_{ext})$ introduced by Plastino & Plastino [@pp] in connection to the Tsallis statistics [@tsallis]. When $D(\rho)=D$ and $\chi(\rho)=\chi$ are constant and $u=-\frac{k}{2\pi}\cos(\theta-\theta')$ we obtain the Brownian Mean Field (BMF) model [@cvb], which is the canonical (fixed $T$) version of the microcanonical (fixed $E$) Hamiltonian Mean Field (HMF) model [@antoni]. This is closely related to the Kuramoto model [@kuramoto] which describes the synchronization of globally coupled nonlinear oscillators. When $D(\rho)=D$ and $\chi(\rho)=\chi$ are constant and $\Phi$ is the gravitational potential satisfying $\Delta \Phi=4\pi G\rho$, Eqs. (\[dde1\])-(\[gle5\]) become the Smoluchowski-Poisson system describing a model of self-gravitating Brownian particles studied by Chavanis & Sire [@crs; @sc; @post; @tcoll; @virial1; @virial2]. When $D(\rho)=D$ and $\chi(\rho)=\chi$ are constant and $\Phi$ is the self-consistent electrostatic potential, we recover the equations introduced by Debye & Hückel [@dh] in their model of electrolytes. Models of the form (\[dde1\])-(\[gle4\]) including a time dependent temperature $\beta(t)$ assuring the conservation of energy have been introduced by Robert & Sommeria [@rs] and Chavanis [@physicaD] to describe the violent relaxation of two-dimensional vortices in geophysical and astrophysical flows. In the Quasi Geostrophic (QG) model, $k^{-1}$ represents the Rossby length. Similar equations have been proposed in [@csr] to model the violent relaxation of stellar systems. On the other hand, for short range interactions, we can expand the potential in the form $\Phi({\bf r},t)=a\rho({\bf
r},t)+\frac{b}{2}\Delta\rho({\bf r},t)$ and Eq. (\[dde1\]) leads to a generalized form of the Cahn-Hilliard equation (see [@lemou; @csbio; @new] for details). As a particular case, for $D=0$ and $u=a \delta$ (leading to $\Phi({\bf r},t)=a\rho({\bf r},t)$), we get $\partial_{t}\rho=\chi a\nabla (g(\rho)\nabla\rho)$ and for $g(\rho)=\rho$, we get the porous medium equation $\partial_{t}\rho=\frac{1}{2}\chi a\Delta\rho^{2}$. Finally, in $d=1$, assuming $D(\rho)=D$, $\chi(\rho)=\chi$ and $u'=-\frac{1}{2\chi}\delta$, we get the Burgers equation $\partial_{t}\rho+\rho\partial_{x}\rho=D\partial_{xx}\rho$ [@burgers]. Therefore, the class of nonlinear mean field Fokker-Planck equations (\[dde1\])-(\[gle2\]) introduced in [@gen] can find physical applications in different areas [@cras].
Generalized Keller-Segel model of chemotaxis {#sec_ks}
--------------------------------------------
In addition to the previous examples, nonlinear mean field Fokker-Planck equations can find important applications in the context of chemotaxis [@murray]. The original Keller-Segel model [@ks] describing the chemotaxis of bacterial populations consists in two coupled differential equations $$\begin{aligned}
\label{ks1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left (D_{2}\nabla\rho)-\nabla\cdot (D_{1}\nabla c\right ),\end{aligned}$$ $$\begin{aligned}
\label{ks2}
\epsilon\frac{\partial c}{\partial t}=D_{c}\Delta c-k(c)c+f(c)\rho, \end{aligned}$$ that govern the evolution of the density of bacteria $\rho({\bf r},t)$ and the evolution of the secreted chemical $c({\bf r},t)$. The bacteria diffuse with a diffusion coefficient $D_{2}$ and they also move in a direction of a positive gradient of the chemical (chemotactic drift). The coefficient $D_{1}$ is a measure of the strength of the influence of the chemical gradient on the flow of bacteria. On the other hand, the chemical is produced by the bacteria with a rate $f(c)$ and is degraded with a rate $k(c)$. It also diffuses with a diffusion coefficient $D_{c}$. In the general Keller-Segel model, $D_{1}=D_{1}(\rho,c)$ and $D_{2}=D_{2}(\rho,c)$ can both depend on the concentration of the bacteria and of the chemical. This takes into account microscopic constraints, like close-packing effects, that can hinder the movement of bacteria. If we assume that the quantities only depend on the concentration of bacteria [^4] and write $D_{2}=Dh(\rho)$, $D_{1}=\chi g(\rho)$, $k(c)=k^{2}$, $f(c)=\lambda$ and $D_{c}=1$, we obtain $$\begin{aligned}
\label{ks3} \frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D
h(\rho)\nabla\rho-\chi g(\rho)\nabla c\right ),\end{aligned}$$ $$\begin{aligned}
\label{ks4}
\epsilon\frac{\partial c}{\partial t}=\Delta c-k^{2}c+\lambda\rho. \end{aligned}$$ These equations are isomorphic to the nonlinear mean field Fokker-Planck equations (\[dde3\])-(\[gle3\]) provided that we make the correspondance $\Phi({\bf r},t)=-c({\bf r},t)$: the potential of interaction is played by the concentration of the secreted chemical.
It is important to note that the Keller-Segel model is a [*mean field*]{} model. If we come back to the exact microscopic equations $$\begin{aligned}
\label{c1}
\frac{d{\bf r}_{i}}{dt}=\chi\nabla c_{d}({\bf r}_{i}(t),t)+\sqrt{2D}{\bf R}_{i}(t),\end{aligned}$$ $$\begin{aligned}
\label{c2}
\epsilon\frac{\partial c_{d}}{\partial t}=\Delta c_{d}-k^2 c_{d}+\lambda \sum_{i=1}^{N}\delta({\bf r}-{\bf r}_{i}(t)),\end{aligned}$$ governing the evolution of each particle (bacteria, cells, social insects,...), the mean field approximation leading to the standard Keller-Segel model (\[ks3\])-(\[ks4\]) with $h(\rho)=1$ and $g(\rho)=\rho$ amounts to neglecting statistical correlations; see [@stevens; @ng; @crrs; @hb; @virial2; @kinbio; @new] for details. Let us first assume that the system is Markovian and possesses no intrinsic memory in the sense that $c_{d}({\bf r},t)=-\int u(|{\bf r}-{\bf
r}'|)\rho_{d}({\bf r}',t)d{\bf r}'$ where $\rho_{d}({\bf r},t)=
\sum_{i=1}^{N}\delta({\bf r}-{\bf r}_{i}(t))$ is the exact distribution of particles. When the particles interact via a weak long-range binary potential $u(|{\bf r}-{\bf r}'|)$, it can be shown that the mean field approximation becomes exact in a proper thermodynamic limit $N\rightarrow +\infty$ [@hb]. For example, in Eq. (\[c2\]) these assumptions correspond to $\epsilon=0$ and $k\ll 1$. By contrast, when memory is not zero and the interaction is short range, corresponding to $\epsilon\neq 0$ and $k\gg 1$ in Eq. (\[c2\]), the mean field approximation can lead to wrong results. In the context of chemotaxis, the differences between mean field and non mean field models have been discussed by Grima [@grima] who showed the failure of the mean field approximation for the prediction of the width of the aggregate sizes (the disagreement becomes very severe close to the critical point where we know that mean field approximation breaks down in general). On the other hand, the mean field approximation assumes that the number of particles $N\rightarrow +\infty$. In stellar systems and plasmas, for example, this is always the case. However, for biological systems, the number of interacting bacteria or cells is frequently less than a few thousands so it may be relevant to return to a microscopic description of the bacteria or cells’ movement in terms of $N$-body stochastic equations like (\[c1\])-(\[c2\]) as discussed in [@kinbio]. In this paper, however, we shall exclusively consider mean field models. As we have seen, mean field approximation works well for Markovian systems ($\epsilon=0$) with weak long-range interactions if (i) $N\rightarrow +\infty$ and (ii) if we are not too close from a critical point. Mean field approximation breaks down: (i) for non-Markovian systems ($\epsilon\neq 0$) (ii) close to a critical point (iii) for small values of $N$.
Generalized free energy and H-theorem {#sec_h}
-------------------------------------
When $\Phi_{ext}$ is an external potential, we define the energy by $$\begin{aligned}
\label{h1}
E=\int \rho \Phi_{ext} \, d{\bf r}.\end{aligned}$$ When $\Phi$ is determined by Eq. (\[gle2\]), the self-interaction energy is given by $$\begin{aligned}
\label{h2}
E=\frac{1}{2}\int \rho \Phi \, d{\bf r}.\end{aligned}$$ Finally, when $\Phi$ is determined by Eq. (\[gle3\]), the energy is $$\begin{aligned}
\label{h3}
E=\frac{1}{2\lambda}\int \left\lbrack (\nabla \Phi)^{2}+k^{2} \Phi^{2}\right \rbrack
\, d{\bf r}+\int \rho \Phi \, d{\bf r}.\end{aligned}$$ For $\epsilon=0$, the expression (\[h3\]) reduces to Eq. (\[h2\]). On the other hand, we define the temperature by $$\begin{aligned}
\label{h4}
T=\frac{D}{\chi}.\end{aligned}$$ Therefore, the Einstein relation is preserved in the generalized thermodynamical framework. We also set $\beta=1/T$. We introduce the generalized entropic functional $$\begin{aligned}
\label{h5}
S=-\int C(\rho)\, d{\bf r},\end{aligned}$$ where $C(\rho)$ is a convex function ($C''\ge 0$) defined by $$\begin{aligned}
\label{h6}
C''(\rho)=\frac{h(\rho)}{g(\rho)}.\end{aligned}$$ Note that the relation (\[h6\]) defines the function $C(\rho)$ up to a term of the form $A\rho+B$. After integration over the domain, the first term is proportional to the mass which is a conserved quantity and the second term is just a constant. Therefore, these terms play no physical role. However, we can adapt the values of the constants $A$ and $B$ in order to obtain convenient expressions of the entropy.
Since the system is dissipative, the energy is not conserved. What is fixed instead of the energy is the temperature defined by the Einstein relation (\[h4\]). Since $D\propto T$, the temperature measures the strength of the stochastic force in Eq. (\[gle1\]). This corresponds to a canonical description where the system is in contact with a heat bath. Note that the heat bath is completely disconnected from the long-range potential of interaction; it corresponds to short-range interactions modelled by the stochastic term in Eq. (\[gle1\]). For an isolated system described by the microcanonical ensemble the proper thermodynamical potential is the entropy. Alternatively, for a dissipative system described by the canonical ensemble, the relevant thermodynamical potential is the free energy. We introduce the generalized free energy $$\begin{aligned}
\label{h7}
F=E-TS.\end{aligned}$$ The definition of the free energy (Legendre transform) is preserved in the generalized thermodynamical framework.
When the energy is given by Eqs. (\[h1\]) or (\[h2\]), a straightforward calculation (see Appendix \[sec\_ht\]) shows that $$\begin{aligned}
\label{h8}
\dot F=-\int \frac{1}{\chi g(\rho)}(Dh(\rho)\nabla\rho+\chi g(\rho)\nabla \Phi)^{2}d{\bf r}.\end{aligned}$$ When the energy is given by Eq. (\[h3\]), we obtain $$\begin{aligned}
\label{h9}
\dot F=-\frac{1}{\lambda\epsilon}\int (\Delta \Phi-k^{2}\Phi-\lambda\rho)^{2} d{\bf r}\nonumber\\
-\int \frac{1}{\chi g(\rho)}(Dh(\rho)\nabla\rho+\chi g(\rho)\nabla \Phi)^{2}d{\bf r}.\end{aligned}$$ Therefore, $\dot F\le 0$. This forms an $H$ theorem in the canonical ensemble. It is sometimes useful to introduce the Massieu function $$\begin{aligned}
\label{h10}
J=S-\beta E,\end{aligned}$$ which is related to the free energy by $J=-\beta F$. Clearly, we have $\dot J\ge 0$. We can now consider particular cases: if $D=0$ (leading to $T=0$), we get $F=E$ so that $\dot E\le 0$. If $\chi=0$ (leading to $\beta=0$), we have $J=S$ so that $\dot S\ge 0$.
Stationary solution {#sec_sta}
-------------------
The steady state of Eq. (\[dde3\]) satisfies $\dot F=0$ leading to ${\bf J}={\bf 0}$ or explicitly $$\begin{aligned}
\label{sta1}
Dh(\rho)\nabla\rho+\chi g(\rho)\nabla \Phi={\bf 0}.\end{aligned}$$ Using Eqs. (\[h4\]) and (\[h6\]), we get $$\begin{aligned}
\label{sta2}
C''(\rho)\nabla\rho+\beta\nabla \Phi={\bf 0},\end{aligned}$$ which can be integrated into $$\begin{aligned}
\label{sta3}
C'(\rho)=-\beta \Phi-\alpha,\end{aligned}$$ where $\alpha$ is a constant of integration. Since $C$ is convex, this equation can be reversed to give $$\begin{aligned}
\label{sta4}
\rho({\bf r})=F(\beta \Phi({\bf r})+\alpha),\end{aligned}$$ where $F(x)=(C')^{-1}(-x)$ is a monotonically decreasing function. Thus, in the steady state, the density is a monotonically decreasing function $\rho=\rho(\Phi)$ of the potential. We have the identity $$\begin{aligned}
\label{sta5}
\rho'(\Phi)=-\frac{\beta}{C''(\rho)}.\end{aligned}$$ Substituting Eq. (\[gle2\]) in Eq. (\[sta3\]), we find that the density profile is determined by an integro-differential equation of the form $$\begin{aligned}
\label{sta6}
C'(\rho)=-\beta\int \rho({\bf r}')u(|{\bf r}-{\bf
r}'|)d{\bf r}' -\alpha.\end{aligned}$$ More specifically, when the potential is given by Eq. (\[gle4\]), we obtain a mean field equation of the form $$\begin{aligned}
\label{sta7}
\Delta \Phi-k^{2}\Phi=\lambda F(\beta \Phi+\alpha).\end{aligned}$$ The constant of integration $\alpha$ is determined by the total mass $M$ (which is a conserved quantity). Finally, we note that the generalized entropy (\[h5\]) is related to the distribution (\[sta4\]) by [@super]: $$\begin{aligned}
\label{sta7b}
C(\rho)=-\int^{\rho}F^{-1}(x)dx.\end{aligned}$$ Equation (\[sta4\]) determines the distribution from the entropy and Eq. (\[sta7b\]) determines the entropy from the distribution.
Minimum of free energy {#sec_min}
----------------------
The critical points of free energy at fixed mass are determined by the variational problem $$\begin{aligned}
\label{min1}
\delta F+T\alpha\delta M=0,\end{aligned}$$ where $\alpha$ is a Lagrange multiplier. We first consider the case where $\Phi$ is given by Eq. (\[gle2\]). Therefore, the free energy that we consider is $$\begin{aligned}
\label{min6} F[\rho]={1\over 2}\int \rho\Phi d{\bf r}+T\int C(\rho) d{\bf r}.\end{aligned}$$ After straightforward calculations, we find that Eq. (\[min1\]) leads to $$\begin{aligned}
\label{min2}
C'(\rho)=-\beta \Phi-\alpha.\end{aligned}$$ Therefore, comparing with Eq. (\[sta3\]), we find that a stationary solution of Eq. (\[dde3\]) is a critical point of $F$ at fixed mass. Furthermore, it is shown in Appendix \[sec\_conn\] that a steady state of Eq. (\[dde3\]) is linearly dynamically stable iff it is a [*minimum*]{} (at least local) of $F$ at fixed mass. This property also results from Lyapunov’s direct method [@frank]. Indeed, we have established that $$\begin{aligned}
\label{min2add}
\dot F\le 0, \qquad \dot F=0 \leftrightarrow \partial_{t}\rho=0.\end{aligned}$$ This implies that $\rho({\bf r})$ is linearly dynamically stable iff it is a (local) minimum of $F$ at fixed mass (maxima or saddle points of $F$ are dynamically unstable). In this sense, dynamical and generalized thermodynamical stability in the canonical ensemble coincide. Furthermore, if $F$ is bounded from below [^5], we can conclude from Lyapunov’s theory that the system will converge to a stable steady state for $t\rightarrow +\infty$ which is a (local) minimum of $F[\rho]$ at fixed mass. If several local minima exist, the choice of the final steady state will depend on a complicated notion of [*basin of attraction*]{}.
In conclusion, a steady solution of the nonlinear mean field Fokker-Planck equation (\[dde3\]) is stable iff it satisfies the minimization problem: $$\begin{aligned}
\label{min3}
\min_{\rho}\quad \lbrace F[\rho]\quad |\quad M[\rho]=M\rbrace.\end{aligned}$$ Taking the second variations of $F$ and using Eq. (\[sta5\]), the condition of stability can be written $$\begin{aligned}
\label{min4} \delta^{2}F[\delta\rho]=-{1\over 2}\biggl\lbrace \int {(\delta
\rho)^{2}\over \rho'(\Phi)}d{\bf r}-\int \delta\rho\delta\Phi
d{\bf r}\biggr\rbrace \ge 0,\end{aligned}$$ for all perturbations $\delta\rho$ that conserve mass. If $\Phi_{ext}({\bf r})$ is an external potential, the second term in Eq. (\[min4\]) vanishes. Therefore, the second variations of the free energy are always positive $\delta^{2}F=-T\delta^{2}S=(1/2)T\int
C''(\rho){(\delta \rho)^{2}}\ge 0$ so that a critical point of $F$ at fixed mass is necessarily a minimum. If the potential is given by Eq. (\[gle3\]), the free energy is a functional $F[\rho,\Phi]$ of the density $\rho$ and potential $\Phi$. The cancellation of the first order variations of $F$ with respect to $\delta\rho$ and $\delta\Phi$ yields Eqs. (\[min2\]) and (\[gle4\]), respectively. The condition of stability can be written $$\begin{aligned}
\label{min5} \delta^{2}F[\delta\rho,\delta\Phi]=-{1\over 2}\int {(\delta
\rho)^{2}\over \rho'(\Phi)}d{\bf r}+\int \delta\rho\delta\Phi
d{\bf r}\nonumber\\
-\frac{1}{2\lambda}\int (\Delta\delta\Phi-k^{2}\delta\Phi)\delta\Phi d{\bf r} \ge 0,\end{aligned}$$ for all perturbations $\delta\rho$ and $\delta\Phi$ that conserve mass. From now on, we shall only consider the case where the potential is given by Eq. (\[gle2\]) since the case of Eq. (\[gle3\]) can be treated similarly by following the lines sketched above.
Particular cases {#sec_pc}
----------------
If we take $h(\rho)=1$ and $g(\rho)=1/C''(\rho)$, we get $$\begin{aligned}
\label{pc1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\rho+
\frac{\chi}{C''(\rho)} \nabla \Phi\right ).\end{aligned}$$ In that case, we have a constant diffusion $D(\rho)=D$ and a variable mobility $\chi(\rho)=\chi/(\rho C''(\rho))$. If we take $g(\rho)=\rho$ and $h(\rho)=\rho C''(\rho)$, we get $$\begin{aligned}
\label{pc2}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D\rho C''(\rho) \nabla\rho+{\chi}\rho \nabla \Phi\right ).\end{aligned}$$ In that case, we have a constant mobility $\chi(\rho)=\chi$ and a variable diffusion $D(\rho)=D\rho \lbrack C(\rho)/\rho\rbrack'$. Note that the condition $D(\rho)\ge 0$ requires that $\lbrack
C(\rho)/\rho\rbrack'\ge 0$. This gives a constraint on the possible forms of $C(\rho)$.
Generalized Smoluchowski equation {#sec_gs}
---------------------------------
The NFP equation (\[pc2\]) can be put in the form of a generalized Smoluchowski (GS) equation $$\begin{aligned}
\label{gs1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left \lbrack \chi (\nabla p+
\rho \nabla \Phi )\right \rbrack,\end{aligned}$$ with a barotropic equation of state $p(\rho)$ given by $$\begin{aligned}
\label{gs2}
p'(\rho)=T\rho C''(\rho).\end{aligned}$$ Since $C$ is convex, we have $p'(\rho)\ge 0$. On the other hand, integrating Eq. (\[gs2\]) twice, we get $$\begin{aligned}
\label{gs3}
TC(\rho)=\rho \int^{\rho}\frac{p(\rho')}{\rho^{'2}}d\rho'.\end{aligned}$$ Therefore, the free energy (\[min6\]) can be rewritten $$\begin{aligned}
\label{gs4}
F[\rho]=\frac{1}{2}\int\rho\Phi d{\bf r}+\int \rho \int^{\rho}\frac{p(\rho')}{\rho^{'2}}d\rho'd{\bf r}.\end{aligned}$$ With these notations, the $H$-theorem (\[h8\]) becomes $$\begin{aligned}
\label{gs5}
\dot F=-\int \frac{\chi}{\rho}(\nabla p+\rho\nabla \Phi)^{2}d{\bf r}\le 0.\end{aligned}$$ The stationary solutions of the GS equation (\[gs1\]) satisfy the relation $$\begin{aligned}
\label{gs6}
\nabla p+\rho\nabla \Phi={\bf 0},\end{aligned}$$ which is similar to a condition of hydrostatic equilibrium. Since $p=p(\rho)$, this relation can be integrated to give $\rho=\rho(\Phi)$ through $$\begin{aligned}
\label{gs7}
\int^{\rho}\frac{p'(\rho')}{\rho'}d\rho'=-\Phi.\end{aligned}$$ This is equivalent to $$\begin{aligned}
\label{gs8}
\frac{p'(\rho)}{\rho}=-\frac{1}{\rho'(\Phi)}.\end{aligned}$$ This relation can also be obtained from Eqs. (\[gs2\]) and (\[sta5\]). Therefore, we recover the fact that, in the steady state, $\rho=\rho(\Phi)$ is a monotonically decreasing function of $\Phi$. We also note the identity $$\begin{aligned}
\label{gs9}
p(\rho)=\frac{1}{\chi}D(\rho)\rho=T\rho^{2} \left \lbrack \frac{C(\rho)}{\rho} \right\rbrack'=T\lbrack C'(\rho)\rho-C(\rho)\rbrack.\nonumber\\ \end{aligned}$$
Finally, we note that the relation (\[gs7\]), equivalent to the condition of hydrostatic equilibrium (\[gs6\]), can also be obtained by extremizing the free energy (\[gs4\]) at fixed mass writing $\delta F-\alpha\delta M=0$. More precisely, we have the important result: [*a steady solution of the generalized Smoluchowski equation (\[gs1\]) is linearly dynamically stable iff it is a (local) minimum of the free energy $F[\rho]$ at fixed mass $M[\rho]=M$.*]{} This corresponds to the minimization problem (\[min3\]) with Eq. (\[gs4\]). Note that the GS equation has been introduced here in a very general context. In this sense, it is valid for arbitrary value of $\chi$ (not necessarily small). Alternatively, we shall see in Sec. \[sec\_strong\] that a generalized Smoluchowski equation with $\chi=1/\xi\rightarrow 0$ can be derived from a generalized Kramers equation in the strong friction limit $\xi\rightarrow +\infty$.
Phenomenological derivation of the nonlinear Fokker-Planck equation {#sec_phen}
-------------------------------------------------------------------
The form of the diffusion current appearing in the NFP equation (\[dde3\]) can be obtained by different phenomenological procedures.
### Functional derivative {#sec_func}
For a given free energy functional $F[\rho]$, we can introduce phenomenologically a dynamical model by writing the evolution of the density as a continuity equation $\partial_{t}\rho=-\nabla \cdot {\bf
J}$ where the current is proportional to the gradient of the functional derivative of the free energy [@frank], i.e. $$\begin{aligned}
\label{func1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left \lbrack \chi({\bf r},t)\rho \nabla\frac{\delta F}{\delta\rho}\right \rbrack.\end{aligned}$$ For the free energy (\[min6\]), we have $$\begin{aligned}
\label{func2}
\frac{\delta F}{\delta\rho}=TC'(\rho)+\Phi,\end{aligned}$$ so that $$\begin{aligned}
\label{func3}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left \lbrack \chi({\bf r},t)\left (T\rho C''(\rho) \nabla\rho+\rho \nabla \Phi\right )\right\rbrack.\end{aligned}$$ This equation is more general than Eq. (\[dde3\]). It shows that, for a given free energy, we can introduce an infinite class of NFP equations where $\chi({\bf r},t)$ is an [*arbitrary*]{} positive function of position and time. In particular, $\chi({\bf r},t)$ can be a positive function of the density $\rho({\bf r},t)$. If we set $\chi({\bf r},t)=\chi g(\rho)/\rho$ we recover Eq. (\[dde3\]) as a particular case. We can also write Eq. (\[func3\]) in the alternative form $$\begin{aligned}
\label{func4}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left \lbrack \tilde{\chi}({\bf r},t)\left (T\nabla\rho+\frac{1}{C''(\rho)} \nabla \Phi\right )\right\rbrack,\end{aligned}$$ where $\tilde{\chi}({\bf r},t)$ is an [*arbitrary*]{} positive function of position and time. If we set $\tilde{\chi}({\bf
r},t)=\chi h(\rho)$ we recover Eq. (\[dde3\]). These two alternative forms (\[func3\]) and (\[func4\]) were given in [@gen].
On the other hand, the formulation (\[func1\]) ensures that the free energy decreases monotonically provided that $\chi({\bf r},t)$ is positive. Indeed, $$\begin{aligned}
\label{func5}
\dot F=\int \frac{\delta F}{\delta\rho}\frac{\partial\rho}{\partial t}\, d{\bf r}=-\int \frac{\delta F}{\delta\rho} \nabla\cdot {\bf J}\, d{\bf r}\nonumber\\
=\int {\bf J}\cdot \nabla \frac{\delta F}{\delta\rho}\, d{\bf r}=-\int \chi({\bf r},t)\rho\left (\nabla \frac{\delta F}{\delta\rho}\right )^{2}\, d{\bf r}\le 0.\end{aligned}$$ Furthermore, a steady state of Eq. (\[func1\]) satisfies $\dot F=0$, i.e $\nabla (\delta F/\delta\rho)=0$ leading to $$\begin{aligned}
\label{func6}
\frac{\delta F}{\delta\rho}=-T\alpha,\end{aligned}$$ where $\alpha$ is a constant of integration. This is equivalent to the extremization of $F$ at fixed mass $M$ if we write the first variations as $\delta F+T\alpha\delta M=0$. Therefore, a steady state extremizes the free energy at fixed mass. Using Eq. (\[func2\]), we find that the equilibrium density profile satisfies $$\begin{aligned}
\label{func7}
C'(\rho)=-\beta \Phi-\alpha.\end{aligned}$$ Finally, using Lyapunov’s direct method, one can show that a steady state $\rho({\bf r})$ of Eq. (\[func1\]) is linearly dynamically stable iff it is a (local) minimum of $F$ at fixed mass. We again emphasize that the stationary states of Eq. (\[func3\]), and the $H$-theorem, only depend on the form of the free energy $F[\rho]$ and are independent on the positive function $\chi({\bf r},t)$. These properties are therefore valid for the whole class of NFP equations associated with a given free energy functional. Therefore, for a given free energy specified by $C(\rho)$, we can construct an infinite class of NFP equations with arbitrary $\chi({\bf r},t)$ that have the same equilibrium states (\[func7\]) but a different dynamics.
### Onsager’s linear thermodynamics {#sec_onsager}
The previous approach is equivalent to Onsager’s linear thermodynamics. Indeed, at equilibrium, we expect that the distribution $\rho({\bf r})$ minimizes the free energy $F$ at fixed mass. This leads to Eq. (\[func6\]) or (\[func7\]). Noting that the chemical potential $$\begin{aligned}
\label{onsager1}
\lambda({\bf r},t)\equiv \frac{\delta F}{\delta\rho}=TC'(\rho)+\Phi,\end{aligned}$$ is uniform at equilibrium, the linear thermodynamics of Onsager suggests that, close to equilibrium, the current is proportional to the gradient of the chemical potential. If we write $$\begin{aligned}
\label{onsager2}
{\bf J}=-\chi({\bf r},t)\rho \nabla\lambda({\bf r},t),\end{aligned}$$ the linear thermodynamics of Onsager leads to Eq. (\[func1\]).
### Maximum Free Energy Dissipation Principle {#sec_mepp}
The same results can be obtained from a variational formulation, called the Maximum Free Energy Dissipation (MFED) principle which is the canonical ensemble version of the Maximum Entropy Production (MEP) principle [@gen]. At equilibrium, the optimal distribution $\rho({\bf r})$ minimizes the free energy $F[\rho]$ at fixed mass $M$. Out of equilibrium, we may expect that the optimal current ${\bf
J}$ maximizes the rate of free energy dissipation $\dot F[{\bf J}]$ under some constraints. This can be viewed as a variational formulation of Onsager’s linear thermodynamics. The rate of dissipation of free energy is given by $$\begin{aligned}
\label{mepp1}
\dot F=\int \frac{\delta F}{\delta\rho}\frac{\partial\rho}{\partial t}\, d{\bf r}=-\int \frac{\delta F}{\delta\rho} \nabla\cdot {\bf J}\, d{\bf r}
=\int {\bf J}\cdot \nabla \frac{\delta F}{\delta\rho}\, d{\bf r}.\nonumber\\\end{aligned}$$ We shall determine the optimal current ${\bf J}_{*}$ which maximizes the rate of dissipation of free energy $\dot F$ under the constraint $J^{2}\le C({\bf r},t)$ putting a physical bound on $|{\bf J}|$. It can be shown that the bound is always reached so that we can replace the inequality by an equality. Thus, we write the variational problem as $$\begin{aligned}
\label{mepp2}
\delta \dot F+\delta \left (\int \frac{{\bf J}^{2}}{2\rho \chi({\bf r},t)}\, d{\bf r}\right )=0,\end{aligned}$$ where $\chi({\bf r},t)$ is a local Lagrange multiplier. Performing the variations on ${\bf J}$, we obtain $$\begin{aligned}
\label{mepp3}
{\bf J}_{*}=-\chi({\bf r},t)\rho \nabla \frac{\delta F}{\delta\rho},\end{aligned}$$ which returns Eq. (\[func1\]). Note that if we introduce the “dissipation” function $$\begin{aligned}
\label{mepp4}
E_{d}\equiv \int \frac{{\bf J}^{2}}{2\rho \chi({\bf r},t)}\, d{\bf r},\end{aligned}$$ we have $$\begin{aligned}
\label{mepp5}
E_{d}[{\bf J}_{*}]=-\frac{1}{2}\dot F[{\bf J}_{*}]. \end{aligned}$$ On the other hand, since $\delta^{2}(\dot F+E_{d})=-\int \frac{(\delta
J)^{2}}{2\rho\chi}d{\bf r}\le 0$, the optimal current (\[mepp3\]) [*maximizes*]{} the dissipation of free energy under the constraint $J^{2}\le C({\bf r},t)$.
Kinetic derivation of the nonlinear Fokker-Planck equation {#sec_kin}
----------------------------------------------------------
Nonlinear Fokker-Planck equation where the diffusion coefficient and the mobility explicitly depend on the local concentration of particles can be derived from a kinetic theory, starting from the master equation, and assuming that the probabilities of transition explicitly depend on the occupation numbers (concentrations) of the initial and arrival states. We briefly summarize the approach developed by Kaniadakis [@k1] and make the link with the phenomenological equations studied previously.
We introduce a stochastic dynamics by defining the probability of transition of a particle from position ${\bf r}$ to position ${\bf
r}'$. Following Kaniadakis [@k1], we assume the following form $$\begin{aligned}
\label{kin1}
\pi({\bf r}\rightarrow {\bf r}')=w({\bf r},{\bf r}-{\bf r}')a\lbrack\rho({\bf r},t)\rbrack b\lbrack\rho({\bf r}',t)\rbrack.\end{aligned}$$ Usual stochastic processes correspond to $a(\rho)=\rho$ and $b(\rho)=1$: the probability of transition is proportional to the density of the initial state and independent on the density of the final state. They lead to the ordinary Fokker-Planck equation (\[sm1\]) as will be shown below. Here, we assume a more general dependence on the occupancy in the initial and arrival states. This can account for microscopic constraints like close-packing effects that can inhibitate the transition. Quite generally, the evolution of the density satisfies the master equation $$\begin{aligned}
\label{kin2}
\frac{\partial\rho}{\partial t}=\int \left\lbrack \pi({\bf r}'\rightarrow {\bf r})-\pi({\bf r}\rightarrow {\bf r}')\right\rbrack d{\bf r}'.\end{aligned}$$ Assuming that the evolution is sufficiently slow, and local, such that the dynamics only permits values of ${\bf r}'$ close to ${\bf r}$, one can develop the term in brackets in Eq. (\[kin2\]) in powers of ${\bf r}-{\bf r}'$. Proceeding along the lines of [@k1], we obtain a Fokker-Planck-like equation $$\label{kin3}
\frac{\partial\rho}{\partial t}=\frac{\partial}{\partial x_{i}}\left\lbrack\left (\zeta_{i}+\frac{\partial\zeta_{ij}}{\partial x_{j}}\right )\gamma(\rho)+\gamma(\rho)\frac{\partial\ln \kappa(\rho)}{\partial\rho}\zeta_{ij}\frac{\partial\rho}{\partial x_{j}}\right\rbrack,$$ with $$\label{kin4}
\gamma(\rho)=a(\rho)b(\rho),\qquad \kappa(\rho)=\frac{a(\rho)}{b(\rho)},$$ and $$\label{kin5}
\zeta_{i}({\bf r})=-\int y_{i}w({\bf r},{\bf y})d{\bf y},$$ $$\label{kin6}
\zeta_{ij}({\bf r})=\frac{1}{2}\int y_{i}y_{j}w({\bf r},{\bf y})d{\bf y}.$$ The moments $\zeta_{i}$ and $\zeta_{ij}$ are fixed by the ordinary Langevin equation $$\begin{aligned}
\label{kin7} \frac{d{\bf
r}}{dt}=-\chi\nabla\Phi+\sqrt{2D}{\bf R}(t),\end{aligned}$$ where $\chi$ and $D$ are constant. Assuming isotropy $\zeta_{i}=J_{i}$, $\zeta_{ij}=D\delta_{ij}$, the kinetic equation (\[kin3\]) becomes $$\label{kin8}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left\lbrack ({\bf J}+\nabla D)\gamma(\rho)+\gamma(\rho)\frac{\partial\ln \kappa(\rho)}{\partial\rho}D\nabla \rho\right\rbrack.$$ Now, according to the Langevin equation (\[kin7\]), $D$ is independent on ${\bf r}$ and ${\bf J}=\chi\nabla \Phi$. Thus, we get $$\label{kin9}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left\lbrack D\gamma(\rho)\frac{\partial\ln \kappa(\rho)}{\partial\rho}\nabla\rho+\chi\gamma(\rho)\nabla \Phi \right\rbrack.$$ If we define $$\label{kin10}
h(\rho)=\gamma(\rho)\frac{\partial\ln \kappa(\rho)}{\partial\rho}, \qquad g(\rho)=\gamma(\rho),$$ the foregoing equation can be written $$\begin{aligned}
\label{kin11a}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left \lbrack Dh(\rho)\nabla\rho+\chi g(\rho)\nabla \Phi\right \rbrack,\end{aligned}$$ and it coincides [^6] with the phenomenological equation (\[dde3\]). We note that $$\begin{aligned}
\label{kin11b}
\ln\kappa(\rho)=C'(\rho).\end{aligned}$$ We also have the relations $$\begin{aligned}
\label{kin12}
a(\rho)=\sqrt{\gamma(\rho)\kappa(\rho)}=\sqrt{g(\rho)}e^{C'(\rho)/2},\end{aligned}$$ $$\begin{aligned}
\label{kin13}
b(\rho)=\sqrt{\frac{\gamma(\rho)}{\kappa(\rho)}}=\sqrt{g(\rho)}e^{-C'(\rho)/2}.\end{aligned}$$ Inversely $$\begin{aligned}
\label{kin14}
g(\rho)=a(\rho)b(\rho), \qquad C'(\rho)=\ln\left\lbrack \frac{a(\rho)}{b(\rho)}\right\rbrack,\end{aligned}$$ $$\begin{aligned}
\label{kin15}
h(\rho)=b(\rho)a'(\rho)-a(\rho)b'(\rho).\end{aligned}$$
It seems natural to assume that the transition probability is proportional to the density of the initial state so that $a(\rho)=\rho$. In that case, we obtain an equation of the form $$\label{kin16}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D\left\lbrack b(\rho)-\rho b'(\rho)\right \rbrack \nabla\rho+\chi\rho b(\rho)\nabla \Phi\right ).$$ Note that the coefficients of diffusion and mobility are not independent since they are both expressed in terms of $b(\rho)$. Choosing $b(\rho)=1$, i.e. a probability of transition which does not depend on the population of the arrival state, leads to the standard Fokker-Planck equation (\[sm1\]). If, now, we assume that the transition probability is blocked (inhibited) if the concentration of the arrival state is equal to $\sigma_0$, then it seems natural to take $b(\rho)=1-\rho/\sigma_{0}$. In that case, we obtain $$\label{kin17}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D\nabla\rho+\chi\rho (1-\rho/\sigma_{0})\nabla \Phi\right ),$$ which will be considered in Sec. \[sec\_ff\]. Inversely, we can wonder what the general form of the mobility will be if we assume a normal diffusion $h(\rho)=1$. This leads to $b(\rho)-\rho b'(\rho)=1$ which is integrated in $b(\rho)=1+K\rho$ where $K$ is a constant. Interestingly, we find that this condition selects the class of fermions ($K=-1$) and bosons ($K=+1$) and intermediate statistics (arbitrary $K$). The corresponding NFP equation is $$\label{kin18}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D\nabla\rho+\chi\rho (1+K\rho)\nabla \Phi\right ),$$ which will be considered in Sec. \[sec\_fbi\].
Examples of nonlinear Smoluchowski equations and generalized Keller-Segel models {#sec_ex}
================================================================================
In this section, we give several explicit examples of nonlinear mean field Fokker-Planck equations. Some correspond to well-known forms of entropies, and others are new. We emphasize that these equations can have applications in different domains of physics as discussed in Sec. \[sec\_known\]. Importantly, they can provide generalized Keller-Segel models of chemotaxis (see Sec. \[sec\_ks\]). Most of these models have not been considered before in biology because the connection with generalized thermodynamics was not made. This is why we give a relatively detailed description of these models since their applications in biology are new.
Standard model: Boltzmann entropy {#sec_sm}
---------------------------------
If we take $h(\rho)=1$ and $g(\rho)=\rho$, we get the ordinary Smoluchowski equation $$\begin{aligned}
\label{sm1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\rho+
\chi \rho \nabla \Phi\right ).\end{aligned}$$ It corresponds to an ordinary diffusion $D(\rho)=D$ and a constant mobility $\chi(\rho)=\chi$. The associated stochastic process is $$\begin{aligned}
\label{sm2}
\frac{d{\bf r}}{dt}=-\chi\nabla \Phi+\sqrt{2D}{\bf R}(t).\end{aligned}$$ The entropy is the Boltzmann entropy $$\begin{aligned}
\label{sm3}
S=-\int \rho\ln\rho\, d{\bf r},\end{aligned}$$ and the stationary solution of Eq. (\[sm1\]) is the Boltzmann distribution $$\begin{aligned}
\label{sm4}
\rho=e^{-\beta \Phi-\alpha-1}.\end{aligned}$$ The pressure law is $$\begin{aligned}
\label{sm5}
p(\rho)=\rho T.\end{aligned}$$ This is similar to the equation of state for an isothermal gas with constant temperature $T$. When the Fokker-Planck equation (\[sm1\]) is coupled to the Poisson equation (\[gle5\]), we obtain the Smoluchowski-Poisson system describing a gas of self-gravitating Brownian particles [@crs; @sc; @post; @tcoll; @virial1; @virial2]. When the Fokker-Planck equation (\[sm1\]) is coupled to the field Eq. (\[gle3\]), we obtain the standard Keller-Segel model describing the chemotactic aggregation of biological populations [@horstmann].
Power law diffusion: Tsallis entropy {#sec_pld}
------------------------------------
If we take $h(\rho)=q\rho^{q-1}$ and $g(\rho)=\rho$, we obtain the NFP equation $$\begin{aligned}
\label{pld1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\rho^{q}+
\chi \rho \nabla \Phi\right ).\end{aligned}$$ It corresponds to a power law diffusion $D(\rho)=D\rho^{q-1}$ and a constant mobility $\chi(\rho)=\chi$. This equation was introduced by Plastino & Plastino [@pp]. The associated stochastic process, introduced by Borland [@borland], is $$\begin{aligned}
\label{pld2}
\frac{d{\bf r}}{dt}=-\chi\nabla \Phi+\sqrt{2D}\rho^{\frac{q-1}{2}}{\bf R}(t).\end{aligned}$$ This model can take into account effects of non-ergodicity and nonextensivity. It leads to a situation of anomalous diffusion related to the Tsallis statistics. For $q=1$, we recover the standard Brownian model with a constant diffusion coefficient, corresponding to a pure random walk. In that case, the sizes of the random kicks are uniform and do not depend on where the particle happens to be. For $q\neq 1$, the size of the random kicks changes, depending on the distribution of the particles around the “test” particle. A particle which is in a region that is highly populated \[large $\rho({\bf r},t)$\] will tend to have larger kicks if $q>1$ and smaller kicks if $q<1$. Since the microscopics depends on the actual density in space, this creates a bias in the ergodic behavior of the system. Then, the dynamics has a fractal or multi-fractal phase space structure [@borland]. The generalized entropy associated to Eq. (\[pld1\]) is the Tsallis entropy $$\begin{aligned}
\label{pld3}
S=-\frac{1}{q-1}\int (\rho^{q}-\rho)\, d{\bf r},\end{aligned}$$ and the stationary solution is the Tsallis distribution $$\begin{aligned}
\label{pld4}
\rho=\left (\frac{1}{q}\right )^{\frac{1}{q-1}}\left\lbrack 1-(q-1)(\beta\Phi+\alpha)\right\rbrack_{+}^{1/(q-1)}.\end{aligned}$$ The pressure law is $$\begin{aligned}
\label{pld5}
p(\rho)=T\rho^{q}.\end{aligned}$$ This is similar to a polytropic gas with an equation of state $p=K\rho^{\gamma}$ where $K=T$ plays the role of a polytropic temperature and $q=\gamma$ is the polytropic index (we also set $\gamma=1+1/n$). Note that the Tsallis entropy can be written $$\begin{aligned}
\label{pld6}
S=-\int \rho \ln_{(q)}\rho \, d{\bf r},\end{aligned}$$ where we have introduced the $q$-logarithm $$\begin{aligned}
\label{pld7}
\ln_{(q)}(x)=\frac{1}{q-1}(x^{q-1}-1).\end{aligned}$$ The stationary solution can be written $$\begin{aligned}
\label{pld8}
\rho=\left (\frac{1}{q}\right )^{\frac{1}{q-1}}e_{(q)}^{-\beta\Phi-\alpha}\end{aligned}$$ with the $q$-exponential $$\begin{aligned}
\label{pld9}
e_{(q)}(x)=\lbrack 1+(q-1)x\rbrack^{\frac{1}{q-1}}.\end{aligned}$$ For $q=1$, we recover the standard model (\[sm1\]). For $q=2$, we have some simplifications. In that case, the NFP equation (\[pld1\]) becomes $$\begin{aligned}
\label{pld10}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\rho^{2}+\chi \rho \nabla \Phi\right ).\end{aligned}$$ The entropy is the quadratic functional $$\begin{aligned}
\label{pld11}
S=-\int \rho^{2} d{\bf r},\end{aligned}$$ and the stationary solution is $$\begin{aligned}
\label{pld12}
\rho=-\frac{1}{2}(\beta \Phi+\alpha),\end{aligned}$$ corresponding to a linear relation between the density and the potential. In that case, the differential equation (\[sta7\]) determining the steady state reduces to the Helmholtz equation. Finally, the pressure is $$\begin{aligned}
\label{pld13}
p(\rho)=T\rho^{2},\end{aligned}$$ corresponding to a polytrope with index $n=1$. When the NFP equation (\[pld1\]) is coupled to the Poisson equation (\[gle5\]), we obtain the polytropic Smoluchowski-Poisson system describing self-gravitating Langevin particles. When the NFP equation (\[pld1\]) is coupled to the field Eq. (\[gle3\]), we obtain a generalized Keller-Segel model of chemotaxis taking into account anomalous diffusion. These models have been introduced by Chavanis & Sire [@lang; @csbio; @csmass].
Logotropic distributions: log-entropy {#sec_log}
-------------------------------------
If we take $h(\rho)=1/\rho$ and $g(\rho)=\rho$, we obtain $$\begin{aligned}
\label{log1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\ln\rho+
\chi \rho \nabla \Phi\right ).\end{aligned}$$ The generalized entropy associated to Eq. (\[log1\]) is the log-entropy $$\begin{aligned}
\label{log2}
S=\int \ln\rho \, d{\bf r},\end{aligned}$$ and the stationary solution is $$\begin{aligned}
\label{log3}
\rho=\frac{1}{\alpha+\beta \Phi}. \end{aligned}$$ For a quadratic potential $\Phi_{ext}=r^{2}/2$, this corresponds to the Lorentzian function. The pressure is $$\begin{aligned}
\label{log4}
p(\rho)=T\ln\rho.\end{aligned}$$ This is similar to a logotropic equation of state [@pudritz]. This is also connected to a polytropic equation of state (or Tsallis distribution) with $\gamma=q=0$. Indeed, the logotropic model (\[log1\]) can be deduced from Eq. (\[pld1\]) by writing $D\nabla\rho^{q}=Dq\rho^{q-1}\nabla\rho$, taking $q=0$ and re-defining $Dq\rightarrow D$. When the NFP equation (\[log1\]) is coupled to the Poisson equation (\[gle5\]), we obtain the logotropic Smoluchowski-Poisson system. When the NFP equation (\[log1\]) is coupled to the field Eq. (\[gle3\]), we obtain a generalized Keller-Segel model of chemotaxis. These models have been introduced by Chavanis & Sire [@logotrope].
Power law diffusion and drift: Tsallis entropy {#sec_pldd}
----------------------------------------------
We introduce here a new model generalizing the polytropic model (\[pld1\]). If we take $h(\rho)=q\rho^{q+\mu-1}$ and $g(\rho)=\rho^{\mu+1}$, we obtain $$\begin{aligned}
\label{pldd1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( Dq\rho^{q+\mu-1}
\nabla\rho+
\chi \rho^{\mu+1} \nabla \Phi\right ).\end{aligned}$$ This corresponds to a power law diffusion $D(\rho)=\frac{Dq}{q+\mu}\rho^{q+\mu-1}$ and a power law mobility $\chi(\rho)=\chi \rho^{\mu}$. The associated stochastic process is $$\begin{aligned}
\label{pldd2}
\frac{d{\bf r}}{dt}=-\chi \rho^{\mu}\nabla\Phi+\sqrt{\frac{2Dq}{q+\mu}}\rho^{\frac{q+\mu-1}{2}}{\bf R}(t).\end{aligned}$$ Since $\rho^{\mu}$ can be put in factor of the diffusion current in Eq. (\[pldd1\]), this model belongs to the infinite family of NFP equations associated to the Tsallis entropy with index $q$ (see discussion in Sec. \[sec\_func\]).
For $\mu=0$, we recover Eq. (\[pld1\]) with a constant mobility and a power law diffusion. For $(\mu,q)=(0,0)$, we recover the logotropic Smoluchowski equation (\[log1\]) provided that we make the transformation $Dq\rightarrow D$. For $\mu=1-q$, we have a normal diffusion and a power law mobility $$\begin{aligned}
\label{pldd3}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( Dq
\nabla\rho+
\chi \rho^{2-q} \nabla \Phi\right ).\end{aligned}$$ For $q=2$, we get $$\begin{aligned}
\label{pldd4}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left (2 D
\nabla\rho+
\chi \nabla \Phi\right ),\end{aligned}$$ which has the same equilibrium states as Eq. (\[pld10\]). If we assume furthermore that $\Phi$ is given by the Poisson equation (\[gle5\]), Eq. (\[pldd4\]) reduces to the linear equation $$\begin{aligned}
\label{pldd5}
\frac{\partial\rho}{\partial t}=2 D
\Delta\rho+
\chi \lambda \rho.\end{aligned}$$ Finally, for $q=0$ (making the transformation $qD\rightarrow D$), we obtain $$\begin{aligned}
\label{pldd6}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D
\nabla\rho+
\chi \rho^{2} \nabla \Phi\right ),\end{aligned}$$ which has the same equilibrium states as Eq. (\[log1\]). When the NFP equation (\[pldd1\]) is coupled to the field equation (\[gle3\]), we obtain a generalized Keller-Segel model of chemotaxis taking into account anomalous diffusion and anomalous mobility. This model will be studied in a forthcoming paper, in continuity with [@lang].
Filling factor: Fermi-Dirac entropy {#sec_ff}
-----------------------------------
If we take $h(\rho)=1$ and $g(\rho)=\rho(1-\rho/\sigma_{0})$, we obtain $$\begin{aligned}
\label{ff1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\rho+
\chi \rho(1-\rho/\sigma_{0}) \nabla \Phi\right ).\end{aligned}$$ This corresponds to a normal diffusion $D(\rho)=D$ and a mobility $\chi(\rho)=\chi(1-\rho/\sigma_{0})$ vanishing linearly when the density reaches the maximum value $\rho_{max}=\sigma_{0}$. The associated stochastic process is $$\begin{aligned}
\label{ff2}
\frac{d{\bf r}}{dt}=-\chi(1-\rho/\sigma_{0})\nabla \Phi+\sqrt{2D}{\bf R}(t).\end{aligned}$$ The generalized entropy associated with Eq. (\[ff1\]) is the Fermi-Dirac entropy in position space $$\begin{aligned}
\label{ff3}
S=-\sigma_{0}\int \left\lbrace \frac{\rho}{\sigma_{0}}\ln\frac{\rho}{\sigma_{0}}+\left (1-\frac{\rho}{\sigma_{0}}\right)\ln \left (1-\frac{\rho}{\sigma_{0}}\right)\right\rbrace d{\bf r},\nonumber\\\end{aligned}$$ and the stationary solution is the Fermi-Dirac distribution in position space $$\begin{aligned}
\label{ff4}
\rho=\frac{\sigma_{0}}{1+ e^{\beta \Phi+\alpha}}.\end{aligned}$$ From Eq. (\[ff4\]), we see that, in the stationary state, $\rho\le \sigma_{0}$. This bound is similar to the Pauli exclusion principle in quantum mechanics. In fact, we can show that $\rho({\bf
r},t)$ remains bounded by $\sigma_{0}$ during the whole evolution. For $\sigma_{0}\rightarrow +\infty$, we recover the standard model (\[sm1\]).
An alternative model, with the same entropy and the same equilibrium states, is obtained by taking $h(\rho)=1/(1-\rho/\sigma_{0})$ and $g(\rho)=\rho$. This leads to $$\begin{aligned}
\label{ff5}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left (-D\sigma_{0} \nabla\ln(1-\rho/\sigma_{0})+
\chi \rho \nabla \Phi\right ).\end{aligned}$$ This corresponds to a nonlinear diffusion with $D(\rho)=-
\sigma_{0}(D/\rho)\ln (1-\rho/\sigma_{0})$ and a constant mobility $\chi(\rho)=\chi$. Equation (\[ff5\]) can be put in the form of a generalized Smoluchowski equation (\[gs1\]) with a pressure law $$\begin{aligned}
\label{ff6}
p(\rho)=-T\sigma_{0} \ln(1-\rho/\sigma_{0}).\end{aligned}$$ For $\rho\ll\sigma_{0}$, we recover the “isothermal” equation of state $p=\rho T$ leading to the standard model (\[sm1\]). However, for higher densities, the equation of state is modified and the pressure diverges when $\rho\rightarrow \sigma_{0}$. This prevents the density from exceeding the maximum value $\sigma_{0}$.
In the context of chemotaxis, the model (\[ff1\]) has been introduced by Hillen & Painter [@hp] and, independently, by Chavanis [@gen; @crrs; @degrad]. It provides a regularization of the standard Keller-Segel model preventing overcrowding, blow-up and unphysical singularities. The filling factor $(1-\rho/\sigma_{0})$ takes into account the fact that the particles cannot interpenetrate because of their finite size $a$. Therefore, the maximum allowable density is $\sigma_{0}\sim 1/a^{d}$. It is achieved when all the cells are packed together. In the model (\[ff1\]), it is assumed that the mobility vanishes when the density reaches the close packing value ($\rho\rightarrow
\sigma_{0}$) while the diffusion is not affected. The alternative model (\[ff5\]) has been introduced in Chavanis [@gen; @degrad]. In that case, the mobility is assumed to be constant and the regularization preventing overcrowding is taken into account in the pressure law (\[ff6\]). As explained in Sec. \[sec\_func\], we can multiply the diffusion term and the mobility term by the [*same*]{} positive function $\chi({\bf r},t)$ in order to obtain a more general model with the same entropy and the same equilibrium states. Note finally that an equation similar to Eq. (\[ff1\]) has been introduced by Robert & Sommeria [@rs] (see also [@csr]) in the statistical mechanics of two-dimensional turbulence for two vorticity levels $0$ and $\sigma_{0}$. In that case, $\rho$ represents the coarse-grained vorticity $\overline{\omega}$ and $\Phi$ plays the role of the stream function $\psi$. The “exclusion principle” leading to the Fermi-Dirac entropy (\[ff3\]) is a consequence of the 2D Euler equation implying that the vorticity levels cannot overlap so that $\overline{\omega}({\bf r},t)\le \sigma_{0}$. These analogies between chemotaxis and 2D turbulence are further discussed in [@degrad].
Fermi, Bose and intermediate statistics {#sec_fbi}
---------------------------------------
If we take $h(\rho)=1$ and $g(\rho)=\rho(1+K\rho)$, we obtain $$\begin{aligned}
\label{fbi1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\rho+
\chi \rho(1+K\rho) \nabla \Phi\right ).\end{aligned}$$ This corresponds to a normal diffusion $D(\rho)=D$ and a variable mobility $\chi(\rho)=\chi(1+K\rho)$. The associated stochastic process is $$\begin{aligned}
\label{fbi2}
\frac{d{\bf r}}{dt}=-\chi(1+K\rho)\nabla \Phi+\sqrt{2D}{\bf R}(t).\end{aligned}$$ Here, $K$ is a real number taking positive or negative values. When $K>0$ the mobility is enhanced in regions of large densities and when $K<0$, it is reduced. This takes into account inclusion ($K>0$) or exclusion ($K<0$) principles. For $K=0$, we recover the standard model (\[sm1\]). The generalized entropy associated with Eq. (\[fbi1\]) is $$\begin{aligned}
\label{fbi3}
S=-\int \left\lbrack {\rho}\ln{\rho}-\frac{1}{K}\left (1+K \rho\right)\ln \left (1+K\rho\right)\right\rbrack d{\bf r},\nonumber\\\end{aligned}$$ and the stationary solution is $$\begin{aligned}
\label{fbi4}
\rho=\frac{1}{e^{\beta \Phi+\alpha}-K}.\end{aligned}$$ For $K=+1$ we obtain the Bose-Einstein statistics and for $K=-1$ we obtain the Fermi-Dirac statistics. For other values of $K$, we obtain intermediate statistics (quons) interpolating between fermions and bosons. For $K=0$, we recover the Boltzmann statistics.
An alternative model with the same entropy and the same equilibrium states is obtained by taking $h(\rho)=1/(1+K\rho)$ and $g(\rho)=\rho$. This leads to $$\begin{aligned}
\label{fbi5}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left (\frac{1}{K} \nabla\ln(1+K\rho)+
\chi \rho \nabla \Phi\right ).\end{aligned}$$ This corresponds to a nonlinear diffusion such that $D(\rho)=
(D/K\rho)\ln(1+K\rho) $ and a constant mobility $\chi(\rho)=\chi$. The pressure law is $$\begin{aligned}
\label{fbi6}
p(\rho)=\frac{T}{K} \ln(1+K\rho).\end{aligned}$$ We recall that Eq. (\[fbi1\]) can be obtained from the master equation (\[kin2\]) when the transition probabilities are of the form (\[kin1\]) with $a(\rho)=\rho$ and $b(\rho)=1+K\rho$. Alternatively, the model (\[fbi5\]) corresponds to $a(\rho)=\rho/\sqrt{1+K\rho}$ and $b(\rho)=\sqrt{1+K\rho}$.
The NFP equations (\[fbi1\]) and (\[fbi5\]) have been introduced by Kaniadakis & Quarati [@kq] (see also [@csr] in the context of the violent relaxation of 2D vortices and stellar systems). For $K=+1$, they can provide a dynamical model of the Bose-Einstein condensation in the canonical ensemble which has been studied in detail in [@bose]. When coupled to the field equation (\[gle3\]), the NFP equations (\[fbi1\]) and (\[fbi5\]) could also provide generalized Keller-Segel models of chemotaxis.
Mixed model: anomalous diffusion and filling factor {#sec_mixed}
---------------------------------------------------
The previous models focus individually on two important effects: anomalous diffusion (see Secs. \[sec\_pld\]-\[sec\_pldd\]) and exclusion constraints when the density becomes too large (see Sec. \[sec\_ff\]). Here we introduce a mixed model which combines these two effects in a single equation. If we take $h(\rho)=q\rho^{q+\mu-1}$ and $g(\rho)=\rho^{\mu+1}(1-\rho/\sigma_{0})$, we obtain $$\begin{aligned}
\label{mixed1}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( Dq \rho^{q+\mu-1}\nabla\rho+
\chi \rho^{\mu+1}(1-\rho/\sigma_{0}) \nabla \Phi\right ).\nonumber\\\end{aligned}$$ This corresponds to a power law diffusion such that $D(\rho)=\lbrack{Dq}/({q+\mu})\rbrack \rho^{q+\mu-1}$ and a mobility $\chi(\rho)=\chi\rho^{\mu}(1-\rho/\sigma_{0})$. The associated stochastic process is $$\begin{aligned}
\label{mixed2}
\frac{d{\bf r}}{dt}=-\chi \rho^{\mu}(1-\rho/\sigma_{0})\nabla \Phi+\sqrt{\frac{2Dq}{q+\mu}}\rho^{\frac{q+\mu-1}{2}}{\bf R}(t).\nonumber\\\end{aligned}$$ The generalized entropy corresponding to Eq. (\[mixed1\]) is obtained by integrating twice the relation $$\begin{aligned}
\label{mixed3}
C''(\rho)=\frac{q\rho^{q-2}}{1-\rho/\sigma_{0}}.\end{aligned}$$ A first integration gives $$\begin{aligned}
\label{mixed4}
C'(\rho)=q\sigma_{0}^{q-1}\Phi_{q-2}\left(\frac{\rho}{\sigma_{0}}\right ),\end{aligned}$$ where $$\begin{aligned}
\label{mixed5}
\Phi_{m}(t)=\int_{0}^{t}\frac{x^{m}}{1-x}dx.\end{aligned}$$ Therefore, the generalized entropy can be expressed as $$\begin{aligned}
\label{mixed6}
C(\rho)=q\sigma_{0}^{q}\int_{0}^{\rho/\sigma_{0}}\Phi_{q-2}(t)dt.\end{aligned}$$ Note that it does not depend on $\mu$ since the term $\rho^{\mu}$ can be put in factor of the diffusion current in Eq. (\[mixed1\]); see the discussion in Sec. \[sec\_func\].
Let us consider some particular cases. (i) For $q=1$, Eq. (\[mixed1\]) has the same entropy and the same equilibrium states as Eq. (\[ff1\]). (ii) For $\sigma_{0}\rightarrow +\infty$, we recover Eq. (\[pldd1\]). (iii) For $\mu=0$ and $q=2$, we have $$\begin{aligned}
\label{mixed7}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\rho^{2}+
\chi \rho (1-\rho/\sigma_{0}) \nabla \Phi\right ).\end{aligned}$$ The generalized entropy is $$\begin{aligned}
\label{mixed8}
S=-2\sigma_{0}^{2}\int \left (1-\frac{\rho}{\sigma_{0}}\right)\ln \left (1-\frac{\rho}{\sigma_{0}}\right) d{\bf r},\end{aligned}$$ and the stationary solution is $$\begin{aligned}
\label{mixed9}
\rho=\sigma_{0}\left\lbrack 1-e^{(\beta \Phi+\alpha)/2\sigma_{0}}\right\rbrack_{+}.\end{aligned}$$ For $\sigma_{0}\rightarrow +\infty$, we recover Eq. (\[pld12\]). We can also consider the alternative model $$\begin{aligned}
\label{mixed10}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left (\frac{2\rho D}{1-\rho/\sigma_{0}}\nabla\rho+
\chi \rho \nabla\Phi\right ),\end{aligned}$$ which has the same entropy and the same equilibrium states as Eq. (\[mixed7\]). The pressure law is $$\begin{aligned}
\label{mixed11}
p(\rho)=-2T\sigma_{0}^{2}\left\lbrack \ln(1-\rho/\sigma_{0})-\rho/\sigma_{0}\right \rbrack.\end{aligned}$$ (iv) For $(\mu,q)=(0,0)$ and performing the transformation $qD\rightarrow D$, or directly taking $h(\rho)=1/\rho$ and $g(\rho)=\rho(1-\rho/\sigma_{0})$, we obtain $$\begin{aligned}
\label{mixed12}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( D \nabla\ln\rho+
\chi \rho (1-\rho/\sigma_{0}) \nabla \Phi\right ).\end{aligned}$$ This corresponds to a logarithmic diffusion and a modified mobility taking into account an exclusion principle through the filling factor. The generalized entropy is obtained from the relation $$\begin{aligned}
\label{mixed13}
C''(\rho)=\frac{1}{\rho^{2}(1-\rho/\sigma_{0})}, \end{aligned}$$ leading to $$\begin{aligned}
\label{mixed14}
C'(\rho)=-\frac{1}{\sigma_{0}}\left\lbrace \ln\left (\frac{\sigma_{0}}{\rho}-1\right )+\frac{\sigma_{0}}{\rho}\right\rbrace,\end{aligned}$$ and finally to the explicit expression $$\begin{aligned}
\label{mixed15}
S=-\int \left (1-\frac{\rho}{\sigma_{0}}\right )\ln \left (\frac{\sigma_{0}}{\rho}-1\right ) d{\bf r}.\end{aligned}$$ We can consider the alternative model $$\begin{aligned}
\label{mixed16}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left \lbrack \frac{D}{\rho(1-\rho/\sigma_{0})}\nabla\rho+
\chi \rho \nabla \Phi\right \rbrack,\end{aligned}$$ with the same entropy and the same equilibrium states. The associated pressure law is $$\begin{aligned}
\label{mixed17}
p(\rho)=-T\ln\left (\frac{\sigma_{0}}{\rho}-1\right ).\end{aligned}$$
$\kappa$-entropy {#sec_kappa}
----------------
We consider the $\kappa$-entropy $$\begin{aligned}
\label{kappa1}
S_{\kappa}=-\frac{1}{2\kappa}\int (\rho^{1+\kappa}-\rho^{1-\kappa})d{\bf r}.\end{aligned}$$ This entropy was introduced by Kaniadakis [@k1]. It can be written $$\begin{aligned}
\label{kappa2}
S_{\kappa}=-\int \rho \ln_{(\kappa)}\rho \, d{\bf r},\end{aligned}$$ with the $\kappa$-logarithm $$\begin{aligned}
\label{kappa3}
\ln_{(\kappa)}(x)=\frac{1}{2\kappa}(x^{\kappa}-x^{-\kappa}).\end{aligned}$$ We have $$\begin{aligned}
\label{kappa4}
C'(\rho)=\frac{1}{2\kappa}\left\lbrack (1+\kappa)\rho^{\kappa}-(1-\kappa)\rho^{-\kappa}\right\rbrack,\end{aligned}$$ $$\begin{aligned}
\label{kappa5}
C''(\rho)=\frac{1}{2\rho}\left\lbrack (1+\kappa)\rho^{\kappa}+(1-\kappa)\rho^{-\kappa}\right\rbrack.\end{aligned}$$ If we take $g(\rho)=\rho$ and $h(\rho)=\rho C''(\rho)$, we obtain the NFP equation $$\begin{aligned}
\label{kappa6}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left ( \frac{D}{2} \nabla (\rho^{1+\kappa}+\rho^{1-\kappa})+
\chi \rho \nabla \Phi\right ).\end{aligned}$$ This corresponds to a power law diffusion $D(\rho)=\frac{D}{2}(\rho^{\kappa}+\rho^{-\kappa})$ and a constant mobility $\chi(\rho)=\chi$. The associated stochastic process is $$\begin{aligned}
\label{kappa7}
\frac{d{\bf r}}{dt}=-\chi\nabla \Phi+\sqrt{D}(\rho^{\kappa}+\rho^{-\kappa})^{1/2} {\bf R}(t),\end{aligned}$$ and the stationary solution of Eq. (\[kappa6\]) can be written $$\begin{aligned}
\label{kappa8}
\rho=\left (\frac{1-\kappa}{1+\kappa}\right )^{\frac{1}{2\kappa}}e_{(\kappa)}^{-(\beta \Phi+\alpha)/\sqrt{1-\kappa^{2}}},\end{aligned}$$ with the $\kappa$-exponential $$\begin{aligned}
\label{kappa9}
e_{(\kappa)}(x)=(\kappa x+\sqrt{1+\kappa^{2}x^{2}})^{1/\kappa}.\end{aligned}$$ Finally, the pressure law is $$\begin{aligned}
\label{kappa10}
p(\rho)=\frac{1}{2}T(\rho^{1+\kappa}+\rho^{1-\kappa}).\end{aligned}$$ For $\kappa=0$, we recover the standard model (\[sm1\]). These results can be generalized to the $(\kappa,r)$ entropy $$\begin{aligned}
\label{kappa11}
S_{\kappa,r}=-\frac{1}{2\kappa}\int \rho^{r}(\rho^{1+\kappa}-\rho^{1-\kappa})d{\bf r},\end{aligned}$$ which reduces in some special cases to the Tsallis [@tsallis], Abe [@abe] and Kaniadakis [@k1] entropies. The corresponding NFP equation can be written $$\begin{aligned}
\label{kappa12}
\frac{\partial\rho}{\partial t}=\nabla\cdot \left \lbrack \frac{D}{2\kappa} \nabla \left (a\rho^{1+a}
+b\rho^{1-b}\right )
+\chi \rho \nabla \Phi\right \rbrack,\end{aligned}$$ with $a=\kappa+r$ and $b=\kappa -r$. Of course, we could give many other examples of generalized Fokker-Planck equations since there exists an infinite number of distributions and entropic functionals. Therefore, we found it more convenient in [@gen] to formulate the problem in a general setting, using an arbitrary entropic functional of the form (\[h5\]).
Nonlinear mean field Fokker-Planck equations in phase space {#sec_phase}
===========================================================
We now describe nonlinear mean field Fokker-Planck equations in phase space taking into account the inertia of the particles. Overdamped models will be recovered in a limit of strong friction.
Generalized Kramers equation {#sec_gk}
----------------------------
We consider a system of $N$ particles in interaction whose dynamics is described by the stochastic Ito-Langevin equations $$\begin{aligned}
\label{gk1} \frac{d{\bf r}_{i}}{dt}={\bf v}_{i},\end{aligned}$$ $$\begin{aligned}
\label{gk2} \frac{d{\bf v}_{i}}{dt}=-\xi(f_i){\bf
v}_{i}-\nabla\Phi_i+\sqrt{2D(f_i)}{\bf R}_{i}(t),\end{aligned}$$ where $\Phi({\bf r},t)$ is a self-consistent potential given by the mean field Eq. (\[gle2\]). In ordinary models, the friction $\xi$ and the diffusion coefficient $D$ are constant. In that case, the statistical equilibrium state is the Maxwell-Boltzmann distribution $f
\sim e^{-\beta \epsilon}$ where $\epsilon=v^{2}/2+\Phi({\bf r})$ is the individual energy and the temperature $T=1/\beta$ is given by the Einstein relation $T=D/\xi$. Here, for sake of generality, the friction coefficient $\xi(f)$ and the diffusion coefficient $D(f)$ are allowed to depend on the distribution function $f({\bf r},{\bf
v},t)=\langle\sum_{i=1}^{N}\delta({\bf r}-{\bf r}_{i}(t))\delta({\bf
v}-{\bf v}_{i}(t))\rangle$. This can take into account microscopic constraints that affect the dynamics and modify the equilibrium distribution. The evolution of the distribution function $f({\bf
r},{\bf v},t)$ is governed by the nonlinear mean field Fokker-Planck equation $$\begin{aligned}
\label{gk3}
\frac{\partial f}{\partial t}+{\bf v}\cdot \frac{\partial f}{\partial {\bf r}}-\nabla\Phi\cdot \frac{\partial f}{\partial {\bf v}} =\frac{\partial}{\partial {\bf v}}\cdot \left\lbrack \frac{\partial}{\partial {\bf v}}(D(f)f)+\xi(f)f{\bf v}\right\rbrack,\nonumber\\\end{aligned}$$ coupled to Eq. (\[gle2\]). We introduce the notations $$\begin{aligned}
\label{gk4} D h(f)=\frac{d}{df}(f D(f)), \qquad \xi g(f)=f\xi(f),\end{aligned}$$ where $D$ and $\xi$ are positive constants and $h(f)$ and $g(f)$ are positive functions. The ordinary model with constant diffusion $D(f)=D$ and constant friction $\xi(f)=\xi$ is recovered for $h(f)=1$ and $g(f)=f$. With these notations, the NFP equation (\[gk3\]) can be rewritten in the form of a generalized Kramers (GK) equation $$\begin{aligned}
\label{gk5} \frac{\partial f}{\partial t}+{\bf v}\cdot
\frac{\partial f}{\partial {\bf r}}-\nabla\Phi\cdot \frac{\partial
f} {\partial {\bf v}} =\frac{\partial}{\partial {\bf v}}\cdot \left
( D h(f)\frac{\partial f}{\partial {\bf v}}+\xi g(f){\bf v}\right
).\nonumber\\\end{aligned}$$ It can be put in the conservative form $$\begin{aligned}
\label{gk6}
\frac{df}{dt}=-\frac{\partial}{\partial {\bf v}}\cdot {\bf J},\end{aligned}$$ where $$\begin{aligned}
\label{gk7} {\bf J}=- \left \lbrack D h(f)\frac{\partial f}{\partial
{\bf v}}+\xi g(f){\bf v}\right \rbrack,\end{aligned}$$ is a diffusion current. This structure guarantees the conservation of mass $M=\int f d{\bf r}d{\bf v}$.
Generalized free energy and H-theorem {#sec_nh}
-------------------------------------
We define the energy by $$\begin{aligned}
\label{nh1}
E=\frac{1}{2}\int f v^{2}\, d{\bf r}d{\bf v}+\frac{1}{2}\int \rho \Phi\, d{\bf r}=K+W,\end{aligned}$$ where $K$ is the kinetic energy and $W$ is the potential energy. We define the temperature by $$\begin{aligned}
\label{nh2}
T=\frac{D}{\xi}.\end{aligned}$$ The Einstein relation is preserved in the generalized thermodynamical framework. We introduce the generalized entropic functional $$\begin{aligned}
\label{nh3}
S=-\int C(f)\, d{\bf r}d{\bf v},\end{aligned}$$ where $C(f)$ is a convex function defined by $$\begin{aligned}
\label{nh4}
C''(f)=\frac{h(f)}{g(f)}.\end{aligned}$$ Finally, we introduce the generalized free energy $$\begin{aligned}
\label{nh5}
F=E-TS.\end{aligned}$$ The definition of the free energy (Legendre transform) is preserved in the generalized thermodynamical framework. Explicitly, $$\begin{aligned}
\label{nh5b}
F[f]=\frac{1}{2}\int f v^{2}\, d{\bf r}d{\bf v}+\frac{1}{2}\int \rho \Phi\, d{\bf r}+T\int C(f)\, d{\bf r}d{\bf v}.\nonumber\\\end{aligned}$$ A straightforward calculation (see Appendix \[sec\_ht\]) shows that $$\begin{aligned}
\label{nh6}
\dot F=
-\int \frac{1}{\xi g(f)}\left (Dh(f)\frac{\partial f}{\partial {\bf v}}+\xi g(f){\bf v}\right )^{2}d{\bf r}d{\bf v}.\end{aligned}$$ Therefore, $\dot F\le 0$. We can also introduce the Massieu function $J=S-\beta E$ which is related to the free energy by $J=-\beta F$. It satisfies $\dot J\ge 0$. If $D=0$ (leading to $T=0$), we get $F=E$ so that $\dot E\le 0$. If $\chi=0$ (leading to $\beta=0$), we get $J=S$ so that $\dot S\ge 0$.
Stationary solution {#sec_nss}
-------------------
The steady states of Eq. (\[gk5\]) must satisfy $\dot F=0$. According to Eq. (\[nh6\]), this implies ${\bf J}={\bf 0}$ or explicitly $$\begin{aligned}
\label{nss1}
Dh(f)\frac{\partial f}{\partial {\bf v}}+\xi g(f){\bf v}={\bf 0}.\end{aligned}$$ Using Eqs. (\[nh2\]) and (\[nh4\]), we get $$\begin{aligned}
\label{nss2}
C''(f)\frac{\partial f}{\partial {\bf v}}+\beta {\bf v}={\bf 0},\end{aligned}$$ which can be integrated into $$\begin{aligned}
\label{nss3} C'(f)=-\beta \left \lbrack \frac{v^{2}}{2}+\lambda({\bf
r})\right \rbrack,\end{aligned}$$ where $\lambda({\bf r})$ is a function of the position. Since $C$ is convex, this relation can be reversed to give $$\begin{aligned}
\label{nss4} f({\bf r},{\bf v})=F\left \lbrack \beta \left (
\frac{v^{2}}{2}+\lambda({\bf r})\right )\right \rbrack,\end{aligned}$$ where $F(x)=(C')^{-1}(-x)$ is a decreasing function. Since ${\bf J}={\bf 0}$ and $\partial f/\partial t=0$, the steady solution of Eq. (\[gk5\]) must also cancel the advective term $$\begin{aligned}
\label{nss5}
{\bf v}\cdot \frac{\partial f}{\partial {\bf r}}-\nabla\Phi\cdot \frac{\partial f}{\partial {\bf v}}=0.\end{aligned}$$ In other words, the steady solution of Eq. (\[gk5\]) is a particular stationary solution of the Vlasov equation (l.h.s.) whose form is selected by the “collision” term (r.h.s.). Substituting Eq. (\[nss4\]) in Eq. (\[nss5\]), we get $$\begin{aligned}
\label{nss6} (\nabla\lambda-\nabla\Phi)\cdot {\bf v}=0,\end{aligned}$$ which must be true for all ${\bf v}$. This yields $$\begin{aligned}
\label{nss7} \lambda({\bf
r})=\Phi({\bf r})+\alpha/\beta,\end{aligned}$$ where $\alpha$ is a constant of integration. Therefore, the stationary solution of Eq. (\[gk5\]) is given by $$\begin{aligned}
\label{nss8} C'(f)=-\beta \epsilon -\alpha,\end{aligned}$$ or, equivalently, $$\begin{aligned}
\label{nss9} f({\bf r},{\bf v})=F\lbrack \beta\epsilon({\bf r},{\bf v})+\alpha\rbrack,\end{aligned}$$ where $\epsilon={v^{2}}/{2}+\Phi({\bf r})$ is the energy of a particle. Thus, in the steady state, the distribution function $f=f(\epsilon)$ is a function of the individual energy. The potential $\Phi$ is determined by an integrodifferential equation obtained by substituting Eq. (\[nss9\]) in Eq. (\[gle2\]), using $\rho=\int f
d{\bf v}$. The constant $\alpha$ is determined by the conservation of mass. On the other hand, differentiating Eq. (\[nss8\]), we obtain $$\begin{aligned}
\label{nss10} \frac{df}{d\epsilon}=-\frac{\beta}{C''(f)}.\end{aligned}$$ Since $C$ is convex, i.e. $C''>0$, the preceding relation implies that $f'(\epsilon)<0$. Therefore, $f(\epsilon)$ is a decreasing function of the energy.
Minimum of free energy {#sec_nmin}
----------------------
The critical points of free energy at fixed mass are determined by the variational problem $$\begin{aligned}
\label{nmin1}
\delta F+T\alpha\delta M=0,\end{aligned}$$ where $\alpha$ is a Lagrange multiplier. These variations give $$\begin{aligned}
\label{nmin2}
C'(f)=-\beta\epsilon-\alpha.\end{aligned}$$ Therefore, comparing with Eq. (\[nss8\]), we find that a stationary solution of Eq. (\[gk5\]) is a critical point of $F$ at fixed mass. Furthermore, it is shown in [@gen] that a steady state of Eq. (\[gk5\]) is linearly dynamically stable iff it is a [*minimum*]{} (at least local) of $F$ at fixed mass. In this sense, dynamical and generalized thermodynamical stability in the canonical ensemble coincide. This property also results from Lyapunov’s direct method [@frank]. Finally, if $F$ is bounded from below, we conclude from the above properties that the system will converge to a stable steady state for $t\rightarrow +\infty$ which is a (local) minimum of $F[f]$ at fixed mass. If several local minima exist, the choice of the final steady state will depend on a complicated notion of basin of attraction. In conclusion, we have the important result: [*a steady solution of the generalized Kramers equation (\[gk5\]) is linearly dynamically stable iff it is a (local) minimum of the free energy $F[f]$ at fixed mass $M[f]=M$.*]{} This corresponds to the minimization problem: $$\begin{aligned}
\label{nmin3}
\min_{f}\quad \lbrace F[f]\quad |\quad M[f]=M\rbrace.\end{aligned}$$ Taking the second variations of $F$ and using Eq. (\[nss10\]), the condition of dynamical stability is $$\begin{aligned}
\label{nmin4} \delta^{2}F[\delta f]=-{1\over 2}\biggl\lbrace \int {(\delta
f)^{2}\over f'(\epsilon)}d{\bf r}d{\bf v}-\int \delta\rho\delta\Phi
d{\bf r}\biggr\rbrace \ge 0,\quad\end{aligned}$$ for all perturbations $\delta f$ that conserve mass.
Particular cases {#sec_npc}
----------------
If we take $h(f)=1$ and $g(f)=1/C''(f)$ we get $$\begin{aligned}
\label{npc1}
\frac{\partial f}{\partial t}+{\bf v}\cdot \frac{\partial f}{\partial {\bf r}}-\nabla\Phi\cdot \frac{\partial f}{\partial {\bf v}}=\frac{\partial}{\partial {\bf v}}\cdot \left (D\frac{\partial f}{\partial {\bf v}}+\frac{\xi}{C''(f)}{\bf v}
\right ).\nonumber\\\end{aligned}$$ In that case, we have a constant diffusion $D(f)=D$ and a variable friction $\xi(f)=\xi/[fC''(f)]$. If we take $g(f)=f$ and $h(f)=fC''(f)$ we get $$\begin{aligned}
\label{npc2}
\frac{\partial f}{\partial t}+{\bf v}\cdot \frac{\partial f}{\partial {\bf r}}-\nabla\Phi\cdot \frac{\partial f}{\partial {\bf v}}=\frac{\partial}{\partial {\bf v}}\cdot \left (DfC''(f)\frac{\partial f}{\partial {\bf v}}+\xi f{\bf v}
\right ).\nonumber\\\end{aligned}$$ In that case, we have a constant friction $\xi(f)=\xi$ and a variable diffusion $D(f)=Df \lbrack C(f)/f\rbrack'$.
Functional derivative {#sec_nfun}
---------------------
For a given free energy functional $F[f]$, we can introduce phenomenologically a dynamical model by writing the evolution of the distribution function as a continuity equation $d_{t}f=-\frac{\partial}{\partial {\bf v}} \cdot {\bf J}$ where the current is proportional to the gradient in velocity space of the functional derivative of the free energy, i.e. $$\begin{aligned}
\label{nfun1}
\frac{\partial f}{\partial t}+{\bf v}\cdot \frac{\partial f}{\partial {\bf r}}-\nabla\Phi\cdot \frac{\partial f}{\partial {\bf v}}=\frac{\partial}{\partial {\bf v}}\cdot \left \lbrack \xi({\bf r},{\bf v},t)f \frac{\partial}{\partial {\bf v}}\frac{\delta F}{\delta f}\right \rbrack.\nonumber\\\end{aligned}$$ For the free energy (\[nh5b\]), we have $$\begin{aligned}
\label{nfun2}
\frac{\delta F}{\delta f}=TC'(f)+\frac{v^{2}}{2}+\Phi,\end{aligned}$$ so that $$\begin{aligned}
\label{nfun3}
\frac{df}{dt}=
\frac{\partial}{\partial {\bf v}}\cdot \left \lbrack \xi({\bf r},{\bf v},t)\left (T f C''(f) \frac{\partial f}{\partial {\bf v}}+f {\bf v}\right )\right\rbrack,\end{aligned}$$ where we have introduced the material derivative $d/dt=\partial/\partial t+{\bf v}\cdot \partial/\partial {\bf
r}-\nabla\Phi\cdot \partial/\partial {\bf v}$ in the advective term. This equation is more general than Eq. (\[gk5\]). It shows that, for a given free energy, we can introduce an infinite class of NFP equations where $\xi({\bf r},{\bf v},t)$ is an [*arbitrary*]{} positive function of position, velocity and time. In particular, it can be a function of $f({\bf r},{\bf v},t)$. If we set $\xi({\bf
r},{\bf v},t)=\xi g(f)/f$ we recover Eq. (\[gk5\]). We can also write Eq. (\[nfun3\]) in the alternative form $$\begin{aligned}
\label{nfun4}
\frac{df}{dt}=\frac{\partial}{\partial {\bf v}}\cdot \left \lbrack \tilde{\xi}({\bf r},{\bf v},t)\left (T \frac{\partial f}{\partial {\bf v}}+\frac{1}{C''(f)} {\bf v}\right )\right\rbrack, \end{aligned}$$ where $\tilde{\xi}({\bf r},{\bf v},t)$ is an [*arbitrary*]{} positive function of position, velocity and time. If we set $\tilde{\xi}({\bf
r},{\bf v},t)=\xi h(f)$ we recover Eq. (\[gk5\]). These two alternative forms (\[nfun3\]) and (\[nfun4\]) were given in [@gen]. On the other hand, the general structure of Eq. (\[nfun1\]) implies an $H$-theorem for the free energy ($\dot F\le 0$). The derivation is similar to that given in Sec. \[sec\_func\].
Equation of state {#sec_eos}
-----------------
The stationary solutions of the nonlinear Kramers equation (\[gk5\]) are of the form $f=f(\epsilon)$ with $f'(\epsilon)<0$ where $\epsilon={v^{2}}/{2}+\Phi({\bf r})$ is the energy of a particle. The function $f$ is determined by the convex function $C$ according to Eq. (\[nss8\]). Therefore, at equilibrium, the density $\rho=\int f
d{\bf v}$ and the pressure $p=\frac{1}{d}\int f v^2 d{\bf v}$ can be expressed as $\rho=\rho(\Phi({\bf r}))$ and $p=p(\Phi({\bf
r}))$. Eliminating the potential $\Phi({\bf r})$ between these expressions, we obtain a barotropic equation of state $p=p(\rho)$ where the function $p(\rho)$ is entirely determined by the convex function $C(f)$. Furthermore, the condition that the distribution function is a function $f=f(\epsilon)$ of the energy alone implies the condition of hydrostatic balance. Indeed, we have $$\begin{aligned}
\label{edd1} \nabla p={1\over d}\int f'(\epsilon)\nabla\Phi v^{2}d{\bf v}={1\over d}\nabla\Phi \int {\partial f\over\partial {\bf v}}\cdot {\bf v}d{\bf v}\nonumber\\
=-\nabla\Phi\int f d{\bf v}=-\rho\nabla\Phi.\end{aligned}$$ The condition of hydrostatic equilibrium can also be written $p'(\Phi)=-\rho(\Phi)$ or $$\begin{aligned}
\label{edd2} p'(\rho)=-\frac{\rho}{\rho'(\Phi)}.\end{aligned}$$ Let us introduce the free energy functional $$\begin{aligned}
\label{strong12ng} F[\rho]=\int \rho\int^{\rho}{p(\rho')\over
\rho'^{2}} \,d \rho'd{\bf r}+{1\over 2}\int\rho\Phi d{\bf r}.\end{aligned}$$ In Appendix \[sec\_passage\], we show that this functional of $\rho$ can be deduced from the free energy functional $F[f]$ given by Eq. (\[nh5\]) by using the relation (\[nss9\]) valid at equilibrium. Furthermore, we show in Appendix \[sec\_equi\] that the minimization problem (\[nmin3\]) is equivalent to the minimization problem $$\begin{aligned}
\label{min3ng}
\min_{\rho}\quad \lbrace F[\rho]\quad |\quad M[\rho]=M\rbrace,\end{aligned}$$ where $\rho({\bf r})$ is the density profile corresponding to the distribution function $f({\bf r},{\bf v})$. This equivalence considerably simplifies the study of the stability of a steady state of the NFP equation (\[gk5\]). A critical point of $F[\rho]$ at fixed mass satisfies the condition of hydrostatic balance. Indeed, writing $\delta F-\alpha\delta M=0$, we have $\int^{\rho}[p'(\rho')/\rho']d\rho'+\Phi-\alpha=0$ implying $\nabla p+\rho\nabla\Phi={\bf 0}$. On the other hand, taking the second variations of $F$ and using Eq. (\[edd2\]), the condition of stability can be written $$\begin{aligned}
\label{min4ng} \delta^{2}F[\delta\rho]=-{1\over 2}\biggl\lbrace \int {(\delta
\rho)^{2}\over \rho'(\Phi)}d{\bf r}-\int \delta\rho\delta\Phi
d{\bf r}\biggr\rbrace \ge 0,\end{aligned}$$ for all perturbations $\delta\rho$ that conserve mass.
The strong friction limit {#sec_strong}
-------------------------
In this section, we shall derive the generalized Smoluchowski equation from the generalized Kramers equation in the strong friction limit $\xi\rightarrow +\infty$. The general case where both the diffusion coefficient and the friction coefficient depend on the distribution function is treated in [@lemou] by using a Chapman-Enskog expansion. Here, we restrict ourselves to the generalized Kramers equation with constant friction coefficient $$\begin{aligned}
\label{strong1}
\frac{\partial f}{\partial t}+{\bf v}\cdot \frac{\partial f}{\partial {\bf r}}-\nabla\Phi\cdot \frac{\partial f}{\partial {\bf v}}=\frac{\partial}{\partial {\bf v}}\cdot \left\lbrack \xi \left (TfC''(f)\frac{\partial f}{\partial {\bf v}}+f{\bf v}
\right )\right\rbrack.\nonumber\\\end{aligned}$$ In that case, it is possible to develop a procedure simpler that the Chapman-Enskog expansion (see [@banach]). Let us derive the hierarchy of hydrodynamic equations associated with Eq. (\[strong1\]). Defining the density and the local velocity by $$\begin{aligned}
\label{strong2} \rho=\int f\,d{\bf v}, \qquad \rho{\bf u}=\int f{\bf
v}\,d{\bf v},\end{aligned}$$ and integrating Eq. (\[strong1\]) on velocity, we get the continuity equation $$\begin{aligned}
\label{strong3} {\partial\rho\over\partial t}+\nabla\cdot (\rho{\bf u})=0.\end{aligned}$$ Next, multiplying Eq. (\[strong1\]) by ${\bf v}$ and integrating on velocity, we obtain the momentum equation $$\begin{aligned}
\label{strong4} {\partial\over\partial t}(\rho
u_{i})+{\partial\over\partial x_{j}}(\rho u_{i}u_{j})= -{\partial
P_{ij}\over\partial x_{j}}-\rho{\partial\Phi\over\partial
x_{i}}-\xi\rho u_i,\nonumber\\\end{aligned}$$ where we have defined the pressure tensor $$\begin{aligned}
\label{strong5}P_{ij}=\int fw_{i}w_{j}\,d{\bf v},\end{aligned}$$ where ${\bf w}={\bf v}-{\bf u}$ is the relative velocity. Using Eq. (\[strong3\]), the momentum equation can be rewritten in the form $$\begin{aligned}
\label{strong6} \rho \left ({\partial u_{i}\over\partial t}+u_{j}{\partial u_{i}\over\partial x_{j}}\right )= -{\partial
P_{ij}\over\partial x_{j}}-\rho{\partial\Phi\over\partial
x_{i}}-\xi\rho u_i.\nonumber\\\end{aligned}$$
We now consider the strong friction limit $\xi\rightarrow +\infty$ with fixed $T$. Since the term in parenthesis in Eq. (\[strong1\]) must vanish to leading order, we find that the out of equilibrium distribution function $f_0({\bf r},{\bf v},t)$ is given by $$\begin{aligned}
\label{strong7} C'(f_0)=-\beta\left\lbrack \frac{v^{2}}{2}+\lambda({\bf
r},t)\right\rbrack+O(\xi^{-1}),\end{aligned}$$ where $\lambda({\bf r},t)$ is a constant of integration that is determined by the density according to $$\begin{aligned}
\label{strong8} \rho({\bf r},t)=\int f_{0}d{\bf v}=\rho[\lambda({\bf
r},t)].\end{aligned}$$ Note that the distribution function $f_0$ is [*isotropic*]{} so that the velocity ${\bf u}({\bf r},t)=O(\xi^{-1})$ and the pressure tensor $P_{ij}=p\delta_{ij}+O(\xi^{-1})$ where $p$ is given by $$\begin{aligned}
\label{strong9} p({\bf r},t)=\frac{1}{d}\int f_{0}v^{2}d{\bf
v}=p[\lambda({\bf r},t)].\end{aligned}$$ Eliminating $\lambda({\bf r},t)$ between the two expressions (\[strong8\]) and (\[strong9\]), we find that the fluid is [*barotropic*]{} with an equation of state $p=p(\rho)$ entirely determined by the function $C(f)$. Of course, this is the same equation of state as the one obtained at equilibrium (see Sec. \[sec\_eos\]). Now, considering the momentum equation (\[strong4\]) in the limit $\xi\rightarrow +\infty$, we find that $$\begin{aligned}
\label{strong10} \rho{\bf u}=-\frac{1}{\xi}(\nabla
p+\rho\nabla\Phi)+O(\xi^{-2}).\end{aligned}$$ Inserting this relation in the continuity equation (\[strong3\]), we obtain the generalized Smoluchowski equation [@gen; @banach; @lemou]: $$\begin{aligned}
\label{strong11}\frac{\partial\rho}{\partial t}=\nabla\cdot
\left\lbrack \frac{1}{\xi}(\nabla p+\rho\nabla\Phi)\right\rbrack.\end{aligned}$$ The free energy associated to this equation is $$\begin{aligned}
\label{strong12} F[\rho]=\int \rho\int^{\rho}{p(\rho')\over
\rho'^{2}} \,d \rho'd{\bf r}+{1\over 2}\int\rho\Phi d{\bf r}.\end{aligned}$$ It can be deduced from the free energy (\[nh5\]) by using Eq. (\[strong7\]) to express $F[f]$ as a functional $F[\rho]=F[f_0]$ of the density (see Appendix \[sec\_passage\]). A direct calculation leads to the $H$-theorem $$\begin{aligned}
\label{strong13} \dot F=
-\int \frac{1}{\xi\rho}(\nabla p+\rho\nabla \Phi)^{2}d{\bf r}\le 0.\end{aligned}$$ The stationary solutions of the generalized Smoluchowski equation (\[strong11\]) are critical points of free energy at fixed mass. They satisfy the condition of hydrostatic balance $$\begin{aligned}
\label{strong14}\nabla p+\rho\nabla\Phi={\bf 0}.\end{aligned}$$ From Lyapunov’s direct method, we conclude that a steady state of the generalized Smoluchowski equation (\[strong11\]) is linearly dynamically stable iff it is a (local) minimum of $F$ at fixed mass $M$. This corresponds to the minimization problem (\[min3ng\]).
The condition of hydrostatic balance (\[strong14\]) only holds at equilibrium. In the strong friction limit $\xi\rightarrow +\infty$, the out-of-equilibrium distribution is of the form $f_{0}=f_{0}(\epsilon')$ with $\epsilon'={v^{2}}/{2}+\lambda({\bf
r},t)$. Taking the gradient of Eq. (\[strong9\]) and using a procedure similar to that followed in Eq. (\[edd1\]) with $\lambda({\bf
r},t)$ in place of $\Phi({\bf r})$, we obtain $$\begin{aligned}
\label{strong15} \nabla p=-\rho\nabla\lambda.\end{aligned}$$ Since $p=p(\rho)$ and $\lambda=\lambda(\rho)$, this can be rewritten $$\begin{aligned}
\label{strong16} \lambda'(\rho)=-\frac{p'(\rho)}{\rho},\end{aligned}$$ so that the out-of-equilibrium chemical potential $\lambda({\bf r},t)$ is given by $$\begin{aligned}
\label{strong17} \lambda(\rho)=-\int^{\rho}\frac{p'(x)}{x}dx.\end{aligned}$$ At equilibrium, comparing Eqs. (\[strong15\]) and (\[strong14\]), we have $\lambda({\bf r})=\Phi({\bf r})+\alpha/\beta$ and Eq. (\[strong7\]) leads to Eq. (\[nss8\]).
The damped Euler equations {#sec_damped}
--------------------------
The generalized Smoluchowski equation (\[strong11\]) can also be obtained formally from the damped Euler [^7] equations [@gen]: $$\begin{aligned}
\label{damped1} {\partial\rho\over\partial t}+\nabla\cdot (\rho{\bf u})=0,\end{aligned}$$ $$\begin{aligned}
\label{damped2}\frac{\partial {\bf u}}{\partial t}+({\bf u}\cdot \nabla){\bf u}=-\frac{1}{\rho}\nabla p-\nabla\Phi-\xi {\bf u}.\end{aligned}$$ The Lyapunov functional associated with the damped Euler equations is the generalized free energy $$\begin{aligned}
\label{ht1grt}
F[\rho,{\bf u}]=\int \rho\int^{\rho}{p(\rho')\over
\rho'^{2}} \,d \rho'd{\bf r}+{1\over 2}\int\rho\Phi d{\bf r}+\int\rho \frac{{\bf u}^{2}}{2}d{\bf r}.\nonumber\\\end{aligned}$$ It satisfies an $H$-theorem of the form (see Appendix \[sec\_ht\]): $$\begin{aligned}
\label{ht1ff}
\dot F=-\int \xi\rho {\bf u}^2 d{\bf r}\le 0.\end{aligned}$$ A steady state of the damped Euler equations (\[damped1\])-(\[damped2\]) satisfies the condition of hydrostatic equilibrium (\[strong14\]). Furthermore, from Lyapunov’s direct method, it is linearly dynamically stable iff it is a minimum of the free energy (\[ht1grt\]) at fixed mass. This corresponds to the minimization problem $$\begin{aligned}
\label{fdr}
\min_{\rho,{\bf u}}\quad \lbrace F[\rho,{\bf u}]\quad |\quad M[\rho]=M\rbrace,\end{aligned}$$ The friction coefficient in Eq. (\[damped2\]) measures the importance of inertial effects. For $\xi=0$, we recover the usual barotropic Euler equations. Alternatively, if we consider the strong friction limit $\xi\rightarrow +\infty$, we can neglect the inertial term in Eq. (\[damped2\]) and we get $$\begin{aligned}
\label{damped3}\xi {\bf u}=-\frac{1}{\rho}\nabla p-\nabla\Phi+O(\xi^{-1}).\end{aligned}$$ Substituting this relation in the continuity equation (\[damped1\]), we obtain the generalized Smoluchowski equation $$\begin{aligned}
\label{damped4}\frac{\partial\rho}{\partial t}=\nabla\cdot
\left\lbrack \frac{1}{\xi}(\nabla p+\rho\nabla\Phi)\right\rbrack.\end{aligned}$$ The physical justification of the damped Euler equations (\[damped1\])-(\[damped2\]) is not clear. They can be obtained from Eqs. (\[strong3\])-(\[strong4\]) if we close the hierarchy by invoking a local thermodynamic equilibrium (L.T.E.) condition [@gen]. However, the rigorous justification of this L.T.E. condition is not established, so this approach remains heuristic. Nevertheless, hydrodynamic equations (hyperbolic models) of the form (\[damped1\])-(\[damped2\]) have been proposed in the context of chemotaxis to describe the organization of endothelial cells [@gamba; @filbet; @csbio]. They lead to the formation of filaments that are interpreted as the beginning of a vasculature. These filaments are not obtained in the Keller-Segel model (parabolic model) which leads to point-wise blow up or round aggregates [@horstmann; @sc].
Explicit examples {#sec_ee}
=================
In this section, we give explicit examples showing the passage from the generalized Kramers equation to the generalized Smoluchowski equation in the strong friction limit.
Isothermal systems: Boltzmann entropy {#sec_is}
-------------------------------------
If we consider the Boltzmann entropy $$\label{is1} S_{B}[f]=-\int f\ln f d{\bf r}d{\bf v},$$ we get the ordinary Kramers equation $$\begin{aligned}
\label{is2}
\frac{df}{dt}=\frac{\partial}{\partial {\bf v}}\cdot\left \lbrack \xi \left (T\frac{\partial f}{\partial {\bf v}}+f{\bf v}
\right )\right\rbrack.\end{aligned}$$ The stationary state is the isothermal (Maxwell-Boltzmann) distribution $$\label{is3} f=A e^{-\beta\epsilon},$$ where $A$ is determined by the conservation of mass. The equation of state is the isothermal one $$\label{is4} p=\rho T.$$ In the strong friction limit, we obtain the ordinary Smoluchowski equation $$\begin{aligned}
\label{is5} {\partial\rho\over\partial t}=\nabla \cdot \biggl\lbrack
{1\over\xi}(T\nabla \rho+\rho\nabla\Phi)\biggr\rbrack.\end{aligned}$$ The corresponding free energy is the Boltzmann free energy $$\begin{aligned}
{F}[\rho]=T\int \rho \ln\rho\ d{\bf r}+{1\over
2}\int\rho\Phi \ d{\bf r},
\label{is6}\end{aligned}$$ and the stationary solution is the Boltzmann distribution $$\label{is7} \rho=A' e^{-\beta\Phi}$$ where $A'=(2\pi/\beta)^{d/2} A$.
Polytropes: Tsallis entropy {#sec_poly}
---------------------------
If we consider the Tsallis $q$-entropy $$\label{poly1} S_{q}[f]=-{1\over q-1}\int (f^{q}-f) d{\bf r} d{\bf
v},$$ we get the polytropic Kramers equation $$\begin{aligned}
\label{poly2}
\frac{df}{dt}=\frac{\partial}{\partial {\bf v}}\cdot\left \lbrack \xi \left (T\frac{\partial f^{q}}{\partial {\bf v}}+f{\bf v}
\right )\right\rbrack.\end{aligned}$$ The stationary state is the polytropic distribution $$\label{poly3} f=\biggl\lbrack \mu-{(q-1)\beta\over
q}\epsilon\biggr\rbrack_{+}^{1\over q-1},$$ where $\mu$ is determined by the conservation of mass. The index $n$ of the polytrope is related to the parameter $q$ by the relation $$\label{poly4} n=\frac{d}{2}+\frac{1}{q-1}.$$ Isothermal distribution functions are recovered in the limit $q\rightarrow 1$ (i.e. $n\rightarrow +\infty$). We shall consider $q>0$ so that $C$ is convex. We have to distinguish two cases. (i) For $q>1$, i.e. $n>d/2$, the distribution has a compact support since $f$ is defined only for $\epsilon\le \epsilon_{m}\equiv \mu
q/[|q-1|\beta]$ (it vanishes at $\epsilon=\epsilon_{m}$). For $\epsilon\ge \epsilon_{m}$, we set $f=0$. For $q\rightarrow +\infty$, i.e. $n=d/2$, $f$ is the Heaviside function. (ii) For $q<1$, the distribution is defined for all energies. For large velocities, it behaves like $f\sim v^{-(d-2n)}$. Therefore, the density and the pressure are finite only for $n<-1$, i.e. $d/(d+2)<q<1$. Therefore the range of allowed parameters are $$\label{poly5} q>1, \qquad n>\frac{d}{2}\quad ({\rm case}\ 1),$$ $$\label{poly6} \frac{d}{d+2}<q<1, \qquad n<-1 \quad ({\rm case}\ 2).$$ The distribution function (\[poly3\]) leads to the polytropic equation of state (see Appendix \[sec\_pol\]) $$\label{poly7} p=K\rho^{\gamma}, \qquad \gamma=1+{1\over n}.$$ For $n>d/2$ the polytropic constant is $$\label{poly8} K=\frac{1}{n+1}\left\lbrack A S_{d}
2^{\frac{d}{2}-1}\frac{\Gamma\left (d/2\right )\Gamma\left
(1-d/2+n\right )}{\Gamma(1+n)}\right \rbrack^{-1/n},$$ and for $n<-1$, we have $$\label{poly9} K=-\frac{1}{n+1}\left\lbrack A S_{d}
2^{\frac{d}{2}-1}\frac{\Gamma\left (d/2\right )\Gamma\left (-n\right
)}{\Gamma(d/2-n)}\right \rbrack^{-1/n},$$ where $A=(\beta |q-1|/q)^{1/(q-1)}$. In the strong friction limit, we get the polytropic Smoluchowski equation $$\begin{aligned}
\label{poly10} {\partial\rho\over\partial t}=\nabla \cdot \biggl\lbrack
{1\over\xi}(K\nabla \rho^{\gamma}+\rho\nabla\Phi)\biggr\rbrack.\end{aligned}$$ The generalized free energy is the Tsallis free energy $$\begin{aligned}
{F}[\rho]={K\over\gamma -1}\int (\rho^{\gamma}-\rho) \ d{\bf
r}+{1\over
2}\int\rho\Phi \ d{\bf r},
\label{poly11}\end{aligned}$$ and the stationary solution is the polytropic distribution $$\label{poly12} \rho=\biggl\lbrack \lambda-{\gamma-1\over
K\gamma}\Phi\biggr\rbrack^{1\over\gamma-1}.$$ Other useful relations valid for polytropic distributions are given in [@cst]. We note that a polytropic distribution with index $q$ in phase space yields a polytropic distribution with index $\gamma=1+2(q-1)/\lbrack 2+d(q-1)\rbrack$ in physical space. In this sense, Tsallis distributions are stable laws since the functions $f(\epsilon)$ and $\rho(\Phi)$ have a similar structure. By comparing Eq. (\[poly3\]) to Eq. (\[poly12\]) or Eq. (\[nh5\]) with Eq. (\[poly1\]) to Eq. (\[poly11\]), we note that $K$ plays the same role in physical space as the temperature $T=1/\beta$ in phase space. It is sometimes called a “polytropic temperature”.
Quite generally, we define the local kinetic temperature $T({\bf r})$ by $\frac{d}{2}T({\bf r})=\frac{1}{2}\langle v^{2}\rangle$ or equivalently $p({\bf r})=\rho({\bf r})T({\bf r})$. It is proportional to the velocity dispersion. Since, at equilibrium, $\rho=\rho(\Phi)$ and $p=p(\Phi)$, we conclude that $T({\bf r})=T[\Phi({\bf r})]$ is a function of the potential $\Phi$. This is true for any barotropic fluid at equilibrium. Now, for a polytropic distribution, using Eq. (\[poly7\]), we have $T({\bf r})=K\rho^{\gamma-1}$. Then, using Eq. (\[poly12\]) we obtain $T({\bf
r})=K\lambda-\frac{\gamma-1}{\gamma}\Phi({\bf r})$ so that $T$ is a linear function of $\Phi$ with a gradient $(\gamma-1)/\gamma=1/(n+1)=2(q-1)/[(d+2)q-d]$ [@cst; @du].
Fermions: Fermi-Dirac entropy {#sec_fermi}
-----------------------------
If we consider the Fermi-Dirac entropy $$\label{fermi1} S_{FD}[f]=-\eta_{0}\int \biggl\lbrace {f\over \eta_{0}}\ln
{f\over \eta_{0}}+\biggl (1-{f\over \eta_{0}}\biggr )\ln \biggl
(1-{f\over \eta_{0}}\biggr ) \biggr\rbrace d{\bf r} d{\bf v},$$ we get the fermionic Kramers equation $$\begin{aligned}
\label{fermi2}
\frac{df}{dt}
=\frac{\partial}{\partial {\bf v}}\cdot\left \lbrack \xi \left
(-T\eta_{0}\frac{\partial}{\partial {\bf v}}\ln\left
(1-\frac{f}{\eta_{0}}\right )+f{\bf v}
\right )\right\rbrack.\end{aligned}$$ The corresponding equation with normal diffusion and nonlinear friction is $$\begin{aligned}
\label{fermi2bis}
\frac{df}{dt}
=\frac{\partial}{\partial {\bf v}}\cdot\left \lbrack D \left
(\frac{\partial f}{\partial {\bf v}}+\beta f(1-f/\eta_0){\bf v}
\right )\right\rbrack.\end{aligned}$$ The stationary state is the Fermi-Dirac distribution function $$\label{fermi3} f={\eta_{0}\over 1+\lambda e^{\beta\epsilon}},$$ where $\lambda>0$ is determined by the conservation of mass. The Fermi-Dirac distribution function (\[fermi3\]) satisfies the constraint $f\le \eta_{0}$ which is related to the Pauli exclusion principle in quantum mechanics. The isothermal distribution function (\[is3\]) is recovered in the non-degenerate limit $f\ll \eta_{0}$. This is the case when $\lambda\rightarrow +\infty$, valid at high temperatures $T\rightarrow +\infty$. On the other hand, in the completely degenerate limit, the distribution is a step function $f=\eta_{0}H(\epsilon-\epsilon_{F})$ where $\epsilon_{F}=-\frac{1}{\beta}\ln\lambda$ is the Fermi energy. This is the case when $\lambda\rightarrow 0$, valid at low temperatures $T\rightarrow 0$. This limiting distribution corresponds to a polytrope with index $n=d/2$. The distribution in physical space, obtained by integrating the Fermi-Dirac statistics (\[fermi3\]) on the velocity, can be written $$\label{fermi4} \rho={\eta_{0}S_{d}2^{{d\over
2}-1}\over\beta^{d/2}}I_{{d\over 2}-1}(\lambda e^{\beta\Phi}),$$ where $I_{n}$ is the Fermi integral $$\label{fermi5} I_{n}(t)=\int_{0}^{+\infty}{x^{n}\over 1+te^{x}}dx.$$ The quantum equation of state for fermions is given in parametric form by $$\label{fermi6} \rho={\eta_{0}S_{d}2^{{d\over
2}-1}\over\beta^{d/2}}I_{{d\over 2}-1}(t),\qquad
p={\eta_{0}S_{d}2^{{d\over 2}}\over d\beta^{{d\over 2}+1}}I_{{d\over
2}}(t).$$ At high temperatures ($t\rightarrow +\infty$) we recover the classical isothermal law $p=\rho T$ and at low temperatures ($t\rightarrow 0$) we get a polytropic equation of state $p=K\rho^{\gamma}$ with $\gamma=(d+2)/2$ (i.e. $n=d/2$) and $K=\frac{1}{d+2}(\frac{d}{\eta_{0}S_{d}})^{2/d}$ [@fermid]. In the strong friction limit, we get the fermionic Smoluchowski equation (\[strong11\]) where the equation of state is given by (\[fermi6\]). The fermionic Smoluchowski-Poisson system has been studied in [@csr; @crrs; @bln].
Conclusion {#sec_conc}
==========
In this paper, we have studied a general class of nonlinear mean field Fokker-Planck equations [@gen] associated with a formalism of effective generalized thermodynamics (E.G.T.). We have given several physical examples of application and we have shown that NFP equations can provide generalizations of the standard Keller-Segel model describing the chemotaxis of biological populations. The main properties of these NFP equations are valid for a large class of entropic functionals, encompassing the Boltzmann, the Fermi-Dirac, the Bose-Einstein and the Tsallis statistics. Indeed, the rich mathematical structure of these equations is almost independent on the precise form of the entropy. These results should therefore evidence which properties in statistical mechanics are specific to the standard Boltzmann entropy and which properties are valid for a larger class of entropies. The distinguished feature of the Boltzmann entropy is that it can be obtained from a combinatorial analysis, assuming that all the microstates are equiprobable. However, equiprobability of the microstates is a strong postulate and it is not clear whether it has a universal scope. The universality of the Boltzmann entropy has been criticized long ago by Einstein [@ein10] who argued that the statistics applicable on a system depends on its underlying dynamics (see discussion in [@cohen]). For example, the Boltzmann distribution can be obtained from a stochastic process describing a classical random walk where the kicks have uniform sizes. However, different distributions emerge when the stochastic process describes a [*biased*]{} random walk where the kicks depend on the region where the particle happens to be. We think that this is the case in many physical systems. This results in a very complex geometrical structure of phase space (fractal, multi-fractal,...) leading to non-Boltzmannian distributions at equilibrium. Indeed, in such circumstances, the microstates are [*not*]{} equiprobable since the system prefers certain regions of phase space rather than others. It would be interesting to derive the corresponding generalized entropies (Tsallis, Abe, Kaniadakis,...) directly from a combinatorial analysis. For example, Tsallis entropy could be the natural entropy on a fractal phase-space. The selection of the entropy demands a complete specification of the microdynamics of the system in agreement with the statement given long ago by Einstein [@ein10].
The isotropic BGK operator {#sec_iso}
==========================
Let us consider a simple kinetic equation where the generalized Fokker-Planck operator in Eq. (\[gk5\]) is replaced by a generalized isotropic BGK operator. This equation has been introduced in Appendix A of [@lemou] and we provide here some complements. The kinetic equation is $$\begin{aligned}
\label{iso1} \frac{\partial f}{\partial t}+{\bf v}\cdot
\frac{\partial f}{\partial {\bf r}}-\nabla\Phi\cdot \frac{\partial
f}{\partial {\bf v}} =-\frac{f-f_0}{\tau},\end{aligned}$$ where $f_{0}({\bf r},{\bf v},t)$ is given by $$\begin{aligned}
\label{iso2} C'(f_0)=-\beta\left\lbrack \frac{v^{2}}{2}+\lambda({\bf
r},t)\right\rbrack,\end{aligned}$$ or $$\begin{aligned}
\label{iso3} f_0=F\left\lbrace \beta\left\lbrack
\frac{v^{2}}{2}+\lambda({\bf r},t)\right\rbrack\right\rbrace,\end{aligned}$$ with $F(x)=(C')^{-1}(-x)$. The function $\lambda({\bf r},t)$ is determined by the density by writing $\rho=\int f_{0}d{\bf
v}=\rho(\lambda)$. First, we show that Eq. (\[iso1\]) admits an H-theorem for the free energy (\[nh5\]). Recalling that the left hand side (Vlasov term) conserves the energy and the Casimirs, hence $F$, we find that $$\begin{aligned}
\label{iso4} \dot F=-\int \left (TC'(f)+\frac{v^{2}}{2}\right
)\frac{f-f_0}{\tau}d{\bf r}d{\bf v}.\end{aligned}$$ Using Eq. (\[iso2\]) and the fact that $\int f_{0}d{\bf v}=\int f d{\bf v}$, we have the identity $$\begin{aligned}
\label{iso5}
\int \left (TC'(f_0)+\frac{v^{2}}{2}\right )\frac{f-f_0}{\tau}d{\bf r}d{\bf v}\nonumber\\
=-\beta\int \lambda({\bf r},t)\frac{f-f_0}{\tau}d{\bf r}d{\bf v}=0.\end{aligned}$$ Therefore, we can rewrite Eq. (\[iso4\]) in the form $$\begin{aligned}
\label{iso6} \dot F=-T\int \left \lbrack C'(f)-C'(f_0)\right \rbrack
\frac{f-f_0}{\tau}d{\bf r}d{\bf v}.\end{aligned}$$ Since $C$ is convex, we have $$\begin{aligned}
\label{iso7} \left \lbrack C'(f)-C'(f_0)\right \rbrack ({f-f_0})\ge 0,\end{aligned}$$ so that $\dot F\le 0$. On the other hand, using a procedure similar to that described in Sec. \[sec\_nss\], it is straightforward to prove that the steady states of Eq. (\[iso1\]) are given by Eq. (\[nss8\]) and that a steady state of Eq. (\[iso1\]) is dynamically stable iff it is a (local) minimum of $F$ at fixed mass. The first two hydrodynamic equations associated with Eq. (\[iso1\]) are $$\begin{aligned}
\label{iso8} {\partial\rho\over\partial t}+\nabla\cdot (\rho{\bf
u})=0,\end{aligned}$$ $$\begin{aligned}
\label{iso9} {\partial\over\partial t}(\rho
u_{i})+{\partial\over\partial x_{j}}(\rho u_{i}u_{j})= -{\partial
P_{ij}\over\partial x_{j}}-\rho{\partial\Phi\over\partial
x_{i}}-\frac{1}{\tau}\rho u_{i}.\nonumber\\\end{aligned}$$ In the limit $\tau\rightarrow 0$, we can repeat the same arguments as in Sec. \[sec\_strong\] and we obtain the generalized Smoluchowski equation $$\begin{aligned}
\label{iso10}\frac{\partial\rho}{\partial t}=\nabla\cdot
\left\lbrack \tau (\nabla p+\rho\nabla\Phi)\right\rbrack.\end{aligned}$$ This equation can also be obtained from a Chapman-Enskog expansion [@lemou]. Rigorous mathematical results have been obtained recently in [@dolbeault].
Connection between dynamical and thermodynamical stability for nonlinear mean field Fokker-Planck equations {#sec_conn}
===========================================================================================================
Let us consider a small perturbation $\delta\rho$ around a stationary solution of the nonlinear mean field Fokker-Planck equation (\[dde3\]). We write the time dependence of the perturbation as $\delta\rho\sim e^{\lambda t}$. It can be shown that $\lambda$ is real, so that the perturbation is damped exponentially for $\lambda<0$ (stable) or increases exponentially for $\lambda>0$ (unstable). The linearized Fokker-Planck equation can be written $$\begin{aligned}
\lambda\delta\rho=-\nabla\cdot \delta{\bf J}.
\label{conn1}\end{aligned}$$ If we multiply Eq. (\[conn1\]) by $C''(\rho)\delta\rho$ and integrate the resulting expression over the volume, we get $$\begin{aligned}
\lambda\int C''(\rho)(\delta\rho)^{2}d{\bf r}=\int \delta{\bf J}\cdot
\left\lbrack C''(\rho)\nabla\delta\rho+C'''(\rho)\delta\rho\nabla\rho\right\rbrack d{\bf r},\nonumber\\
\label{conn2}\end{aligned}$$ where we have used an integration by parts in the r.h.s. Now, the linear variation of the current (\[dde5\]) around equilibrium can be written $$\begin{aligned}
-\frac{\delta {\bf J}}{D}=h'(\rho)\delta\rho\nabla\rho+h(\rho)\nabla\delta\rho+\beta g'(\rho)\delta\rho\nabla\Phi+\beta g(\rho)\nabla\delta\Phi.\nonumber\\
\label{conn3}\end{aligned}$$ Using $h(\rho)=C''(\rho)g(\rho)$ and the relation $$\begin{aligned}
C''(\rho)\nabla\rho=-\beta\nabla\Phi,
\label{conn4}\end{aligned}$$ resulting from Eq. (\[sta3\]), we obtain $$\begin{aligned}
-\frac{\delta {\bf J}}{Dg(\rho)}=C'''(\rho)\delta\rho\nabla\rho+C''(\rho)\nabla\delta\rho+\beta\nabla\delta\Phi.\nonumber\\
\label{conn5}\end{aligned}$$ Then, Eq. (\[conn2\]) can be rewritten $$\begin{aligned}
\lambda\int C''(\rho)(\delta\rho)^{2}d{\bf r}+\beta\int \delta{\bf J}\cdot \nabla\delta\Phi d{\bf r}=-\int \frac{(\delta{\bf J})^{2}}{Dg(\rho)}d{\bf r}.\nonumber\\
\label{conn6}\end{aligned}$$ Now, multiplying Eq. (\[conn1\]) by $\delta\Phi$ and integrating over the volume we find that $$\begin{aligned}
\lambda\int \delta\rho \delta\Phi d{\bf r}=\int \delta{\bf J}\cdot \nabla\delta\Phi d{\bf r},
\label{conn7}\end{aligned}$$ where we have used an integration by parts in the r.h.s. Inserting Eq. (\[conn7\]) in Eq. (\[conn6\]), we obtain $$\begin{aligned}
-\lambda\beta \int \frac{(\delta\rho)^{2}}{\rho'(\Phi)}d{\bf r}+\lambda\beta\int \delta\rho \delta\Phi d{\bf r}=-\int \frac{(\delta{\bf J})^{2}}{Dg(\rho)}d{\bf r}.\nonumber\\
\label{conn8}\end{aligned}$$ On the other hand, the second variations of the rate of free energy dissipation (\[h8\]) around equilibrium are $$\begin{aligned}
\delta^{2}{\dot F}=-\int \frac{(\delta{\bf J})^{2}}{\beta D g(\rho)}d{\bf r},
\label{conn9}\end{aligned}$$ and they are clearly negative. Inserting Eqs. (\[conn9\]) and (\[min4\]) in Eq. (\[conn8\]), we finally obtain the relation $$\begin{aligned}
2\lambda \delta^{2}{F}=\delta^{2}{\dot F}\le 0.
\label{conn10}\end{aligned}$$ This relation shows that a steady state of the nonlinear mean field Fokker-Planck equation (\[dde3\]) is linearly dynamically stable ($\lambda<0$) iff it is a minimum (at least local) of the free energy at fixed mass ($\delta^{2}F>0$). Therefore, dynamical and generalized thermodynamical stability coincide.
Stability of the homogeneous phase and critical point {#sec_bif}
=====================================================
We consider a homogeneous stationary solution $\rho({\bf r})=\rho$ of the nonlinear mean field Fokker-Planck equation (\[dde3\]). The corresponding potential is $\Phi({\bf r})=\Phi=U\rho$ where $U\equiv
\int u({\bf x})d{\bf x}$. The dynamical evolution of a small perturbation around equilibrium is given by the linearized equation $$\begin{aligned}
\label{bif1}
\frac{\partial\delta\rho}{\partial t}=\nabla\cdot \left\lbrack Dh(\rho)\nabla
\delta\rho+\chi g(\rho)\nabla\delta\Phi\right\rbrack\nonumber\\
=Dg(\rho)\left\lbrack C''(\rho)\Delta\delta\rho+\beta\Delta\delta\Phi\right\rbrack,\end{aligned}$$ where we have used Eqs. (\[h4\]) and (\[h6\]) to get the second line. The perturbations $\delta\rho$ and $\delta\Phi$ can be decomposed in Fourier modes of the form $f({\bf r},t)=\int
\hat{f}({\bf k})e^{i{\bf k}\cdot {\bf r}}e^{\lambda({\bf k})t}d{\bf
k}$ leading to the dispertion relation $$\begin{aligned}
\label{bif2}
\lambda({\bf k})=-Dg(\rho)k^{2}\left\lbrack C''(\rho)-\beta \hat{v}(k)\right\rbrack,\end{aligned}$$ where we have used $\delta\hat{\Phi}=(2\pi)^{d}\hat{u}(k)\delta\hat{\rho}$ (convolution) and set $\hat{v}(k)=-(2\pi)^{d}\hat{u}(k)$. The system is stable ($\lambda<0$) if $$\begin{aligned}
\label{bif3}
C''(\rho)-\beta \hat{v}(k)>0,\end{aligned}$$ for all $k$ and unstable (to some wavenumbers) otherwise. For a potential satisfying $\hat{v}<0$, the homogeneous phase is always stable. Otherwise, there exists a critical point in the problem. The homogeneous phase is stable for $$\begin{aligned}
\label{bif4}
T>T_{c}\equiv \frac{\hat{v}(k)_{max}}{C''(\rho)},\end{aligned}$$ where $\hat{v}(k)_{max}$ is the largest value of $\hat{v}(k)$ achieved for $k=k_{*}$. On the other hand, for $T<T_{c}$ the homogeneous phase is unstable to wavenumbers satisfying $\hat{v}(k)>TC''(\rho)$. The growth rate of the mode $k$ is given by Eq. (\[bif2\]). This stability analysis has been explicited for particular potentials of the form (\[gle2\]) in [@hb; @csjeans]. It has also been generalized to potentials of the form (\[gle3\]) in [@chemojeans].
The condition of generalized thermodynamical stability demands that $\rho$ is a minimum of the free energy $F[\rho]$, given by Eq. (\[min6\]), at fixed mass. Using Eq. (\[sta5\]), the stability criterion (\[min4\]) can be rewritten $$\begin{aligned}
\label{bif5}
\delta^{2}F=\frac{1}{2\beta}\left\lbrace \int C''(\rho)(\delta\rho)^{2}
d{\bf r}+\beta\int \delta\rho\delta\Phi d{\bf r}\right\rbrace \ge 0,\nonumber\\\end{aligned}$$ for all perturbations $\delta\rho$ that conserve mass. We need therefore to investigate the eigenvalue equation $$\begin{aligned}
\label{bif6}
C''(\rho)\delta\rho+\beta\delta\Phi=\lambda\delta\rho.\end{aligned}$$ The homogeneous phase is a minimum of free energy at fixed mass (stable) iff all the eigenvalues $\lambda$ are positive. If at least one eigenvalue is negative, the homogeneous phase is an unstable saddle point of free energy at fixed mass. Solving the eigenvalue Eq. (\[bif6\]) in Fourier space, we get $$\begin{aligned}
\label{bif7}
C''(\rho)-\beta\hat{v}(k)=\lambda.\end{aligned}$$ The spectrum of eiganvalues is continuous and Eq. (\[bif7\]) determines the eigenmode $k$ corresponding to the eigenvalue $\lambda$. Therefore, the homogeneous phase is a minimum of free energy at fixed mass (stable) iff the l.h.s. of Eq. (\[bif7\]) is positive for all $k$. This returns the condition (\[bif3\]), i.e. $T>T_{c}$. Alternatively, for $T<T_{c}$, there exists modes $k$ such that $\lambda<0$ implying $\delta^{2}F<0$ for these modes. In that case, the homogeneous phase is a maximum or a saddle point of free energy at fixed mass (unstable). Using the Parseval theorem, the second variations of free energy (\[bif5\]) can be written $$\begin{aligned}
\label{bif8}
\delta^{2}F=\frac{1}{2\beta}(2\pi)^{d}\int \left\lbrack C''(\rho)-\beta \hat{v}(k)\right\rbrack |\delta\hat{\rho}|^{2}d{\bf k},\end{aligned}$$ leading directly to the preceding results. We also check on this explicit example (homogeneous state) that the conditions of dynamical and generalized thermodynamical stability coincide. As shown in Appendix \[sec\_conn\], this is also true for inhomogeneous equilibrium distributions.
Finally, let us show that $T_{c}$ corresponds to a bifurcation point. The general steady state of the nonlinear mean field Fokker-Planck equation (\[dde3\]) is given by Eq. (\[sta3\]). Close to the bifurcation point, the inhomogeneous equilibrium density profile can be written $\rho({\bf
r})=\rho+\epsilon({\bf r})$ where $\rho$ is the homogeneous solution and $\epsilon({\bf r})\ll \rho$. Substituting this relation in Eq. (\[sta3\]) and expanding the equation to first order in $\epsilon$, we get $$\begin{aligned}
\label{bif9}
C''(\rho)\epsilon({\bf r})=-\beta\int\epsilon({\bf r}')u(|{\bf r}-{\bf
r}'|)d{\bf r}'.\end{aligned}$$ In Fourier space, this relation becomes $$\begin{aligned}
\label{bif10}
C''(\rho)\hat{\epsilon}({\bf k})=\beta \hat{v}(k)\hat{\epsilon}({\bf k}).\end{aligned}$$ This equation has a non-zero solution $\hat{\epsilon}({\bf k})\neq 0$ provided that there exists a mode $k=k_*$ such that $C''(\rho)=\beta
\hat{v}(k_*)$. This precisely corresponds to the condition $T=T_{c}$. Therefore, stable stationary inhomogeneous solutions appear for $T<T_{c}$ precisely when the homogeneous phase becomes unstable. For the BMF model [@cvb], the stability analysis can be carried out explicitly and we find that the phase transition at $T=T_c$ is second order. More generally, there may exist other systems where inhomogeneous solutions are stable for $T>T_{c}$. In that case, they are in “competition” with the homogeneous solution. One of these solutions is stable (global minimum of free energy) and the other metastable (local minimum of free energy). This is usually associated with a first order phase transition.
Passage from $F[f]$ to $F[\rho]$ {#sec_passage}
================================
We assume that the distribution function $f({\bf r},{\bf v},t)$ is given by an expression of the form $$\label{passage1} C'(f)=-\beta\biggl\lbrack {v^{2}\over
2}+\lambda({\bf r},t)\biggr\rbrack.$$ This expression appeared at several occasions in our analysis (see Secs. \[sec\_nss\], \[sec\_eos\], \[sec\_strong\] and Appendix \[sec\_equi\]). Since $C$ is convex, the foregoing relation can be reversed to give $$\label{passage2} f=F\biggl \lbrack \beta\biggl ({v^{2}\over
2}+\lambda({\bf r},t)\biggr )\biggr\rbrack,$$ where $F(x)=(C')^{-1}(-x)$. Since the distribution function is isotropic, the local velocity vanishes: ${\bf u}={\bf 0}$. The density and the pressure are then given by $$\label{passage3}\rho=\int f d{\bf v}=\rho[\lambda],\qquad p={1\over
d}\int f v^{2}d{\bf v}=p[\lambda].$$ The first relation determines $\lambda({\bf r},t)$ as a function of the density $\rho({\bf r},t)$. Substituting Eq. (\[passage2\]) in Eq. (\[passage3\]), we find that $$\label{passage4}\rho={1\over \beta^{d/2}}g(\beta\lambda), \qquad p={1\over \beta^{d+2\over 2}}h(\beta\lambda),$$ with $$\label{passage5} g(x)=2^{d-2\over 2}S_{d} \int_{0}^{+\infty} F(x+t)\ t^{d-2\over 2} dt,$$ $$\label{passage6} h(x)={1\over d}2^{d\over 2}S_{d} \int_{0}^{+\infty} F(x+t)\ t^{d\over 2} dt,$$ where $S_{d}$ is the surface of a unit sphere in $d$ dimensions. Eliminating $\lambda$ between the foregoing expressions, we find that the fluid is [*barotropic*]{}, in the sense that $p=p(\rho)$ where the equation of state is entirely specified by $C(f)$. We can now express the free energy (\[nh5\]) as a functional of $\rho$ by writing $F[\rho]\equiv F[f]$. The energy (\[nh1\]) is simply given by $$\label{passage7}
E={d\over 2}\int p\ d{\bf r}+{1\over 2}\int\rho\Phi d{\bf
r}.$$ On the other hand, the entropy (\[nh3\]) can be written $$\begin{aligned}
{S}=-{2^{d-2\over 2}S_{d}\over\beta^{d/2}}\int d{\bf
r}\int_{0}^{+\infty}C\lbrack F(t+\beta\lambda)\rbrack \ t^{d-2\over 2}dt.
\label{passage8}\end{aligned}$$ Integrating by parts and using $C'\lbrack F(x)\rbrack=-x$, we find that $$\begin{aligned}
{S}=-{2^{d/2}S_{d}\over d\beta^{d/2}}\int d{\bf r}\int_{0}^{+\infty}
F'(t+\beta\lambda)(t+\beta\lambda)t^{d/2}dt.\nonumber\\
\label{passage9}\end{aligned}$$ Integrating by parts one more time and using Eqs. (\[passage4\]), (\[passage5\]) and (\[passage6\]), we finally obtain $$\begin{aligned}
{S}={d+2\over 2}\beta\int p d{\bf r}+\beta\int \lambda\rho d{\bf r}.
\label{passage10}\end{aligned}$$ Collecting all the previous results, the free energy (\[nh5\]) becomes $$\begin{aligned}
{F}[\rho]=-\int\rho \biggl (\lambda+{p\over\rho}\biggr )d{\bf
r} +{1\over 2}\int\rho\Phi d{\bf r}. \label{passage11}\end{aligned}$$ Finally, using the relation $h'(x)=-g(x)$ obtained from Eqs. (\[passage5\]) and (\[passage6\]) by a simple integration by parts, it is easy to check that Eq. (\[passage4\]) implies $$\begin{aligned}
\lambda'(\rho)=-\frac{p'(\rho)}{\rho},
\label{passage12}\end{aligned}$$ so that $$\begin{aligned}
\lambda+{p\over\rho}=-\int^{\rho}{p(\rho')\over\rho'^{2}}d\rho'.
\label{passage13}\end{aligned}$$ Hence, the free energy can be written more explicitly as $$\begin{aligned}
{F}[\rho]=\int \rho \int^{\rho}{p(\rho')\over\rho'^{2}}d\rho'
d{\bf r}+{1\over
2}\int\rho\Phi d{\bf r}.
\label{passage14}\end{aligned}$$
Equivalence between the stability criteria of the generalized Kramers and Smoluchowski equations {#sec_equi}
================================================================================================
Let us consider the minimization problem (\[nmin3\]). We want to determine the distribution $f_*({\bf r},{\bf v})$ which minimizes the free energy $F[f]$ at fixed mass $M[f]=M$. To solve this minimization problem, we can proceed in two steps. [*First step:*]{} we determine the distribution $f_1({\bf r},{\bf v})$ which minimizes $F[f]$ at fixed density profile $\rho({\bf r})=\int fd{\bf v}$. This gives a distribution $f_1[\rho({\bf r}),{\bf v}]$ depending on $\rho({\bf r})$ and ${\bf v}$. Substituting this distribution in the functional $F[f]$, we obtain a functional $F[\rho]\equiv F[f_1]$ of the density. [*Second step:*]{} we determine the density $\rho_*({\bf r})$ which minimizes $F[\rho]$ at fixed mass $M[\rho]=M$. Finally, we have $f_*({\bf r},{\bf v})=f_1[\rho_*({\bf r}),{\bf v}]$.
Let us be more explicit. If we fix the density profile $\rho({\bf
r})$, the potential energy $W[\rho]$ is automatically determined. Therefore, minimizing $F[f]=E[f]-TS[f]$ at fixed density profile is equivalent to minimizing $\tilde{F}[f]=K[f]-TS[f]$ at fixed density profile, where $K[f]$ is the kinetic energy. The distribution $f_1({\bf r},{\bf v})$ that extremizes $\tilde{F}[f]$ with the constraint $\int f \, d{\bf v} =\rho({\bf r})$ satisfies the first order variations $\delta F+\int \lambda({\bf r}) \delta (\int f d{\bf
v}) d{\bf r}=0$, where $\lambda({\bf r})$ is a Lagrange multiplier. This leads to $$\label{equi1} C'(f_{1})=-\beta\biggl\lbrack {v^{2}\over
2}+\lambda({\bf r})\biggr\rbrack,$$ where $\lambda({\bf r})$ is related to $\rho({\bf r})$ by writing $\rho=\int f_{1}d{\bf v}$. Since $\delta^{2}F=-T\delta^{2}S=\frac{1}{2}T\int C''(f_{1})(\delta f)^{2}d{\bf r}d{\bf v}\ge 0$ the distribution $f_{1}$ is a [*minimum*]{} of $F[f]$ at fixed density profile. Now, we remark that Eq. (\[equi1\]) has the form (\[passage1\]) so that the functional $F[\rho]\equiv F[f_{1}]$ is explicitly given by $$\begin{aligned}
{F}[\rho]=\int \rho \int^{\rho}{p(\rho')\over\rho'^{2}}d\rho'
d{\bf r}+{1\over
2}\int\rho\Phi d{\bf r}.
\label{equi3}\end{aligned}$$ Therefore, we conclude that [*$f_{*}({\bf r},{\bf
v})=f_1[\rho_*({\bf r}),{\bf v}]$ is a minimum of $F[f]$ at fixed mass iff $\rho_{*}({\bf r})$ is a minimum of $F[\rho]$ at fixed mass*]{}. Thus, the variational problems (\[nmin3\]) and (\[min3ng\]) are equivalent for global minimization.
We shall now show that they are equivalent for local minimization. A critical point of (\[nmin3\]) is a local minimum of $F[{f}]$ at fixed mass iff inequality (\[nmin4\]) is satisfied for all perturbations $\delta{f}$ that conserve mass. A critical point of (\[min3ng\]) is a local minimum of $F[\rho]$ at fixed mass iff inequality (\[min4ng\]) is satisfied for all perturbations $\delta{\rho}$ that conserve mass. In order to make the connection between the second order variations (\[nmin4\]) and (\[min4ng\]), the idea is to project the perturbation $\delta f$ on a suitable space and write $\delta f=\delta f_{\|}+\delta f_{\perp}$ where $\delta
f_{\perp}$ is the orthogonal perturbation (this is a relatively general method that has been applied in different contexts; see [@frank; @bouchet; @assise]). We can always write the perturbation in the form $$\begin{aligned}
\delta f=\delta f_{\|}+\delta f_{\perp}=\frac{\delta\rho}{\int f'(\epsilon)d{\bf v}}f'(\epsilon)+\delta f_{\perp},
\label{app9}\end{aligned}$$ where $\delta f_{\perp}\equiv \delta f-\delta f_{\|}$ ensures that all the perturbations are considered. By construction, we have $\int \delta f
d{\bf v}=\int \delta f_{\|} d{\bf v}=\delta\rho$ so that $\int
\delta f_{\perp} d{\bf v}=0$. Therefore, $\delta f_{\perp}$ is orthogonal to $\delta f_{\|}$ in the sense that $$\begin{aligned}
\int \delta f_{\|} \delta f_{\perp} \frac{1}{f'(\epsilon)}d{\bf v} \propto \int \delta f_{\perp} d{\bf v}=0.
\label{app10}\end{aligned}$$ Then, we readily obtain $$\begin{aligned}
\int \frac{(\delta f)^{2}}{f'(\epsilon)}d{\bf v}=\int \frac{(\delta f_{\perp})^{2}}{f'(\epsilon)} d{\bf v}+\int \frac{(\delta f_{\|})^{2}}{f'(\epsilon)}d{\bf v}\nonumber\\
=\int \frac{(\delta f_{\perp})^{2}}{f'(\epsilon)} d{\bf v}
+\frac{(\delta\rho)^2}{\int f'(\epsilon)d{\bf v}}.
\label{app11}\end{aligned}$$ Now, a critical point of (\[nmin3\]) is of the form $f=f(\epsilon)$ with $\epsilon=v^2/2+\Phi({\bf r})$ and $f'(\epsilon)<0$. This implies that $\rho=\int f(\epsilon)d{\bf v}=\rho(\Phi)$ and $\rho'(\Phi)=\int
f'(\epsilon)d{\bf v}$. Therefore, Eq. (\[app11\]) can be rewritten $$\begin{aligned}
\int \frac{(\delta f)^{2}}{f'(\epsilon)}d{\bf v}
=\int \frac{(\delta f_{\perp})^{2}}{f'(\epsilon)} d{\bf v}+\frac{(\delta\rho)^2}{\rho'(\Phi)}.
\label{app11bis}\end{aligned}$$ Combining Eqs. (\[app11bis\]), (\[nmin4\]) and (\[min4ng\]) we finally obtain $$\begin{aligned}
\delta^{2}F[\delta f]=-\frac{1}{2}\int \frac{(\delta f_{\perp})^{2}}{f'(\epsilon)} d{\bf r}d{\bf v} +\delta^{2}F[\delta \rho].
\label{app14}\end{aligned}$$ If $\delta^{2}F[\delta \rho]\ge 0$ for all perturbations $\delta \rho$ that conserve mass, then $\delta^{2}F[\delta f]\ge 0$ for all perturbations $\delta f$ that conserve mass. Alternatively, if there exists a perturbation $\delta\rho_{*}$ such that $\delta^{2}F[\delta\rho_{*}]< 0$, by taking $\delta f_*$ in the form (\[app9\]) with $\delta\rho=\delta\rho_{*}$ and $\delta
f_{\perp}=0$, we get $\delta^{2}F[\delta f_{*}]=\delta^{2}F[\delta
\rho_{*}]< 0$. We conclude that $f({\bf r},{\bf v})$ is a local minimum of $F[f]$ at fixed $M$ iff $\rho({\bf r})$ is a local minimum of $F[\rho]$ at fixed $M$. Thus: (\[nmin3\]) $\Leftrightarrow$ (\[min3ng\]) for local and global minimizations.
There are several consequences to this result (see also the more detailed discussion in [@assise]):
\(i) We have seen that $f$ is a linearly dynamically stable steady state of the generalized Kramers equation (\[gk5\]) iff it is a (local) minimum of $F[f]$ at fixed mass $M[f]=M$. On the other hand, we have seen that $\rho$ is a linearly dynamically stable steady state of the generalized Smoluchowski equation (\[gs1\]) iff it is a (local) minimum of $F[\rho]$ at fixed mass $M[\rho]=M$. According to the above-mentioned result (\[nmin3\]) $\Leftrightarrow$ (\[min3ng\]), we conclude that: [*$f_{*}({\bf r},{\bf
v})=f_1[\rho_*({\bf r}),{\bf v}]$ is a linearly dynamically stable steady state of the generalized Kramers equation iff $\rho_{*}({\bf
r})$ is a linearly dynamically stable steady state of the generalized Smoluchowski equation*]{}.
\(ii) It can be shown that a distribution function which minimizes a functional of the form $F[f]=E[f]-TS[f]$ (where $T$ is a constant and $S[f]$ is given by Eq. (\[nh3\]) where $C(f)$ is an arbitrary convex function) at fixed mass $M[f]=M$ is a nonlinearly [^8] dynamically stable stationary solution of the Vlasov equation (see, e.g., the case of stellar systems [@canto]). Therefore, the generalized mean field Kramers equation (\[gk5\]), which precisely solves this minimization problem, can be used as a [*numerical algorithm*]{} to construct nonlinearly dynamically stable stationary solutions of the Vlasov equation. Indeed, if a distribution is linearly dynamically stable with respect to the generalized Kramers equation then it is nonlinearly dynamically stable with respect to the Vlasov equation (but the converse may be wrong, see below).
\(iii) It can be shown that a density profile is a nonlinearly dynamically stable stationary solution of the Euler equation with a barotropic equation of state iff it minimizes a functional of the form $F[\rho]$ (where $F[\rho]$ is given by Eq. (\[gs4\])) at fixed mass $M[\rho]=M$ (see, e.g., the case of barotropic stars [@canto]). Therefore, the generalized mean field Smoluchowski equation (\[gs1\]), which precisely solves this minimization problem, can be used as a numerical algorithm to construct nonlinearly dynamically stable stationary solutions of the Euler equation. Indeed, a density profile is linearly dynamically stable with respect to the generalized Smoluchowski equation iff it is nonlinearly dynamically stable with respect to the barotropic Euler equation.
\(iv) According to the above mentioned results, a distribution function $f=f(\epsilon)$ with $f'(\epsilon)<0$ is a nonlinearly dynamically stable stationary solution of the Vlasov equation if the corresponding barotropic gas with equation of state $p=p(\rho)$ is a nonlinearly dynamically stable stationary solution of the Euler equation. In astrophysics, this corresponds to the nonlinear Antonov first law [@canto]. However, the minimization of $F[f]=E[f]-TS[f]$ at fixed mass $M[f]=M$ (problem with one constraint) is just a [*sufficient*]{} condition of nonlinear dynamical stability with respect to the Vlasov equation. Thus, this minimization problem does not allow to construct all the nonlinearly dynamically stable stationary solutions of the Vlasov equation. A larger class of nonlinearly dynamically stable stationary solutions is obtained by maximizing a Casimir functional of the form $S[f]$ (where $S[f]$ is given by Eq. (\[nh3\])) at fixed mass $M[f]=M$ and energy $E[f]=E$ (problem with two constraints) [@canto]. A numerical algorithm solving this maximization problem is proposed in [@gen]. In case of “ensemble inequivalence”, these solutions cannot be obtained by minimizing $F[f]=E[f]-TS[f]$ at fixed mass $M[f]=M$. Therefore, the problem with two constraints (“microcanonical”) provides a [*refined*]{} condition of nonlinear dynamical stability with respect to the problem with one constraint (“canonical”).
Extension of the Eddington formula {#sec_eddington}
==================================
In Sec. \[sec\_eos\], we have seen that a distribution function $f=f(\epsilon)$ determines a barotropic equation of state $p=p(\rho)$ and we have explained how to obtain it (some explicit examples have been given in Sec. \[sec\_ee\]). We shall now consider the inverse problem: find the distribution function $f=f(\epsilon)$ leading to the equation of state $p(\rho)$. This problem was first encountered in astrophysics and solved by Eddington [@eddig]. In astrophysics, a distribution function of the form $f=f(\epsilon)$ describes a particular class of spherical stellar systems that are stationary solutions of the Vlasov-Poisson system. To any such stellar system, we can associate a corresponding barotropic star with an equation of state $p=p(\rho)$ which is a stationary solution of the Euler-Poisson system. The problem is to find the equation of state $p=p(\rho)$ corresponding to the distribution function $f=f(\epsilon)$ and [*vice et versa*]{}. A similar problem arises in the context of nonlinear mean field Fokker-Planck equations. Thus, we can adapt many results of astrophysics to the present situation. For sake of generality, we shall consider the Eddington inverse problem in a space of arbitrary dimension $d$.
Knowing the equation of state $p=p(\rho)$, we can obtain the equilibrium density $\rho(\Phi)$ by integrating the condition of hydrostatic equilibrium (\[edd2\]). The problem now is to determine $f(\epsilon)$ from the knowledge of $\rho(\Phi)$. Let us rewrite the density in the form $$\begin{aligned}
\label{edd3}
\rho(\Phi)=\int_{0}^{+\infty}f(\epsilon)S_{d}v^{d-1}dv.\end{aligned}$$ We shall consider two cases:
\(i) We first assume that the distribution function has a compact support so that $f=0$ if $\epsilon\ge \epsilon_{m}$. In that case, the range of integration in Eq. (\[edd3\]) is restricted to $v\le
\sqrt{2(\epsilon_{m}-\Phi)}$ so that $$\begin{aligned}
\label{edd4}
\rho(\Phi)=\int_{0}^{\sqrt{2(\epsilon_{m}-\Phi)}}f(\epsilon)S_{d}v^{d-1}dv.\end{aligned}$$ Taking $\epsilon=\frac{v^{2}}{2}+\Phi$ as a variable of integration instead of $v$ we obtain $$\begin{aligned}
\label{edd5}
\rho(\Phi)=\int_{\Phi}^{\epsilon_{m}}f(\epsilon)S_{d}\left\lbrack
2(\epsilon-\Phi)\right\rbrack^{\frac{d-2}{2}}d\epsilon.\end{aligned}$$ It is convenient at this stage to define $\psi=\epsilon_{m}-\Phi$ and $x=\epsilon_{m}-\epsilon$. In terms of these variables, we get $$\begin{aligned}
\label{edd6} \rho(\psi)=2^{\frac{d-2}{2}}S_{d}\int_{0}^{\psi}f(x)(\psi-x)^{\frac{d-2}{2}}dx.\end{aligned}$$ In $d=3$, taking the derivative of Eq. (\[edd6\]) we find $$\begin{aligned}
\label{edd7} \frac{1}{\sqrt{8}\pi}\frac{d\rho}{d\psi}=\int_{0}^{\psi}\frac{f(x)}{\sqrt{\psi-x}}dx.\end{aligned}$$ This is an Abel integral whose solution is $$\begin{aligned}
\label{edd8} f(x)=\frac{1}{\sqrt{8}\pi^{2}}\frac{d}{dx}\int_{0}^{x}\frac{d\rho}{d\psi} \frac{d\psi}{\sqrt{x-\psi}}\nonumber\\
=\frac{1}{\sqrt{8}\pi^{2}}\left\lbrack
\int_{0}^{x}\frac{d^{2}\rho}{d\psi^{2}}
\frac{d\psi}{\sqrt{x-\psi}}+\frac{1}{\sqrt{x}}\left
(\frac{d\rho}{d\psi}\right )_{\psi=0}\right\rbrack.\end{aligned}$$ This is the Eddington formula [@eddig]. In $d=1$, Eq. (\[edd6\]) becomes $$\begin{aligned}
\label{edd9} \frac{1}{\sqrt{2}}\rho(\psi)=\int_{0}^{\psi}\frac{f(x)}{\sqrt{\psi-x}}dx.\end{aligned}$$ Comparing with the previous case, we immediately have $$\begin{aligned}
\label{edd10} f(x)=\frac{1}{\sqrt{2}\pi}\frac{d}{dx}\int_{0}^{x}\rho(\psi) \frac{d\psi}{\sqrt{x-\psi}}\nonumber\\
=\frac{1}{\sqrt{2}\pi}\left\lbrack \int_{0}^{x}\frac{d\rho}{d\psi} \frac{d\psi}{\sqrt{x-\psi}}+\frac{1}{\sqrt{x}}\rho(\psi=0)\right\rbrack.\end{aligned}$$ Finally, in $d=2$, Eq. (\[edd5\]) reduces to $$\begin{aligned}
\label{edd11}
\rho(\Phi)=2\pi\int_{\Phi}^{\epsilon_{m}}f(\epsilon)d\epsilon,\end{aligned}$$ and we get the very simple result $$\begin{aligned}
\label{edd12}
f(\epsilon)=-\frac{1}{2\pi}\frac{d\rho}{d\Phi}(\epsilon).\end{aligned}$$
\(ii) We now consider the case where the distribution function takes strictly positive values for all energies so that $\epsilon_{m}\rightarrow +\infty$. Then Eq. (\[edd3\]) becomes $$\begin{aligned}
\label{edd13}
\rho(\Phi)=\int_{\Phi}^{+\infty}f(\epsilon)S_{d}\left\lbrack
2(\epsilon-\Phi)\right\rbrack^{\frac{d-2}{2}}d\epsilon.\end{aligned}$$ In $d=3$, taking the derivative of Eq. (\[edd13\]) we obtain $$\begin{aligned}
\label{edd14}
-\frac{1}{\sqrt{8}\pi}\frac{d\rho}{d\Phi}=\int_{\Phi}^{+\infty}\frac{f(\epsilon)}{\sqrt{\epsilon-\Phi}}d\epsilon.\end{aligned}$$ This is an Abel integral whose solution is $$\begin{aligned}
\label{edd15} f(\epsilon)=\frac{1}{\sqrt{8}\pi^{2}}\frac{d}{d\epsilon}\int_{\epsilon}^{+\infty}\frac{d\rho}{d\Phi} \frac{d\Phi}{\sqrt{\Phi-\epsilon}}\nonumber\\
=\frac{1}{\sqrt{8}\pi^{2}}\left\lbrack
\int_{\epsilon}^{+\infty}\frac{d^{2}\rho}{d\Phi^{2}}
\frac{d\Phi}{\sqrt{\Phi-\epsilon}}-\lim_{\Phi\rightarrow
+\infty}\frac{\frac{d\rho}{d\Phi}}{\sqrt{\Phi-\epsilon}}\right\rbrack.\end{aligned}$$ In $d=1$, Eq. (\[edd13\]) becomes $$\begin{aligned}
\label{edd16}
\frac{1}{\sqrt{2}}\rho(\Phi)=\int_{\Phi}^{+\infty}\frac{f(\epsilon)}{\sqrt{\epsilon-\Phi}}d\epsilon,\end{aligned}$$ and we get $$\begin{aligned}
\label{edd17} f(\epsilon)=-\frac{1}{\sqrt{2}\pi}\frac{d}{d\epsilon}\int_{\epsilon}^{+\infty}\rho(\Phi) \frac{d\Phi}{\sqrt{\Phi-\epsilon}}\nonumber\\
=-\frac{1}{\sqrt{2}\pi}\left\lbrack
\int_{\epsilon}^{+\infty}\frac{d\rho}{d\Phi}\frac{d\Phi}{\sqrt{\Phi-\epsilon}}-\lim_{\Phi\rightarrow
+\infty}\frac{\rho}{\sqrt{\Phi-\epsilon}}\right\rbrack.\end{aligned}$$ For $d=2$, Eq. (\[edd12\]) remains unchanged. For example, the distribution function associated with the Fermi-Dirac statistics in physical space $$\begin{aligned}
\label{edd18}\rho(\Phi)=\frac{\sigma_{0}}{1+ e^{\beta\Phi+\alpha}},\end{aligned}$$ is $$\begin{aligned}
\label{edd19}f(\epsilon)=\frac{\sigma_{0}\beta}{8\pi\cosh^{2}\left\lbrack\frac{1}{2}(\beta\epsilon+\alpha)\right\rbrack}, \qquad (d=2).\end{aligned}$$
Derivation of the $H$-theorems {#sec_ht}
==============================
Let us derive the $H$-theorem (\[h8\]) for the NFP equation (\[dde3\])-(\[gle2\]). The time variations of the entropy (\[h5\]) and of the energy (\[h1\]) associated with an external potential are $$\begin{aligned}
\label{ht1}
\dot S=-\int C'(\rho)\frac{\partial\rho}{\partial t}d{\bf r}, \qquad \dot E=\int\Phi_{ext}\frac{\partial\rho}{\partial t}d{\bf r}.\end{aligned}$$ On the other hand, using Eqs. (\[h2\]) and (\[gle2\]), the energy associated with a binary potential of interaction can be written $$\begin{aligned}
\label{ht2}
E=\frac{1}{2}\int \rho({\bf r},t)u(|{\bf r}-{\bf r}'|)\rho({\bf r}',t)d{\bf r}d{\bf r}'.\end{aligned}$$ Its time derivative is $$\begin{aligned}
\label{ht3}
\dot E=\frac{1}{2}\int \frac{\partial\rho}{\partial t}({\bf r},t)u(|{\bf r}-{\bf r}'|)\rho({\bf r}',t)d{\bf r}d{\bf r}'\nonumber\\
+\frac{1}{2}\int \rho({\bf r},t)u(|{\bf r}-{\bf r}'|)\frac{\partial\rho}{\partial t}({\bf r}',t)d{\bf r}d{\bf r}'.\end{aligned}$$ Interchanging the dummy variables ${\bf r}$ and ${\bf r}'$ and using Eq. (\[gle2\]), we finally obtain $$\begin{aligned}
\label{ht4}
\dot E=\int\Phi\frac{\partial\rho}{\partial t}d{\bf r}.\end{aligned}$$ Therefore, the time variation of the free energy (\[h7\]) is given by $$\begin{aligned}
\label{ht5}
\dot F=\dot E-T\dot S=\int (\Phi+TC'(\rho))\frac{\partial\rho}{\partial t} d{\bf r}.\end{aligned}$$ Using Eq. (\[dde4\]) and integrating by parts, we obtain $$\begin{aligned}
\label{ht6}
\dot F
=\int {\bf J}\cdot (\nabla\Phi+TC''(\rho)\nabla\rho ) d{\bf r}. \end{aligned}$$ Inserting Eqs. (\[dde5\]), (\[h4\]) and (\[h6\]) in Eq. (\[ht6\]), we finally obtain the $H$-theorem (\[h8\]).
Let us derive the $H$-theorem (\[h9\]) for the NFP equation (\[dde3\])-(\[gle3\]). The time variation of the energy given by Eq. (\[h3\]) is $$\begin{aligned}
\label{ht7}
\dot E=\frac{1}{\lambda}\int \left (\nabla\Phi\cdot \nabla\frac{\partial\Phi}{\partial t}+k^{2}\Phi\frac{\partial\Phi}{\partial t}\right )d{\bf r}\nonumber\\
+\int \frac{\partial\rho}{\partial t}\Phi d{\bf r}+\int\rho\frac{\partial\Phi}{\partial t}d{\bf r}.\end{aligned}$$ Integrating the first term by parts and using Eq. (\[gle3\]), it can be rewritten $$\begin{aligned}
\label{ht8}
\dot E=-\frac{1}{\lambda\epsilon}\int (\Delta\Phi-k^{2}\Phi-\lambda\rho)^{2} d{\bf r}+\int \frac{\partial\rho}{\partial t}\Phi d{\bf r}.\end{aligned}$$ Inserting Eq. (\[dde3\]) and proceeding as above, we obtain the $H$-theorem (\[h9\]).
Consider now the generalized Smoluchowski equation (\[gs1\]). The time variation of the free energy (\[gs4\]) is $$\begin{aligned}
\label{ht9}
\dot F=\int \left (\Phi+\int^{\rho}\frac{p(\rho')}{\rho^{'2}}d\rho' d{\bf r}+\frac{p}{\rho}\right )\frac{\partial\rho}{\partial t}d{\bf r}\end{aligned}$$ Inserting Eq. (\[gs1\]) in Eq. (\[ht9\]) and integrating by parts we obtain the $H$-theorem (\[gs5\]).
Let us derive the $H$-theorem (\[nh6\]) for the NFP equation (\[gk5\])-(\[gle2\]). The time variations of the entropy (\[nh3\]) and of the energy (\[nh1\]) are $$\begin{aligned}
\label{ht10}
\dot S=-\int C'(f)\frac{\partial f}{\partial t}d{\bf r}d{\bf v}, \quad \dot E=\int \left (\frac{v^{2}}{2}+\Phi\right )\frac{\partial f}{\partial t}d{\bf r}d{\bf v}.\nonumber\\\end{aligned}$$ For $D=\xi=0$, the NFP equation (\[gk5\])-(\[gle2\]) reduces to the Vlasov equation. The Vlasov equation conserves the energy and the Casimirs. Indeed, using integrations by parts, we have $$\begin{aligned}
\label{ht11}
\dot E=\int \left (\frac{v^{2}}{2}+\Phi\right )\left (-{\bf v}\cdot \frac{\partial f}{\partial {\bf r}}+\nabla\Phi\cdot \frac{\partial f}{\partial {\bf v}}\right )d{\bf r}d{\bf v} \nonumber\\
=\int (\nabla\Phi\cdot {\bf v}-{\bf v}\cdot \nabla\Phi)f d{\bf r}d{\bf v}=0, \qquad\end{aligned}$$ and $$\begin{aligned}
\label{ht12}
\dot S=-\int C'(f)\left (-{\bf v}\cdot \frac{\partial f}{\partial {\bf r}}+\nabla\Phi\cdot \frac{\partial f}{\partial {\bf v}}\right )d{\bf r}d{\bf v} \nonumber\\
=\int \left\lbrack {\bf v}\cdot \frac{\partial C(f)}{\partial {\bf r}}-\nabla\Phi\cdot\frac{\partial C(f)}{\partial {\bf v}}\right\rbrack d{\bf r}d{\bf v}\nonumber\\
=\int \left\lbrack \frac{\partial}{\partial {\bf r}}\cdot (C(f){\bf v})-\frac{\partial}{\partial {\bf v}}\cdot (C(f)\nabla\Phi)\right\rbrack d{\bf r}d{\bf v}=0. \qquad\end{aligned}$$ Therefore, coming back to the NFP equation (\[gk5\])-(\[gle2\]), the only contribution to the time variation of the free energy (\[nh5\]) comes from the Fokker-Planck current. Using Eqs. (\[ht10\]) and (\[gk6\]), we have $$\begin{aligned}
\label{ht13}
\dot F=\int \left (\frac{v^{2}}{2}+\Phi+TC'(f)\right )\frac{\partial f}{\partial t} d{\bf r}d{\bf v}\nonumber\\
=-\int \left (\frac{v^{2}}{2}+\Phi+TC'(f)\right )\cdot \frac{\partial {\bf J}}{\partial {\bf v}} d{\bf r}d{\bf v}\nonumber\\
=\int \left ({\bf v}+TC''(f)\frac{\partial f}{\partial {\bf v}}\right )\cdot {\bf J} d{\bf r}d{\bf v}. \qquad\end{aligned}$$ Inserting Eqs. (\[gk7\]), (\[nh2\]) and (\[nh4\]) in Eq. (\[ht13\]) and integrating by parts we obtain the $H$-theorem (\[nh6\]).
Let us finally derive the $H$-theorem (\[ht1ff\]) for the damped Euler equations (\[damped1\]), (\[damped2\]) and (\[gle2\]). The time variation of the free energy (\[ht1grt\]) is $$\begin{aligned}
\label{ht14}
\dot F=\int \left (\Phi+\int^{\rho}\frac{p(\rho')}{\rho^{'2}}d\rho' d{\bf r}+\frac{p}{\rho}+\frac{{\bf u}^{2}}{2}\right )\frac{\partial\rho}{\partial t}d{\bf r}\nonumber\\
+\int \rho {\bf u}\cdot \frac{\partial {\bf u}}{\partial t}d{\bf r}.\end{aligned}$$ Substituting Eqs. (\[damped1\]) and (\[damped2\]) in Eq. (\[ht14\]) and integrating by parts, we get $$\begin{aligned}
\label{ht15}
\dot F=\int \rho {\bf u}\cdot \left\lbrack \nabla\left (\frac{{\bf u}^{2}}{2}\right )-\xi {\bf u}-({\bf u}\cdot \nabla){\bf u}\right \rbrack d{\bf r}.\end{aligned}$$ Using $({\bf u}\cdot\nabla){\bf u}=\nabla({\bf u}^{2}/2)-{\bf u}\times (\nabla\times {\bf u})$ we finally obtain the result (\[ht1ff\]).
Polytropic equation of state {#sec_pol}
============================
For $n>d/2$ (case 1), the polytropic DF can we written $$\begin{aligned}
\label{pol1}
f=A(\epsilon_{m}-\epsilon)_{+}^{n-d/2}.\end{aligned}$$ The density and the pressure can be expressed as $$\begin{aligned}
\label{pol2}
\rho=AS_{d}Q_{0}(\Phi), \qquad p=\frac{1}{d}AS_{d}Q_{2}(\Phi),\end{aligned}$$ with $$\begin{aligned}
\label{pol3}
Q_{k}=\int_{0}^{\sqrt{2(\epsilon_{m}-\Phi)}}\left (\epsilon_{m}-\Phi-\frac{v^{2}}{2}\right )^{n-d/2}v^{k+d-1}dv.\quad \end{aligned}$$ Setting $x=v^{2}/\lbrack 2(\epsilon_{m}-\Phi)\rbrack$, we obtain $$\begin{aligned}
\label{pol4}
Q_{k}=2^{(k+d-2)/2}(\epsilon_{m}-\Phi)^{n+k/2}\nonumber\\
\times\int_{0}^{1}(1-x)^{n-d/2}x^{(k+d-2)/2}dx. \end{aligned}$$ The integral can be expressed in terms of Gamma functions leading to $$\begin{aligned}
\label{pol5}
Q_{k}=2^{(k+d-2)/2}(\epsilon_{m}-\Phi)^{n+k/2}\nonumber\\
\times\frac{\Gamma((d+k)/2)\Gamma(1-d/2+n)}{\Gamma(1+k/2+n)}.\end{aligned}$$ Then, the density and the pressure can be expressed in terms of the potential $\Phi$ as $$\begin{aligned}
\label{pol6}
\rho=A S_{d}(\epsilon_{m}-\Phi)^{n}2^{d/2-1}\frac{\Gamma(d/2)\Gamma(1-d/2+n)}{\Gamma(1+n)},\qquad \end{aligned}$$ $$\begin{aligned}
\label{pol7}
p=\frac{A S_{d}}{n+1}(\epsilon_{m}-\Phi)^{n+1}2^{d/2-1}\frac{\Gamma(d/2)\Gamma(1-d/2+n)}{\Gamma(1+n)},\qquad \end{aligned}$$ where we have used the identity $\Gamma(n+1)=n\Gamma(n)$ to simplify the second expression. Eliminating the potential $\Phi$ between these equations, we obtain the polytropic equation of state (\[poly7\]) with $K$ given by Eq. (\[poly8\]).
For $n<-1$ (case 2), the polytropic DF can we written $$\begin{aligned}
\label{pol8}
f=A(\epsilon_{m}+\epsilon)^{n-d/2}.\end{aligned}$$ The density and the pressure can be expressed as $$\begin{aligned}
\label{pol9}
\rho=AS_{d}R_{0}(\Phi), \qquad p=\frac{1}{d}AS_{d}R_{2}(\Phi),\end{aligned}$$ with $$\begin{aligned}
\label{pol10}
R_{k}=\int_{0}^{+\infty}\left (\epsilon_{m}+\Phi+\frac{v^{2}}{2}\right )^{n-d/2}v^{k+d-1}dv.\quad \end{aligned}$$ Setting $x=v^{2}/\lbrack 2(\epsilon_{m}+\Phi)\rbrack$, we obtain $$\begin{aligned}
\label{pol11}
R_{k}=2^{(k+d-2)/2}(\epsilon_{m}+\Phi)^{n+k/2}\nonumber\\
\times\int_{0}^{+\infty}(1+x)^{n-d/2}x^{(k+d-2)/2}dx. \end{aligned}$$ The integral can be expressed in terms of Gamma functions leading to $$\begin{aligned}
\label{pol12}
R_{k}=2^{(k+d-2)/2}(\epsilon_{m}+\Phi)^{n+k/2}\nonumber\\
\times\frac{\Gamma((d+k)/2)\Gamma(-k/2-n)}{\Gamma(d/2-n)}.\end{aligned}$$ Then, the density and the pressure can be expressed in terms of the potential $\Phi$ as $$\begin{aligned}
\label{pol13}
\rho=A S_{d}(\epsilon_{m}+\Phi)^{n}2^{d/2-1}\frac{\Gamma(d/2)\Gamma(-n)}{\Gamma(d/2-n)},\qquad \end{aligned}$$ $$\begin{aligned}
\label{pol14}
p=-\frac{A S_{d}}{n+1}(\epsilon_{m}+\Phi)^{n+1}2^{d/2-1}\frac{\Gamma(d/2)\Gamma(-n)}{\Gamma(d/2-n)},\qquad \end{aligned}$$ where we have used the identity $\Gamma(n+1)=n\Gamma(n)$ to simplify the second expression. Eliminating the potential $\Phi$ between these equations, we obtain the polytropic equation of state (\[poly7\]) with $K$ given by Eq. (\[poly9\]).
Note, as a final remark, that spatially homogeneous polytropic distributions are obtained by taking $\Phi({\bf r})=0$ in the above expressions.
[99]{}
[^1]: We recall that the Boltzmann entropy can be obtained from a [*combinatorial analysis*]{} assuming that all the accessible microstates are equiprobable. This is the basic postulate of statistical mechanics. Non-standard entropies can be relevant for complex systems where this postulate breaks down, i.e. when the accessible microstates are [*not*]{} equiprobable. This happens when the system prefers some regions of phase space better than others or when the particles are subjected to exclusion/inclusion principles or fine-grained constraints [@cras].
[^2]: A generalized Fokker-Planck equation leading to the Fermi-Dirac statistics has also been introduced by Chavanis [*et al.*]{} (1996) [@csr] in the context of the violent relaxation of collisionless stellar systems described by the Vlasov equation. This is based on the Lynden-Bell’s form of entropy (1968) [@lb] which becomes similar to the Fermi-Dirac entropy in the two-levels approximation of the theory.
[^3]: Generalized Kramers and Smoluchowski equations describe [*dissipative*]{} systems where the temperature is fixed instead of the energy. They are therefore associated with the canonical ensemble. The appropriate thermodynamical potential is the free energy $F=E-TS$ which decreases monotonically with time at fixed mass. Generalized Boltzmann and Landau equations describing [*conservative*]{} systems where the energy is fixed have been introduced by Kaniadakis (2001) [@k1] and Chavanis (2004) [@gen2]. They are associated with the microcanonical ensemble. The proper thermodynamical potential is the entropy $S$ which increases monotonically with time at fixed mass and energy.
[^4]: In this paper, we shall restrict ourselves to this situation. The original Keller-Segel model (\[ks1\])-(\[ks2\]) where $D_{1}$ and $D_{2}$ depend on both $\rho({\bf r},t)$ and $c({\bf r},t)$ does not seem to possess the nice “thermodynamical” properties of the reduced Keller-Segel (\[ks3\])-(\[ks4\]) such as Legendre structure of the free energy functionals, canonical $H$-theorem, Einstein relation etc.
[^5]: There are important cases, like the system of self-gravitating Brownian particles, where the free energy is not bounded from below. In that case, the system can either relax towards a local minimum of $F$ at fixed mass (when it exists) or collapse to a Dirac peak [@post], leading to a divergence of the free energy $F(t)\rightarrow -\infty$.
[^6]: In Sec. \[sec\_gle\], we have obtained generalized Fokker-Planck equations by using ordinary Master equations (based on usual transition probabilities $a(\rho)=\rho$ and $b(\rho)=1$) and generalized Langevin equations where the diffusion coefficient and the mobility depend on the density. In this section, we have obtained generalized Fokker-Planck equations by using generalized Master equations (based on density dependent transition probabilities) and ordinary Langevin equations with constant coefficients.
[^7]: Hydrodynamical Euler equations involving a friction force $-\xi(|{\bf u}|) {\bf u}$ have appeared in various contexts for different reasons. We may mention, for example, bottom-wall friction in 2D turbulence [@pt], frictional force of air on a turbulent soap film [@rw], effective dynamical friction in the process of violent relaxation for collisionless stellar systems [@csr] and Epstein or Stokes friction laws for the dynamics of dust particles in the solar nebula [@aa].
[^8]: We implicitly consider here the [*formal*]{} nonlinear dynamical stability in the sense of Holm [*et al.*]{} [@holm].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'M. Zuin$^{1}$, A. Carati$^{2}$, M. Marino$^{2}$, E. Martines$^{1}$ and L. Galgani$^{2}$'
title: A first principles explanation for the density limit in magnetized plasmas
---
**Fusion research on magnetic confinement is confronted with a severe problem concerning the electron densities $n_e$ to be used in fusion devices. Indeed, high densities are mandatory for obtaining large efficiencies, whereas it is empirically found that catastrophic disruptive events occur for densities exceeding a maximal one $n_e^M$. On the other hand, despite the large theoretical work “there is no widely accepted, first principles model for the density limit” (see [@green2], abstract). Here, we propose a simple microscopic model of a magnetized plasma suited for a tokamak, for which the existence of a density limit is proven. This property turns out to be a general collective feature of electrodynamics of point charges, which is lost in the continuum approximation. The law we find is $$\label{legge}
n_e^M = 1.74\, \frac{1} {{m_e}c^2}\, \frac{B^2} {\mu_0}$$ where $\mu_0$ is the vacuum permeability, $c$ the speed of light, ${m_e}$ the electron mass, and $B$ the magnetic field. As shown in Fig. 1, the theoretical limit (big circles) is in rather good agreement with the empirical data, actually a surprisingly good one for a model based on first principles, with no adjustable parameter.**
The way in which law (\[legge\]) was established is an interesting example of an encounter between fundamental and applied research. Indeed three of the present authors are involved since some time in studies of a general character concerning the microscopic electrodynamics of systems of point particles (see [@cg] and [@mcg]), in which both the mutual retarded forces, and the radiation reaction force of Abraham Lorentz and Dirac [@dirac] (see also [@massimodirac] and [@jackson]) are taken into account. One of the results obtained is the proof of an identity conceived by Wheeler and Feynman [@wf], and an appreciation of the role the latter plays in allowing for the very existence of a dispersion relation. In particular, some examples of dispersion relations were given (see [@cg], Fig. 1), which exhibit, as the matter density is increased, a bifurcation of a topological character, entailing an instability. But the physical relevance of this fact was not emphasized. Such a density controlled bifurcation impressed instead very much those of the present authors who deal with plasma physics, who suggested it may be relevant for fusion plasmas. To this end, the simplest possible model was formulated, that should capture, within the frame of the foundational works mentioned, the essential physics of a magnetized plasma, confined in a tokamak configuration (the most studied one for fusion plasmas). The model is presented here, together with the deduction of law (\[legge\]). Preliminarily, the main evidence for the existence of a density limit in tokamaks is recalled, and it is discussed how well does law (\[legge\]) fit the data.
For the purposes of the present paper, all is needed to know about tokamaks is essentially that they are toroidal devices in which the confining magnetic field $B$ is the vector sum of a strong toroidal field $B_t$ produced by a set of coils wound around a torus, and of a much smaller poloidal field $B_p$ generated by a toroidal plasma current $I_p$. A few more details will be mentioned later.
![Greenwald plot for three tokamak devices (data extracted from [@green2]). Dotted line is the empirical Greenwald limit (\[legge\_greenwald\]). Big circles are the theoretical predictions (\[legge\]) for the same devices. []{data-label="fig:1"}](figura_1.eps){width="1.\columnwidth"}
The empirical data which show the existence of a density threshold beyond which tokamaks cannot operate, were collected by Greenwald [@green2] in a classical figure, the data of which are reported here in Fig.1. In the figure, the electron densities $n_e$ at which three different tokamak devices could actually be operated are reported versus the so-called Greenwald parameter $n_G \equiv I_p/\pi {a}^2$ (where ${a}$ is the minor radius of the torus), which is presumed to be the relevant control parameter. Indeed, the law proposed by Greenwald for the maximal density (dotted line in the figure) is $$\label{legge_greenwald}
n_e^M = \alpha_G\, I_p/\pi{a}^2 \ ,$$ where $\alpha_G$ is a constant with suitable dimensions, such that $\alpha_G = 1$ in the units indicated in the figure. The theoretical predictions given by (\[legge\]) are also reported as big circles.
{width=".7\columnwidth"}
Thus there naturally arises the question, how is it possible that two analytically different predictions, (\[legge\]) and (\[legge\_greenwald\]), happen to agree with each other, at least in a few definite cases. The reason is that one has $$\label{relazione}
\frac{B}{\mu_0} \approx \, b \, \frac{I_p}{\pi a^2} \ , \quad b\,=\,q(a)\, R\, \frac{1}{\kappa^{1/2}(1+\kappa^2)} \ ,$$ where the dimensional coefficient $b$ (a length) is not a universal one, but depends parametrically on geometric factors and operative conditions characterising each experiment. These are the major and the minor radii $R$ and $a$, the plasma elongation $\kappa$ (equal to 1 for circular plasmas, see [@Miyamoto] pg. 277), and the edge safety factor $q(a)$ defined below.
This is seen as follows. One has $B_p=\mu_0 I_p/(2 \pi a \kappa^{1/2})$. Furthermore, in the approximation $B \approx B_t$, one has $B \approx (B_t/B_p)B_p$, while the operative parameter $B_t/B_p$ is determined by the edge safety factor $q(a)$ which, for the simplified case of an elliptical plasma (see [@Miyamoto]), is defined as $q(a)= (a/R) (B_t/ B_p) (1+ \kappa^2)\, /2$. This gives relation (\[relazione\]). The formula for $b$ in the general case is also easily established.
Now, magnetohydrodynamic stability requires $q(a)>2$, but the actual value at which each experimental data point of Fig. \[fig:1\] was taken is not given in the literature. So we assumed $q(a)=4$, which is a typical operational value, and this introduces an uncertainty in the theoretical points reported in Fig. \[fig:1\]. Thus, we decided to look directly at the experimental values available in the literature, from which a definite estimate of $B$ could be obtained, and this we did for conventional tokamaks more recent than those in Fig. \[fig:1\]. Such values are plotted in Fig. \[fig:2\], where they are compared to laws (\[legge\]) and (\[legge\_greenwald\]). The agreement with law (\[legge\]) is perhaps a little better. This fact might have relevant implications for future tokamaks, as it implies a favorable density scaling for machines with large values of the product $R\, B$. For example, let us consider the international thermonuclear experimental reactor (ITER) [@Iter], which should operate at a toroidal field of 5.3 T. According to law (\[legge\]) it would be able to operate at densities up to $n_e^M= 4.7\times 10^{20}
m^{-3}$. This is a value more than three times larger than that expected according to Greenwald law (\[legge\_greenwald\]) for a plasma current of 15 MA, which is the corresponding value of $I_p$ for a $q(a)=3$ scenario.
It is worth mentioning that a $B^2$ dependence of the density limit in the ALCATOR C experiment was noted in the past by Granetz [@Granetz], although such a clear dependence was not observed on other experiments (see for example [@Petrie] for the DIII-D tokamak). Also, one should point out that tokamaks with very low aspect ratio $R/a$ (spherical tokamaks) and reversed field pinches (RFP) [@Lorenzini], for which the validity of the Greenwald scaling has been proposed [@Valisa], seem not to fit well into the proposed $B^2$ scaling. This, perhaps, suggests that at low magnetic fields other effects, not considered in the present simple model, might come into play.
We show now how law (\[legge\]) was obtained, in the frame of microscopic electrodynamics of point particles (see [@cg] and [@mcg]), rather than of magnetohydrodynamics (see for example [@gold], chapter 17), or of the mean field theories of the Vlasov approach. We describe the plasma as constituted of point particles obeying Newton equations, with both the retarded electromagnetic interactions among all particles and the radiation reaction force taken into account. We then concentrate on the role played by the gyration of the electrons around the magnetic field lines, and so ignore their motions along the field lines, and also ignore the electrostatic part of the problem. Finally we also limit ourselves to the extremely simplified case of a one–dimensional array.
So we introduce the following model. Given a constant magnetic field $B$, which we take oriented along the ${z}$ axis of a cartesian coordinate system, we constrain each electron, say the $n$-th one, to move on a plane parallel to the ($x,y$) plane, so that its $z_n$ coordinate is fixed. The simplest choice is to take $z_n=n {l}$, with $n \in Z$, for a given positive step ${l}$. Each electron, say the $n$–th one, is subjected to the external magnetic field $B$, and also to the electromagnetic field created by the electrons themselves. Namely, the sum of the Liénard–Wiechert fields, which are determined as the retarded solutions of Maxwell equations having as sources the charge and the current densities of each other electron $m\neq n$, and the Abraham–Lorentz–Dirac radiation reaction force, due to the motion of the $n$–th electron itself, The latter is given, in the nonrelativistic approximation, by $(2/3) e^2/(4\pi \epsilon_0
c^3)\, {{\hskip2pt\ddot{\null}\hskip2.5pt
\dot{\null}\kern -5pt {{\mathbf{x}}}}}_n$, where $e$ is the electron charge and ${{\mathbf{x}}}_n$ the position vector of the electron.
We then perform the dipole approximation. Thus we neglect the magnetic field due to the $m$–th electron, and for the electric field created by it we take the well known expression for a dipole. Finally, we approximate the distance between electrons $n$ and $m$ by $r_{n,m}={l}|n-m|$. The system of equations of motion defining the model is then
$$\begin{split}
{\ddot x}_n- {\omega_c}{\dot y}_n -\frac 23\, \frac{e^2 }
{4\pi\epsilon_0 {m_e}c^3}\, {{{\hskip2pt\ddot{\null}\hskip2.5pt
\dot{\null}\kern -5pt {x}}}}_n &= -
\frac{e^2 }{4\pi\epsilon_0 {m_e}} \sum_{m\neq n}\Big[ \frac{x_m (t-r_{nm}/c)}{r^3_{nm}} +
\frac{1}{c}\frac{{\dot x}_m (t-r_{nm}/c)}{r^2_{nm}} + \frac{1}{c^2}
\frac{{\ddot x}_m(t-r_{nm}/c)}{r_{nm}}\Big]\\ {\ddot y}_n+ {\omega_c}{\dot x}_n -\frac 23\,
\frac{e^2 }{4\pi\epsilon_0 {m_e}c^3}\, {{{\hskip2pt\ddot{\null}\hskip2.5pt
\dot{\null}\kern -5pt {y}}}}_n &= -
\frac{e^2 }{4\pi\epsilon_0 {m_e}} \sum_{m\neq
n}\Big[ \frac{y_m(t-r_{nm}/c)}{r^3_{nm}} + \frac{1}{c}\frac{{\dot
y}_m(t-r_{nm}/c)}{r^2_{nm}} + \frac{1}{c^2} \frac{{\ddot
y}_m(t-r_{nm}/c)}{r_{nm}}\Big]\\
\end{split}$$
for $n\in Z$, where $\epsilon_0$ is the vacuum permittivity, and ${\omega_c}=eB/{m_e}$ the Larmor or cyclotron frequency of the electrons in the external magnetic field $B$. This is an infinite system of linear equations with delay, which is just a simple variant of the system considered in [@cg].
Our aim is now to investigate the stability properties of the system, as the control parameters $n_e=1/{l}^3$ and $B$ (or equivalently ${\omega_c}$) are varied. Following a completely standard procedure (see for example [@chandra]), we compute the normal modes of the system and determine the values of the parameters for which the frequencies become complex. So we look for normal mode solutions with wavenumber $k$ and angular frequency $\omega$, i.e., of the form $$x_j=A_x e^{i(k{l}j+\omega t)}\ , \quad y_j=A_ye^{i(k{l}j+\omega t)}\ .
\label{}$$ This leads to a linear system in the unknowns $A_x$, $A_y$, from which the dispersion relation between $\omega$ and $k$ is found by equating the determinant to zero. This gives two real equations in the two unknowns $\omega$ and $k$, namely, $$\label{eq:4}
\Big(\frac {\omega}{{\omega_c}}\Big)^2 \pm \, \frac\omega{\omega_c}\ + p\,
F(k{l},{l}\omega/c) = 0 ,
\quad p= \frac{{\omega_p}^2}{{\omega_c}^2}\ .$$ $$\label{eq:2}
\frac 23\, \frac{e^2}{4\pi\epsilon_0 {m_e}c^3} \, \omega^3 -
G(k{l},{l}\omega/c) = 0\ .$$ Here, ${\omega_p}$ is the familiar plasma frequency defined by $${\omega_p}^2=e^2/\epsilon_0 {m_e}{l}^3\ =n_e \, e^2/\epsilon_0 {m_e}\ ,$$ while $F$ and $G$, as functions of the variables $\alpha=k{l}$, $\beta={l}\omega/c $, are defined by $$\label{eq:3}
\begin{split}
F(\alpha,\beta) &=\frac 1{4\pi} \Big[ \beta^2\log \big(2\,
|\cos\beta-\cos\alpha| \,
\big)- f(\alpha,\beta)\Big]\\ G(\alpha,\beta)&=\beta^3 -g(\alpha,\beta) \ ,
\end{split}$$ the functions $f$ and $g$ being the ones already introduced in [@cg], namely, $$\begin{split}
f({\alpha},\beta)=\sum_{n\neq 0}(\frac{\cos(n{\alpha}-|n|
\beta)}{|n^3|}-\beta \frac{\sin(n{\alpha}-|n| \beta)}{|n^2|}) \\
g({\alpha},\beta)=\sum_{n\neq 0}(\frac{\sin(n{\alpha}-|n| \beta)}{|n^3|} +
\beta \frac{\cos(n{\alpha}-|n| \beta)}{|n^2|})\ .
\end{split}$$ Some details concerning the summation of the series leading to the term $\beta^2\log \big(2\, |\cos\beta-\cos\alpha|\big) $ entering the function $F$ are here omitted.
Now, one meets here with a deep question of principle. Indeed, for fixed values of the parameters ${l}$ and ${\omega_c}$ one has two equations in two unknowns ($\omega$ and $k$), and this would not allow for the existence of a dispersion relation, i.e., of a function $\omega=\omega(k)$ for a continuous range of values of $k$. However, the existence of a dispersion relation is guaranteed by the fact that equation (\[eq:2\]) actually is an identity. In fact, this is a particular case of a general identity, conceived by Wheeler and Feynman [@wf] and first proven in [@cg] (see section 6) for a one-dimensional case and in [@mcg] for a three-dimensional one.
So, the problem of obtaining the dispersion relation is reduced to solving (\[eq:4\]) in the unknown $\omega=\omega(k)$, in which $p=\omega_p^2/\omega_c^2$ plays the role of a parameter. In Fig. \[fig:3\] the dispersion relations are shown for a cyclotron frequency ${\omega_c}=3.8\cdot 10^{11} Hz$, and for several values of the parameter $p=\omega_p^2/{\omega_c}^2$ (or of the corresponding electron density $n_e/ n_G$, normalized to the Greenwald density $n_G\equiv I_p/\pi {a}^2$).
![The dispersion curves, solutions of equation (\[eq:4\]) (with the plus sign), in the plane ($k{l}$, $\omega/{\omega_c}$) for ${\omega_c}=3.8 \times 10^{11}$ Hz, and for several values of the parameter $p=\omega_p^2/{\omega_c}^2$ (or equivalently of the electron density $n_e$, normalized to the empirical $n_G$ limit).[]{data-label="fig:3"}](figura_3.eps){width="1.\columnwidth"}
The most important qualitative result is that normal modes are found to exist (for all $k$) only below a critical value of $p$, i.e., below a certain threshold of plasma density. Indeed, starting up from low densities, at a certain critical density a bifurcation is seen to occur, characterized by the fact that the curves no more intersect the vertical axis $k{l}=\pi$. This means that for values of $k$ just below $\pi/{l}$ equation (\[eq:4\]) does not admit a real solution, so that the corresponding frequencies acquire an imaginary part, and the whole system becomes unstable. Numerical computations not reported here show that the characteristic time of the instability is of the order of $2\pi/{\omega_c}$ and that above the critical density the Wheeler and Feynman identity is no more satisfied.
Notice that this phenomenon of the existence of a maximal allowed density is obviously lost if one introduces the continuum approximation, i.e., is a characteristic feature of the discrete structure of matter. Indeed, following [@mcg], the continuum approximation corresponds to deal with wavelengths much larger than the step ${l}$, i.e., to assume $k\ll \pi/{l}$, whereas the existence of a density limit depends on the behavior of the system for $k{l}\simeq \pi$. We have now to determine the bifurcation value of the parameter $p$. As the bifurcation occurs for $k{l}=\pi$ and for values of $\omega/{\omega_c}\leq 1$, i.e., for ${l}\omega/c\simeq 0$, one can just limit oneself to study equation (\[eq:4\]) for a fixed value of the function $F$, namely $F(\pi, 0)$, so that one is simply reduced to deal with an algebraic equation of second degree. One computes $F(\pi,
0)\simeq0.14$, and so real values of $\omega$ are found to exist only for $p\lesssim1.74$. This, together with the definition of $p$ in (\[eq:4\]) and $\epsilon_0 \mu_0=1/c^2 $, gives law (\[legge\]).
Notice that law (\[legge\]) has the same form of the Brillouin limit [@Brillouin], which is known to apply to the case of nonneutral plasmas [@Davidson]. The main difference with respect to our procedure is that in the case of the Brillouin limit the electric field acting on each electron is introduced within a mean field approach, whereas here it is computed in the frame of a many–body microscopic theory. Correspondingly, we find that the instability involves normal modes with wavelengths of the order of the mean electron distance, so that it escapes a mean field approach. In particular, such an instability is found to occur in neutral plasmas, for which the mean charge density vanishes, and the Brillouin approach cannot be used.
A final comment concerns the possibility of dealing with the other main magnetic configuration studied for the confinement of fusion-relevant plasmas, i.e., the Stellarator [@Boozer]. The present model does not directly apply. Indeed, in the Stellarator a large amount of power is typically transferred to the electrons through electron cyclotron resonance heating (ECRH), and this requires to add in our model a forcing term.
In conclusion, through an extremely simplified model of a magnetized plasma suited for a tokamak, based on first principles, we have proved the existence of a density limit, beyond which the system becomes unstable. The law thus found differs from the usually accepted one, and this fact might have relevant implications for future tokamaks.
*The authors wish to thank Dr. Nicola Vianello for fruitful discussions.\
This work, supported by the European Communities under the contract of Association between EURATOM/ENEA, was carried out within the framework the European Fusion Development Agreement.*
Greenwald, M., Density limits in toroidal plasmas, *Plasma Phys. Control. Fusion* **44**, R27–R80 (2002). Carati, A. and Galgani, L., Nonradiating normal modes in a classical many-body model of matter-radiation interaction, *Nuovo Cim.* **118 B**, 839-849 (2003). Marino, M., Carati, A, Galgani, L., Classical Light dispersion theory in a regular lattice, *Annals of Physics* **322**, 799-823 (2007). Dirac, P.A.M., Classical Theory of Radiating Electrons, *Proc. Royal Soc. London* **167**, 148–169 (1938). Marino, M., Classical electrodynamics of point charges, *Annals of Physics* **301**, 85–127 (2002). Jackson, J.D.,*Classical Electrodynamics*, New York, J. Wiley and Sons (1975). Wheeler, J.A. and Feynman, R.P., Interaction with the absorber as the mechanism of radiation, *Rev. Mod. Phys.* **17**, 157-181 (1945). Miyamoto, K., *Plasma Physics and Controlled Nuclear Fusion*, Berlin, Springer (2005). Stabler, Ä. *et al.*, Density limit investigations on ASDEX, *Nucl. Fusion* **32**, 1557 (1992). de Vries, P.C., Rapp, J., Schüller, F.C. and Tokar, M.Z., Influence of recycling on the density limit in TEXTOR-94, *Phys. Rev. Lett.* **80**, 3519-3522 (1998). Frigione, D. *et al.*, High density operation on Frascati Tokamak Upgrade, *Nucl. Fusion* **36**, 1489 (1999). Merezhkin, V.G., Electron energy balance near the density limit in T-10 and FTU OH regimes, *33rd EPS Conference on Plasma Phys.,* Rome, 19-23 June 2006 ECA Vol. 30I, P-4.085 (2006). Merthens, V. *et al.*, High density operation close to Greenwald limit and H Mode limit in ASDEX UPGRADE, *Nucl. Fusion* , 1607-1614 (1997). Howard, J. and Person, M., Cold bubble formation during tokamak density limit disruptions, *Nucl. Fusion* **32**, 361-377 (1992). Dyabilin, K. S. *et al.*, Global energy balance and density limit on CASTOR tokamak, *Czech. J. Phys.* **B 37**, 713-724 (1987). Asif, M. *et al*, Study of recycling and density limit in the HT-7 superconducting tokamak, *Phys. Lett. A* **336**, 61-65 (2005). LaBombard, B. *et al.*, Particle transport in the scrape-off layer and its relationship to discharge density limit in Alcator C-MOD, *Phys. Plasmas* **8**, 2107-2117 (2001). Takenaga, H. *et al.*, Compatibility of advanced tokamak plasma with high density and high radiation loss operation in JT-60U, *Nucl. Fusion* **45**, 1618-1627 (2005). Saibene, G. *et al.*, The influence of isotope mass, edge magnetic shear and input power on high density ELMy H modes in JET, *Nucl. Fusion* **39**, 1133-1156 (1999). Petrie, T.W., Kellman, A.G. and Mahdavi, M.Ali, Plasma density limits during ohmic L mode and elming H Mode operation in DIII-D, *Nucl. Fusion* , 929-952 (1993). The ITER physics basis, *Nucl. Fusion* **47**, S1-S413 (2007). Granetz, R.S., Density threshold for magnetohydrodynamic activity in Alcator C, *Phys. Rev. Lett.* **49**, 658-661 (1982). Lorenzini, R. *et al*, Self-organized helical equilibria as a new paradigm for ohmically heated fusion plasmas, *Nature Phys.* **5**, 570-574 (2009). Valisa, M. *et al.*, The Greenwald density limit in the Reversed Field Pinch, *IAEA-CN-116/EX/P4-13, 20th IAEA Fusion Energy Conference*, 1-6 November 2004, Vilamoura, Portugal. Goldston, R.J. and Rutherford, P.H., *Introduction to Plasma Physics*, Bristol, IOP Publishing (1995). Chandrasekhar, S., *Hydrodynamic and Hydromagnetic Stability*, Oxford, Clarendon Press (1961). Brillouin, L., A theorem of Larmor and its importance for electrons in magnetic fields, *Phys. Rev.* **67**, 260-266 (1945). Davidson, R.C., *Physics of Nonneutral Plasmas*, Redwood City, Addison–Wesley (1990). Boozer, A.H., What is a stellarator?, *Phys. Plasmas* **5**, 1647-1655 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper considers a generalized framework to study OSNR optimization-based end-to-end link level power control problems in optical networks. We combine favorable features of game-theoretical approach and central cost approach to allow different service groups within the network. We develop solutions concepts for both cases of empty and nonempty feasible sets. In addition, we derive and prove the convergence of a distributed iterative algorithm for different classes of users. In the end, we use numerical examples to illustrate the novel framework.'
author:
-
-
- 'Quanyan Zhu, Lacra Pavel, [^1]'
bibliography:
- 'IEEEabrv.bib'
- 'Xbib.bib'
title: Enabling Differentiated Services Using Generalized Power Control Model in Optical Networks
---
Introduction
============
Reconfigurable optical Wavelength-Division Multiplexing (WDM) communication networks with arbitrary topologies are currently enabled by technological advances in optical devices such as optical add/drop MUXes (OADM), optical cross connects (OXC) and dynamic gain equalizer (DGE). It is important that channel transmission performance and quality of service (QoS) be optimized and maintained after reconfiguration. At the physical transmission level, channel performance and QoS are directly determined by the bit-error rate (BER), which in turn depends on optical signal-to-noise ratio (OSNR), dispersion and nonlinear effects, [@ARG05]. Thus, OSNR is considered as the dominant performance parameter in link-level optimization. Conventional off-line OSNR optimization is done by adjusting channel input power at transmitter (Tx) to equalize the dominant impairment of noise accumulation in chains of optical amplifiers. However, for reconfigurable optical networks, where different channels can travel via different optical paths, it is more desirable to implement on-line decentralized iterative algorithms to accomplish such adjustment.
Recently, this problem is addressed in many research works [@PAV06b],[@PAV06a],[@PAN05], and two optimization-based approaches are prevalently used: the central cost and the non-cooperative game approach. The goals and models of the two approaches are inherently different. Central cost approach satisfies the target OSNR with minimum total power consumption. The model embeds the OSNR requirements in its constraints and indirectly optimizes a certain design criterion. Such model yields a relatively simple closed-form solution; however, it doesn’t optimize OSNR in a direct fashion, and thus, channel performance can be potentially improved for users who need higher quality of transmission. On the other hand, the game approach is a naturally distributed model which directly optimizes OSNR based on a payoff function in a non-cooperative manner. Each user optimizes her own utility to achieve the best possible OSNR. The solution from this approach is given by Nash equilibrium. As a result, this solution concept yields best achievable OSNR levels for each user. Since the game approach involves a cost function arising from pricing, it gives an over-allocation of resources. Some users may wish to avoid such cost and only demand a basic level of transmission. Apparently, these two approaches are for two different type of users and different transmission purposes.
To make use of the advantages from each approach, we propose a generalized model that combines their features. Such a generalization allows to accommodate different types of users and also provides a novel mixed framework to study OSNR power control problem. We separate users into two different categories. One type of users are those who are willing to pay a price to fully optimize their transmission performance. Another type of users are those who are content with basic transmission quality, or OSNR level, set by the network. The quality of service (QoS) can be met for the former by a game-theoretically based optimization approach; and for the later by a mechanism similar to central cost approach.
The contribution of this paper lies in the capability of service differentiation of the generalized model. For simplicity, total capacity constraints are not considered. The paper is organized as follows. In section 2, we review the network OSNR model and the basic concepts about the two optimization-based approaches. In section 3, we establish a general framework and propose two solution concepts for two different cases of feasible sets. Section 4 gives an iterative algorithm to achieve such solutions in the framework. This is illustrated in section 5 by numerical examples. Section 6 concludes the paper and points out future directions of research.
Background
==========
Review of Optical Network Model
-------------------------------
Consider a network with a set of optical links $\mathcal{L}=\{1,2,..,L\}$ connecting the optical nodes, where channel add/drop is realized. A set $\mathcal{N}=\{1,2,...,N\}$ of channels are transmitted, corresponding to a set of multiplexed wavelengths. Illustrated in Figure \[oplink\], a link $l$ has $K_l$ cascaded optically amplified spans. Let $N_l$ be the set of channels transmitted over link $l$. For a channel $i \in \mathcal{N}$, we denote by $\mathcal{R}_i$ its optical path, or collection of links, from source (Tx) to destination (Rx). Let $u_i$ be the $i$th channel input optical power (at Tx), and $\textbf{u}=[u_1,...,u_N]^T$ the vector of all channels’ input powers. Let $s_i$ be the $i$th channel output power (at Rx), and $n_i$ the optical noise power in the $i$th channel bandwidth at Rx. The $i$th channel optical OSNR is defined as $OSNR_i=\frac{s_i}{n_i}$. In [@PAV06b], some assumptions are made to simplify the expression for OSNR, typically for uniformly designed optical links:
1. (A1) Amplified spontaneous emission (ASE) noise power does not participate in amplifier gain saturation.
2. (A2) All the amplifiers in a link have the same spectral shape with the same total power target and are operated in automatic power control mode.
Under A1 and A2, dispersion and nonlinearity are considered to be limited, and ASE noise accumulation will be the dominant impairment. The OSNR for the $i$th channel is given as $$\label{OSNR}
OSNR_i=\frac{u_i}{n_{0,i}+\sum_{j\in
\mathcal{N}}\Gamma_{i,j}u_j}, i\in\mathcal{N}$$ where $\mathbf{\Gamma}$ is the full $n \times n$ system matrix which characterizes the coupling between channels. $n_{0,i}$ denotes the $i$th channel noise power at the transmitter. System matrix $\mathbf{\Gamma}$ encapsulates the basic physics present in optical fiber transmission and implements an abstraction from a network to an input-output system. This approach has been used in [@SAAB02] for the wireless case to model CDMA uplink communication. Different from the system matrix used in wireless case, the matrix $\mathbf{\Gamma}$ given in (\[eqnGamma\]) is commonly asymmetric and is more complicatedly dependent on parameters such as spontaneous emission noise, wavelength-dependent gain, and the path channels take. $$\label{eqnGamma}
\Gamma_{i,j}=\sum_{i\in\mathcal{R}_i}\sum_{k=1}^{K_l}\frac{G_{l,j}^{k}}{G_{l,i}^{k}}
\left(\prod_{q=1}^{l-1}\frac{\mathbf{T}_{q,j}}{\mathbf{T}_{q,i}}\right)\frac{ASE_{l,k,i}}{P_{o,l}},
\forall j\in \mathcal{N}_l.$$ where $G_{l,k,i}$ is the wavelength dependent gain at $k$th span in $l$th link for channel $i$; $\mathbf{T}_{l,i}=\prod_{q=1}^{K_l}G_{l,k,i}L_{l,k}$ with $L_{l,k}$ being the wavelength independent loss at $k$th span in $l$th link; $ASE_{l,k,i}$ is the wavelength dependent spontaneous emission noise; $P_{0,l}$ is the output power at each span.
![A Typical Optical Link in DWDM Optical Networks []{data-label="oplink"}](oplink.eps "fig:")\
Central Cost Approach
---------------------
Similar to the SIR optimization problem in the wireless communication networks [@SMG02; @SG01], OSNR optimization achieves the target OSNR predefined by each channel user by allowing the minimum transmission power. Let $\gamma_i, i\in \mathcal{N}$ be the target OSNR for each channel. By setting the OSNR requirement as a constraint, we can arrive at the following central cost optimization problem (CCP): $$\label{CCAOpt}
\begin{tabular}{ll}
$\textrm{(CCP) }$&$\min_\mathbf{u} \sum_{i\in \mathcal{N}}u_i$ \\
$\textrm{subject to }$& $OSNR_i \ge \gamma_i \texttt{ } \forall
i\in \mathcal{N}.$ \\
\end{tabular}$$ Under certain conditions, it has been shown in [@PAV06b] that the feasible set of (CCP) is nonempty and the optimal solution is achievable at the boundary of the feasible set.
The formulated optimization problem can be extended to incorporate more constraints such as $$\label{CCAConst3}
u_{i,\min} \leq u_i
\leq u_{i,\max},$$ where $u_{i,\min}$ is minimum threshold power required for transmission for channel $i$ and $u_{i,\max}$ is maximum power channel $i$ can attain. In the central cost approach, power $u_i$ are the parameters to be minimized and the objective function is linearly separable. In addition, the constraints are linearly coupled. These nice characteristics in central cost approach leads to a relatively simple optimization problem.
Non-cooperative Game Approach
-----------------------------
Let’s review the basic game-theoretical model for power control in optical networks without constraints. Consider a game defined by a triplet $\langle
\mathcal{N},(A_i),(J_i)\rangle$. $\mathcal{N}$ is the index set of players or channels; $A_i$ is the strategy set $\{u_i\mid u_i\in
[u_{i,\min},u_{i,\max}]\}$; and, $J_i$ is the cost function. It is chosen in a way that minimizing the cost is related to maximizing OSNR level. In [@PAV06a], $J_i$ is defined as $$\label{GAUtil}
J_i(u_i,u_{-i})=\alpha_i u_i - \beta_i \ln
\left(1+a_i\frac{u_i}{X_{-i}}\right), i\in\mathcal{N}$$ where $\alpha_i,\beta_i$ are channel specific parameters, that quantify the willingness to pay the price and the desire to maximize its OSNR, respectively, $a_i$ is a channel specific parameter, $X_{-i}$ is defined as $X_{-i}=\sum_{j \neq
i} \Gamma_{i,j}u_j+n_{0,i}$. This specific choice of utility function is non-separable, nonlinear and coupled. However, $J_i$ is strictly convex in $u_i$ and takes a specially designed form such that its first-order derivative is linear with respect to $\mathbf{u}$.
The solution from the game approach is usually characterized by Nash equilibrium (NE). Provided that $\sum_{j\neq i}\Gamma_{i,j}< a_i$, the resulting NE solution is uniquely determined in a closed form by $$\label{GASoln}
\mathbf{\widetilde{{\Gamma}} u^*=\widetilde{b}},$$ where $\widetilde{\Gamma}_{i,j}=a_i,$ for $j=i$; $\widetilde{\Gamma}_{i,j}=\Gamma_{i,j},$ for $j \neq i$ and $\widetilde{b}=\frac{a_i\beta_i}{\alpha_i}-n_{0,i}$. Similar to the wireless case [@SAAB02], we are able to construct iterative algorithms to achieve the Nash equilibrium. A simple deterministic first order parallel update algorithm is: $$\label{GAAlg2}
u_i(n+1)=\frac{\beta_i}{\alpha_i}-\frac{1}{a_i}\left(\frac{1}{OSNR_i(n)}-\Gamma_{i,i}\right)u_i(n).$$ As proved in [@PAV06a], the algorithm (\[GAAlg2\]) converges to Nash equilibrium $\mathbf{u}^*$ provided that $\frac{1}{a_i}\sum_{j\neq i}\Gamma_{i,j} < 1, \forall i$.
Generalized Model
=================
In this section, we consider a game designed to allow service differentiation by separating users into two groups: one group seeking a minimum OSNR target and another group participating in a game setting for OSNR optimization. The minimum OSNR for target seekers is set by the network to ensure the minimum quality of service. However, the game players can submit their parameters and optimize their service accordingly, but they have to pay a price set by the network for unit power consumption. This concept is illustrated in Figure \[GPTS\]. Let’s denote set $\mathcal{N}_1=\{1,2,...,N_1\}$ as the set of competitors, i.e. users who wish to compete for an optimal OSNR. Let set $\mathcal{N}_2=\{N_1+1,\cdots,N_2\}$ be the group of users with target OSNR given by $\gamma_i, i\in \mathcal{N}_2$. Let $\mathcal{N}=\mathcal{N}_1\cup\mathcal{N}_2$, $m=|\mathcal{N}_1|=N_1$, $n=|\mathcal{N}_2|$, $N=|\mathcal{N}|=m+n$ and $\mathbf{u}=[u_1,\cdots,u_{N_1},u_{N_1+1},\cdots,u_{N_2}]^T$.
![Game players and target seekers in the network[]{data-label="GPTS"}](GPTS2.eps "fig:")\
For the game-theoretical players, using the cost function given in (\[GAUtil\]), we can form a system of equations given by $$a_iu_i+X_{-i}=\frac{a_i\beta_i}{\alpha_i}, \forall i\in\mathcal{N}_1$$ and thus, $\widetilde{\mathbf{\Gamma}}\mathbf{u}=\widetilde{\mathbf{b}},$ where $\widetilde{\mathbf{\Gamma}}\in \mathcal{R}^{m\times N}$ and $\widetilde{\mathbf{b}}\in \mathcal{R}^{m}$ are defined as in (\[GASoln\]). Users with target OSNR shall have $\mathbf{u}$ satisfy $OSNR_i\geq \gamma_i, \forall i\in \mathcal{N}_2,$ or equivalently from (\[OSNR\]), $$\frac{u_i}{\Gamma_{i,i}u_i+\sum_{j\neq i}\Gamma_{i,j}u_j+n_{0,i}}\geq \gamma_i$$ and thus in a matrix form, $\widehat{\mathbf{\Gamma}}\mathbf{u}\geq\widehat{\mathbf{b}},$ where $\widehat{\mathbf{b}}=[\gamma_1n_{0,1},\cdots,\gamma_Nn_{0,N}]^T\in \mathcal{R}^{n}$, $\widehat{\mathbf{\Gamma}}\in \mathcal{R}^{n\times N}$ and is given in (\[GammaN2\]).
$$\label{GammaN2}
\widehat{\mathbf{\Gamma}}=\left[
\begin{array}{ccccc}
-\gamma_{N_1+1}\Gamma_{N_1+1,1} & \cdots & 1-\gamma_{N_1+1}\Gamma_{N_1+1,N_1+1} & \cdots & -\gamma_{N_1+1}\Gamma_{N_1+1,N} \\
\vdots & \ddots & \ddots & \ddots & \vdots \\
-\gamma_{N-1}\Gamma_{N-1,1} & -\gamma_{N-1}\Gamma_{N-1,2} & \cdots & 1-\gamma_{N-1}\Gamma_{N-1,N-1} & -\gamma_{N-1}\Gamma_N \\
-\gamma_N\Gamma_{N,1} & -\gamma_N\Gamma_{N,2} & \cdots & \cdots & 1-\gamma_N\Gamma_{N,N} \\
\end{array}
\right].$$
Let $F_1=\{\mathbf{u}\in\mathcal{R}^N\mid \widetilde{\mathbf\Gamma}\mathbf{u}=\widetilde{\mathbf{b}}\}$ and $F_2=\{\mathbf{u}\in\mathcal{R}^N\mid\widehat{\mathbf{\Gamma}}\mathbf{u}\geq\widehat\mathbf{b}\}$. In summary, we have a problem formulated as in (DS), where we find solutions that satisfy $F_1$ subject to the constraint described by $F_2$. $$\label{DS1}
\begin{array}{cc}
\textrm{(DS)} & \widetilde{\mathbf{\Gamma}}\mathbf{u}=\widetilde{\mathbf{b}}\\
\textrm{s.t.} &\widehat{\mathbf{\Gamma}}\mathbf{u}\geq\widehat{\mathbf{b}} \end{array}$$
In the following discussion, we separate (DS) into two cases: (1) $F=F_1\cap F_2\neq\emptyset$, (2)$F=F_1\cap F_2=\emptyset$, which require different techniques to find appropriate solutions.
Non-empty Feasible Set
----------------------
A non-empty $F$ may give rise to multiple points that solve (DS). We may impose some design criteria, or, objective function to reformulate DS for finding an appropriate solution that solves DS and meet the design criteria at the same time.
We can use the following result to ensure the nonempty feasible set $F$.
\[nonemptyF\] If $\overline\mathbf{\Gamma}=\left[
\begin{array}{c}
\widetilde\mathbf{\Gamma} \\
\widehat\mathbf{\Gamma} \\
\end{array}
\right]
$ is nonsingular, the feasible set $F=F_1\cap F_2$ is non-empty.
Let $\mu\in\mathcal{R}^n_+$ a nonnegative vector. Equivalently, we can express $F_2$ into $F_2=\{\mathbf{u}\in\mathcal{R}^n\mid \widehat\mathbf\Gamma \mathbf{u}=\widehat\mathbf{b}+\mu,\textrm{~for~some~}\mu\in\mathcal{R}^n_+\}$. The set $F$ is thus equivalently $F=\{\mathbf{u}\in\mathcal{R}^N\mid \overline{\mathbf\Gamma}\mathbf{u}=\mathbf{\phi},\textrm{~for~some~}\mu\in\mathcal{R}^n_+\}$, where $\overline\mathbf{\Gamma}=\left[
\begin{array}{c}
\widetilde\mathbf{\Gamma} \\
\widehat\mathbf{\Gamma} \\
\end{array}
\right]
$ and $\mathbf\phi=\left[
\begin{array}{c}
\widetilde\mathbf{b} \\
\widehat\mathbf{b}+\mu \\
\end{array}
\right]
$. If $\overline\mathbf{\Gamma}$ is nonsingular, there exist a unique $\mathbf{u}\in\mathcal{R}^N$ for every nonnegative $\mu$. Therefore $F$ is non-empty.
Suppose conditions in Theorem \[nonemptyF\] hold and $F$ is nonempty. We consider an appropriate solution in $F$ that satisfies a certain design criteria. Thus, we formulate (DSNP[^2]) in which we minimize total power consumption subject to the conditions arising from the different service requirements.
$$\label{DSNP}
\begin{array}{cc}
\textrm{(DSNP)} & \min \sum_iu_i\\
\textrm{s.t.} & \widetilde{\mathbf{\Gamma}}\mathbf{u}=\widetilde{\mathbf{b}},\widehat{\mathbf{\Gamma}}\mathbf{u}\geq\widehat{\mathbf{b}}
\end{array}$$
The constraints of (DSNP) can be relaxed and augmented into $$\label{relaxedEq}
\overline\mathbf{\Gamma}\mathbf{u}\geq\overline\mathbf{b}.$$ where $\overline{\mathbf{\Gamma}}=\left[
\begin{array}{c}
\widetilde{\mathbf{\Gamma}} \\
\widehat{\mathbf{\Gamma}} \\
\end{array}
\right]
\in\mathcal{R}^{N\times N}$ and $\overline{\mathbf{b}}=\left[
\begin{array}{c}
\widetilde{\mathbf{b}} \\
\widehat{\mathbf{b}} \\
\end{array}
\right]
\in\mathcal{R}^{N}.$
According to the fundamental theorem of linear programming [@BER03], if (DSNP) is realistic, the solution is obtained at the extreme point of the feasible set $F$. Since $F$ has only one extreme point when $\overline\mathbf\Gamma$ is non-singular, the solution is uniquely given by $$\label{eqSoln}
\mathbf{u}=\overline\mathbf{\Gamma}^{-1}\overline\mathbf{b}.$$
To further characterize the solution $\mathbf{u}$, we assume strict diagonal dominance of matrix $\overline\mathbf{\Gamma}$ [@Horn90], which leads to non-singularity of the matrix and uniqueness of the solution.
\[StrictDom\] Suppose OSNR targets $\gamma_i, i\in\mathcal{N}_2$ are chosen such that $\gamma_i<\frac{1}{\sum_{j\in\mathcal{N}}\Gamma_{i,j}}, i\in\mathcal{N}_2$. In addition, parameters $a_i$ are chosen as $a_i>\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{ij}, \forall i\in\mathcal{N}_1.$ The matrix $\overline\mathbf{\Gamma}$ is strictly diagonally dominant. And thus, a unique solution to (DSNP) is given by (\[eqSoln\]).
From the assumption that $\gamma_i\sum_{j\in\mathcal{N}}\Gamma_{ij}<1,i\in\mathcal{N}_2$, it is apparent that $\gamma_i<\frac{1}{\Gamma_{ii}}$ and $\left|1-\gamma_i\Gamma_{ii}\right|>\gamma_i\sum_j\Gamma_{ij}, \forall i\in\mathcal{N}_2$. In addition, $a_i>\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}, \forall i\in\mathcal{N}_1$. Therefore, matrix $\overline\mathbf{\Gamma}$ is strictly diagonally dominant. Using Gershgorin theorem in [@Horn90], we conclude that there exists a unique solution to (DSNP).
The assumption of strict diagonal dominance in Theorem \[StrictDom\] is reasonable because typical values of $\Gamma_{ij}$ are found to be on the order of $10^{-3}$ and desirable levels of OSNR are 20-30dB.
(DSNP) can be seen as a generalized approach that combines central cost approach in [@PAV06b] and non-cooperative game approach in [@PAV06a]. When $N_1=\emptyset, N_2\neq\emptyset$, (DSNP) reduces to the central cost approach. Similarly, when $N_1\neq\emptyset, N_2=\emptyset$, (DSNP) reduces to the game-theoretical approach and the given solution is Nash equilibrium accordingly. This framework allows to study two different types of users at the same time.
We illustrate a two-person (DSNP), where player 1 chooses to compete and optimize his utility and player 2 chooses to meet a certain OSNR target $\gamma_2$. We form the 2-by-2 matrix $\overline{\mathbf{\Gamma}}$ and $\overline\mathbf{b}$ as follows. $$\overline{\mathbf{\Gamma}}=\left[
\begin{array}{cc}
a_1 & \Gamma_{12} \\
-\Gamma_{21}\gamma_2 & 1-\Gamma_{22}\gamma_2 \\
\end{array}
\right],
\overline{\mathbf{b}}=\left[
\begin{array}{c}
\frac{a_1\beta_1}{\alpha_1}-n_{0,1} \\
n_{0,2}\gamma_2 \\
\end{array}
\right]$$
The feasible set $F=F_1\cap F_2$ is shown in Figure \[DiffServGamma2\] by a dotted line. The relaxed (DSNP) has its relaxed feasible depicted in the shaded region. The solution is given by $\mathbf{u}^*=\overline{\mathbf{\Gamma}}^{-1}\overline{\mathbf{b}},$ which is illustrated by the dark point in Figure \[DiffServGamma2\]. $\mathbf{u}^*$ is nonnegative componentwise if network price $\alpha_1$ is set such that $s_2>\frac{n_{0,2}}{1-\Gamma_{22}}$.
![The feasible set of two-person (DSNP). $s_1=\frac{\tilde{b}_1}{a_1}$; $s_2=\frac{\tilde{b}_1}{\Gamma_{12}}$[]{data-label="DiffServGamma2"}](DiffServGamma2new.eps "fig:")\
Based on Theorem \[StrictDom\], we can further investigate how parameters chosen by game players and target seekers influence the outcome of the allocation. The result is summarized in Theorem \[varControl\].
\[varControl\] Let $\kappa$ be the condition number of $\overline{\mathbf{\Gamma}}$, $T_i=a_i+\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{ij}, \forall i\in\mathcal{N}_1$ and $S_k=2-2\gamma_k\Gamma_{kk}, \forall k\in\mathcal{N}_2$. Suppose $\overline\mathbf{\Gamma}$ is strictly diagonally dominant by satisfying conditions in Theorem \[StrictDom\]. In addition, $T_i>S_k$ and $\tilde{b}_i>\hat{b}_k, \forall i\in\mathcal{N}_1, \forall k\in\mathcal{N}_2.$ The maximum allocated power allocated to users are bound as follows. $$\frac{\max_{i\in\mathcal{N}_2}\gamma_in_{0,i}}{\max_{i\in\mathcal{N}_1}2a_i}\leq\|\mathbf{u}\|_\infty\leq\kappa\max_{i\in\mathcal{N}_1}\frac{\beta_i}{\alpha_i}$$
Let $R_i$ denote the i-th row absolute sum of matrix $\overline\mathbf{\Gamma}$, i.e., $$\label{Row}
R_i=\sum_{j\in\mathcal{N}}\left|\overline{\Gamma}_{ij}\right|.$$ Using conditions from Theorem \[StrictDom\], we arrive at $$\label{Ri}
R_i=\left\{
\begin{array}{ll}
1+\gamma_i\sum_{j\in\mathcal{N}}\Gamma_{ij}-2\gamma_i\Gamma_{ii}<2-2\gamma_i\Gamma_{ii}, & {i\in\mathcal{N}_2;} \\
a_i+\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{ij}<2a_i., & {i\in\mathcal{N}_1.}
\end{array}
\right.$$
With the assumption that $a_i+\sum_{j\neq i}\Gamma_{ij}> 2-2\gamma_k\Gamma_{kk}, \forall i\in\mathcal{N}_1, \forall k\in\mathcal{N}_2,$ we obtain $\|\overline\mathbf{\Gamma}\|_\infty=\max_{i\in\mathcal{N}}R_i=\max_{i\in\mathcal{N}}a_i+\sum_{j\neq i}\Gamma_{ij}.$ Using (\[Ri\]) and the fact that $\Gamma_{ij}\geq0$, we obtain an upper and lower bound on $\|\overline{\mathbf{\Gamma}}\|_\infty$, i.e., $$\label{ublbOGammaNorm}
\max_{i\in\mathcal{N}_1}a_i\leq \|\overline\mathbf{\Gamma}\|_\infty \leq\max_{i\in\mathcal{N}_1}2a_i.$$
In addition, from $\tilde{b}_i>\hat{b}_k, \forall i\in\mathcal{N}_1, \forall k\in\mathcal{N}_2,$ we obtain an upper bound and lower bound for $\|\overline\mathbf{b}\|_\infty$, given by $$\label{ublbObnorm}
\max_{i\in\mathcal{N}_2}\gamma_in_{0,i}\leq\|\overline\mathbf{b}\|_\infty=\max_{i\in\mathcal{N}}\overline{b}_i\leq\max_{i\in\mathcal{N}_1}\tilde{b}_i
=\max_{i\in\mathcal{N}_1}\frac{a_i\beta_i}{\alpha_i}$$ Since $\overline{\mathbf{\Gamma}}$ is strictly diagonally dominant, using matrix norm sub-multiplicativity, we obtain from (\[eqSoln\]) $$\label{ublbu}
\frac{\|\overline{\mathbf{b}}\|_\infty}{\|\overline{\mathbf{\Gamma}}\|_\infty}\leq\|\mathbf{u}\|_\infty\leq\frac{\kappa\|\overline{\mathbf{b}}\|_\infty}{\|\overline\mathbf{\Gamma}\|_\infty},$$ where $\kappa$ is the condition number of $\overline{\mathbf{\Gamma}}$ given by $\kappa=\|\overline{\mathbf{\Gamma}}\|_\infty\|\overline{\mathbf{\Gamma}}^{-1}\|_\infty\geq1.$
Using (\[ublbOGammaNorm\]), (\[ublbObnorm\]) and (\[ublbu\]), we obtain $$\begin{aligned}
\label{ublbu2}
\nonumber \frac{\max_{i\in\mathcal{N}_2}\gamma_in_{0,i}}{\max_{i\in\mathcal{N}_1}2a_i}\leq\|\mathbf{u}\|_\infty
&\leq&\frac{\kappa\max_{i\in\mathcal{N}_1}a_i\beta_i/\alpha_i}{\max_{i\in\mathcal{N}_1}a_i}\\
\nonumber &\leq&\frac{\kappa\max_{i\in\mathcal{N}_1}a_i\max_{i\in\mathcal{N}_1}\beta_i/\alpha_i}{\max_{i\in\mathcal{N}_1}a_i}\\
&\leq&\kappa\max_{i\in\mathcal{N}_1}\frac{\beta_i}{\alpha_i}.\end{aligned}$$
It is easy to observe that the upper bound is dependent on the parameters of the game players and the lower bound is dependent on the OSNR levels of target seeker and parameter $a_i$ of the game players. In essence, game players control the outcome of the model and the choice of OSNR target can only affect the lower bound. Such relation describes a fair scenario in which game players, who pay for their power at $\alpha_i$, have their choices of parameters $a_i,\beta_i$ to influence the network allocation.
Since $\|\mathbf{u}\|_\infty\leq\|\mathbf{u}\|_2\leq \sqrt{N}\|\mathbf{u}\|_\infty$, we can translate the result obtained in (\[ublbu2\]) directly into Euclidean norm, i.e., $$\label{ublbu2Euclid}
B_\infty^L\leq\|\mathbf{u}\|_2\leq \sqrt{N}B^U_\infty$$ where $B_\infty^U=\kappa\max_{i\in\mathcal{N}_1}\frac{\beta_i}{\alpha_i}$ and $B_\infty^L=\frac{\max_{i\in\mathcal{N}_2}\gamma_in_{0,i}}{\max_{i\in\mathcal{N}_1}2a_i}$.
By (\[ublbu2Euclid\]), we can see that the network can encourage uniform channel power distribution by letting $B_\infty^U$ close to $\sqrt{N}B_\infty^L$ and provide incentive for differentiated services by letting them far apart. It can be implemented by the network by adjusting OSNR level $\gamma_i$ and pricing $\alpha_i$. Decreasing $\alpha_i$ encourages more users to be game players, giving rise to more competitions or service differentiation as a result of higher upper bound. On the other hand, increasing $\gamma_i$ raises the lower bound and encourages more users being target-seekers.
Empty Feasible Set
------------------
In this section, we consider the second case where feasible set $F$ is empty. Instead of finding an appropriate feasible solution, we find the closest points between set $F_1$ and $F_2$. We use a quadratic program (DS2) to minimize the error norm subject to the constraint described by $F_2$.
$$\label{DS2}
\begin{array}{cc}
\textrm{(DS2)} & \min_{\mathbf{u}}\|\widetilde{\mathbf{\Gamma}}\mathbf{u}-\widetilde{\mathbf{b}}\|_2\\
\textrm{s.t.} &\widehat{\mathbf{\Gamma}}\mathbf{u}\geq\widehat{\mathbf{b}} \end{array}$$
We can turn the constrained problem (\[DS2\]) into an unconstrained problem by studying its corresponding dual problem. Since $\|\widetilde{\mathbf{\Gamma}}\mathbf{u}-\widetilde{\mathbf{b}}\|_2
=\mathbf{u}^T\widetilde{\mathbf{\Gamma}}^T\widetilde{\mathbf{\Gamma}}\mathbf{u}
-2(\widetilde{\mathbf{b}}^T\widetilde{\mathbf{\Gamma}})\mathbf{u}+\widetilde{\mathbf{b}}^T\widetilde{\mathbf{b}}
$, we denote $\mathbf{H}=\frac{1}{2}\widetilde\mathbf{\Gamma}^T\widetilde\mathbf{\Gamma}, \mathbf{d}=-2(\widetilde\mathbf{\Gamma}^T\widetilde\mathbf{b})$, $\mathbf{D}=-\widehat\mathbf{\Gamma}(\mathbf{H}^T\mathbf{H})^{-1}\mathbf{H}^T\widehat\mathbf\Gamma^{T}$, $\mathbf{c}=\widehat\mathbf{b}+\widehat\mathbf{\Gamma}(\mathbf{H}^T\mathbf{H})^{-1}\mathbf{H}^T\mathbf{d}$; and form a Lagrangian from the original problem (DS2). $$\begin{aligned}
\label{dualL}
% \nonumber to remove numbering (before each equation)
D(\mu) &=& \min_{\mathbf{u}}\mathcal{L}(\mathbf{u},\mu) \\
\nonumber &=& \min_{\mathbf{u}}\left(
\frac{1}{2}\mathbf{u}^T\mathbf{H}\mathbf{u}+\mathbf{d}^T\mathbf{u}+\widetilde\mathbf{b}^T\widetilde\mathbf{b}
+\mu^T(-\widehat\mathbf\Gamma \mathbf{u}+\widetilde\mathbf{b})
\right)\end{aligned}$$
Since the objective function is convex, the necessary and sufficient condition for a minimum is that the gradient must vanish,i.e., $$\label{DualImplicit}
\mathbf{H}\mathbf{u}+\mathbf{d}-\hat\mathbf{\Gamma}^T\mathbf{\mu}=0.$$
For $n<N$, $\widetilde{\mathbf{\Gamma}}$ is not full rank. Therefore, $\mathbf{H}$ is singular and there exist multiple solutions to (\[DualImplicit\]). Using pseudoinverse [@Horn90], we can find a solution to (\[DualImplicit\]) given by $$\mathbf{u}=-(\mathbf{H}^T\mathbf{H})^{-1}\mathbf{H}^T\left(\mathbf{d}-\hat\mathbf{\Gamma}^T\mu\right).$$ Thus, after replacing into (\[dualL\]), we obtain $\mu$ as a solution to the dual problem (DDS2). $$\label{DDS2}
\textrm{(DDS2)} \max_{\mu\geq0} \frac{1}{2}\mu^T\mathbf{D}\mu+\mu^T\mathbf{c}-\frac{1}{2}\mathbf{d}^T(\mathbf{H}^T\mathbf{H})^{-1}\mathbf{H}^T\mathbf{d}+\mathbf{b}^T\mathbf{b}$$
The problem (LDS2) and dual problem (DDS2) can be solved using unconstrained optimization algorithms in [@BAZ93], [@BER03].
Iterative Algorithm
===================
In this section, we develop algorithm for the case of nonempty $F$ set. Let $u_i(n)$ denote the power at channel $i$ at step $n$. An iterative algorithm is given as follows. $$\label{GenItAlg1}
\left\{
\begin{array}{ll}
u_i(n+1)=\frac{\beta_i}{\alpha_i}-\frac{1}{a_i}\left(\frac{1}{OSNR_i(n)}-\Gamma_{i,i}\right)u_i(n), & \forall i\in\mathcal{N}_1; \\
u_i(n+1)=\frac{\gamma_i}{1-\gamma_i\Gamma_{i,i}}\left(\frac{1}{OSNR_i(n)}-\Gamma_{i,i}\right)u_i(n), & \forall i\in\mathcal{N}_2.
\end{array}
\right.$$
Algorithm (\[GenItAlg1\]) converges provided that $a_i>\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}$ and $\gamma_i$ is chosen such that $\gamma_i<\frac{1}{\sum_{j\in\mathcal{N}}\Gamma_{i,j}}$.
We use a similar approach from [@PAV06a] to show the convergence of (\[GenItAlg1\]). Let’s define $e_i(n)=u_i(n)-u_i^*$, where $u_i^*$ is given in (\[eqSoln\]). Since $\overline\mathbf{\Gamma}\mathbf{u}^*=\overline\mathbf{b}$, $\widetilde{\Gamma}_{i,i}u_i^*+\sum_{j\neq i}\widetilde{\Gamma}_{i,j}u_j^*=\tilde{b}_i$, for $i\in\mathcal{N}_1$; and, $\widehat{\Gamma}_{i,i}u_i^*+\sum_{j\neq
i}\widehat{\Gamma}_{i,j}u_j^*=\hat{b}_i$, for $i\in\mathcal{N}_2$.
Substitute the expression for $u_{i}^*$ into $e_i(n+1)$, and we obtain $e_i(n+1)=u_i(n+1)-u_i^*=-\frac{1}{{a_i}}\left[\sum_{j\neq i}{\Gamma}_{i,j}(u_j(n)-u_j^*)\right]$, for $i\in\mathcal{N}_1$; and $e_i(n+1)=u_i(n+1)-u_i^*=\frac{1}{1-{\Gamma}_{i,i}\gamma_i}\left[\sum_{j\neq i}{\Gamma}_{i,j}\gamma_i(u_j(n)-u_j^*)\right]$, for $i\in\mathcal{N}_2$. Let $\mathbf{e}=[e_i(n)],i\in\mathcal{N}$. Therefore, for $i\in\mathcal{N}_1$, $$\begin{aligned}
% \nonumber to remove numbering (before each equation)
|e_i(n+1)|
&=& \left|\frac{1}{a_i}\left[\sum_{j\neq
i,j\in\mathcal{N}}\Gamma_{i,j}(e_j(n))\right]\right| \\
&\leq & \frac{1}{a_i}\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}\max_{j\in\mathcal{N}}|e_j(n)|\\
&\leq & \frac{1}{a_i}\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}\|\mathbf{e}(n)\|_\infty.\end{aligned}$$ and similarly, for $i\in\mathcal{N}_2$, $$\begin{aligned}
% \nonumber to remove numbering (before each equation)
\nonumber |e_i(n+1)| &=& \left|\frac{1}{1-\Gamma_{i,i}\gamma_i}\left[\sum_{j\neq
i,j\in\mathcal{N}}\Gamma_{i,j}\gamma_i(e_j(n))\right]\right| \\
&\leq & \frac{\gamma_i}{|1-\Gamma_{i,i}\gamma_i|}\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}\max_{j\in\mathcal{N}}|e_j(n)|.\\
&\leq & \frac{\gamma_i}{|1-\Gamma_{i,i}\gamma_i|}\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}\|\mathbf{e}(n)\|_\infty.\end{aligned}$$ Since we assumed that $a_i>\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}$ and $\gamma_i$ is chosen such that $\gamma_i<\frac{1}{\sum_{j\in\mathcal{N}}\Gamma_{i,j}}\leq\frac{1}{\Gamma_{i,i}}$, we can conclude that $\|\mathbf{e}(n)\|\rightarrow 0$ from the contraction mapping theorem. As a result, we have $u_i(n)\rightarrow u_i^*$ as $n\rightarrow \infty$, for $i\in\mathcal{N}$.
From the proof, we note that the rate of convergence of \[GenItAlg\] is determined by $$\sigma=\max\left\{\max_{i\in\mathcal{N}_1}\frac{\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}}{a_i},\max_{i\in\mathcal{N}_2}\frac{\sum_{j\neq i,j\in\mathcal{N}}\Gamma_{i,j}\gamma_i}{1-\Gamma_{i,i}\gamma_i}\right\}.$$ In addition, it is easy to observe that the OSNR target-seeking users are algorithmically equivalent to competition seeking users by letting $\beta_i/\alpha_i=0$ and $a_i=\Gamma_{i,i}-\frac{1}{\gamma_i}$, $i\in\mathcal{N}_2$. This is because no notion of pricing is used for the OSNR target seekers and they just have a utility target to meet or equivalently optimize by letting $a_i=\Gamma_{i,i}-\frac{1}{\gamma_i}$.
Numerical Examples
==================
In this section, we illustrate the concept by a MATLAB simulation. We consider an end-to-end link described in Figure \[oplink\] with 5 amplified spans. We assume channels are transmitted at wavelengths distributed centered around 1555nm with channel separation of 1nm. Suppose input noise power is 0.5 percent of the input signal power. The gain profile for each amplifier is identically assumed to be parabolic as in Figure \[GainProfile\], which is normalized with respect to $G_{\max}=30.0$dB. Suppose 20dB is the target OSNR level for users who just want to meet a satisfactory level of transmission. We first show the case of 3 users, in which 2 users need better quality of service and one user is simply interested in meeting 20dB as a target. From Figure \[3userDS\], we can observe that users who need better services reach an OSNR of 26.33dB and 29.20dB, respectively. With an appropriate choice of initial conditions, the algorithm quickly converges in 1-2 steps. In Figure \[30userDS\], we similarly show the case of 30 users, in which 20 are game players and 10 are target seekers.
![Optical Amplifier Spectral Profile[]{data-label="GainProfile"}](GainProfile.eps "fig:")\
![OSNR simulation with 3 users in time steps[]{data-label="3userDS"}](3userDS.eps "fig:")\
![OSNR simulation with 30 users in time steps[]{data-label="30userDS"}](30userDS.eps "fig:")\
Conclusion
==========
In this paper, we examined a generalized power control model in optical networks, which combines features of central cost approach and game-theoretical approach. It enables two major service types in the network. One is game player, who pays for his power consumption and the other is target seeker, who is satisfied with a minimum service level set by the network. We discussed two different solutions concepts for nonempty and empty feasible set respectively and specifically designed an iterative algorithm that converges to a unique solution for the case of nonempty feasible set. The convergence of the algorithm was proved and illustrated by numerical examples of a WDM end-to-end optical link.
In this work, we didn’t include capacity constraints for the sake of simplicity. We hope this work will lead to future investigations of more complicated cases where network constraints and nonlinear effects are considered. In addition, we expect this framework to be used to solve similar problems in other types of networks, for example, wireless networks.
[^1]: Quanyan Zhu is with the Department of Electrical and Computer Engineering, University of Illinois at Urbana Champaign, IL, 61801, USA email: [email protected]; L. Pavel is with the Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON, M5S 3L1 Canada e-mail:[email protected].
[^2]: DSNP stands for “Differentiated Service N-person Problem”.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Ferroelectric relaxors are complex materials with distinct properties. The understanding of their dielectric susceptibility, which strongly depends on both temperature and probing frequency, have interested researchers for many years. Here we report a macroscopic and phenomenological approach based on statistical modeling to investigate and better understand how the dielectric response of a relaxor depends on temperature. Employing the Maxwell-Boltzmann distribution and considering temperature dependent dipolar orientational polarizability, we propose a minimum statistical model and specific equations to understand and fit numerical and experimental dielectric responses versus temperature. We show that the proposed formula can successfully fit the dielectric response of typical relaxors, including Ba(Zr,Ti)O$_{3}$, Pb(Zn$_{1/3}$Nb$_{2/3}$)$_{0.87}$Ti$_{0.13}$O$_{3}$, and Pb(Mg$_{1/3}$Nb$_{2/3}$)O$_{3}$-0.05Pb(Zr$_{0.53}$Ti$_{0.47}$)O$_{3}$, which demonstrates the general applicability of this approach.'
author:
- 'J. Liu'
- 'F. Li'
- 'Y. Zeng'
- 'Z. Jiang'
- 'D. Wang'
- 'Z.-G. Ye'
- 'C.-L. Jia'
title: Insight into dielectric response of ferroelectric relaxors by statistical modeling
---
Introduction
============
Relaxor ferroelectrics are materials that exhibit interesting dielectric responses different from normal ferroelectrics. For instance, they often possess relaxation modes at low frequency ($<1$GHz). Relaxors ferroelectrics have been exploited in many applications such as actuators due to their giant electromechanical couplings [@Uchino1996], and their properties extensively investigated, including structural properties (e.g., polar nanoregions or PNRs) using neutron scattering [@PhononLocalization], dielectric responses [@Bokov2006a; @Nuzhnyy2012; @Petzelt2014; @Kleeman2014; @Wang2014; @Wang2016], the crossover from ferroelectrics to relaxor [@Kleeman2014]. To understand such systems, many theories have been proposed [@Pirc1999; @Bokov2012; @Akbarzadeh2012; @Sherrington2013a; @Uchino2014; @Sherrington2014; @PhononLocalization; @Kleemann2015]. Relaxors are complex systems, to some extent similar to spin glasses [@Sherrington2013a; @Sherrington2014a; @Sherrington2014], in that their compositions are, without exception, made of complex oxides containing different ions and inevitably inhomogeneous. For instance, the B-site ions of typical relaxors Ba(Zr,Ti)O$_{3}$ (BZT) and Pb(Mg$_{1/3}$Nb$_{2/3}$)O$_{3}$(PMN) are randomly distributed.
The dielectric response of ferroelectric relaxors is the defining feature that differentiates them from normal ferroelectrics: (i) large susceptibilities at low frequency (GHz or lower) ; (ii) even more unusual, the characteristic temperature $T_{m}$, at which the susceptibility peaks strongly depends on the frequency of the probing *ac* electric field. In other words, the susceptibility, $\chi$, depends on both temperature, $T$, and the probing frequency, $\nu$. While such phenomena are well known experimentally [@Bokov2007BZT; @Bokov2012; @Gridnev2004; @Bokovo2006; @BZT-susceptibility; @Tagantsev1999; @PMN-BZT], numerical generation of relaxor’s dielectric response from model-based simulations has been a challenging work. For instance, the shift of $T_{m}$ of the lead free relaxor BZT was only numerically achieved recently[@Wang2016]. Since numerous ferroelectric relaxors exist, numerically treating each of them remains a daunting task. One way to mitigate this difficulty is to resort to statistical modeling [@Sethna2010]. For a complex system, a statistical approach can provide intuitive understanding by capturing dominant factors, provide equations to understand experimental results. and help extracting useful information. In the present work, we adopt this approach to treat the dielectric response of relaxors and show that such a statistical model can indeed be applied to understanding how the dielectric constants change with temperatures and probing frequency.
While the susceptibility of relaxors, $\chi\left(T,\nu\right)$, depends on both temperature and frequency, theoretical models are often proposed to treat $\nu$ and $T$ separately [@Uchino2014; @Bokov2012; @Cross1987a; @UniversalRelaxation; @DielRelaxSolids; @Nuzhnyy2012]. For instance, at a given temperature, two processes are employed in the fitting of $\chi\left(\nu\right)$ of Ba (Ti$_{0.675}$Zr$_{0.325}$ )O$_{3}$: the universal relaxor process and the conventional relaxor process, which have different relaxation characteristics employing the Curie-Von Schweidler law at low frequency and the Kohlrausch-Williams-Watts law at higher frequency [@Bokov2007BZT]. Other formula, such as the Cole-Cole and the Havriliak-Negami equations, are also employed to model the dielectric response with respect to frequency at given temperatures. When phonon modes are close or interacting with the relaxation modes, it becomes necessary to use coupled modes to model the dielectric response [@Wang2014; @Wang2016]. On the other hand, there are also many investigations on how the dielectric response, $\chi$, depends on the temperature, $T$, at given frequencies. In addition to the well known Curie law for $\chi\left(T\right)$ at high temperature, most useful equation for fitting around the dielectric peak appears to be the square law. In particular, the formula $$\begin{aligned}
\frac{1}{\varepsilon\left(T\right)}= & \frac{1}{\varepsilon_{A}}+\frac{\left(T-T_{A}\right)^{^{\eta}}}{B}\label{eq:square-law}\end{aligned}$$ was proposed to describe the permittivity at $T>T_{m}$ [@Smolenskii1970; @Kirillov1973]. Initially, $\eta$ was found to be 2, but later was shown to be between 1 and 2 [@Clarke1974; @Uchino1982; @Santos2001; @Bokov2006a; @Bokov2012]. Here, we further the investigation in this direction and attempt to address some important questions regarding relaxor behaviors. We will explain why the dielectric constant has a peak value at $T_{m}$, and what causes the asymmetry around the peak. Moreover, by constructing a statistical model that properly describes how dipoles behave in relaxors, we propose formula to fit experimental results, which further illuminates the physics behind relaxation behavior.
This paper is organized as the follows. In Sec. \[sec:Statistical-modeling\], we introduce the statistical model. In Sec. \[sec:Results\], we apply this model to both lead-free and lead-based relaxors. In Sec. \[subsec:Discussion\], we discuss the implication and limitation of this approach. Finally, in Sec. \[sec:Conclusion\], we present a brief conclusion.
Statistical modeling\[sec:Statistical-modeling\]
================================================
The statistical model starts by considering a critical difference between ferroelectric relaxors and normal ferroelectrics. One crucial observation is that all relaxor ferroelectrics discovered so far are inevitably disordered and inhomogeneous systems. For instance, in BZT Zr and Ti ions are distributed randomly, so are the Mg and Nb ions in PMN, when the samples are treated macroscopically. In addition, PMN possesses the electric field arising from heterovalent Mg and Nb ions, which affects dipole distribution. It is important to further note that well known relaxors can become non-relaxor if their ions are perfectly ordered [@Setter1980; @Bokov1984; @WangJAD].
Individual dipoles
------------------
The randomness of ions and the ensuing lack of long-range correlation has the important consequence that phonon modes may not be the best description to understand relaxor. This fact is evidenced by the effective Hamiltonian that describes the BZT relaxor [@Akbarzadeh2012; @Sherrington2013a; @Sherrington2014] $$\begin{aligned}
E= & \sum_{i}\left(\kappa_{i}\left|u_{i}\right|^{2}+\lambda_{i}\left|u_{i}\right|^{4}\right)+\dots,\label{eq:effHBZT}\end{aligned}$$ where $i$ labels the sites occupied randomly by Zr or Ti, and $\kappa_{i}$ ($\lambda_{i}$) are the second (fourth) order coefficients in the Taylor expansion of energy with respect to $u_{i}$, which is the local dipole on site $i$. For a homogeneous system, where $k_{i}$ and $\lambda_{i}$ are constants, we can usually first consider the harmonic term and construct phonon modes, which are then used to describe the system, especially in low temperature when the system condense to particular phonon modes [@CondenedPhononMode]. In contrast, with the loss of periodicity in relaxors, this approach is no longer profitable. One can insist on using averaged atoms (e.g., replacing Zr and Ti atoms with their average in BZT) to retain the use of phonon modes. However, it is then necessary to consider defect-pinned intrinsic localized modes [@PinnedPhonon] and phonon localization [@PhononLocalization].
The inhomogeneity also has important consequences on ferroelectric phase transitions. In the typical ferroelectric material BaTiO$_{3}$, we may ascribe the temperature-driven phase transition to the condense of phonons to a particular phonon mode [@CondenedPhononMode]. At high temperature, many phonons modes are occupied (occupancy obeying the Bose-Einstein distribution); at low temperatures, due to mode softening, certain mode (often corresponding to the well-known in perovskites) has essentially zero energy, which dominates the system and induce phase transitions. Unlike BaTiO$_{3}$, there is no global phase transition due to the existence of PNRs and/or random electric fields, which eliminates the mode softening phenomenon and renders a global dipole pattern difficult to achieve [@Akbarzadeh2012; @PMNeffHami2015]. In addition, relaxors exhibit strange phonon behavior, such as the “waterfall” effect [@Gehring2000; @Hlinka2003; @PMNWaterFall] and the localization[@PhononLocalization], further showing their difference from normal ferroelectrics. In this work, we focus on individual dipoles and statistically model their dielectric response. This change of view point implies that phonon modes are less important in our analysis. We will show in the following that such change leads to fruitful results, and better understanding of relaxors.
Statistics of individual dipoles
--------------------------------
Individual dipoles can be categorized into different groups based on their dynamics, and each group shall have different contribution to susceptibility. We proceed simplify the interaction between dipoles, assuming that the interaction effectively introduce a potential well of *average* depth, $E_{b}$. We may relate $E_{b}$ to the size of PNRs arising from the clustering of same-type ions [@Akbarzadeh2012] and/or random electric field caused by heterovalent ions [@PMNeffHami2015; @Kleemann2014; @PMNWaterFall].
Since the kinetic energy of individual dipoles obeys the Maxwell-Boltzmann distribution, at temperature $T$, the number of dipoles with kinetic energy $E_{\textrm{kin}}$ is given by $$\begin{aligned}
f\left(E_{\textrm{kin}}\right)= & 2N\sqrt{\frac{E_{\textrm{kin}}}{\pi}}\left(\frac{1}{k_{B}T}\right)^{3/2}\exp\left(-\frac{E_{\textrm{kin}}}{k_{B}T}\right),\label{eq:kinetic-energy}\end{aligned}$$ where $k_{B}$ is the Boltzmann constant, $N$ is the total number of dipoles, and $f\left(E_{\textrm{kin}}\right)dE_{\textrm{kin}}$ is the number of dipoles having a kinetic energy between $E_{\textrm{kin}}$ and $E_{\textrm{kin}}+dE_{\textrm{kin}}$. With this distribution function, we can calculate the number of dipoles with kinetic energy that exceeds the potential well $E_{b}$, which is given by $$\begin{aligned}
N_{1}\left(E_{b},T\right) & =\int_{E_{b}}^{\infty}dE_{\textrm{kin}}f\left(E_{\textrm{kin}}\right)\nonumber \\
= & N\sqrt{\frac{4}{\pi}}\sqrt{\frac{E_{b}}{k_{B}T}}\exp\left(-\frac{E_{b}}{k_{B}T}\right)+N\text{erfc}\left(\sqrt{\frac{E_{b}}{k_{B}T}}\right),\label{eq:particles-above-well}\end{aligned}$$ where erfc is the complementary error function. The number of dipoles confined to the potential well is then given by $$\begin{aligned}
N_{2}\left(E_{b},T\right)= & N-N_{1}\left(E_{b},T\right).\label{eq:particles-inside-well}\end{aligned}$$
The next step is to treat the two sets of dipoles ($N_{1}$ versus $N_{2}$) separately, assigning different susceptibility to them. The total susceptibility is then given by $$\begin{aligned}
\chi(T,\nu)= & \chi_{1}\left(T,\nu\right)P_{1}\left(E_{b},T\right)+\chi_{2}\left(T,\nu\right)P_{2}\left(E_{b},T\right),\label{eq:total-susceptibility}\end{aligned}$$ where $\chi_{1}\left(T,\nu\right)$ and $\chi_{2}\left(T,\nu\right)$ describes the dielectric responses of each dipole group, whose form will be specified later. We also define $$\begin{aligned}
P_{1}\left(E_{b},T\right)\equiv & \frac{1}{N}N_{1}\left(E_{b},T\right),\\
P_{2}\left(E_{b},T\right)\equiv & \frac{1}{N}N_{2}\left(E_{b},T\right),\end{aligned}$$ to normalize the dipoles to unit volume. Equation (\[eq:total-susceptibility\]) is the centerpiece of this work and will be demonstrated to be useful for investigation of various relaxors.
Results \[sec:Results\]
=======================
We now apply Eq. (\[eq:total-susceptibility\]) to fit various $\chi$ versus $T$ obtained experimentally or numerically. The relaxors shown here include both lead-basd (e.g., PMN) and lead-free relaxors (e.g., BZT).
Susceptibility of BZT \[sec:lead-free-relaxor\]
-----------------------------------------------
For the static susceptibility of lead-free relaxor BZT[@BZT-susceptibility; @Prosandeev; @dielectric-susceptibility; @Petzelt2014; @Wang2016], we assume (i) dipoles with kinetic energy that overcomes potential well can be treated as free dipoles, subject only to thermal excitation; (ii) dipoles inside the potential well only contribute a constant susceptibility, $\chi_{2}$. The total susceptibility is thus given by $$\begin{aligned}
\chi(T)= & \chi_{1}\mathcal{L}\left(\frac{\theta}{T}\right)P_{1}\left(E_{b},T\right)+\chi_{2}P_{2}\left(E_{b},T\right)\label{eq:BZT-susceptibility}\end{aligned}$$ where $\mathcal{L}\left(x\right)=\coth\left(x\right)-1/x$ is the Langevin function, known for depicting orientational polarization under thermal fluctuations [@Kasap2006; @CurieLaw]. $E_{b}$, $\chi_{1}$, $\chi_{2}$ and $\theta$ are constants, which will be determined by fitting experimental or numerical data. It can be inferred from equation \[eq:BZT-susceptibility\] that $\chi_{1}$ is the susceptibility of the material at zero Kevin, $\chi_{1}\mathcal{L}\left(\frac{\theta}{T}\right)$ is essentially the Curie law at high temperature, and $\theta$ is proportional to the magnitude of the low-frequency electric field used in experimental measurements.
![Fitting the static susceptibility of Ba(Zr$_{0.5}$,Ti$_{0.5}$)O$_{3}$ obtained from Monte-Carlo simulation using Eq. (\[eq:BZT-susceptibility\]). The blue dots are from Monte-Carlo simulation [@Akbarzadeh2012] and the red solid line is the fitting curve using Eq. (\[eq:BZT-susceptibility\]). \[fig:Fitting-susceptibility-of\]](Fig1){width="8cm"}
We first test Eq. (\[eq:BZT-susceptibility\]) against the static susceptibility versus temperature obtained with Monte-Carlo (MC) simulation in a previous work [@Akbarzadeh2012]. \[fig:Fitting-susceptibility-of\] shows the overall fitting is good enough to reproduce results from MC simulations with parameters shown in and Tab. \[tab:parameters-BZT-static\]. The closeness of $E_{b}$ and $T_{m}$ indicates the average depth of potential wells plays a dominant role in determining $T_{m}$. Close examination of Fig. \[fig:Fitting-susceptibility-of\] also indicates the fitting at the lowest temperature ($\lesssim25$ K) is not as good as the rest. To address this issue, we tried adding a Gaussian distribution to $E_{b}$ and remedied the minor problem. However, the resulting equation is quite complicated and deviates from our original goal of proposing simple analytical formula to fit susceptibility. Therefore this additional step is not adopted here.
$\chi_{1}$ $\theta$ (K) $\chi_{2}$ $E_{b}$ (K)
-------- ------------ -------------- ------------ -------------
Values 741.6 220.5 64.7 159.1
: Fitting parameters for the Ba(Zr$_{0.5}$,Ti$_{0.5}$)O$_{3}$ static susceptibility. \[tab:parameters-BZT-static\]
![Maxwell-Boltzmann distribution [\[]{}Panel (a)[\]]{} and the Langevin function, $\mathcal{L}$, [\[]{}Panel (b)[\]]{} versus temperature. Parameters from Tab. \[tab:parameters-BZT-static\] are used in plot each function. \[fig:components\]](Fig_0){width="6cm"}
In order to have a good understanding of BZT’s susceptibility, we show each component of Eq. (\[eq:BZT-susceptibility\]) in Fig. \[fig:components\]. Figure \[fig:components\] (a) shows that $P_{1}\left(E_{b},T\right)$ and $P_{2}\left(E_{b},T\right)$ have opposite trends as temperature increases. The number of dipoles that can overcome the potential confinement ($P_{1}$) steadily increases with temperature, while the number of dipoles inside ($P_{2}$) continuously becomes smaller. Figure \[fig:components\] (b) shows that the Lagevin function is normalized at $T=0$, and decreases with temperature. Such a feature describes the ability of the free dipoles to respond to an external *dc* electric field. Moreover, Fig. \[fig:components\] (b) also shows the product of the Langevin function and $P_{1}$, which already exhibits some resemblance to BZT’s susceptibility [\[]{}Fig. (\[fig:Fitting-susceptibility-of\])[\]]{}.
![Fitting the susceptibility of Ba(Zr$_{0.5}$,Ti$_{0.5}$)O$_{3}$ at $f=1,10,100,100$ GHz, obtained from molecular dynamics simulation using Eq. (\[eq:BZT-frequency\]).\[fig:Fitting-BZT-frequency\]](Fig2){width="8cm"}
Having examined the static susceptibility, we now move to the frequency-dependent dielectric response, which is often taken as a characteristic property of relaxors [@Kleeman2014; @Sommer]. We propose another equation to fit the susceptibility versus temperature:
$$\begin{aligned}
\chi\left(T\right)= & \frac{\chi_{1}}{1+b\exp\left(-\theta/T\right)}P\left(E_{b},T\right)+\chi_{2}\left[1-P\left(E_{b},T\right)\right],\label{eq:BZT-frequency}\end{aligned}$$
where $\chi_{1}$ , $\chi_{2}$, $b$ and $\theta$ are constants at a given frequency (but may change when the frequency changes). The dielectric contribution from dipoles with kinetic energy higher than the potential well is given by $$\begin{aligned}
w_{1}\left(T\right)= & \frac{1}{1+b\exp\left(-\theta/T\right)}.\label{eq:h1T}\end{aligned}$$ which, similar to the Langevin function, monotonically decreases with temperature $T$. The choice of this function reflects two considerations: (i) at very low temperature ($T$ close to 0), such dipoles shall follow the probing $ac$ electric field closely, leading $w_{1}\left(T\right)$ to its maximum; and (ii) at higher temperature, thermal motions of these dipoles hamper their ability to follow the *ac* electric field, leading to smaller $w_{1}\left(T\right)$. We will discuss this equation further in Sec. \[subsec:Discussion\]. With one more parameter ($b$), this function may be taken as an extension to the Langevin function.
1 GHz 10 GHz 100 GHz 1000 GHz
-------------- ------- -------- --------- ----------
$\theta$ (K) 579.6 762.6 1128.4 2158
$\chi_{1}$ 406.5 312.6 209.3 99.9
$\chi_{2}$ 73.1 63.5 56.2 57.9
$b$ 10.2 9.9 9.5 7.7
: Fitting parameters of numerically simulated Ba(Zr$_{0.5}$,Ti$_{0.5}$)O$_{3}$ susceptibility at various frequency [@Wang2016] using Eq. (\[eq:BZT-frequency\]). \[tab:Fitting-parameters-of-BZT\]
![Analysis of fitting parameters versus probing frequency. (a) $\chi_{1}$vs $\log\left(f\right)$; (b) $w_{1}\left(T\right)$. \[fig:w1T\]](Fig3_ab){width="8cm"}
We use Eq. (\[eq:BZT-frequency\]) to fit BZT’s susceptibility versus temperature at different frequency and show the results in Fig. \[fig:Fitting-BZT-frequency\]. Since $E_{b}$ is a material parameter, we use the same value ($E_{b}=159.1$ K) obtained by fitting the static susceptibility (cf. Fig. \[fig:Fitting-susceptibility-of\]). In Fig. \[fig:Fitting-BZT-frequency\], the numerical results are obtained from molecular dynamics simulations reported in Ref. [@Wang2016]. As the figure shows, satisfactory fittings are achieved for frequencies between 1 and 1000 GHz. Table \[tab:Fitting-parameters-of-BZT\] shows all the parameters. Among them, $\theta_{1}$ and $\chi_{1}$ have substantial changes over the specified frequency range as shown in Fig \[fig:w1T\](a). Figures \[fig:w1T\] (a) shows that $\theta$ depends on $\log\left(f\right)$ quadratically while $\chi_{1}$ linearly depends on it, and Fig. \[fig:w1T\](b) shows $w_{1}\left(T\right)$. At low frequency ($\lesssim10$ GHz), $w_{1}\left(T\right)$ resembles the Fermi-Dirac function, that is, below$\sim250$ K, its value is close to one but becomes close zero for $T$ above $\sim250$ K. At a higher frequency (e.g., 1000 GHz), however, this function strongly deviates from the Fermi-Dirac function, with a long tail extending to high temperature.
![Experimental susceptibility of Ba(Ti$_{0.675}$Zr$_{0.325}$)O$_{3}$ ceramics versus temperature fitted with Eq. (\[eq:BZT-frequency\]). \[fig:real-bzt\]](Fig4_v2){width="6cm"}
To further verify the suitability of this equation for experimental data, we also fit the result shown in Fig. 1 of Ref. [@Bokov2007BZT], where Ba(Ti$_{0.675}$Zr$_{0.325}$)O$_{3}$ ceramics is measured at $10^{-2}$ and $10^{5}$ Hz. Figure \[fig:real-bzt\] shows that satisfactory fittings are achieved.
Pb-based relaxors \[subsec:Lead-based-relaxors\]
------------------------------------------------
Unlike the lead-free BZT, which possesses PNRs that separate dipole clusters, lead-based ferroelectrics [@Kirillov1973; @Gridnev2004; @PMNeffHami2015; @PNR-PMN] have the Pb-driven dipoles across the system [@Yk], which cause phase transitions in systems such as PbTiO$_{3}$ and Pb(Zr,Ti)O$_{3}$ [@PZTPhase; @PTOeffHam; @JiangPTOFilm-1]. Due to heterovalent ions inside, typical lead-based relaxors are subject to random electric fields, which distort the orientation of dipoles. While the precise consequence of the random field is not all clear [@Sherrington2014; @Kleemann2015], such distracting effect on dipoles appears to lead to a strongly modified phase transition with diffused and smeared peak, in sharp contrast to that of normal ferroelectrics [@Pirc2007; @PMNeffHami2015].
To model such a system and account for the moderate phase transition, we need a function that properly describes the dielectric constant versus temperature. Here, we propose to use the slightly modified well known quadratic relation $\frac{\varepsilon_{A}}{\varepsilon^{\prime}}-1=\frac{\left(T-T_{A}\right)^{2}}{2\delta^{2}}$ proposed by Bokov *et al* [@Bokov2006a] to relate relaxor’s permittivity to temperature [@Smolenskii1970; @Kirillov1973] for dipoles above the average potential well (also see Eq. (\[eq:square-law\]) ). This equation can be rearranged to give the following expression $$\begin{aligned}
w_{2}\left(T\right)= & \frac{1}{1+\left|\frac{T-T_{O}}{\theta}\right|^{\gamma}},\label{eq:w2t}\end{aligned}$$ where $\gamma$ is a critical exponent, $T_{o}$ is associated with the peak position of the moderate phase transition, $\theta$ and $\gamma$ are parameters describing the peak. We note that such choice of $w_{2}\left(T\right)$ also agree with the analysis recently given by Uchino [@Uchino2014]. Combining Eqs. (\[eq:total-susceptibility\]) and (\[eq:w2t\]), we obtain the following equation to fit lead-based relaxors, $$\begin{aligned}
\chi\left(T\right)= & \frac{\chi_{1}}{1+\left|\frac{T-T_{O}}{\theta}\right|^{\gamma}}P\left(E_{b},T\right)+\chi_{2}\left[1-P\left(E_{b},T\right)\right],\label{eq:PZN}\end{aligned}$$ where $E_{b}$, $\chi_{1}$, $\chi_{2}$, $T_{o}$, $\theta$, and $\gamma$ are fitting parameters. The meaning of $\chi_{1}$, $\chi_{2}$, and $E_{b}$ are the same as discussed in Sec. \[sec:lead-free-relaxor\].
![Fitting the relative permittivity of PZN-0.13PT at $f=1,10,100$ kHz using Eq. (\[eq:PZN\]). Note the abnormal decrease in the range below $\sim250$ K. The second panel is $w_{2}\left(T\right)$. \[fig:PZN-relaxor\]](Fig5_v2){width="6cm"}
To verify that Eq. (\[eq:PZN\]) indeed works, we experimentally obtained the permittivity of Pb(Zn$_{1/3}$Nb$_{2/3}$)$_{0.87}$Ti$_{0.13}$O$_{3}$ (PZN-0.13PT) versus temperature at frequencies $f=1,10,100$ kHz. As Fig. \[fig:PZN-relaxor\](a) shows, all three fittings are satisfactory. With Eq. (\[eq:PZN\]) and the fittings, we are able to single out $w_{2}\left(T\right)$, which shows a slight increase of the peak temperature ($T_{o}$, around $220$ K) with frequency [\[]{}Fig. \[fig:PZN-relaxor\](b)[\]]{}. We note that in this fitting there is no need to have two $\gamma$ values above and below $T_{o}$ [@Uchino2014]. The asymmetric peak shown in Fig. \[fig:PZN-relaxor\](a) is naturally caused by the Maxwell-Boltzmann distribution function.
1 kHz 10 kHz 100 kHz
----------------- -------- -------- ---------
$\gamma$ 2.02 1.82 1.63
$T_{O}$ (K) 219.6 229.0 240.6
$E_{b}$ (K) 22.3 22.9 23.5
$\chi_{1}$ [^1] 56601 55781 53529
$\chi_{2}$ 1320.9 1284.2 1238.0
$\theta$ (K) 102.4 103.1 104.5
: Fitting parameters of PZN-0.13PT’s permittivity measured at different frequencies. As the table demonstrates, $\gamma$ and $T_{C}$ are the most important variable that changes a lot with frequency. \[tab:Fitting-parameters-of-PZN\]
Table \[tab:Fitting-parameters-of-PZN\] summarizes fitting parameters of the permittivity measured at various frequencies. Among all the parameters, the critical component $\gamma$ changes most (19.3% from 1 kHz to 100 kHz), and decreases with increasing frequency; similarly, $T_{O}$ also changes by 9.5%. On the other hand, $E_{b}$, $\chi_{1}$, $\chi_{2}$, $\theta$ are relatively constant, which are independent of the frequency, and may be taken as material parameters. Such results hints towards the following formula that describes the dependence of PZN-0.13PT on both temperature and the probing frequency $$\begin{aligned}
\chi\left(T,\nu\right)= & \frac{\chi_{1}}{1+\left|\frac{T-T_{O}\left(\nu\right)}{\theta}\right|^{\gamma\left(\nu\right)}}P\left(E_{b},T\right)+\chi_{2}\left[1-P\left(E_{b},T\right)\right],\label{eq:PZN-1}\end{aligned}$$ where the two functions $T_{O}\left(\nu\right)$ and $\gamma\left(\nu\right)$ are frequency dependent while other parameters are constants for a given material. It is also worth noting that for PZN-0.13PT $\chi_{2}\ll\chi_{1}$, which indicates that dipoles with kinetic energy above the potential well play a more important role, in contrast to the case of BZT (see Tab. \[tab:Fitting-parameters-of-BZT\]).
![Fitting the relative permittivity of Pb(Mg$_{1/3}$Nb$_{2/3}$)O$_{3}$-0.05Pb(Zr$_{0.53}$Ti$_{0.47}$)O$_{3}$ measured at $800$ Hz using Eq. (\[eq:PZN\]). The second panel is $w_{2}\left(T\right)$. \[fig:PMN-PZT\]](Fig6_new_v3){width="6cm"}
To further verify the proposed formula, we also fit the permittivity versus temperature of another lead-based relaxor, Pb(Mg$_{1/3}$Nb$_{2/3}$)O$_{3}$-0.05Pb(Zr$_{0.53}$Ti$_{0.47}$)O$_{3}$ [@Gridnev2004]. It can be seen from Fig. \[fig:PMN-PZT\](a) that the overall fitting is satisfactory. Figure \[fig:PMN-PZT\](b) shows $w_{2}\left(T\right)$ with fitting parameters $\gamma=1.66$ and $T_{O}=256.7$ K. Similar to the PZN-0.13PT case, the results here also shows $\chi_{2}\ll\chi_{1}$.
Discussion\[subsec:Discussion\]
===============================
In the statistical model we divide the dipoles inside ferroelectrics relaxors into two groups, one group being confined in potential wells, while the other having can overcome the potential confinement and exhibiting a more vibrant dynamics. It has been demonstrated that the Maxwell-Boltzmann distribution plays a significant role in determining the profile of $\chi\left(T\right)$. To address a particular type of relaxor, one may only need to adjust the dielectric response function associated with each group of dipoles, while keeping the rest unchanged. In this section, we discuss a few issues of this approach and its limitations.
Characteristic temperature $T_{m}$
----------------------------------
The present analysis helps us to understand why the susceptibility of a relaxor reaches its peak value at some temperature, $T_{m}$. For BZT, the function $\chi_{1}\mathcal{L}\left(\frac{\theta}{T}\right)$ [\[]{}see Eq. (\[eq:BZT-susceptibility\]) and Fig. \[fig:components\][\]]{} or $\chi_{1}/\left[1+b\exp\left(-\theta/T\right)\right]$ [\[]{}see Eq. (\[eq:BZT-frequency\])[\]]{} describes the contribution to susceptibility from dipoles with kinetic energy higher than $E_{b}$. These two functions are both monotonically decreasing with $T$, reflecting the fact that thermal motions prevent dipoles from aligning with the applied electric field. On the other hand, the number of dipoles above potential wells increases with $T$ as governed by the function $P\left(E_{b},T\right)$ (see Fig. \[fig:components\]). The combined effects of the two factors give rise to the susceptibility peak at $T_{m}$. However, the situation for lead-based relaxors is different. The function $\chi_{1}/\left(1+\left|\frac{T-T_{O}}{\theta}\right|^{\gamma}\right)$ [\[]{}see Eq. (\[eq:PZN\])[\]]{}, which largely determines the value of $T_{m}$, manifests the vestige of a true phase transition in normal ferroelectrics, which is torn down by random electric fields and/or PNRs in relaxors.
Rationale for Eq. (\[eq:h1T\])
------------------------------
![The energy well for dipoles is shifted and lowered when an electric field is applied .\[fig:falling\]](Fig7){width="8cm"}
For lead-free BZT, we propose Eq. \[eq:h1T\] to describe the susceptibility of dipoles with kinetic energy higher than $E_{b}$. This choice follows the Debye relaxation, i.e., $\chi\sim1/\left(1+\omega^{2}\tau^{2}\right)$ [@Kasap2006], where $\omega$ is a constant (the probing frequency), and $\tau$ is temperature-dependent relaxation time. For a thermally activated process, the relation between $\tau$ and $T$ is often specified by the Arrhenius law, i.e., $1/\tau=A\exp\left(-E_{a}/T\right)$, where $E_{a}>0$ is the activation energy [@PinnedPhonon; @negativeActivation; @Pirc2007]. In this case, the susceptibility will be $\chi\sim1/\left[1+A^{2}\omega^{2}\exp\left(\frac{2E_{a}}{T}\right)\right]$, which is discussed by Jonscher [@Jonscher1981]. However, the dynamic process considered here describes dipoles falling to a state of lower energy, which temporarily created by the probing electric field (see. Fig. \[fig:falling\]). Therefore the activation energy in this process shall be *negative*, i.e., $\chi\sim1/\left[1+A^{2}\omega^{2}\exp\left(-\frac{2\left|E_{a}\right|}{T}\right)\right]$, which is the form adopted in Eq. (\[eq:h1T\]).
We note that negative activation energy is known in some chemical reactions [@negativeActivation]. Negative activation energy appears here because when an *ac* electric field perturbs dipoles and tilts the relative energy of potential wells, dipoles outside potential wells will move towards the temporary potential minimum. However, the drifting to the potential minimum is hindered by thermal fluctuations of such dipoles. In fact, higher temperature (corresponding to larger kinetic energy) results in a slower relaxation to the energy minimum (corresponding to larger $\tau$), leading to negative activation energy. We also note that since the applied *ac* electric field is responsible for shifting the energy minimum and causing dipoles to drift, the change of its frequency may well alter the negative activation energy, explaining why $\theta$ in Eq. (\[eq:BZT-frequency\]) is dependent on the probing frequency. Similar arguments explains why $T_{O}$ also depends on the probing frequency.
In the above analysis, we focus on dipoles with kinetic energy higher than $E_{b}$. These dipoles are able to drift from one energy minimum to another when an *ac* electric field perturbs the system. It has been proposed that the drifting/hopping of dipoles from one potential well to another causes relaxations . However, without distinguishing dipoles inside and outside the potential well, such proposal seems to have a tendency of confusing the wait time before hopping, $t$ (which reflects the distribution of dipoles at a given temperature) with the relaxation time, $\tau$ (which reflects how fast dipoles drift to the transient energy minimum and relates to the loss peak frequency in the Debye function), leading to some difficulties.
Limitations
-----------
In previous studies [@Wang2014; @Wang2016; @PMNeffHami2015], *ab initio* calculation was used, which prescribes all important interactions between dipoles and other degrees of freedom in relaxors, and MC or MD was used to numerically work out the consequences. In the present work we do not start from *ab initio* calculation, instead, employs statistical and phenomenological arguments. Having shown results and insights obtained with this approach, we now discuss possible limitations to the present approach with respect to treatment of details, accuracy, and prediction power.
First, the proposed equations for lead-free [\[]{}Eq. (\[eq:BZT-frequency\])[\]]{} and lead-based relaxors [\[]{}Eq. (\[eq:PZN-1\])[\]]{} have five and seven parameters, respectively. Ideally, one hopes to be able to reduce this large number and use as few parameters as possible. However, it shall be noted that, among these parameters, many are only material dependent (i.e., they do not depend on frequency or temperature). For instance, for lead-free relaxor, $E_{b}$ is a constant; for lead-based relaxor, $E_{b}$, $\chi_{1}$, $\chi_{2}$, $\theta$ are close to constants (see Tab. \[tab:Fitting-parameters-of-PZN\]). For a given sample, these parameters may only need to be calibrated once. In this way, the number of parameters will be significantly reduced.
Second, in this work we have focused on the temperature dependence of susceptibility. The dependence on frequency needs further investigation. For instance, $T_{C}\left(\nu\right)$ and $\gamma\left(\nu\right)$ in Eq. (\[eq:PZN-1\]) need to be specified explicitly to address this issue. We note that results shown in Tab. \[tab:Fitting-parameters-of-PZN\] will provide clues to $T_{C}$ and $\gamma$’s dependence on $\nu$, and eventually help finding analytical expressions for $\chi\left(T,\nu\right)$. In addition, we generally ignored the long-range correlation of dipoles, which is another limitation to this approach. While such long-range correlation makes relaxor physics so rich, it will bring back Bose-Einstein statistics and make the current formulation more complicated. To what extent the Bose-Einstein and the Maxwell-Boltzmann distribution shall be used for ferroelectric relaxors remains an open question.
Third, at high temperature, Curie law is observed in many ferroelectrics. For the static susceptibility of BZT [\[]{}Fig. \[fig:Fitting-susceptibility-of\][\]]{}, this law can be recovered from the proposed equation, Eq. (\[eq:BZT-susceptibility\]). On the other hand, for Eqs. (\[eq:BZT-frequency\]) and (\[eq:PZN\]), the Curie law cannot be directly recovered. For Eq. (\[eq:PZN\]), we have the asymptotic relation $\text{\ensuremath{\chi}}\sim A/\left(T-T_{C}\right)^{\gamma}+B/T^{3/2}$ at very large $T$. It is unclear how good this relation can fit the Curie law. Therefore, in fitting experimental data at high temperatures, one needs to bear in mind that Eqs. (\[eq:BZT-frequency\]) and (\[eq:PZN\]) should be used with care.
Finally, with Eqs. (\[eq:BZT-susceptibility\]) and (\[eq:PZN\]), in principle we can obtain the relation between $T_{m}$ and $\nu$, which can then be compared to the well-known Vogel-Fulcher law [@Wang2016]. However, we have failed to obtain analytical expressions for $T_{m}\left(\nu\right)$ and believe numerical calculation seems to be the only feasible way to establish the relation between $T_{m}$ and $\nu$.
Conclusion\[sec:Conclusion\]
============================
Instead of working on the atomic level, the present work employs a macroscopic statistical approach to help understanding dielectric properties of relaxors. The effects of disorder, PNRs, and random electric fields are considered statistically by introducing the average potential well, which can trap dipoles of low kinetic energy. An external electric field will mostly increase the magnitude **** of trapped dipoles, but rotate to align dipoles free from such trapping, giving rise to two different types of dielectric responses as shown in Eqs. (\[eq:BZT-susceptibility\]), \[eq:BZT-frequency\], and (\[eq:PZN\]). This approach, by proposing analytical equations, provides insights to experimental and numerical results of relaxors. **** Among other things, it shows that the characteristic temperature, $T_{m}$, is determined by the Maxwell-Boltzmann distribution of dipoles’ kinetic energy, as well as their ability to respond to the applied electric field. We can also conclude that lead-free relaxors (e.g., BZT) are different from lead-based relaxors (e.g., PZN-0.13PT) in that (i) The mechanisms determining $T_{m}$ are different. For lead-based relaxors, it appears $T_{O}$ alone in able to determine $T_{m}$, while for BZT, both $w_{1}\left(T\right)$ and $P\left(E_{b},T\right)$ are important; and (ii) For BZT, $\chi_{1}$ and $\chi_{2}$ are on the same order, in contrast to the fact while $\chi_{1}\gg\chi_{2}$ for the Pb-based relaxors, indicating that dipoles outside the average potential well dominate dielectric response of Pb-based relaxors. With these results, we hope this statistical approach can help better understanding important relaxor systems and the proposed equations be adopted in fitting experimental data.
We thank Drs C.-L. Wang, A. A. Bokov and L. Bellaiche for fruitful discussion. This work is financially supported by the National Natural Science Foundation of China (NSFC), Grant No. 51390472, 11574246, U1537210, and National Basic Research Program of China, Grant No. 2015CB654903. F.L. acknowledges NSFC Grant No. 51572214. Z.J. acknowledges the support from China Scholarship Council (CSC No. 201506280055). We also acknowledge the “111 Project of China (Grant No. B14040), the Natural Science and Engineering Council of Canada (NSERC) and the United States Office of Naval Research (ONR Grants No. N00014-12-1-1045 and N00014-16-1-3106).
urlstyle\[1\][doi: \#1]{}
[Petzelt et al.(2014)Petzelt, Nuzhnyy, Savinov, Bovtun, Kempa, Ostapchuk, Hlinka, Canu, and Buscaglia]{} K. Uchino, . (Springer, 1996).
M. E. Manley, J. W. Lynn, D. L. Abernathy, E. D. Specht, O. Delaire, A. R. Bishop, R. Sahul, and J. D. Budai, Phonon localization drives polar nanoregions in a relaxor ferroelectric. *Nat. Commun.* **5**, 3683 (2014).
A. A. Bokov and Z.-G. Ye, Recent progress in relaxor ferroelectrics with perovskite structure. *J. Mater. Sci.* **41**, 31 (2006).
D. Wang, J. Hlinka, A. A. Bokov, Z. G. Ye, P. Ondrejkovic, J. Petzelt, and L. Bellaiche. Fano resonance and dipolar relaxation in lead-free relaxors. *Nat. Commun.* **5**, 5100 (2014).
D. Wang, A. A. Bokov, Z.-G. Ye, J. Hlinka, and L. Bellaiche, Subterahertz dielectric relaxation in lead-free Ba(Zr,Ti)O$_{3}$ relaxor ferroelectrics. *Nat. Commun.* **7**, 11014 (2016).
D. Nuzhnyy, J. Petzelt, and M. Savinov, Broadband dielectric response of Ba(Zr,Ti)O$_{3}$ ceramics: from incipient via relaxor and diffuse up to classical ferroelectric behavior. *Phys. Rev. B* **86**, 1, (2012).
J. Petzelt, D. Nuzhnyy, M. Savinov, V. Bovtun, M. Kempa, T. Ostapchuk, J. Hlinka, G. Canu, and V. Buscaglia, Broadband dielectric spectroscopy of Ba(Zr,Ti)O$_{3}$: dynamics of relaxors and diffuse ferroelectrics.*Ferroelectrics* **469**, 14 (2014).
W. Kleemann, S. Miga, Z. K. Xu, S. G. Lu, and J. Dec, Non-linear permittivity study of the crossover from ferroelectric to relaxor and cluster glass in BaTi$_{1-x}$Sn$_{x}$O$_{3}$. *Appl. Phys. Lett.* **104**, 182910 (2014).
K. Uchino, Fractal Phenomena in Ferroelectrics. *J. Nanotech. Mater. Sci.* **1**, 1 (2014).
A. A. Bokov and Z.-G. Ye, Dielectric relaxation in relaxor ferroelectrics. *J. Adv. Dielectr.* **2**, 1241010 (2012).
A. Akbarzadeh, S. Prosandeev, E. Walter, A. Al-Barakaty, and L. Bellaiche, Finite-temperature properties of Ba(Zr, Ti)O$_{3}$ relaxors from first principles. *Phys. Rev. Lett.* **108**, 257601 (2012).
D. Sherrington, BZT: a soft pseudospin glass. *Phys. Rev. Lett.* **111**, 227601 (2013).
D. Sherrington, Pb(Mg$_{1/3}$Nb$_{2/3}$)O$_{3}$: A minimal induced-moment soft pseudospin glass perspective. *Phys. Rev. B* **89**, 064105 (2014).
W. Kleemann, J. Dec, and S. Miga, The cluster glass route of relaxor ferroelectrics, *Phase Trans.* **88**, 234 (2015).
R. Pirc, R. Blinc, Spherical random-bond-random-field model of relaxor ferroelectrics. *Phys. Rev. B* **60**, 13470 (1999).
D. Sherrington, A spin glass perspective on ferroic glasses.*Phys. Status Solidi* **251**, 1967 (2014).
A. A. Bokov, M. Maglione, and Z.-G. Ye, Quasi-ferroelectric state in Ba (Ti$_{1-x}$Zr$_{x}$) O3 relaxor: dielectric spectroscopy evidence. *J. Phys.: Condens. Matter* **19**, 092001 (2007).
S. A. Gridnev, A. A. Glazunov, and A. N. Tsotsorin, Nonlinear dielectric response of relaxor PMN-PZT ceramics under dc electric field. *Ferroelectrics*, **307**, 151 (2004).
A. A. Bokov and Z.-G. Ye, ** Double freezing of dielectric response in relaxor Pb (Mg$_{1\slash3}$Nb$_{2\slash3}$)O$_{3}$ crystals. *Phys. Rev. B* **74**, 132102 (2006).
H. N. Tailor, A. A. Bokov, Z. G. Ye, **Freezing of polarization dynamics in relaxor ferroelectric (1x)Pb(Mg$_{1/3}$Nb$_{2/3}$)O$_{3}$xBi(Zn$_{1/2}$Ti$_{1/2}$)O$_{3}$ solid solution.** *Curr. Appl. Phys.* **11**, 175 (2011).
A. Dixit, S. B. Majumder, R. S. Katiyar, A. S. Bhalla, **Studies on the relaxor behavior of sol-gel derived Ba(Zr$_{x}$Ti$_{1-x}$)O$_{3}$ (0.30$\le$x$\le$0.70) thin films*. J. Mater. Sci.* **41**, 87 (2006).
A. K. Tagantsev, A. E. Glazounov, ** **Does freezing in Pb(Mg$_{1/3}$Nb$_{2/3}$)O$_{3}$ relaxor manifest itself in nonlinear dielectric susceptibility?** *Appl. Phys. Lett.* **74**, 1910 (1999).
J. P. Sethna, . (Clarendon Press, Oxford, 2010).
L. Cross. . *Ferroelectrics*, **76**,0241 (1987).
A. K. Josnscher, Universal relaxation law (Chelesea Dielectrics Press, 1996).
A. K. Jonscher, Dielectric relaxation in solids (Chelesea Dielectrics Press, 1983).
G. A. Smolenskii, Physical phenomena in ferroelectrics with diffused phase transition. *J. Phys. Soc. Jpn***.** **28**, 26 (1970).
V. V. Kirillov, V. A. Isupov, **Relaxation polarization of PbMg$_{1/3}$Nb$_{2/3}$O$_{3}$ (PMN)-A ferroelectric with a diffused phase transition*. Ferroelectrics*, **5**, 3 (1973).
K. Uchino, S. Nomura, Critical exponents of the dielectric constants in diffused-phase-transition crystals. *Ferroelectrics* **44**, 55 (1982).
R. Clarke and J. C. Burfoot, The diffuse phase transition in potassium strontium niobate. *Ferroelectrics*, **8**, 505 (1974).
I. A. Santos, J. A. Eiras, Phenomenological description of the diffuse phase transition in ferroelectrics. *J. Phys. Condens. Matter* **13**, 11733 (2001).
D. Wang and Z. Jiang, Dielectric response of BaZrO$_{3}$/BaTiO$_{3}$ superlattice, *J. Adv. Dielectr* **6**, 1650015 (2016).
A. A. Bokov, I. P. Raevskii, and V. G. Smotrakov, Composition, ferroelectric and antiferroelectric ordering in Pb$_{2}$InNbO$_{6}$ crystals, *Sov. Phys. Sol. Stat.* **26**,1708 (1984).
N. Setter and L. Cross, The role of B-site cation disorder in diffuse phase transition behavior of perovskite ferroelectrics. *J. Appl. Phys.* **51**, 4356 (1980).
N. Nakanishi, A. Nagasawa, Y. Murakami, Lattice stability and soft modes.****
A. Bussmann-Holder, A. R. Bishop, T. Egami, Relaxor ferroelectrics and intrinsic inhomogeneity. *Europhys. Lett*. **71**, 249 (2005).
W. Zhong, D. Vanderbilt, K. M. Rabe, Phase Transitions in BaTiO$_{3}$ from First Principles. Phys. Rev. Lett. **73**, 1861 (1994).
****
R. A. Cowley, Structural phase transitions I. Landau theory. *Adv. Phys.* **29**, 1-110 (1980).
A. Al-Barakaty, S. Prosandeev, D. Wang, B. Dhkil, and L. Bellaiche, ** Finite-temperature properties of the relaxor PbMg$_{1/3}$Nb$_{2/3}$O$_{3}$ from atomistic simulations. *Phys. Rev. B* **91**, 214117 (2015).
D. Phelan, C. Stocka, J. A. Rodriguez-Riveraa, S. Chia, J. Leãoa , X. Long, Y. Xie, A. A. Bokov, Z.-G. Ye, P. Ganeshe, and P. M. Gehring, Role of random electric fields in relaxors, *PNAS* **111**, 1754 (2014).
J. Hlinka, S. Kamba, J. Petzelt, J. Kulda, C. A. Randall, S. J. Zhang, Origin of the Waterfall Effect in Phonon Dispersion of Relaxor Perovskites. *Phys. Rev. Lett.* **91**, 107602 (2003).
P. M. Gehring, S.-E. Park, and G. Shirane, Soft phonon anomalies in the relaxor ferroelectric Pb(Zr$_{1/3}$Nb$_{2/3}$)$_{0.92}$Ti$_{0.08}$O$_{3}$. *Phys. Rev. Lett.* **84**, 5216 (2000).
W. Kleemann, Cluster glass ground state via random fields and random bonds. *Phys. Status Solidi* **251**, 1993 (2014). .
S. Prosandeev, D. Wang, A. R. Akbarzadeh, L. Bellaiche, **First-principles-based effective Hamiltonian simulations of bulks and films made of lead-free Ba(Zr,Ti)O$_{3}$ relaxor ferroelectrics*. J. Phys . Condens. Matter*. **27**, 223202 (2015).
S. O. Kasap, Principles of Electronic Materials and Devices, Third Edition (McGraw-Hill, 2006), pp612.
We also verified that replacing the Langevin function with the simpler Curie law ($\sim1/T$) will result in similar fitting results and analysis.
R. Sommer, N. K. Yushin, J. J. Van Der klink. **Dielectric susceptibility of PMN under DC bias*. Ferroelectrics* **127**, 235(1992).
I. K. Jeong, T. W. Darling, J. K. Lee, T. Proffen, R. H. Heffner, Direct observation of the formation of polar nanoregions in Pb(Mg$_{1/3}$Nb$_{2/3}$)O$_{3}$ using neutron pair distribution function analysis. *Phys. Rev. Lett.* **94**, 147602 (2005).
Y. Kuroiwa, S. Aoyagi, A. Sawada, J. Harada, E. Nishibori, M. Takata, M. Sakata, Evidence for Pb-O covalency in tetragonal PbTiO$_{3}$. *Phys. Rev. Lett.* **87**, 217601 (2001).
I. A. Kornev, L. Bellaiche, P. E. Janolin, B. Dkhil, and E. Suard, Phase diagram of Pb(Zr,Ti)O$_{3}$ solid solutions from first principles. *Phys. Rev. Lett.* **97**, 157601 (2006).
D. Sichuga and L. Bellaiche, ** Epitaxial Pb(Zr,Ti)O$_{3}$ ultrathin films under open-circuit electrical boundary conditions. *Phys. Rev. Lett*. **106**, 196102 (2011).
Z. Jiang, R. Zhang, D. Wang, D. Sichuga, C.-L. Jia, and L. Bellaich, Strain-induced control of domain wall morphology in ultrathin PbTiO$_{3}$ films. *Phys. Rev. B* **89**, 214113 (2014).
R. Pirc and R. Blinc, Vogel-Fulcher freezing in relaxor ferroelectrics*. Phys. Rev. B* **76**, 020101 (2007).
J. M. Valverde, On the negative activation energy for limestone calcination at high temperatures nearby equilibrium. *Chem. Eng. Sci.* **132**, 169 (2015).
A. K. Jonscher, A new understanding of the dielectric relaxation of solids. *J. Mater. Sci.* **16**, 2037 (1981).
[^1]: Here, $\chi_{1}$ and $\chi_{2}$ shall be understood as relative permittivity, not susceptibility.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The measurement of the CP violation in the b-meson system and of b-meson rare decays provides information about the electroweak symmetry breaking in terms of flavour structure of the CKM matrix and flavour changing neutral currents. The deviation of the experimental observations from the Standard Model predictions allows to constrain new physics. In this paper recent results from Tevatron are reported.'
author:
- |
N. D’Ascenzo\
[*for the CDF and D0 collaborations*]{}\
\
LPNHE University Pierre et Marie Curie CNRS/IN2P3, Paris, France\
Nuclear Research National University, Russia\
E-mail: [email protected]
title: Recent results on B physics at Tevatron
---
CDF and D0 experiments at Tevatron
==================================
Tevatron is a $p\bar{p}$ collider with a center of mass energy of 1.96 TeV. A total luminosity of 8.8 $fb^{-1}$ was delivered to the experiments and 7.3 $fb^{-1}$ was recorded up to July 2010.
The two experiments at Tevatron, CDF and D0, are suitable for the analysis of b-mesons properties.
The CDF detector relies on an excellent tracking system, with momentum resolution $\sigma_{p_{t}}/p_{t}^{2}=0.07\%$\[GeV/c\]$^{-1}$ and impact parameter resolution $\delta D_{0}=40 \mu$m. A dedicated trigger allows to provide the track displacement information from the silicon detectors with several 10 kHz rates. The electron identification is performed in the calorimeters and muons are identified in the muon system with a coverage up to $\eta=1$. The ionization measurement in the drift chambers combined with time of flight measurement in the TOF detector allow an efficient $\pi/K$ separation used for the identification of kaons from b-quark fragmentation and b-mesons decay.
The key-point of the D0 detector is the excellent coverage of the tracker and of the muon system up to $\eta>2$. The vertex detector was updated in 2006 with an additional silicon layer L0 near the beam pipe (R=1.6 cm). The momentum resolution is $\sigma_{p_{t}}/p_{t}^{2}=0.14\%$\[GeV/c\]$^{-1}$ and the impact parameter resolution is $\delta D_{0}=13\mu\rm{m}\oplus 50 /p_{t}$.
New physics in $B_{s}$ oscillation and CP violation
===================================================
The $B_{s}\bar{B}_{s}$ system is a two states quantum system with flavour eigenstates (b,$\rm{\bar{s}}$),($\rm{\bar{b}}$,s).
The flavour eigenstates are not CP eigenstates ($\rm{CP}\psi_{B_{s}^{0}}=-\psi_{\bar{B}_{s}^{0}}$) and mixing between the two states can be mediated by the weak interaction through the flavour-changing transition $b\rightarrow s$.
Flavour-changing processes involving quarks are described in the SM by the lagrangian: $$\mathcal L=\frac{g}{\sqrt{2}}\left(\hat{u},\hat{c},\hat{t} \right)_{L}\gamma_{\mu}\left(\begin{array}{ccc}V_{ud}&V_{us}&V_{ub}\\V_{cd}&V_{cs}&V_{cb}\\V_{td}&V_{ts}&V_{tb}\\ \end{array} \right) \left( \begin{array}{c}d\\s\\b\end{array}\right)_{L}W_{\mu}^{+}$$ The matrix $V_{ij}$ is the Cabibbo-Kobayashi-Maskawa (CKM) matrix [@CKM].
The CKM matrix is unitary. The elements of the CKM matrix satisfy the relations $\sum_{i}{V_{ij}V_{ik}^{*}}=\delta_{jk}$ and $\sum_{i}{V_{ij}V_{kj}^{*}}=\delta_{ik}$. The CKM matrix is parameterized by three mixing angles $\theta_{ij}$ and one CP violating phase $\delta$. The CP violation is generated in the SM only by the phase $\delta$ for flavour changing processes involving quarks. CP violating sources are necessary to explain the matter-antimatter asymmetry observed in nature [@Sak]. Flavour-changing processes in NP models affect the size of the CKM matrix elements and of the CP violating phase.
The $B_{s}^{0}\bar{B}_{s}^{0}$ mixing is described by the effective Hamiltonian $$i\frac{\partial}{\partial t} \psi_{B_{s}^{0}}\left(t\right)= \left(\begin{array}{cc}M-i\frac{\Gamma}{2}& M_{12}-i\frac{\Gamma_{12}}{2}\\M_{12}^{\ast}-i\frac{\Gamma_{12}^{\ast}}{2}& M-i\frac{\Gamma}{2}\\ \end{array} \right)\psi_{B_{s}^{0}}\left(t\right)
\label{Heff}$$ The off-diagonal terms depend on the elements of the CKM matrix and are sensitive to NP contributions [@Lenz]. $M_{12}$ receives contribution from massive internal particles. The main SM contribution is the t-quark and NP models predict highly massive particles which can affect sensitively $M_{12}$. $\Gamma_{12}$ receives contribution from light internal particles, as c- or u- quarks. It is not sensitive to NP which has contribution at higher mass scale. The phase $\phi_{s}=\rm{arg}\left( -\frac{M_{12}}{\Gamma_{12}}\right)$ is also studied. As the SM expectation is $\phi_{s}=\left(0.0041\pm0.0008\right)$, $\phi_{s}$ is a very sensitive quantity to new physics effects.
Three experimental observables can be constructed after diagonalization of the Hamiltonian \[Heff\] and determination of the two $B_{s,L}^{0}$ and $B_{s,H}^{0}$ mass eigenstates.
The mass difference between the two mass eigenstates $\Delta M_{s}=\left(M_{H}-M_{L}\right)\approx 2 \left|M_{12}\right|$ is sensitive to NP due to the proportionality to $M_{12}$. It was measured by CDF and D0 collaboration using 1 $fb^{-1}$ of data [@DMs]. The CDF and D0 experiment report respectively $\Delta M_{s}=17.77\pm 0.10\hspace{2mm} \rm{(stat)} \pm 0.07 \hspace{2mm} \rm{(syst)} \rm~{ps}^{-1}$ and $\Delta M_{s}=18.53\pm 0.93 \hspace{2mm} (stat) \pm 0.30\hspace{2mm} (syst) \rm~{ps}^{-1}$. The experimental value of $\Delta M_{s}$ and of $\Delta M_{d}$ [@Dmb] are combined to determine the CKM matrix element $V_{td}$, which is used to prove the unitarity condition $V_{ud} V^{\ast}_{ub} + V_{cd} V^{\ast}_{cb} + V_{td} V^{\ast}_{tb} = 0$. The experimental result was found in agreement with the SM expectation [@DmbDms].
The width difference between the two mass eigenstates $\Delta \Gamma_{s}=\left(\Gamma_{L}-\Gamma_{H}\right)\approx 2 \left|\Gamma_{12} \right|\rm{cos}\phi_{s}$ is very sensitive to NP due to the correlation with $\phi_{s}$. The golden mode to study this observable is the measurement of the CP violation in the interference between mixing and decay in $B_{s}^{0}\rightarrow J/\psi \phi$. In this decay the relation $V_{us} V^{\ast}_{ub} + V_{cs} V^{\ast}_{cb} + V_{ts} V^{\ast}_{tb} = 0$ is studied from which the angle $\beta_{s}$ is defined as $2\beta_{s}=2arg\left[-\frac{\left(V_{ts}V_{tb}^{\ast}\right)^{2}}{\left(V_{cs}V_{cb}^{\ast}\right)^{2}} \right]$. The SM expectation is $\beta_{s}=0.04\pm 0.01$ rad. It is hence very sensitive to effects of NP[^1] (section 2.1).
The flavour specific or semi-leptonic CP asymmetries are also sensitive probe of NP. In case $\bar{B}_{s}^{0}\rightarrow f$ and $B_{s}^{0}\rightarrow \bar{f}$ are forbidden the asymmetry is defined as $a_{sl}^{s}=\frac{\Gamma\left(\bar{B}_{s}\rightarrow f \right)-\Gamma\left(B_{s}\rightarrow \bar{f} \right)}{\Gamma\left(\bar{B}_{s}\rightarrow f \right)+\Gamma\left(B_{s}\rightarrow \bar{f} \right)}=\frac{\Delta \Gamma_{s}}{\Delta M_{s}}\rm{tan}\phi_{s}$. The same relation yields for $B_{d}^{0}$ mesons: $a_{sl}^{d}\equiv \frac{\Gamma\left(\bar{B}_{d}\rightarrow f \right)-\Gamma\left(B_{d}\rightarrow \bar{f} \right)}{\Gamma\left(\bar{B}_{d}\rightarrow f \right)+\Gamma\left(B_{d}\rightarrow \bar{f} \right)}=\frac{\Delta \Gamma_{s}}{\Delta M_{d}}\rm{tan}\phi_{d}$. The relation to $\phi_{q}$ makes this observables sensitive to NP. These observables are studied in the measurement of the di-muon charge asymmetry[^2] $A^{b}_{sl}\equiv\frac{N_{b}^{++}-N_{b}^{--}}{N_{b}^{++}+N_{b}^{--}}$. $A^{b}_{sl}$ receives contribution from both $B_{s}$ and $B_{d}$ mesons: $A_{sl}^{b}=\left(0.506\pm0.043 \right)a_{sl}^{d}+\left(0.494\pm0.043 \right)a_{sl}^{s} $. The SM expectation is $A_{sl}^{b}=\left(-2.3^{+0.5}_{-0.6}\right)\times 10^{-4}$. NP contributions to the Feynman box diagrams responsible for $B_{q}^{0}$ mesons mixing affect significantly the value of $A_{sl}^{b}$ (section 2.2).
CP violation in $B_{s}\rightarrow J/\psi \left(\mu\mu \right)\phi \left(KK \right)$
-----------------------------------------------------------------------------------
The analysis is performed on the decay $B_{s}^{0}\rightarrow J/\psi \phi$ with the subsequent two body decays $\phi\rightarrow KK$ and $J/\psi \rightarrow \mu\mu$.
The final state of the decay $B_{s}\rightarrow J/\psi \phi$ is composed of a mixture of CP eigenstates. As the $J/\psi$ and $\phi$ are spin-1 particles, the allowed CP states are defined by the relation $CP \psi_{J/\psi \phi} = \eta_{J\/psi}\eta_{\phi}\left(-1 \right)^{L}$, with $L=0,1,2$ and $\eta$ intrinsic CP of the two particles. Two CP-even (L=0,2) and one CP-odd (L=1) states are formed. The total decay amplitude is described in the transversity base [@trans] and is decomposed in the polarization amplitudes $A_{0}$, $A_{\parallel}$ and $A_{\bot}$.
Five parameters defined in QCD are used to describe the polarization amplitudes [@dec]. They are the strong phase and the amplitude of $A_{\parallel}$, the strong phase and the amplitude of $A_{\bot}$ and the real-valued amplitude $A_{0}$.
The time evolution of each polarization amplitude is calculated from the Schrödinger equation \[Heff\]. The angular distribution of the particles in the final states is calculated through the decomposition of the $J/\psi\rightarrow \mu\mu$ and $\phi\rightarrow KK$ decay amplitudes in partial amplitudes [@ang].
The analysis consists of a likelihood fit to the angular and time distributions in order to extract the value of $\Delta \Gamma$ and $\beta_{s}$. This report refers to the recent CDF update with a data sample of $5.2$ fb$^{-1}$ [@bsjpsiphi]. The analysis relies on:
- The event reconstruction and event selection which yield $6504\pm85$ $B_{s}^{0}\rightarrow J/\psi \phi$ signal events.
- Two b-tagging methods. The opposite side tagging combines the information of jet charge and lepton identification to reconstruct the flavour of the b quark at the moment of production. It was calibrated using $B^{+}\rightarrow \Psi K^{+}$ decay and has an efficiency of 94.2 $\%$ and dilution $11.5 \%$. The same side tagging identifies a $K^{\pm}$ from fragmentation of the b-quark. It was calibrated with a simultaneous measurement of $\Delta M_{s}$ and of a dilution scale factor $A$. It has a total tagging power of $3.2 \%$.
- Angular efficiency function used to take into account the detector angular acceptance.
The resulting value of $\Delta\Gamma_{s}$ and $\beta_{s}$ (Fig. \[CPV\]a) is found in agreement with the SM prediction at 0.8 $\sigma$. The possibility of the extension of this analysis including data collected with a two displaced tracks trigger is being studied.
CP violation in di-muon charge asymmetry measurement
----------------------------------------------------
The analysis consists of the reconstruction of events with two same charge muons in the final state.
In this paper the results are reported from the D0 experiment obtained with a total data set of $6.1$ fb$^{-1}$ [@bsasym].
The analysis is performed on two data samples. The inclusive muon sample is composed of events with at least one single muon candidate passing the muon selection and at least one single muon trigger. The like-sign di-muon sample consists of all events with at least two muons candidates of the same charge that pass the dimuon selection and at least one dimuon trigger. This analysis relies on:
- Reversing of the polarity of the toroidal ad solenoidal magnet every two weeks, in order to cancel first order effects related with the instrumental asymmetry.
- Measurement of background contribution to the asymmetry. Muons are considered from the decay of charged kaons and pions and punch-through kaons, pions and protons. The contribution of these backgrounds to the asymmetry $a$ and $A$ are extracted directly from data.
The quantities $a\equiv\frac{n^{+}-n^{-}}{n^{+}+n^{-}}$, where $n^{\pm}$ is the number of positive and negative identified muons, and $A\equiv\frac{N^{++}-N^{--}}{N^{++}+N^{--}}$ are extracted respectively from the inclusive and like-sign sample. The value of $A_{sl}^{b}$ is extracted independently from $A$ and $a$ according to the relations $a=k\times A_{sl}^{b}+a_{bkg}$ and $A=K\times A_{sl}^{b}+A_{bkg}$ where $a_{bkg}$ and $A_{bkg}$ are the background contributions extracted from data and $k$ and $K$ are scale factors estimated in MC. It was found that $a_{bg}\sim A_{bg}$ so that a subtraction of the two relations reduces the systematic error on $A_{sl}^{b}$.
The measured like-sign dimuon asymmetry is $A_{sl}^{b}=-0.00957\pm 0.00251\rm{(stat)}\pm0.00146\rm{(syst)}$ (Fig. \[CPV\]b). This asymmetry is in disagreement with the prediction of the SM by 3.2$\sigma$ deviation. This is the first evidence of anomalous CP-violation in the mixing of neutral B-mesons.
New physics in rare B meson decays
==================================
$B_{s}^{0}\rightarrow \mu^{+}\mu^{-}$
-------------------------------------
The decay $B_{s}\rightarrow \mu\mu$ is a Flavour Changing Neutral Current (FCNC) process forbidden at tree level. It occurs through GIM suppressed box or penguin diagrams. The expected branching ratio of this decay in the SM is $BR\left(B_{s}\rightarrow \mu\mu\right)=\left(3.42\pm0.54 \right)\times 10^{-9}$. Decay amplitudes can be enhanced of few orders of magnitude in extensions of the SM. This rare decay is hence very sensitive to effects of new physics.
The analysis is performed on a data sample composed of two identified muons, satisfying the di-muon trigger and the di-muon selection criteria. The number of events passing the selection is normalized to the yield of the well studied normalization channel $B^{+}\rightarrow J/\psi K^{+}$.
The upper limit on $BR\left(B_{s}\rightarrow \mu\mu\right)$ measured by the two experiments is $4.3\times 10^{-8}$ (CDF [@CDFbsmumu]) and $5.1\times 10^{-8}$ (D0 [@D0bsmumu]) at $95\%$ C.L. obtained analysing respectively a data sample of $3.2$ fb$^{-1}$ (CDF) and $6.1$ fb$^{-1}$ (D0). The result is in agreement with the SM.
$B_{q}^{0}\rightarrow h \mu\mu$
-------------------------------
The $B\rightarrow h \mu\mu$ decay is a FCNC process, mediated by the quark transition $b\rightarrow s l l$. It is described by three dominant Wilson coefficients $C_{7}^{eff},C_{9}^{eff},C_{10}^{eff}$, where $C_{7}^{eff}$ gets contribution from the photon penguins, $C_{9}^{eff}$ comes from the vector and $C_{10}^{eff}$ form the axial component of the weak diagrams [@Wilson]. The BR and the front-backward asymmetry $A_{FB}$ depend on the Wilson coefficients, which indicate if the underlying dynamics is described by SM or NP physics like supersymmetry, Technicolor or fourth generation.
In the CDF analysis performed on a data sample of about $4.4$ fb$^{-1}$ [@bkmumu] the decay modes $B^{+}\rightarrow K^{+} \mu^{+}\mu^{-}$, $B^{0}\rightarrow K^{0} \mu^{+}\mu^{-}$ are searched and fully reconstructed. The branching ratio is calculated respect to the normalization channel $B\rightarrow J/\psi h$. Both BR and $A_{FB}$ for the $B^{+}\rightarrow K^{+} \mu^{+}\mu^{-}$, $B^{0}\rightarrow K^{0} \mu^{+}\mu^{-}$ are measured in agreement with the SM and with the current best results at BaBar and Belle (Fig. \[rare\]b).
A first observation with $6.3\sigma$ evidence of the decay mode $B_{s}\rightarrow \phi \mu \mu$ is performed in the CDF experiment on a data sample of about $4.4$ fb$^{-1}$ (Fig. \[rare\]a). The measured BR is $BR\left(B_{s}\rightarrow \phi \mu \mu \right)=\left(1.44\pm0.33\rm{(stat)}\pm0.03\rm{(syst)} \right)\times 10^{-6}$ [@bkmumu].
Conclusions
===========
The high luminosity reached at Tevatron and the interplay between CDF and D0 experiments allow high accuracy in the measurement of CP violation in the b-mesons mixing and of rare b-meson decays. Both measurements are important for the constrain of new physics processes and for the understanding of the Standard Model.
[0]{}
N. Cabibbo, Phys. Rev. Lett. B [**10**]{} (1963) 531;\
M. Kobayashi, T. Masawa, Progr. Theor. Phys. [**49**]{} (1973) 652.
A. Lenz, hep-ph/0705.3802 (2007). The CDF Collaboration: T. Aaltonen, et al., Phys. Rev. Lett. [**97**]{} (2006) 242003\
D0 Collaboration, D0 note [**5474-CONF**]{} (2007). A.D. Sakarov, Pisma Zh. Eksp. Teor. Fiz. [**5**]{} (1967) 32. HFAG, http://www.slac.stanford.edu/xorg/hfag/ and references therein. O. Dechamps, CKMFitter group, hep-ph/0810.3139 (2009). K.Abe, M.Satpathy, H.Yamamoto, arXiv:hep-ex/0103002v1 (2001) J. L.Rosner, Phys. Rev. D [**42**]{} (1990) 3732. A.S. Dighe, I.Dunietz, J.Lipkin, L.Rosner, arxiv:[**9511363**]{} (1995). The CDF Collaboration, CDF note [**10206**]{} (2010). The D0 Collaboration, arxiv:1005.2757, submitted to Phys.Rev.D (2010). The CDF Collaboration, CDF note [**9892**]{} (2009). The D0 Collaboration, arxiv:1006.3469, submitted to Phys.Lett.B (2010). K.G. Wilson, Phys.Rev. [**179**]{}, 1499 (1969) The CDF Collaboration, CDF note [**10047**]{} (2010).
[^1]: A relation exists between the angle $\beta_{s}$ and the parameter $\phi_{s}$. NP alters the phase $2\beta_{s}$ to $\phi^{\Delta}_{s}-2\beta_{s}$ and the phase $\phi_{s}$ to $\phi_{s}+\phi^{\Delta}_{s}$. As both $\phi_{s}$ and $2\beta_{s}$ are expected to be very small in the SM, the effects of NP are dominant and of the same size $\phi^{\Delta}_{s}$ for both parameters.
[^2]: $N_{b}^{++}$ and $N_{b}^{--}$ are the number of events in which two selected muons in the final state have the same charge
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Einan Gardi
- Matteo Cacciari
title: 'Perturbative and non-perturbative aspects of heavy–quark fragmentation[^1]'
---
‘=11 versim\#1\#2 ‘=12
Introduction {#intro}
============
The heavy–quark fragmentation function $D(z,m^2,\mu^2)$ is the probability distribution to produce a heavy meson of a heavy quark. It depends on $z$, the momentum fraction of the meson, on the quark mass $m^2$ and on the factorization scale $\mu^2$. The fragmentation function has a formal definition [@CS] as the Fourier transform $$D(z;\mu^2)\equiv
\frac{1}{2\pi\,z} \int_{-\infty}^{\infty} \frac{dy_{-}}{y_{-}}\,
\exp(i{p}y/z)\, F(p y;\mu^2),
\label{D_def}$$ of the hadronic matrix element of a non-local operator on the light-cone ($y^2=0$): $$\begin{aligned}
\label{F_def}
&&F(p y;\mu^2) \equiv \\ \nonumber
&&\frac{1}{4\, N_c}\,\sum_{X}\,
{\rm Tr}\left\{ \langle 0 \vert {\mbox{$y$\hspace{-0.5em}\raisebox{0.1ex}{$/$}}}\Psi(y)\vert H(p)+X\rangle
\langle H(p)+X\vert \overline{\Psi}(0)\vert0\rangle_{\mu^2}
\right\}.\end{aligned}$$ Here the final state is composed of the measured heavy meson ($H$) carrying momentum $p$ plus anything else ($X$).
We concentrate here on inclusive observables, the prime example being the single B–meson inclusive cross section in $e^+e^-$ annihilation, shown in Fig. \[nlo\].
Cross sections of this sort can be written as a convolution $$\frac{d\sigma(x,Q^2)}{dx} = \int_x^1\frac{dz}{z}\,{C}(x/z,Q^2;{\mu^2})\,{D}(z;{\mu^2})\,
\label{conv}$$ between a process–specific coefficient function $C$, describing the hard interaction where the heavy quark is produced, and the process–independent fragmentation function $D$, defined in (\[D\_def\]), describing the hadronization stage.
The common practice is to bridge the gap between experimental data and fixed–order calculations in QCD by means of a fragmentation model, i.e. a given functional form for ${D}(z;{\mu^2})$ in (\[conv\]) with one or more free parameters. For heavy–quark fragmentation the most famous examples are [@kart; @peterson]. Upon excluding the more difficult $z\longrightarrow 1$ region such fits can indeed be performed. However, the gain is limited: the relation between the parameters in these models and the matrix elements cannot be made precise. The models provides no information about the underlying hadronization dynamics. Moreover, the universality of the extracted parameters is unclear. At large $z$ fragmentation models simply fail to bridge the gap between the resummed perturbative calculation and the data. This has been recently demonstrated in a clear way [@Ben_haim] by directly extracting the “non perturbative fragmentation component” from $e^+e^-$ data in moment space and then comparing the resulting distribution in $z$ space to models.
An important application of the heavy–quark fragmentation function which demonstrates these problems is in the description of B production in hadron colliders. The CDF collaboration found [@CDF] an alarming discrepancy (a factor of 3) between the transverse–momentum distribution of B$^+$ hadroproduction data and the standard treatment of this cross section, where a NLO calculation is convoluted with a Peterson model [@peterson] for the fragmentation function. In the latter the free parameter was set to a standard value based on $e^+e^-$ annihilation data. Ref. [@CN] applied a resummed perturbative calculation for the coefficient function and combined it with the relevant fragmentation effect extracted from $e^+e^-$ data in moment space, concluding that the discrepancy is much smaller. This shows that the separation between the perturbative and non-perturbative ingredients of (\[conv\]) is very delicate. A naïve application of (\[conv\]) simply fails: if the perturbative ingredient $C$ in (\[conv\]) is taken at fixed order in $\alpha_s$, the required “non-perturbative” ingredient $D$ appears not to be the same in different processes.
As heavy–quark production in hadron colliders becomes increasingly important experimentally, it is evermore urgent to correctly apply perturbative QCD to such cross sections, to separate in a systematic way between the perturbative and the non-perturbative ingredients, and finally, to understand hadronization in a quantitative way. In particular, the parametrization of the fragmentation function $D$ must eventually be understood in terms of its field theoretic definition (\[D\_def\]).
Our approach to heavy–quark fragmentation is primarily a perturbative one: we start off with a perturbative calculation of the matrix element in (\[F\_def\]), replacing the outgoing meson by an on on-shell heavy quark, and treat non-perturbative effects, which make for the difference between the quark and the meson, as [*corrections*]{}. Hadronization corrections are power-suppressed: they are inversely proportional to the mass of the heavy quark $m$. The perturbative approach is appropriate so long as $m \gg \Lambda$. Thus it is definitely applicable to bottom, and probably, with some care, also to charm.
It should be kept in mind that a perturbative calculation is at all possible owing to two properties: (1) the presence of the quark mass regulating collinear divergences; and (2) the inclusive nature of the observable, which guarantees the cancellation of infrared singularities between real and virtual diagrams at any order in perturbation theory. This cancellation does leave, however, a significant trace in the expansion: Sudakov logarithms of $1-z$. This is why the ${\cal O}(\alpha_s)$ result shown in Fig. \[nlo\] diverges at $z\longrightarrow 1$, whereas the physical cross section vanishes at this limit. It is only upon summing the $z\longrightarrow 1$ singular terms in the perturbative series to all orders (exponentiation) that the vanishing of the cross section is recovered.
Asymptotic Scaling {#asym}
==================
Let us first see what can be deduced on the fragmentation function from general considerations. If the quark mass $m$ is infinitely large, hadronization effects are negligible, and the fragmentation function is just $\delta(1-z)$. Taking a large but finite ratio $m/\Lambda$, one would expect the function to be somewhat smeared towards smaller $z$. This smearing is proportional to $m/\Lambda$, as expressed by the following scaling law (see e.g. [@Buras:qm]): $D(z)=(m/\Lambda) f((1-z)m/\Lambda)$.
This property can be formulated more precisely upon taking moments, $$\tilde{D}(N,m^2) \equiv\int_0^1 dz\; z^{N-1} D(z,m^2),
\label{mom}$$ and it can be explicitly derived [@CG] from the field–theoretic definition (\[D\_def\]). One can consider two limits, one where the mass becomes large and the other where the moment index $N$ gets large. For large $m$ one can match the matrix element (\[F\_def\]) onto the heavy–quark effective theory, getting [@JR]: $$\frac{F(p y, m^2)}{p y}\,\,\exp\left(i{p}y\right)
\longrightarrow \,{{\cal F}(p y \,\bar{\Lambda}/m)}+{\cal O}(\bar{\Lambda}/m),
\label{large-m}$$ namely, at the leading order in the large–$m$ expansion the dependence on $m$ and on the light–cone separation $y_{-}$ (i.e. on $py$) is coupled: the matrix element becomes a function of a single argument $p y \,\bar{\Lambda}/m$. Here $\bar{\Lambda}$ is the difference between the heavy–meson mass $M$ and the heavy–quark mass $m$. For large $N$ it follows from the definition (\[F\_def\]) and from (\[mom\]) that $$\tilde{D}(N,m^2) \longrightarrow \left.{\frac{F(p y,m^2)}{p
y} \exp\left(i{p}y\right)}\right\vert_{py=-iN}\!\!\!\!\!\!\!+\,{\cal
O}\left(\frac1N\right),
\label{large-N}$$ namely that to leading order in $1/N$ the $N$-th moment of the fragmentation function can be obtained by analytically continuing the matrix element as a function of the light–cone separation to the complex plane and evaluating it at $py=-iN$. From (\[large-m\]) and (\[large-N\]) together it follows that upon taking the simultaneous limit $m \longrightarrow \infty$ and $N \longrightarrow \infty$ with a fixed ratio $m/N$, $$\tilde{D}(N,m^2)\simeq
\left. \,{\cal F}(p y \,\bar{\Lambda}/m) \,
\right\vert_{py=-iN}+{\cal O}\left(\frac1N\right),
\label{large-m-and-N}$$ so the fragmentation function becomes a function of a single argument $N\bar{\Lambda}/m$.
In Sec. \[dge\] we shall see how the dependence on $m$ and $N$ through the combination $N\bar{\Lambda}/m$ follows from the large–order behaviour of the perturbative expansion in the large–$\beta_0$ limit. Having established Eq. (\[large-m-and-N\]) non-perturbatively, we know that this is indeed the leading behaviour at large $N$ and that corrections to this behaviour are suppressed by a power of $1/N$.
We see that the scale which characterizes the fragmentation process in the large $z$ region is $m(1-z)$ or, in moment space, $m/N$. This scale has a clear meaning when considering the bremsstrahlung off a heavy quark. Let us examine the emission in a frame where the quark energy $E$ is much larger than its mass. The radiation pattern (to ${\cal O}(\alpha_s)$) is $$\frac{dD}{dz\,d\sin^2\theta}\simeq\frac{C_F\,
\alpha_s}{\pi}\,\frac{1}{1-z}\,\frac{\sin^2\theta}{(\sin^2\theta+m^2/E^2)^2},
\label{rad_pattern}$$ where only the leading term in the limit $z\longrightarrow 1$ was kept and the angle of emission $\theta$ is related to the gluon transverse momentum by $\sin^2\theta={k_{\perp}^2}/\left({E^2z^2(1-z)^2}\right)$. As discussed in [@Dokshitzer:fd], the radiation vanishes in the exact forward direction, but it peaks close to the forward direction at $\theta\,\simeq\, m/E$ (the ‘dead cone’), or in a boost-invariant formulation at $\vert k_{\perp}\vert \simeq m(1-z)$. So $m(1-z)$ is the typical transverse momentum of radiated gluons. The scaling law (\[large-m-and-N\]) can be understood in physical terms as the observation that the hadronization effects ($\tilde{D}(N,m^2)$ at large $N$ and $m$) are dominated by interaction with gluons of transverse momentum $\bar{\Lambda}=M-m$.
Factorization {#fact}
=============
Factorization is based on the fact that dynamical processes taking place on well–separated physical scales are quantum-mechanically incoherent. This allows one to treat different subprocess independently of one another and to resum large corrections.
Eq. (\[conv\]) is often regarded as the separation between perturbative and non-perturbative contributions to the cross section. However, factorization can be a much stronger tool upon considering separately the dynamics taking place on different physical scales. Consider, for example, the case of bottom production in $e^+e^{-}$ annihilation, shown in Fig.\[nlo\]. Referring to (\[conv\]) one can naïvely interpret the gap between the data and some perturbative calculation as the “non-perturbative fragmentation function” and then try to bridge this gap using a model. As stressed above this interpretation leads to much confusion. Instead, the reasons for having large (perturbative and non-perturbative) corrections need to be identified and the corrections be resummed.
The first step is to separate the scales involved. Upon neglecting higher order corrections which are suppressed by powers of $m^2/q^2$ the moments of the cross section can be written as [@MN; @CC] $$\begin{aligned}
\tilde{\sigma}(N,q^2,m^2) =
{\tilde{C}(N,q^2;\mu_F^2)}
{\tilde{E}(N,\mu_F^2,\mu_{0F}^2)}
\tilde{D}(N,m^2;\mu_{0F}^2).
\nonumber
$$ Choosing $\mu_F^2\sim q^2$ and $\mu_{0F}^2\sim m^2$, the coefficient function $\tilde{C}$ and the fragmentation function $\tilde{D}$ depend only on scales of order $q^2$ and $m^2$, respectively. The evolution factor $\tilde{E}$ can be obtained solving the Dokshitzer–Gribov–Lipatov–Altarelli–Parisi (DGLAP) equation. This factor then resums corrections depending on $\alpha_s\ln m^2/q^2$ to all orders. Resummation of this kind was implemented in computing the full line in Fig. \[nlo\]. Clearly, this is insufficient.
Next, one observes that the subprocesses $\tilde{C}$ and $\tilde{D}$ may contain additional large corrections. One generic source of large corrections (see [@Beneke]) are running coupling (renormalon) effects, which induce factorial growth of the coefficient at high orders owing to the increasing sensitivity to extreme ultraviolet or infrared scales. Infrared renormalons in particular are non-summable and introduce a power–suppressed ambiguity in the perturbative definition of any quantity. Since for observable quantities this ambiguity must cancel it can serve as a probe of non–perturbative contributions.
Another source of large corrections develops at large $N$: the Sudakov logs [@Dokshitzer:1995ev; @CC]. As stressed above the fragmentation process is dominated at large $N$ by momenta of order $m/N$. When $m$ and $m/N$ become far apart the concept of factorization applies again, and can be used to resum logs of $N$ into a Sudakov form factor. This resummation takes the form of exponentiation in moment space. A similar situation occurs in the coefficient function $\tilde{C}$, as is demonstrated in Fig. \[large\_N-fact\].
$\tilde{C}$ is dominated at large $N$ by the invariant mass $q^2/N$ of the unresolved jet which recoils against the measured heavy meson. The fact that this jet was also initiated by a heavy quark plays no role at this level [@CG]: the relevant scale here is the total invariant mass of the jet. The same jet function dominates deep inelastic structure functions at large $N$ [@DIS; @Gardi:2002xm].
It should be emphasized that factorization (contrary to its diagrammatic proofs) is a non-perturbative concept. One should therefore expect that non-perturbative corrections on a certain scale would factorise together with the corresponding perturbative sum. In particular, this must apply to renormalon–related power corrections. In the case of Sudakov logs factorization leads to exponentiation. Going beyond the logarithmic level, one finds that power corrections on the corresponding scale exponentiate as well. This is the conceptual basis for the “shape function” approach to hadronization corrections, which has been developed in the context of event–shape distributions [@KS; @Korchemsky:2000kp; @Gardi:2001ny] (see also [@DW]). This is also the basis of the approach of [@DIS; @Gardi:2002xm] to higher twist in deep inelastic structure functions at large $N$ and of our approach [@CG] to heavy–quark fragmentation.
Dressed Gluon Exponentiation {#dge}
============================
In order to deal with heavy–quark fragmentation at large $N$ both Sudakov logs and renormalons need to be taken into account. At large $N$, the perturbative coefficients are dominated by Sudakov logs. However, the resummation of the leading logarithms alone does not provide any information on power corrections. It is the subleading logs generated by the running of the coupling which produce the renormalon ambiguity [@Gardi:2001ny; @DGE; @CG]. Their resummation is therefore essential to probe the non-perturbative regime.
From these considerations it follows that the Sudakov exponent needs to be computed to all orders rather than to some fixed logarithmic accuracy. Clearly, the full calculation cannot be done. However, relevant all–order information can be obtained from the large–$\beta_0$ limit corresponding to a single dressed gluon. Calculating the Sudakov exponent in this way is referred to as “Dressed Gluon Exponentiation” (DGE) [@Gardi:2001ny; @DGE; @CG].
A process–independent calculation of the fragmentation function (\[D\_def\]) in the large–$\beta_0$ limit was performed in [@CG]. In the light-cone axial gauge $A\cdot y=0$ where the path–ordered exponential is 1, there is just one diagram – see Fig. \[frag\_ren\].
This diagram was computed using an off-shell gluon splitting function, which was derived identifying the limit where the massive quark propagator prior to the emission of the gluon is singular[^2].
The result for the logarithmic derivative of the fragmentation function, written as a scheme invariant Borel transform, is: $$\begin{aligned}
\label{dD_dm}
\nonumber
&&\frac{d \tilde{D}(N,m^2)}{d\ln m^2}\!=\! -\frac{ C_F}{\beta_0}\int_0^{\infty}\!\!\!du
{\left(\frac{\Lambda^2}{m^2}\right)}^{u}\!{\rm e}^{\frac53 u}\int_{0}^1 dz \left(z^{N-1}-1\right)
\\
&& \left(\frac{z}{(1-z)^2}\right)^u \left[\frac{z}{1-z}\,(1-u)+\frac12(1-z)\,(1+u)\right],\end{aligned}$$ where $\Lambda$ is in the ${\overline {\rm MS}}$ scheme. A generalization of this result beyond the large $\beta_0$ limit which fully captures the next–to–leading logarithms (NLL) was constructed in [@CG].
Eq. (\[dD\_dm\]) takes into account the cancellation between real ($z^{N-1}$) and virtual (1) corrections. In the square brackets we distinguish between $z=1$ singular and regular terms. The former lead to logarithmically enhanced contributions in the perturbative expansion, and therefore need to be exponentiated.
According to (\[dD\_dm\]) the natural scale for the renormalization of the coupling at fixed $z$ is $(1-z)^2m^2/z$. Thus, integrating over the Borel variable $u$ first is not possible for $(1-z)m\lsim \Lambda$. As expected, perturbation theory breaks down when the gluon virtuality or its transverse momentum become comparable to the QCD scale. This constraint takes a completely different form when considered in moment space: infrared renormalons show up.
We proceed to compute the Sudakov exponent in the large–$\beta_0$ limit by isolating the $z=1$ singular terms, performing the $z$-integration and then integrating over $m^2$. The result is: $$\begin{aligned}
\label{D_N_MSbar}
&&\ln \tilde{D}(N,m^2;\mu_{F0}^2)
=\frac{ C_F}{\beta_0}\int_0^{\infty}\!\frac{du}{u}
\left({\frac{{\Lambda^2}}{m^2}}\right)^{u} \times \\ \nonumber
&&\qquad \qquad
\bigg[ \left(\frac{m^2}{\mu_{F0}^2}\right)^{u} B_{\cal A}(u) \ln N
-{B_{\tilde{D}}^{\DGE}(u,N)} \bigg],\end{aligned}$$ where $$\label{BD}
{ B_{\tilde{D}}^{\DGE}(u,N)}=-\,{\rm e}^{\frac53 u}\,
{ (1-u)}\,\Gamma({-2u})\,\left({N}^{2u}-1\right).$$ The $m^2$ integration requires to introduce an ultraviolet subtraction: a $\mu_{F0}^2$–dependent counter term which cancels the $u=0$ singularity of the fragmentation function. This term is the well-known cusp anomalous dimension [@KR; @KM], given by $B_{\cal A}(u)\ln N$, where $B_{\cal A}(u)=1+\frac53u+\ldots$ (we use the $\overline {\rm MS}$ factorization scheme). Note that contrary to $B_{\tilde{D}}^{\DGE}(u,N)$ this subtraction term has just a single $\ln N$ to any order in $u$ and it is also free of infrared renormalon singularities.
According to Eq. (\[BD\]), renormalons in the Sudakov exponent (\[D\_N\_MSbar\]) appear at all integer and half integer $u$ values with the exception of $u=1$. It is clear from Eq. (\[dD\_dm\]) that these renormalons are exclusively related to the $z\longrightarrow 1$ limit. To define the perturbative sum corresponding to $\ln \tilde{D}(N,m^2)$ one needs to integrate over $u$ with some prescription that avoids the poles. The natural choice is the principal–value (PV) prescription (it was implemented numerically in [@CG]). The ambiguity in choosing a prescription is compensated by power corrections corresponding to the residues. Introducing a free parameter for each singularity one ends up with an additive correction to the perturbative Sudakov exponent having the form: $$\ln \tilde{D}_{\NP}({ N\Lambda/m})=
-\epsilon_1{\frac{N\Lambda}{m}}-\epsilon_3
\left(\frac{N\Lambda}{m}\right)^3-\epsilon_4
\left(\frac{N\Lambda}{m}\right)^4+\cdots.
\label{NP}$$ Finally, exponentiating the result to compute $\tilde{D}(N,m^2)$ the perturbative and non-perturbative contributions appear as two factors: $$\tilde{D}(N,m^2;\mu_{F0}^2)={\tilde{D}_{\PT}(N,m^2;\mu_{F0}^2)}\,
\tilde{D}_{\NP}(N\Lambda/m).$$ The leading power correction of the form $N\Lambda/{m}$ predicted in [@Webber_Nason] is readily obtained from (\[NP\]) upon expanding the exponent.
It should be stressed that in both the perturbative (\[D\_N\_MSbar\]) and the non-perturbative (\[NP\]) contributions to the fragmentation function we considered here only the leading terms at large $N$. At the perturbative level the result is improved [@CG] by matching it with the full NLO coefficient. At the non-perturbative level, there may be additional ${\cal O}(\Lambda/m)$ terms which we do not parametrize, and consequently the description of the first few moments is of limited accuracy. In practice, to deal with low moments, it is useful to modify the parametrization (\[NP\]) replacing $N\longrightarrow N-1$ such that the $N=1$ moment is exactly $1$, as it must be by definition.
The perturbative PV–regulated DGE result of Eq. (\[D\_N\_MSbar\]), matched to the NLL and the NLO expressions and combined with the proper coefficient function, is compared as a function of $N$ with the ALEPH data in Fig. \[DGE\_N\] (full line). In contrast with the NLL result of Ref. [@CC] (dashed line), the DGE one does not have a Landau singularity [@CG] and thus it extrapolates smoothly towards the values of $N\gsim m/\Lambda$ which are beyond perturbative reach.
Also shown in Fig. \[DGE\_N\] is the ambiguity (band shown by two dot-dashed lines) corresponding to the residue of the first renormalon pole located at $u=1/2$. The lower edge of the band just matches the data, indicating that the power correction of the form and magnitude(!) expected based on the renormalon analysis is supported by the data.
The different perturbative results are converted to $x$ space in Fig. \[DGE\_x\]. Here the significant impact of Sudakov resummation to NLL as well as that of the additional renormalon resummation achieved by DGE on the shape of the distribution is evident. Note that the shape of the DGE curve resembles that of the data but it is centered at larger $x$. Indeed, the leading effect of the non-perturbative function (assuming in (\[NP\]) that only $\epsilon_1\neq 0$) is a shift of the entire perturbative distribution, very much the same as the leading corrections in event–shape distributions [@KS; @DW; @Gardi:2001ny]. Finally, regarding the non-perturbative parameters $\epsilon_n$ as free parameters in a fit, the data can be well described. The result of a fit in moment space where the only non-perturbative correction is $\epsilon_1$ is shown in Fig. \[fit\].
Upon using more non-perturbative parameters the details of the prediction (\[NP\]) can be confronted with data. The analysis in [@CG] shows that subleading non-perturbative corrections at the exponent are rather small, and the absence of a correction of the form $N^2\Lambda^2/m^2$ can be consistent with the data.
Conclusions {#concl}
===========
We described here a new approach to the QCD description of heavy–quark fragmentation concentrating on the $z\longrightarrow 1$ limit. It was first rigorously demonstrated that the non-perturbative dynamics is dominated by the scale $m(1-z)$. This scale corresponds in perturbation theory to the transverse momentum of gluons radiated from the heavy quark. Based on a renormalon analysis we extended the perturbative technique for resumming soft gluon radiation to the non-perturbative regime, identified power–like effects and separated them from the perturbative fragmentation function by means of a PV prescription. The non-perturbative contribution was then parametrized based on the renormalon ambiguity. We found that the simplest possible parametrization of power corrections which follows from renormalons, namely a shift of the perturbative distribution, is sufficient to describe the data on B production in $e^+e^-$ annihilation. This way phenomenological models for the non-perturbative fragmentation function are not needed.
The fragmentation function was treated, based on its definition (\[D\_def\]), in a process independent way. The results are thus applicable independently of the production process, given that the corresponding coefficient function in the ${\overline{\rm MS}}$ scheme is known. Universality of the leading power corrections at large $z$ can now be tested experimentally.
M. Cacciari and E. Gardi, “Heavy-quark fragmentation,” [*Nucl. Phys.*]{} [**B664**]{} (2003) 299 \[hep-ph/0301047\]. J. C. Collins and D. E. Soper, [*Nucl. Phys.*]{} [**B194**]{} (1982) 445.
V.G. Kartvelishvili, A.K. Likehoded and V.A. Petrov, [*Phys. Lett.*]{} [**B78**]{} (1978) 615.
C. Peterson, D. Schlatter, I. Schmitt and P.M. Zerwas, [*Phys. Rev.*]{} [**D27**]{} (1983) 105.
E. Ben-Haim, P. Bambade, P. Roudeau, A. Savoy-Navarro and A. Stocchi, “Extraction of the $x$-dependence of the non-perturbative QCD b-quark fragmentation distribution component,” \[hep-ph/0302157\]. D. Acosta [*et al.*]{} \[CDF Collaboration\], [*Phys. Rev.*]{} [**D65**]{} (2002) 052005 \[hep-ph/0111359\]. M. Cacciari and P. Nason, [*Phys. Rev. Lett.*]{} [**89**]{} (2002) 122003 \[hep-ph/0204025\]. P. Nason, in [*Heavy Flavors*]{}, eds. A.J. Buras and M. Lindner, (World Scientific, Singapore, 1992) Advanced series on directions in high energy physics, vol. 10.
R. L. Jaffe and L. Randall, [*Nucl. Phys.*]{} [**B412**]{} (1994) 79 \[hep-ph/9306201\]. Y. L. Dokshitzer, V. A. Khoze and S. I. Troian, [*J. Phys.*]{} [**G17**]{} (1991) 1602. B. Mele and P. Nason, [*Phys. Lett.*]{} [**B245**]{} (1990) 635. M. Cacciari and S. Catani, [*Nucl. Phys.*]{} [**B617**]{} (2001) 253 \[hep-ph/0107138\]. M. Beneke, [*Phys. Rept.*]{} [**317**]{} (1999) 1 \[hep-ph/9807443\]; M. Beneke and V. M. Braun, “Renormalons and power corrections,”, in the Boris Ioffe Festschrift, [*At the Frontier of Particle Physics / Handbook of QCD*]{}, ed. M. Shifman (World Scientific, Singapore, 2001), vol. 3, p. 1719 \[hep-ph/0010208\]. Y. L. Dokshitzer, V. A. Khoze and S. I. Troian, [*Phys. Rev.*]{} [**D53**]{} (1996) 89 \[hep-ph/9506425\]. E. Gardi, G. P. Korchemsky, D. A. Ross and S. Tafat, [*Nucl. Phys.*]{} [**B636**]{} (2002) 385 \[hep-ph/0203161\]. E. Gardi and R. G. Roberts, [*Nucl. Phys.*]{} [**B653**]{} (2003) 227 \[hep-ph/0210429\]. G. P. Korchemsky and G. Sterman, [*Nucl. Phys.*]{} [**B437**]{} (1995) 415 \[hep-ph/9411211\]; in [*Moriond 1995*]{}: Hadronic: 0383-392, \[hep-ph/9505391\]; [*Nucl. Phys.*]{} [**B555**]{} (1999) 335 \[hep-ph/9902341\]. G. P. Korchemsky and S. Tafat, [*JHEP*]{} [**0010**]{} (2000) 010 \[hep-ph/0007005\]. E. Gardi and J. Rathsman, [*Nucl. Phys.*]{} [**B609**]{} (2001) 123 \[hep-ph/0103217\]; [*Nucl. Phys.*]{} [**B638**]{} (2002) 243 \[hep-ph/0201019\]. Y. L. Dokshitzer and B. R. Webber, [*Phys. Lett.*]{} [**B404**]{} (1997) 321 \[hep-ph/9704298\]. E. Gardi, [*Nucl. Phys.*]{} [**B622**]{} (2002) 365 \[hep-ph/0108222\]. S. Catani, S. Dittmaier and Z. Trocsanyi, [*Phys. Lett.*]{} [**B500**]{} (2001) 149 \[hep-ph/0011222\]. G. P. Korchemsky and A. V. Radyushkin, [*Phys. Lett.*]{} [**B279**]{} (1992) 359 \[hep-ph/9203222\]. G. P. Korchemsky and G. Marchesini, [*Nucl. Phys.*]{} [**B406**]{} (1993) 225 \[hep-ph/9210281\]. P. Nason and B. R. Webber, [*Phys. Lett.*]{} [**B395**]{} (1997) 355 \[hep-ph/9612353\].
[^1]: Invited talk (E.G.) at the HEP2003 Europhysics Conference in Aachen, Germany. These proceedings are based on [@CG].
[^2]: In this limit the gluon virtuality $k^2$, its transverse momentum $k_{\perp}^2$ and the quark mass $m^2$ are taken to be small simultaneously keeping the ratios between them fixed. This is a generalization of the quasi–collinear limit discussed in [@Catani:2000ef; @CC].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cosmological perturbation theory is the theory of fluctuations (scalar as well as tensor) around the inflationary cosmological background solution. It is important to understand the details of the process of renormalization in this theory. In more familiar applications of quantum field theory, the dependence on the external momenta of the dimensionally regulated expression of the one-loop contribution to a correlator determines the number of counter terms (and their forms) required to renormalize it. In this work, it is pointed out that in cosmological perturbation theory, though this still happens, it happens in a completely different way such that in the late time limit, the information about the number and forms of counter terms required gets erased. This is to be compared with what happens in spontaneous symmetry breaking where the use of fluctuation fields around a chosen vacuum seems to suggest that more counter terms shall be needed to renormalize the theory than are actually required. We also comment on how the field strength of curvature perturbation, $\zeta$, could get renormalized.'
author:
- Gaurav Goswami
bibliography:
- 'Bibliography\_QFT.bib'
- 'biblio\_inf\_tmp.bib'
- 'QFT-ubiquity.bib'
title: On counterterms in cosmological perturbation theory
---
Introduction
============
The methods of quantum field theory (QFT) are some of the most universally applicable techniques in all of physics. The same techniques apply to (among other things) elementary particle physics [@quantum], to statistical and condensed matter physics [@tsvelik2007quantum], to the non-linear theory of structure formation in the universe [@2002PhR...367....1B; @2013arXiv1301.7182P] , to the theory of turbulence in fluids and plasmas [@Zhou20101], to the calculation of spectrum of gravitational waves from inspiralling binaries [@2006IJMPD..15.2293G], to quantum optics [@gerry2005introductory], to black hole thermodynamics [@2010arXiv1011.5875K] as well as to the calculation of correlations of primordial metric perturbations in inflationary cosmology [@2009arXiv0907.5424B]. In this light, it is not surprising that the details of calculations in these varied scenarios, though similar, are not exactly identical. Thus, it is important to find out which ideas apply universally to all these problems and which ones are specific to the applications we are most familiar with (e.g. scattering problems). To aid the discussion, in this work, the familiar version of QFT shall be referred to as usual QFT.
It is a well known fact that inflationary cosmological perturbation theory is almost like a QFT of fluctuations around a time-dependent background solution [@Maldacena-2003]. Consequently, cosmological perturbation theory derives most of its calculational machinery from the usual QFT. Still, there are many differences, and in this work, we look at some of these differences. The interactions of metric perturbations cause corrections to cosmological correlations calculated in linear perturbation theory. This can cause not only primordial non-Gaussianity [@Maldacena-2003], but also, higher order corrections to e.g. the two-point function [@SZ_all_loops].
Just like in the usual QFT, in cosmological perturbation theory too, while calculating correlations at sufficiently high order, we encounter expressions which are ultraviolet (UV) divergent. In usual QFT, we cancel these infinities by (i) collecting the divergent terms with similar dependence on external momenta, (ii) finding which terms (allowed by the symmetries of the theory) in the action can lead to the terms in the correlations with the said dependence on external momenta, and finally, (iii) adding these counterterms to the old action (which is now called the renormalized action while the sum of the renormalized action and the counterterm is the bare action) in order to get correlations which are UV finite at every step of the calculation.
Does the same procedure work in cosmological perturbation theory too? As the analysis of this work illustrates, there are subtleties associated with this, one thus has to be extremely careful (in this context, see [@PhysRevD.72.043514; @PhysRevD.74.023508; @SZ2010]). In cosmological perturbation theory, one is studying fluctuations around a time dependent background solution, thus, the action of these fluctuations is not Lorentz invariant. Thus, we can not simply list all the possible terms in the action (moreover, since the typical interactions are irrelevant, infinite couterterms shall be required). All this ensures that it is difficult to spot the couterterms in cosmological perturbation theory. Given this situation, one could ask, given a correlator (e.g. the two-point function), which counterterms shall we need to renormalize it at one loop?
From our experience in usual QFT, we are used to spotting the counterterms by looking at the expression for dimensionally regulated correlators. This is because the expression for the correlator (with external lines amputed) is of the form $\sum_n c_n (k^2)^n$ (where, $k$ is the Lorentzian momentum of the external line) i.e., a polynomial in the external momenta with divergent coefficients.
Unlike in usual QFT, in cosmological perturbation theory, the dimensionally regulated logarithmically divergent two-point correlator (for external momentum $k$) is (at late times) of the form $$\left[ \sum_n c_n (|{\bf k}|)^n \right] \left[\sum_m a_m (-k \eta)^m \right] \; ,$$
where, $k$ is the 3-momentum of the external line and $\eta$ is the conformal time at which the correlator is evaluated. We thus have an additional factor involving a polynomial in $-k \eta$: it is this extra piece which determines the forms and number of counterterms needed to renormalize the theory. During inflation, we are typically interested in the ${\eta \rightarrow 0^-}$ limit of the correlations. If this limit is taken before regularizing the UV divergent momentum integrals, the information about the forms of counterterms needed and the number of counterterms needed to get rid of UV divergences gets erased. Since the number of counterterms also determines the number of observations actually needed to fix the renormalized parameters, this is a bad news. On the other hand, regularization of momentum integrals at non-zero $\eta$ is a much more involved task to perform and hence the results are much less transparent. All this illustrates that many properties of the familiar usual QFT calculations do not apply to cosmological perturbation theory. Thus, the connection between QFT and cosmological perturbation theory is quite subtle and continues to offer surprises.
While calculating the loop corrections to cosmological correlations, it is often assumed that the process of renormalization can be performed and the emphasis usually is on the non-trivial logarithmic running which may turn up (see, e.g., [@PhysRevD.72.043514]). E.g. Senatore and Zaldarriaga [@SZ2010] have studied renormalization of two-point function in cosmological perturbation theory. Their main focus has been on the nature of logarithmic running. In contrast, in the present work, we look at the actual process of renormalization and the associated subtleties.
We begin in Sec. \[sec:usual\_qft\] by recalling how counterterms are found in usual QFT. Then, in sec. \[sec:cosmo\_pert\], after introducing the particular regime of effective field theory of inflation for which we present the arguments about the UV divergent two-point correlator, we shall describe how the case of cosmological perturbation theory is so different from the usual QFT in so peculiar a way. We summarize the results in Sec. \[sec:summary\]. We have set $\hbar = c = 1$.
Counterterms in usual QFT: {#sec:usual_qft}
==========================
In usual QFT, LSZ reduction formula ensures that the most relevant quantity to evaluate is the vacuum expectation value of time ordered product of the Heisenberg picture fields. While evaluating the Fourier transforms of such correlators, at sufficiently high order in perturbation theory one encounters UV divergences. E.g. one could encounter quadratically divergent integrals of the form (for illustrative purposes, we work with Euclidean integrals) $$I_Q = \int \frac{d^4 \ell}{(\ell^2 + \Delta)} \; ,$$ which, on dimensional regularization gives $I_Q = \Delta \left( \frac{\mu^2}{\Delta} \right)^{\delta/2} F(\delta)$, where, $\delta = 4-d$, $\mu$ is the fake renormalization scale (which inevitably gets introduced while performing dimensional regularization) and $F(\delta)$ is a dimensionless function of $\delta$ which contains poles of $\delta$. If $F$ has a simple pole and its Laurent series expansion of is $F = F_{-1} \delta^{-1} + F_0 + F \delta + \cdots $, then,
$$\label{eq:usual_qft_quad}
I_Q = \Delta \left( \frac{F_{-1}}{\epsilon} + \frac{F_{-1}}{2} \log \left( \frac{\mu^2}{\Delta} \right) + F_0 + \cdots \right) \; .$$
Typically, $\Delta$ is a polynomial in the external momentum $k^2$($=-E^2+{\bf p}^2$) or the masses (and often, the Feynman parameters). The presence of $\Delta$ in the above expression causes the dimensionally regulated UV divergent integral to be a sum of two parts; the first part is polynomial in the external momentum with divergent coefficients and the second one is a finite function of the external momenta, the masses and the fake renormalization scale $\mu$. E.g. while evaluating the two-point function in $\phi^4$ theory in $d=4$, the corresponding $\Delta$ turns out to be simply $m^2$, so this is a trivial example. On the other hand, while renormalizing the two-point function for $\phi^3$ in $d=6$, we find that the Fourier transform of the two-point function is given by an expression of the form $$\label{eq:2_pt_usual_QFT}
{\cal G}(k^2) = {\cal G}_{\rm Free}(k^2) +
{\cal G}_{\rm Free}(k^2) \left[ \Pi(k^2) \right] {\cal G}_{\rm Free}(k^2) + \cdots \; ,$$ where ${\cal G}_{\rm Free}(k^2)$ is the Feynman propagator. $\Pi(k^2)$ is the contribution of the one-loop diagram (with external lines amputated) and is given by a quadratically divergent integral and the expression for dimensionally regulated $\Pi(k^2)$ is of the form $$\begin{aligned}
\label{Pi_usual_QFT}
\Pi(k^2) = A_2\left( \frac{1}{\epsilon} \right) k^2 + A_0 \left( \frac{1}{\epsilon} \right) m^2
+ {\cal F}(k^2,m,\mu) \; .\end{aligned}$$
The following points are important to notice
1. $\Pi(k^2)$ has two kinds of contributions: (1) a polynomial in the external momentum $k^2$ (the $m^2$ term can be thought of as the term $(k^2)^0$), (2) another function ${\cal F}$, which depends on, among other things, the fake scale $\mu$.
2. The coefficients of the polynomial in the external momentum $k^2$, the $A_i$, are divergent, they are functions of $1/\epsilon$, on the other hand, the function ${\cal F}$ is finite.
3. The forms of the terms in the polynomial i.e. how they depend on the external momenta, dictate the form and number of counter-terms needed to be introduced in the Lagrangian in order to cancel the UV divergences of the theory. E.g. in the above case, they tell us that we need two counter-terms to renormalize the two-point function and they shall be of the form $- \frac{A}{2} (\partial \phi)^2 $ and $- \frac{B}{2} m^2 \phi^2$. In fact, this is how we get convinced that the correlators can be renormalized at all.
4. In contrast, for a logarithmically divergent integral $$I_L = \int \frac{d^4 \ell}{(\ell^2 + \Delta)^2 } \; ,$$ following the above procedure shall simply give $$\label{eq:log_usual}
I_L = \left( \frac{F_{-1}}{\epsilon} + \frac{F_{-1}}{2} \log \left( \frac{\mu^2}{\Delta} \right) + F_0 + \cdots \right) \; .$$ It is clear that for a logarithmically divergent integral, the dependence on external momenta does not turn up. We could say that for this case, the polynomial is just one. This causes the counterterm to be just a factor such as $Z_g g$ (where $g$ is the coupling constant and $Z_g$ is the renormalizing $Z$ factor).
5. The function ${\cal F}$ has this property that in the limit $k^2 \gg m^2$ and $k^2 \gg \mu^2$, it takes up the form ${\cal F} \sim k^2 \log \left( \frac{k^2}{\mu^2} \right)$.
It is important to notice that all this holds good even when we regularize the UV divergent integrals in some other way. Similarly, it is not just a property of this theory: it can be easily seen for other theories (e.g. Yukawa theory, $\phi^4$ theory, QED): in general, the pole part of any sub-divergence-free diagram is a polynomial in its external momenta. It has been argued that this holds good also in other theories, where the residues of the poles always contain the external momenta and masses as low-order polynomials (see [@'tHooft1972189], [@PhysRevD.10.1201], ch. 9 of [@kleinert2001critical], pg. 148 of [@hooft1994under]). Thus, a polynomial in external momenta plus a function which is logarithmic in external momenta is a generic feature of Green’s functions in usual QFT.
At this point it is worth re-emphasizing that if we wish to find the number of counter terms needed to renormalize a correlator or the form of these counter terms (things we wish to find, even for a non-renormalizable theory), the dependence of the Green’s function on external momenta is a very important tool.
Counterterms in cosmological perturbation theory {#sec:cosmo_pert}
================================================
In inflationary cosmological perturbation theory, the most relevant correlator is the late time limit of $n-$point function of (Heiseberg picture) comoving curvature perturbation, $\zeta$ on a constant time hyper-surface $$\lim_{\eta \rightarrow 0^-}
\langle \Omega | \zeta(\eta,{\bf x}_1) \zeta(\eta,{\bf x}_2) \cdots \zeta(\eta,{\bf x}_n) | \Omega \rangle \; .$$ Homogeneity of the inflationary background implies that all such correlators shall be invariant under translations in space at a fixed time, which implies that the Fourier transform of the above correlator shall be of the form $$\label{eq:def_G}
F(\eta, {\bf k}_1, \cdots, {\bf k}_n) = (2 \pi)^{3}~ \delta^3({\bf k}_1 + \cdots + {\bf k}_n)~
\tilde{\cal G}(\eta, {\bf k}_1, \cdots, {\bf k}_n) \; .$$ It is worth mentioning that whenever we talk about the usual QFT, we shall be dealing with four dimensional Lorentzian momenta while whenever we talk about cosmological perturbation theory, we shall be dealing with three dimensional Euclidean momenta. At sufficiently high order in perturbation theory, one expects to encounter Feynman diagrams with loops. This issue, in cosmological perturbation theory has been studied in great detail in the last few years. Beginning with [@PhysRevD.72.043514], there was a debate about whether these loop corrections to the cosmological correlations shall freeze at late times. Recently (see [@SZ_all_loops; @2013JHEP...02..151A]), it is claimed to be shown that this shall surely happen at all loops. The familiar primordial power spectrum $\Delta^2_{\zeta}(k)$ is defined by the Eq. $$\lim_{\eta \rightarrow 0^{-}} \langle \zeta(\eta,{\bf k}) \zeta(\eta,{\bf k}') \rangle =
\frac{2 \pi^2}{k^3}
~ \delta^3({\bf k} + {\bf k}') ~ \Delta^2_{\zeta}(k) \; ,$$ for slow roll inflation with a single (canonical) scalar field, the lowest order (i.e. tree level) contributions to power spectrum, assuming Bunch-Davies vacuum (in the limit $\eta \rightarrow 0^{-}$), is given by (the classic result) $$\label{eq:PS:free}
\Delta^2_{\zeta}(k)|_{\rm tree} = \frac{1}{2 \epsilon(\eta_k)} \left(\frac{H(\eta_k)}{2 \pi M_{\rm Pl}} \right)^2 \; ,$$ where $\eta_k$ is the conformal time when the mode in question crosses Hubble radius (i.e. when $k = a H$). Notice that, on comparing with Eq (\[eq:def\_G\]), it becomes clear that for two-point function, $$\lim_{\eta \rightarrow 0} \tilde{\cal G}(\eta, {\bf k})|_{\rm tree} =
\frac{1}{4 \pi k^3} \Delta_{\zeta}^2 (k) = \frac{1}{32 \pi^3 k^3} \left( \frac{H^2}{\epsilon M_{\rm Pl}^2 } \right)\; .$$ Since the result in Eq (\[eq:PS:free\]) is obtained in linear theory, it corresponds to a free theory calculation. A well known fact is that Eq (\[eq:PS:free\]) implies that, the dimensionless Primordial Power spectrum is a power law $$\Delta_{\zeta, {\rm Free}}^2 (k) = A_s \left( \frac{k}{k_0} \right)^{n_s - 1} \; ,$$ where $A_s$ is the spectral index while $n_s$ is the spectral amplitude and $k_0$ is a pivot scale, see [@2009arXiv0907.5424B]. This implies that $$\label{eq:G:zeta:free}
\tilde{\cal G}_{\rm Free}(k) = \frac{A_s}{4 \pi k_0^3} \left( \frac{k}{k_0} \right)^{n_s - 4} \; .$$
The leading interactions of the metric fluctuations are typically due to cubic operators, so, it is expected that the loop correction to $\tilde{\cal G}$ shall be of the form (notice that the mass dimension of $\tilde{\cal G}$ is $-3$) $$\tilde{\cal G}_{\rm 1-loop} = k^3 (\tilde{\cal G}_{\rm tree})^2 \times ({\rm factors}) \; .$$ The factors on the RHS can involve various non-trivial logarithmic runnings e.g. $\log a \sim Ht$ or $\log kL$ [@2011JCAP...01..023G; @2010JCAP...08..006B] or $\log (H/\mu)$ [@SZ2010], while a running of the form $\log k/\mu$ is not possible since this shall not leave $\tilde{\cal G}$ invariant under the transformation $x \rightarrow \Lambda x$, $k \rightarrow k/\Lambda$ and $a \rightarrow a/\Lambda$ [@SZ2010].
The $\dot{\pi}^3$ theory
------------------------
While evaluating correlations for an interacting quantum field on an accelerating universe, one can encounter, apart from the familiar UV divergences, relatively unfamiliar divergences too. To avoid having to deal with these unfamiliar divergences, and still deal with a realistic model of cosmological perturbations which has simple interactions, we work with the theory of large $\dot{\pi}^3$ interactions [@SZ2010]. The Effective Field Theory (EFT) [@EFT2008] of inflation provides the most general framework for systematically studying the dynamics of fluctuations around an inflationary background solution. The action of the theory of fluctuations can be expanded in powers of the relevant fluctuation field (and also in powers of the slow-roll parameters such as the Hubble flow functions). The theory is first formulated in unitary slicing of the perturbed spacetime in which $\delta \phi$ vanishes (and all the dynamics lies in the metric) and then general covariance is restored by introducing the Stueckelberg field (denoted by $\pi$ in the following). It turns out that if one chooses to ignore ${\cal O}(\epsilon^2)$ terms in the action (which also corresponds to the interaction terms which give rise to primordial non-Gaussianity for a canonical scalar field, see [@Maldacena-2003]), and one chooses to fix the sound speed of fluctuations to unity (i.e. we wish to only consider the cases in which $c_s \rightarrow 1$), [^1] the leading order interactions for the Stueckelberg field shall be captured by terms of the form (ignoring ${\cal O}(\pi^5)$ terms, see [@EFT2008]) $$\frac{M_3^4}{6} \left(g^{00} + 1 \right)^3 = - \frac{2}{3} M_3^4 \left[ 2 \dot{\pi}^3 + 3 \dot{\pi}^4
- \frac{3}{a^2} \dot{\pi}^2 (\partial_i \pi)^2 \right] \; .$$
It is important to recognize that, thanks to the EFT formulation, it is very easy to identify a regime in which the Stueckelberg field has non-negligible self-interactions without violating the slow-roll nature of the background solution. Since the symmetry arguments can not fix the value (or sign) of the coefficient $M_3$ (which should be determined from observations, see [@Planck-2013-NG] for the latest limits), one can write $M_3^4 (t) = - c_3 (t) M^4$ where $M$ is a mass scale characterizing the interaction. An extra shift symmetry can be imposed requiring that the time dependence of $c_3$ is negligibly weak.
In this regime, the action of the $\pi$ field becomes (see [@SZ2010]) $$\begin{aligned}
\label{S_SZ2010}
S=\int d^4x \; a^3 \left[ -\dot{H} M_{\rm Pl}^2 \left(\dot\pi^2-\frac{1}{a^2}(\partial_i\pi)^2\right) + \right. \\ \nonumber
\left. \frac{2}{3}c_3 M^4\left(2\dot\pi^3+3\dot\pi^4-3\frac{1}{a^2}\dot\pi^2(\partial_i\pi)^2\right)\right]\ ,\end{aligned}$$
This is perhaps the simplest possible interacting cosmological perturbation theory. It is also observationally interesting (see [@Planck-2013-NG]). This is the most general kind of interactions which are not slow roll suppressed and which are leading order when we impose the requirement that $c_s \rightarrow 1$. To connect to the usual perturbation variables, one can make a gauge transformation to comoving gauge (see [@Cheung-consistency]) and find that $\zeta = - H \pi + {\cal O}(\pi^2)$.
We would like to emphasize again that for the inflationary background caused by canonical scalar fields, the interaction terms in the action of $\zeta$ are ${\cal O}(\epsilon^2)$ so that the interactions we are dealing with are different from those. Moreover, as is obvious from Eq (\[S\_SZ2010\]), in this regime the interaction terms are much simpler (and fewer) as compared to e.g. those in [@Maldacena-2003]. Thus, if we wish to try anything new e.g. loop corrections in cosmological perturbation theory and we want a regime which is realistic but which is also simple, this theory of large $\dot{\pi}^3$ interactions is the best possible choice [@SZ2010].
Given the action (Eq (\[S\_SZ2010\])), the Hamiltonian can be readily worked out and then one can use the in-in formalism (see [@PhysRevD.72.043514; @Contour_rotation] and references therein) to evaluate the two-point function at one-loop. In the rest of the present work, we shall focus on only the $\dot{\pi}^3$ term in the action given by Eq. (\[S\_SZ2010\]). This interaction term leads to two contributions to the two-point function only one of which is UV divergent (see [@SZ2010]). On canonical normalization, it becomes apparent that the $\dot{\pi}^3$ interaction is of mass dimension $+6$. In the action, a dimension six operator is expected to be accompanied with a factor of $1/{\Lambda_U}^2$, where $\Lambda_U$ is the unitarity bound of the theory. This is what happens, on canonical normalization, $\pi_c \equiv \sqrt{- 2 \dot{H} M_{\rm Pl}^2 } ~ \pi$, and the coefficient of $\dot{\pi}_c^3$ operator turns out to be $$\frac{4}{3} \cdot \frac{c_3}{(2 \epsilon)^{3/2}} \cdot \frac{M^4}{H^3 M_{\rm Pl}^3}
= \frac{4 c_3}{3} \cdot \frac{1}{\Lambda_U^2}$$ where $\Lambda_U$ is the energy scale at which this (non-renormalizable) theory becomes strongly coupled (i.e. the perturbative calculations are valid only at energy scales much smaller than this scale). We have, $$\frac{1}{{\Lambda_U}^2} = \frac{M^4}{(2 \epsilon)^{3/2} H^3 M_{\rm Pl}^3}$$ In correlators, it is expected that the (three-line) interaction vertex of $\dot{\pi}^3$ theory is always going to be accompanied with factors of the form $\frac{4 c_3}{3} \frac{H^2}{\Lambda_U^2}$ since $H$ is the energy scale of the inflationary experiment. This suggests that for $\dot{\pi}^3$ theory, $$\label{gen_exp}
\tilde{\cal G}_{\rm 1-loop} = k^3 (\tilde{\cal G}_{\rm tree})^2 \times \left( \frac{H^2}{\Lambda_U^2} \right)^2
({\rm factors}) \; .$$
Dimensional analysis and homogeneity of the background suggest that the correlator $\langle\zeta_{\vec k}(\eta)\zeta_{\vec k'}(\eta)\rangle$ shall be of the form $$\langle\zeta_{\vec k}(\eta)\zeta_{\vec k'}(\eta)\rangle =
\frac{\delta^3(\vec k + \vec k')}{k^3} \times \left({\rm rest} \right) \; ,$$ where the rest terms have to be dimensionless. In the one loop calculation, the rest terms shall contain UV divergent momentum integral. Thus, apart from a few numerical factors and a factor of $(c_3^2 M^8/\epsilon^4 M_{\rm Pl}^8)$, the rest terms shall be of the form $$\label{eq:loop_SZ_gen}
\int d^3{\vec k}_1 d^3{\vec k}_2 \delta^3( {\vec k} + {\vec k}_1 + {\vec k}_2 )
f(\eta, H, {\vec k}, {\vec k}_1, {\vec k}_2) \; .$$ Power counting makes it clear that the mass dimension of $f$ is $-3$. Since this integral is dimensionless, naively, we’d expect that on dimensional regularization it would give $$\left( \frac{k}{\mu} \right)^\delta F(\delta) \; ,$$ where $F(\delta)$ is a dimensionless function (which contains poles of $\delta = D-3$). This result shall not leave $\tilde{\cal G}$ invariant under the transformation $x \rightarrow \Lambda x$, $k \rightarrow k/\Lambda$ and $a \rightarrow a/\Lambda$. Thus, it is not correct and a detailed calculation (by Senatore and Zaldarriaga [@SZ2010]) shows that in fact the UV divergent momentum integral gives $$\left( \frac{k}{\mu} \right)^\delta F(\delta) G(\delta,k,-\eta, H) \; ,$$ and where $G$ is another dimensionless function and when $\eta = 0$, it is of the form $$G = 1 + \delta \log (-c H \eta_k) + \cdots \; ,$$ where $c$ is an ${\cal O}(1)$ constant. This changes the logarithmic running to $\log (H/\mu)$ [@SZ2010].
Counterterms
------------
A careful look at the argument presented by Senatore and Zaldarriaga [@SZ2010] (to establish that the logarithmic running is $\log (H/\mu)$) also tells that when $\eta = 0$, only one kind of divergence is present. This may suggest from the arguments in familiar applications of QFT (Sec. \[sec:usual\_qft\]) that we need just one counterterm to cancel the UV divergences in the two-point function in this theory. Thus, when $\eta = 0$, we have (see [@SZ2010]), $$\tilde{\cal G}_{\rm 1-loop} \sim \frac{1}{k^3}
\left( \frac{F_{-1}}{\epsilon} + \frac{F_{-1}}{2} \log \left( \frac{H(\eta_k)}{\mu} \right) + F_0 + \cdots \right) \; ,$$ where a factor of $(c_3^2 M^8/\epsilon^4 M{\rm Pl}^8)$ is understood to sit in the front, apart from some numerical factors. This implies that $$\label{G_loop_asym}
\tilde{\cal G}_{\rm 1-loop} \sim k^3 \tilde{\cal G}_{\rm tree}^2
\left( \frac{F_{-1}}{\epsilon} + \frac{F_{-1}}{2} \log \left( \frac{H(\eta_k)}{\mu} \right) + F_0 + \cdots \right) \; ,$$ with a factor of $(c_3^2 M^8/\epsilon^2 M_{\rm Pl}^4 H^4)$ in the front. This should be compared with Eq. (\[Pi\_usual\_QFT\]) and (\[eq:usual\_qft\_quad\]).
At this point it is worth reminding ourselves that the theory that we are dealing with is a non-renormalizable theory so that the countererm (CT) needed is not necessarily one of the terms we have already written down in the action Eq. (\[S\_SZ2010\]). From Eq. (\[gen\_exp\]), it is clear that the loop correction shall have a factor of $(H/\Lambda_U)^4$. Again, by dimensional analysis, it is clear that a single vertex of dimension $+8$ operator can give this factor. But since we want the CT to renormalize the two-point function, it better have two external lines. Thus, counterterm shall be a dimension $+8$ quadratic operator. On canonical normalization, $\pi_c$ has dimension $+1$ and shift symmetry forbids any polynomials in $\pi_c$ to be present in the action. Thus, we can only take derivatives, hence CTs can only be operators of the form $(\partial^3 \pi_c)^2$. The derivatives that we can take are either w.r.t. time or w.r.t. space, since we want to write a rotationally invariant action, the only options are the square of $\partial_t \partial_t \partial_t \pi $, the square of $\partial_t \partial_i \partial_i \pi$, and $\partial_t \partial_t \partial_i \pi \partial_t \partial_t \partial_i \pi$. This means that by dimensional analysis, there are three possible candidates for the CTs. On the other hand, we have only one kind of divergent term present in the dimensionally regulated expression for the $\langle\zeta_{\vec k}(\eta)\zeta_{\vec k'}(\eta)\rangle$ when $\eta = 0$ since there is just a monomial of $k^3$ in the front (see Eq. (\[G\_loop\_asym\]) as compared to Eq. (\[Pi\_usual\_QFT\])). In this case, we cannot determine the coefficients of the operators in the CT Lagrangian in any unique way.
When $\tilde{\cal G}_{\rm 1-loop}$ is worked out for $\eta$ which is non-zero but still such that $-k\eta \ll 1$, then we get an expression of the form (save for some numerical factors) $$\begin{aligned}
\tilde{\cal G}_{\rm 1-loop} &\sim& k^3 ~\tilde{\cal G}_{\rm tree}^2 ~\frac{c_3^2 M^8}{\epsilon^2 M_{\rm Pl}^4 H^4} \times \\ \nonumber
&& \left( \frac{F_{-1}}{\epsilon} + \frac{F_{-1}}{2} \log \left( \frac{H(\eta_k)}{\mu} \right) + F_0 + \cdots \right) \times \\ \nonumber
&& \left( \sum_{i=0}^{n} c_i (- k \eta )^{i} \right) \; ,\end{aligned}$$ i.e., we get an extra term multiplied to the $\eta = 0$ result which is a polynomial, not in the external momentum $k$ but in $-k\eta$. Thus, in this case, even for a dimensionless integral (see the discussion after Eq.(\[eq:loop\_SZ\_gen\])) we get an additional polynomial, but it is an effect of having non-zero $\eta$. This is to be compared with Eq.(\[eq:log\_usual\]) in flat spacetime, where a dimensionless integral which leads to a logarithmic divergence gives no polynomials in external momenta and hence the correspondng CTs are trivial.
What is important is the fact that the quadratic CTs (of the form required) also end up giving a polynomial in $-k\eta$ and hence renormalization can be performed in this limit. But in order to renormalize, we need to determine the (unknown) coefficients of the operators in the CT Lagrangian. In the case of familiar field theory (see pt. (4) after Eq. (\[Pi\_usual\_QFT\])), each individual CT gives a different kind of dependence on the external momenta, on the other hand here, that is not the case, we get a polynomial in external momenta whose coefficients are linear combinations of the coefficients of the operators in the CT Lagrangian. Thus, unlike the case of familiar field theory, even in the case with $-k\eta \ll 1$ but $\eta$ being non-zero, we cannot just look at the expression for dimensionally regulated correlator and read-off the number of CTs required or their forms. Thus, to determine the (unknown) coefficients of the operators in the CT Lagrangian, we need to solve a set of linear equations, it so happens that there are four linear Eqs. in the three unknowns and there still is a unique solution. Geometrically, this is like having four planes in the three dimensional Euclidean space and still they all passing through one common point. The choice of the renormalization conditions shall determine this point. This may raise the following concern: could it be that if we find $\tilde{\cal G}_{\rm 1-loop}$ for $\eta$ such that $-k \eta \gtrsim 1$, we end up having many more equations and only three unknowns? Would the solution be guaranteed to exist in that case?
But most importantly, in the limit $\eta \rightarrow 0$, only one divergent term is left and thus in taking this limit, we end up erasing the information about the form of CTs or their number completely. This is similar to what happens in e.g. spontaneous symmetry breaking. If we consider the $Z(2)$ symmetric renormalizable scalar field theory ${\cal L} = - \frac{1}{2} (\partial \phi)^2 - m^2 \phi^2 - \lambda \phi^4/24$. When $m^2 > 0$, three CTs are enough to absorb all the infinites in the theory. The same is true when $m^2 < 0$, but in that case, we can also write the same theory as ${\cal L} = - \frac{1}{2} (\partial \rho)^2 - \lambda v^2 \rho^2/6 - \lambda v \rho^3/6 - \lambda \rho^4/24$ (where $v = + (6 |m^2|/\lambda)^{1/2}$ and $\rho = \phi - v$) and looking at this Lagrangian, it may appear that we shall need more CTs for the $\rho^3$ term as well as to cancel the tadpole (i.e. divergent one-point function) it shall cause. In this case, the change of variables from $\phi$ to $\rho$ seems to suggest that we shall need more CTs while in reality, we do not. In contrast for the case we are dealing with, the process of taking $\eta \rightarrow 0$ limit suggests that we shall need fewer CTs, while in reality we do not.
Remarks on renormalization
--------------------------
In $\dot{\pi}^3$ theory, the field is $\pi$ while the various parameters appearing in the action are $\epsilon, H/M_{\rm Pl}$ and $c^3 M^4$. The parameter $c^3 M^4$ can be constrained from the observations of Primordial Non-Gaussianity in the CMB sky (and it has been constrained by the Planck collaboration [@Planck-2013-NG]). There also are (upper) limits on the values of $\epsilon$ and $H/M_{\rm Pl}$ [@2013arXiv1303.5082P]. If we knew the actual values of these parameters (rather than just the limits), we expect that we could work in the On-shell (OS) renormalization scheme to perform the actual renormalization of the theory. In usual QFT, while renormalizing the two-point function in OS scheme, we choose the counter-terms such that (1) the divergent part in the loop integral (the $1/\epsilon$ term) gets canceled, (2) the dependence on the fake renormalization scale gets cancelled, and, (3) the rest of the part of the counterterm is chosen such that $\Pi(-m^2) = 0$ and $\Pi'(-m^2) = 0$, see [@quantum], these conditions ensure that the parameter $m$ appearing in the Lagrangian is the physical mass and the field strength is normalized (thus, at $k^2=-m^2$, the propagator has a simple pole with unit residue). In Eq. (\[G\_loop\_asym\]), we can use $$\log \left( \frac{H(\eta_k)}{\mu}\right) =
\log \left( \frac{H(\eta_k)}{H(\eta_{k_0})}\right) + \log \left( \frac{H(\eta_{k_0})}{\mu}\right) \; ,$$ where $k_0$ is a pivot scale to absorb the $\mu$ dependent term in the CT, but instead of the two conditions (given by Lehmann-Källén form of the exact propagator), in the usual flat spacetime field theory, we shall need three (since we have three CTs) and we have no equivalent of the Lehmann-Källén spectral representation of the propagator. Thus, in cosmological perturbation theory, we seem to have no straightforward way to apply the OS scheme since we can not naturally relate the measured correlations to the physical values of the parameters of the Lagrangian of the theory. At this stage, we can again go back to usual field theory to seek inspiration about how to perform renormalization. In usual QFT, in any chosen renormalization scheme, when we have the expression for e.g. ${\cal G}(k^2)$ (the notation of Eq. (\[eq:2\_pt\_usual\_QFT\])), we find $M$ such that at $k^2 = - M^2$, ${\cal G}(k^2)$ has a simple pole, we realize that $M$ is the physical mass (as opposed to the parameter which merely turns up in the renormalized part of the Lagrangian). At $k^2 = - M^2$, if the residue of ${\cal G}(k^2)$ is $R$ (and in general, it is not unity), we define the renormalized field as $\phi_{\rm ren} \equiv \phi/{\sqrt R}$, so that the residue of the pole for renormalized field is unity.
In cosmological perturbation theory, we could proceed in the following way: if loop corrected power spectrum is $$\Delta^2_{\zeta} (k) = A_s \left( \frac{k}{k_0} \right)^{n_s - 1} \left( 1 + g(k,\mu,c_3 M^4, H) \right) \; ,$$ then, in general at $k = k_0$, $\Delta^2_{\zeta} (k) \neq A_s$. But if $\Delta^2_{\zeta} (k_0) = R(\mu)$, then, let us redefine $\zeta$ such that $$\zeta_{\rm ren} (\mu) \equiv \sqrt{ \frac{A_s}{R(\mu)} } \zeta \; ,$$ and this is how field strength renormalization could be done in cosmological perturbation theory. Moreover, one could use the observations of $A_s$, $n_s$ and $dn_s/d \log k$ (the running of the spectral index) to fix the finite parts of the three CTs of this theory. At this stage however, the observational constraints on most parameters: $M_3$ [@Planck-2013-NG], $\epsilon$, $H$ (during inflation), $dn_s/d \log k$ [@2013arXiv1303.5082P] are not good enough to perform this procedure. Notice that had we taken the $\eta \rightarrow 0$ limit before renormalizing, we could not have known that we need three CTs and the above would not have been possible.
Summary {#sec:summary}
=======
In this work, we explored issues of renormalization in cosmological perturbation theory. In the more familiar applications of QFT, a logarithmically divergent loop integral has a trivial polynomial dependence on external momenta e.g. $${\cal I}_1(k) = \int_0^{\infty} \frac{d \ell}{\ell + k} = C - \log k \; ,$$ (where $C$ is divergent), on the other hand, for a quadratically divergent integral ${\cal I}_2$, one gets, $${\cal I}_2(k) = \int_0^{\infty} \frac{\ell d \ell}{\ell + k} = a + b k + k \log k \; ,$$ where $a$ and $b$ are divergent. In general, in usual QFT, every diagram with external lines amputated and with no sub-divergences is of the form $$G = \sum_{i = 0}^{n} A_i (1/\epsilon) (k^2)^i + G_{\log} (k^2,m,\mu) \; ,$$ (with $k^2$ being Lorentz invariant) where, for the case of a logarithmically divergent diagram, only the the $i=0$ term is present. In cosmological perturbation theory, in contrast, even for a logarithmically divergent diagram, one gets, when $-k\eta \ll 1$, $$G = \left[ A_0 (1/\epsilon) k^3 + G_{\log} (k,m,\mu) \right] \left[ \sum_{i=0}^{n} c_i (- k \eta )^{i} \right] \; .$$ We argued that if one intends to perform renormalization, one needs to identify the CTs from the expression of dimensionally regulated correlators. Unlike the case of usual QFT, here, when $-k\eta \ll 1$, the expression for dimensionally regulated correlator is a polynomial in $-k\eta$, and taking the limit $\eta \rightarrow 0$ erases information about the forms and number of CTs required to renormalize the theory. We thus realized that one should be very careful in taking the limit in which the external time $\eta$ goes to zero. We also explored how the process of renormalization could be performed in cosmological perturbation theory (e.g. how $\zeta$ shall undergo field strength renormalization). This illustrates the many subtleties and surprises associated with field theoretic aspects of cosmological perturbation theory.
[**Acknowledgment:**]{} The author thanks T. Padmanabhan (IUCAA, Pune), U. Yajnik (IIT-Bombay, Mumbai) and R. Rangarajan (PRL, Ahmedabad) for comments on a previous draft of the manuscript. \[Bibliography\]
[^1]: Here, the decoupling limit has already been taken and so the terms in the action of EFT which cause Stueckelberg field to mix with gravity already vanish, see [@EFT2008] for details.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The massive exploitation of cosmic voids for precision cosmology in the upcoming dark energy experiments, requires a robust understanding of their internal structure, particularly of their density profile. We show that the void density profile is insensitive to the void radius both in a catalogue of observed voids and in voids from a large cosmological simulation. However, the observed and simulated voids display remarkably different profile shapes, with the former having much steeper profiles than the latter. Sparsity can not be the main reason for this discrepancy, as we demonstrate that the profile can be recovered with reasonable accuracy even with very sparse samples of tracers. On the other hand, the observed profile shows a significant dependence on the galaxy sample used to trace the matter distribution. Samples including low-mass galaxies lead to shallower profiles with respect to the samples where only massive galaxies are used, as faint galaxies live closer to the void centre. We argue that galaxies are biased tracers when used to probe the matter distribution within voids.'
author:
- |
E. Ricciardelli$^{1}$[^1], V. Quilis$^{1}$, J. Varela$^{2}$\
$^{1}$Departament d’Astronomia i Astrofisica, Universitat de Valencia, c/ Dr. Moliner 50, E-46100 - Burjassot, Valencia, Spain\
$^{2}$Centro de Estudios de Física del Cosmos de Aragón (CEFCA), Plaza San Juan 1, 44001 Teruel, Spain
date: 'Accepted ... Received ...; in original form ... '
title: On the universality of void density profiles
---
\[firstpage\]
cosmology: dark matter – cosmology: observations– large-scale structure of Universe – methods: numerical
Introduction {#intro}
============
Large redshift surveys [@York00; @Colless01] and cosmological simulations [@Bond96] have revealed that galaxies are distributed inside a cosmic web of walls, filaments and compact clusters. Such a web encloses large underdense regions, referred to as cosmic voids.
Voids were first recognized in the earliest redshift surveys [@Gregory78; @Kirshner81] as huge empty holes in the galaxy distribution. Nowadays, there is a general consensus in that voids occupy most of the volume of the Universe [@Sheth04; @Vandew11; @Pan12], although there is not yet an agreement on how a genuine void should be defined. Several void finders, which are based on different principles, have been developed. Voids can be identified as spherical regions devoid of galaxies/haloes [@Gottlober03; @Patiri06b; @Varela12] or underdense regions, relying on the continuous density field [@Plionis02; @Colberg05]. More complex algorithms able to capture the complex morphology of voids also exist [@Platen07; @Neyrinck08; @AragonCalvo13]. Despite their different definition of voids, all these void finders agree in that voids ere extremely empty in the centre and show a sharp increase in the density towards the voids edges (e.g. @Colberg08).
Voids are believed to originate from negative density fluctuations in the primordial density field. As a result of their underdensity, they are subject to an effective repulsive peculiar gravity, causing their expansion. As a consequence of such an expansion, the matter within the voids evacuates from the interior and accumulates to the boundaries. This leads to void density profiles that evolve towards a reverse top-hat shape [@Sheth04].
A considerable appeal of cosmic voids is their potential in probing cosmological parameters. In particular, being almost devoid of matter, they are extremely sensitive to the nature of dark energy. Indeed, the void ellipticity and its evolution through cosmic time are intimately connected with the local tidal tensor, which, in turn, depends on the dark energy content [@Park07; @Lavaux10; @Bos12]. Voids are also the ideal candidate for probing the expansion history of the Universe through the Alcock-Paczynski test [@AP79], using the average shape of stacked voids [@Lavaux12; @Sutter12a]. The application of such a test to the voids that will be identified in the future Euclid survey [@Laureijs11] promises to outperform Baryonic Acoustic Oscillation by an order of magnitude in accuracy.
The huge potentiality of voids for precision cosmology requires a robust knowledge of their internal structure, particularly of the density profiles. Works based on cosmological simulations [@Colberg05; @Ricciardelli13] indicate that the void density profile is universal. As such, it does not depend on void size. On the observational side, the ideal approach to directly constrain the void density profile is through the weak lensing signal of stacked voids [@Krause13]. However, the number of voids available from spectroscopic catalogues is still limited to provide a robust measurement of the signal [@Melchior13]. At present, we can only rely on the galaxy distribution to trace the density within voids [@Sutter12b]. Thus, to robustly assess a void model to describe the universal density profile, one also needs to assess the systematic effects arising from the use of the sparse galaxy sampling.
In a previous work (@Ricciardelli13, hereafter RQP13), we have shown, by means of a cosmological simulation, that a two parameters law can be used to fit the density profile of voids of any size, density, morphology and redshift. The best-fit parameters show some dependence on redshift, density, and, on a less degree, morphology, but they are almost independent on the void size, although the limited statistics prevented us to draw robust conclusions. In this work, we want to test this model and its dependence on void radius, against an observed catalogue of voids and a larger simulation, thus dramatically increasing the statistics. In doing so, we provide a robust determination of the systematic effects arising when using the sparse distribution of void galaxies as density tracers.
The structure of the paper is as follows. In Section \[sdss\] we introduce our catalogue of observed voids, in Section \[sim\] we describe the simulation used and our void identification procedure. The results on the void density profiles are discussed in Section \[profiles\]. We conclude in Section \[conclu\].
The SDSS void catalogue {#sdss}
=======================
The catalogue of cosmic voids used for the present analysis has been described in @Varela12. Here we only give a brief description of the main features and the changes with respect to that work.
The galaxy sample used for void identification has been extracted from the New York University Value-Added Galaxy Catalog[^2] (NYU-VACG; @Blanton05), based on the photometric and spectroscopic catalog of SDSS/DR7[^3], complete down to $r\sim 17.8$. These authors also provide stellar masses computed with the code kcorrect (version 4.1.4) following the prescriptions of @Blanton07. Stellar masses have been computed assuming $h=1$. To guarantee the homogeneity of the sample and avoid the detection of spurious voids, a complete catalogue up to redshift $0.12$ and down to magnitude $M_r-5logh=-20.17$ has been used. Using this galaxy sample, voids are defined as spherical regions devoid of galaxies. In the original catalogue of @Varela12 only voids larger than $10 \, h^{-1}\, Mpc$ were considered for the analysis. In this work, in order to increase the number of voids, we have extended the original catalogue to include voids down to $7 \,
h^{-1}\, Mpc$. Moreover, we relax the assumption on void overlapping of @Varela12 and consider as separate voids all overlapping voids whose distance between centers is larger than the radius of the largest void. The final catalogue contains $4453$ voids, with radius as large as $18.7 \, h^{-1}\, Mpc$. We find a total of $44617$ void galaxies, which are, by definition, fainter than $M_r-5logh=-20.17$.
Figure \[completeness\] shows how our void galaxies populate the redshift stellar mass plane. For each redshift bin we compute a stellar mass threshold (coloured points), above which the sample can be considered complete. To choose this mass, we have computed the number counts in mass bins and considered the threshold mass as the central mass of the bin having the largest counts.
![Upper panel: void galaxies (black points) in the redshift stellar mass plane. For the sake of clearness only a randomly selected subsample, including 15% of the galaxies, has been plotted. The coloured points indicate the stellar mass threshold adopted for each redshift bin. Lower panel: number of galaxies with redshift less than $z$ and more massive than the threshold mass at $z$. []{data-label="completeness"}](fig1.eps){width="\columnwidth"}
Voids in simulations {#sim}
====================
The simulation used in this work has been performed with the hydrodynamical code MASCLET (Quilis 2004). MASCLET couples an Eulerian approach for describing the gaseous component with an N-body scheme for treating the dark-matter, collisionless component. Gas and dark matter are coupled by the gravity solver. To gain spatial and temporal resolution an adaptive mesh refinement (AMR) scheme is implemented.
The numerical simulation was run assuming a spatially flat $\Lambda
CDM$ cosmology, with the following cosmological parameters: matter density parameter, $\Omega_m=0.27$; cosmological constant, $\Omega_{\Lambda}=\Lambda/{3H_o^2}=0.73$; baryon density parameter, $\Omega_b=0.045$; reduced Hubble constant, $h=H_o/100 km\, s^{-1}\,
Mpc^{-1}=0.71$; power spectrum index, $n_s=1$; and power spectrum normalisation, $\sigma_8=0.8$.
The initial conditions are set up at z = 100, using a CDM transfer function from @EiHu98, for a cube of comoving side length $512 \, h^{-1}\, Mpc$. The computational domain is discretized with $512^3$ cubical cells. The mass resolution is thus $\sim6\times
10^{10}\,h^{-1}\,M_{\odot}$ and the coarse spatial resolution is $1 \, h^{-1}\, Mpc$.
Following the philosophy of the simulation presented in RQP13, we have designed the simulation to follow the formation and evolution of low-density regions. Contrary to the common practice in AMR simulations, where the high density regions are refined, we use more resolution in low density regions. During the evolution, the regions in the coarse grid are refined based on the local density, when $\rho/\rho_B<10$, being $\rho$ and $\rho_B$ the total density and the background density, respectively. The ratio between the cell sizes for a given level ($l+1$) and its parent level ($l$) is, in our AMR implementation, $\Delta
x_{l+1}/\Delta x_{l}=1/2$. Since in this work we are not interested in the study of the void sub-structures and void galaxies, in this simulation we have only used one level of refinement. The best spatial resolution is therefore $0.5 \, h^{-1}\, Mpc$.
Voids are identified in the simulated volume using the void finder algorithm described in RQP13. This algorithm relies on the continuous density field (including dark matter and gas) to identify the low density regions, that we define as voids. It is based on the basic assumptions that the velocity divergence of the gas within the void is always positive (as a result of void expansion) and that the density at the void edges has a sharp increase. Broadly speaking, it performs the following steps. It first marks cells as candidate for being centers of voids when their overdensity is below a threshold limit and the velocity divergence is positive. It then expands these volumes by adding cells on each coordinate directions until one of the conditions that define the void edge is reached. Void edges are reached when the velocity divergence becomes negative or the density gradient exceeds a threshold value. The procedure thus provides the protovoid, the minimum rectangle parallelepiped contained within a void. To build the actual void, protovoids are allowed to merge with each other when the ratio between the overlapping volume and the largest void is within $0.5$ and $0.6$. The free parameters involved in the procedure have been set by means of extensive tests of the code on a set of Montecarlo mock voids, as well as on the voids in the simulation. We adopt the same reference values as in RQP13. The density and the velocity divergence used are those defined in the base level (l=0) grid.
The final sample of simulated voids includes a total of $\sim 35000$ voids, filling 60% of the simulated volume and with typical overdensity $\rho/\rho_B=0.2$. For the analysis of the density profiles, we restrict the sample only to large voids, with effective radius[^4] $R_e>7\,h^{-1}\,Mpc$. We also exclude voids having too large porosity and ellipticity, as they are the most affected by contamination from non-void regions. We end up with 3186 voids.
Void density profiles {#profiles}
=====================
![Stacked density profiles for all voids larger than $7\,h^{-1}\, Mpc$ identified in the SDSS database (black diamonds) and the best-fit model (black line). As a comparison, the coloured symbols indicate the mean profiles of MASCLET voids larger than $7\,h^{-1}\, Mpc$. The simulated profiles are computed using various density tracers: total - dark matter plus gas - density field (blue circles), gas density field (orange squares) and dark matter particles (green triangles). The solid coloured lines indicate the best-fits for each curve. $R_e$ refers to the radius of the voids. For the simulated voids, which have arbitrary shape, $R_e$ is defined as an equivalent spherical radius, i.e. the radius of the sphere having the same volume of the void. The discrepancy between observed and simulated profiles is addressed in Section 4.2.[]{data-label="profiles_all"}](fig2.eps){width="\columnwidth"}
To derive the void density profiles for the simulated voids, we can rely on two different tracers: the continuous density field and the dark matter particles. In both cases, we compute the density in spherical apertures, thus discarding the information about void shape. When using the continuous density field, we compute the profiles for the individual voids and then use the bi-weight estimator at any given aperture for getting the stacked profile. We have restricted the analysis of the profiles to $r \geqslant 0.3R_e$, as in the observed voids we can not reach regions at smaller radii, due to the paucity of galaxies. In Fig. \[profiles\_all\] we show the profile computed in this way for all simulated voids larger than $7\,h^{-1}\,
Mpc$ (blue line).
The stacked profile is fitted by the two-parameters law proposed in RQP13: $$\label{myfit}
\frac{\rho(<r)}{\rho_e}=\Big(\frac{r}{R_e}\Big)^{\alpha}\exp\Big[ \Big(\frac{r}{R_e}\Big)^{\beta}-1\Big]$$ where $\rho(<r)$ is the density enclosed within the void-centric distance $r$, $\rho_e$ is the density enclosed within the void effective radius $R_e$ and $\alpha$ and $\beta$ are the best-fit parameters to be obtained from the fit. We notice that in order to avoid a divergent profile for $r=0$, we should require $\alpha, \beta \geqslant 0$. However, since we are applying Eq. \[myfit\] to a limited radial range ($0.3-1 R_e$), we allow $\alpha$ and $\beta$ to assume any value. This is particularly useful to quantify the behavior of the inner part of the profile in very different situations. In fact, cases where $\alpha$ takes negative values do exist, as we show in the following sections. The best-fit parameters that we obtain for the stacked void of Fig. \[profiles\_all\] are: $\alpha=0.06$ and $\beta=1.76$, compatible with those determined in RQP13, using a sample of smaller voids. We also show the profile when only the density of the baryonic component is considered (orange symbols). As shown in RQP13, the distribution of the gas within low density regions closely follows that of dark matter. Indeed, the density profiles for the two components are in a remarkable agreement and the best-fit parameters for the gas only profile are: $\alpha=0.01$ and $\beta=1.65$.
The second method we use, relies on the dark matter particles within the void regions. In building the stack, we include, for each void, all dark matter particles within the radius limit, and rescale their void-centric distance to the radius of the void hosting the particle. The particles are ranked according to their rescaled distance $r$, from the smallest to the largest, and the stacked profile at $r$ is computed with the following expression: $$\frac{\rho(<r)}{\rho_e}=\frac{1}{N_{void}}\sum_{i=1}^{N(<r)}\frac{m_iw_i}{(4/3)\pi(rR_{ei})^3\rho(<R_{ei})}$$ where the summation is intended over all the particles contained within the rescaled radius $r$, $N_{void}$ is the total number of voids entering in the stack, $m_i$ is the mass of the [*i-th*]{} particle, $R_{ei}$ is the effective radius of the void containing the [*i-th*]{} particle and $\rho(<R_{ei})$ is the density enclosed within $R_{ei}$, computed with all the particles within the void containing the [*i-th*]{} particle[^5]. The weights $w_i$ are intended to give low weights to the massive particles falling too close to the void centers, which otherwise would bias the inner profile towards high values. They are defined as: $$\label{w1}
w_i=(1-u^2)^2$$ with: $$\label{w2}
u=min\left( \left|\frac{(\rho\rho(r)-\overline{\rho\rho}}{N_{\sigma}\sigma} \right|,1\right)$$ being $\rho\rho=\rho(r)/\rho(R_e)$ the density at the location $r$, computed with the mass and location of the [*i-th*]{} particle, rescaled to the density at the void radius $\rho(R_e)$; $\overline{\rho\rho}$ is the median density computed with 20 neighbour particles and $\sigma$ is the median absolute deviation, $N_{\sigma}$ is an adjustable parameter, that in our configuration is chosen to be 6[^6]. This profile is shown by the green symbols in Fig. \[profiles\_all\]. It slightly deviates from the one computed with the continuous density field, because the density in the inner part of voids can be seriously affected by the sparsity of the particle distribution. The bi-weight estimator of the individual profiles turns out to be a far more robust method when dealing with very noisy data, such as the density in the very inner part of voids. However, with dark matter particles we want to adopt the same method that can be used with the SDSS voids, where individual profiles are difficult to be obtained, given the paucity of galaxies living in them.
To derive the void density profiles of the SDSS voids, we need to rely on the luminous galaxies as density tracers. We therefore adopt the same stacking procedure used for stacking the dark matter particles in the simulated voids, by considering all the void galaxies in the parent sample. The observed void density profile is shown in Fig. \[profiles\_all\] as black symbols. The functional form expressed in Eq. \[myfit\] turns out to be adequate in reproducing also the observed profile, with $\alpha=0.50$ and $\beta=4.15$. We find nevertheless that the observed profile is much steeper than the simulated one. Given the close agreement between gas and dark matter density profiles, we can not ascribe such a steepness to a baryonic bias. We investigate the origin of such difference in the following sections.
Impact of the undersampling {#undersampling}
---------------------------
A possible reason for the steepness of the observed density profile could lie in the sparsity of the galaxy distribution. The paucity of galaxies within the observed voids, especially at small void-centric distances, makes difficult to reconstruct the underlying distribution of matter. Indeed, in our catalogue of SDSS voids, the typical number of galaxies populating the voids is 10, and several voids contain only 1-2 galaxies. Therefore, it is important to assess whether the limited statistics or the low density of tracers (either dark matter particles in the simulation or galaxies in the SDSS catalogue) could affect the shape of the void density profile.
In order to test the undersampling, we have used the simulated voids. The undersampling has been tested against both the number of voids used in the stacking and the number of density tracers. Since the resolution of the simulation used for this work is quite modest, we do not find haloes and galaxies within the voids. Therefore, to study the effect of the sparsity of the density tracers we rely on the dark matter particles.
In assessing the effect of the limited statistics, we have randomly extracted subsamples of voids from the parent catalogue. The resulting density profiles are shown in Fig. \[profiles\_nv\] for subsamples populated with an increasing number of voids. The dependence of the best-fit parameters on the number of voids is shown in the small panels on the right-hand side of Fig. \[profiles\_nv\]. The errors on $\alpha$ and $\beta$ have been estimated by means of a bootstrap resampling. For each void subsample, we have generated 100 resamplings with replacement and computed the stacked profile with the relative best-fit values. Their standard deviations give the errors on the measured $\alpha$ and $\beta$. The effect of the limited statistics is to increase the noise in samples where very few voids are stacked, but no systematic effect is observed. As shown by the best-fit parameter panels, $\alpha$ and $\beta$ converge to the reference values when at least 100 voids are stacked.
We have also analysed the undersampling effect by computing the profiles with voids populated with an increasing number of dark matter particles. The resulting profiles are shown in Fig. \[profiles\_ngv\]. The most important conclusion is that the paucity of tracers does not bias the recovered profiles. As inferred from the behavior of the best-fit parameters of Fig. \[profiles\_ngv\], voids populated with less than 10 particles show some deviations from the reference values, because the profiles are particularly noisy, but no systematic effect is observed.
To conclude, we can consider the void density profiles as reliable when more than 100 voids are used in the stack and when they are populated with at least 10 tracers. In the SDSS stack void shown in Fig. \[profiles\_all\], both conditions are satisfied, hence we can not ascribe the steepness of the observed profile to undersampling effects.
Impact of the mass tracers
--------------------------
In this section, we study the impact of the sample of galaxies adopted on the resulting density profiles of the observed voids. This is particularly important if one wants to study the void density profiles as a function of the radius of the voids. In fact, the largest voids are more likely observed in the higher redshift bins, as the sampled volume is larger. However, as a consequence of the Malmquist bias, at high redshift only the brightest galaxies are observed (see Fig. \[completeness\]). Therefore, to compare voids located at different redshifts, we need to know whether the different tracers adopted can affect the resulting profile.
To do this, we have built volume limited samples of galaxies up to a given redshift and complete down to the corresponding threshold mass limit. The redshift and mass limits are those illustrated in Fig. \[completeness\]. It is worth to emphasize that the choice of the galaxy sample used only affects the recovered density profiles, leaving the sample of voids unchanged. We do not consider samples at $z<0.04$ as the number of voids is too limited and the profiles can be affected by undersampling. In Fig. \[profiles\_cuts\] we show the recovered density profiles using the different galaxy samples. The profiles appear to steepen as galaxies at higher redshift and higher stellar mass are used. Interestingly, the profiles traced with the faintest galaxy samples approach the simulated profiles shown in Fig. \[profiles\_all\]. The steepening is particularly evident in the evolution of $\alpha$, that becomes progressively higher as more massive galaxies are concerned. On the other hand, $\beta$ does not show any clear dependence on the tracers, as the $\beta$ beta values are just scattered around the reference values.
We exclude redshift evolution as the reason for the profile steepening observed in Fig. \[profiles\_cuts\]. Indeed, the evolution of voids in such a narrow redshift range is expected to be negligible and should go in the opposite sense, i.e. steeper profiles at lower redshift (see Figure 9 in RQP13). To understand how the choice of the mass tracers affects the profiles, we show in Fig. \[mass\_segr\] the distribution of void-centric distances of void galaxies of different masses. As expected, in all the samples the number of galaxies is extremely scarse in the inner part and then it rapidly grows towards the edge, reaching a maximum at $r/Re \sim 0.9$. Low mass galaxies appear to live closer to the void centre, in particular there is an excess of dwarf galaxies at $r/Re \sim 0.5$. Conversely, the massive galaxies are more concentrated towards the void edge. A similar effect for the dwarf systems has been also observed by @Hoyle12. We argue that the steepness of the observed profile, with respect to the simulated ones, can be explained by the absence of tracers in the innermost regions of the observed voids. It is not clear however whether such absence could be solved by using deeper data or it is just a consequence of the galaxy bias.
We note that a similar comparison of density profiles measured with different samples of galaxies has been shown by @Nadathur13. They did not find any bias on the profile when considering tracers of different magnitude. However, their galaxy samples are relatively bright ($Mr<-18.16+5log(h)$). Void galaxies in our SDSS catalogue, are, by definition, fainter than $Mr=-20.17+5log(h)$, hence allowing us to probe the profiles using also galaxies with very low mass. In fact, if only the highest mass bins were concerned, we would not observe such dependence of the profile on the galaxy mass.
![Fraction of galaxies as a function of their distance to the void centre, normalized to the void radius. The different lines show galaxies within different mass ranges, as indicated. []{data-label="mass_segr"}](fig6.eps){width="0.95\columnwidth"}
Dependence on void radius
-------------------------
To assess the dependence of the void profiles on the void radius, we use both the observed and simulated voids. In the observed voids, in order not to be affected by the bias described in the previous section, we rely on a homogenous sample of galaxies. We focus on voids located at $z<0.08$ and use only galaxies more massive than the threshold mass at this redshift ($10^{9.9}\,M_{\odot}$) for the stacking. We choose this couple of redshift and mass because is the one maximizing the number of galaxies (see lower panel of Fig. 1). We find 1539 voids and 7725 galaxies satisfying this criteria.
We divide the void sample in equi-populated subsamples, having $\sim300$ voids each. This should limit the effect of noise at small radii, shown in Section \[undersampling\]. The profiles for different void radii are shown in Fig. \[profiles\_rad\]. All the best-fit parameters, $\alpha$ and $\beta$, fall within 1-2 $\sigma$ of the reference values, derived by fitting the profile of the parent sample at $z<0.08$, without any dependence on the radius. Indeed, the best-fit profile derived for the parent sample (dotted line) is compatible with the profile shape in all the size bins.
We also probe the dependence of void profiles on radius by means of the simulated voids. Fig. \[profiles\_rad\_sim\] shows the density profiles for different void sizes. The density here refers to the total density, including dark matter and gas. A similar result is obtained when using the dark matter particles as density tracers. The void sample is divided in different size bins, containing more than $300$ voids. Given the large box of the simulation, we are able to probe even the largest voids, having radii up to $\sim 50\,h^{-1}Mpc$. The profiles of each size bins are in good agreement with that of the parent sample (dotted line). However, in the trend of the best-fit parameters with radius (right-hand upper panels), we observe a positive correlation of $\alpha$ with $R_e$ and a negative correlation of $\beta$ with $R_e$. This is due to some degeneracy in the fit, as the two parameters are not completely independent. Therefore, we have fitted the two parameters separately. Hence $\alpha$ ($\beta$) has been fitted by keeping $\beta$ ($\alpha$) fixed and equal to its reference value. The results are shown in the lower panels on the right-hand side of Fig. \[profiles\_rad\_sim\]. In this case, we do not see any correlation with $R_e$. We argue that the profile is independent on the void radius.
We notice that the void density profiles have also been tackled in RQP13, using an analogous simulation as the one presented in this work, though with a much smaller volume. However, in that work the statistics of voids larger than $8\,h^{-1}Mpc$ were too limited, due to the small volume of the simulation, and it was not possible to draw robust conclusions.
Conclusions {#conclu}
===========
We have robustly assessed the universality of void density profiles, by means of a catalogue of observed voids and a large cosmological simulation.
The observed void catalogue has been drawn from the SDSS database, and includes spherical voids whose radius is larger that $7 \, h^{-1}\, Mpc$ [@Varela12]. To measure the density profiles in these voids, we rely on the luminous galaxies. As a matter of comparison, we have performed a large cosmological simulation with the code MASCLET, devoted to follow the formation and evolution of the low-density regions. This simulation has been designed to target, with sufficient statistics, voids spanning a wide range of radius. To this aim, we have simulated a large volume, having a comoving side length of $512 \, h^{-1}\, Mpc$, with only one level of refinement in the AMR grid, reaching the spatial resolution of $0.5 \, h^{-1}\, Mpc$. Since this modest resolution does not allow to follow the formation of structures in the simulated box, void galaxies in our simulation do not form. Therefore, we adopt as density tracer the continuous density field or, where a sparse distribution of tracers is needed, the dark matter particles within the void regions.
The void density profiles recovered by means of the observed and simulated voids share the same qualitative shape, showing a significant underdensity in the centre and a sharp density increase approaching the void edges. Both profiles can be well described by the functional form proposed in RQP13. However, the observed profile is significantly steeper than the simulated one.
To figure out the reasons for the steepness of the observed profiles, we have assessed the impact of the number and type of tracers on the resulting density profile. The sparsity of the density tracers has been investigated by means of subsamplings of the simulated voids, populated with an increasing number of particles. We have shown that even in the less populated void samples, the original density profile can be recovered with reasonable accuracy. Stacks built with a limited number of voids or sparsely populated present a significant noise at small radii, but no systematic effect with the number of voids/tracers is observed. The low impact of the sparsity of the tracers on the internal void density profiles has been pointed out also by @Sutter13, using both dark matter particles and haloes as density tracers.
Nevertheless, we observe that the profile shape can have a significant dependence on the type of galaxies used to trace the matter distribution. Within the observed voids, the density profiles recovered by means of faint samples of galaxies are shallower than those determined through the brighter galaxies. The reason for that lies in the galaxy mass segregation within voids. In fact, faint galaxies are those living closer to the void centre and, thus, allow to probe the matter distribution even in the innermost part of the voids.
The strong impact of the type of galaxies chosen to trace the density, forces us to use an homogenous sample of galaxies and voids, limited in volume and magnitude, to assess the dependence of the void density profile on the void radius. With such a sample, we have demonstrated the insensitivity of the observed void profile on void radius. Likewise, by using our simulated sample of voids, we do not observe any dependence of the profile shape on the void size, and the same best-fit can correctly describe voids whose size ranges from 7 to $\sim 50 \, h^{-1}\, Mpc$.
Finally, we note that the difference in profile between the observed and simulated voids can not be driven by the different algorithms used to identify voids. Indeed, the density profile of our SDSS stack is very similar to the profile published in Pan et al. (2012), using the same SDSS DR7 dataset, albeit with a completely different void finder. Moreover, our simulated void density profiles are in remarkable agreement with the simulations of @Colberg05, where voids are identified through spherical underdensities. Therefore, we argue that the difference between observed and simulated void density profiles is a robust result and is due to the biased tracers used, when relying on the observed galaxies. To corroborate this hypothesis, we definitely need high resolution simulations, capable to follow structure formation in the most rarefied regions of the Universe.
Acknowledgements {#acknowledgements .unnumbered}
================
We are grateful to Ignacio Trujillo for useful discussions. This work was supported by the Spanish Ministerio de Economía y Competitividad (MINECO, grants AYA2010-21322-C03-01) and the Generalitat Valenciana (grant PROMETEO-2009-103). J.V. did part of the work thanks to a post-doc fellowship from the former Spanish Ministry of Science and Innovation under programs 3I2005 and 3I2406. J.V. also acknowledges the financial support from the FITE (Fondos de Inversión de Teruel) and the Spanish grant AYA2012-30789.
Alcock C., Paczynski B., 1979, Nature, 281, 358
Aragon-Calvo M. A., Szalay A. S., 2013, MNRAS, 428, 3409
Blanton M. R., et al., 2005, AJ, 129, 2562
Blanton M. R., Roweis S., 2007, AJ, 133, 734
Bond J. R., Kofman L., Pogosyan D., 1996, Nature, 380, 603
Bos E. G. P., van de Weygaert R., Dolag K., Pettorino V., 2012, MNRAS, 426, 440
Park D., Lee J., 2007, PhRvL, 98, 081301
Colberg J. M., et al., 2008, MNRAS, 387, 933
Colberg J. M., Sheth R. K., Diaferio A., Gao L., Yoshida N., 2005, MNRAS, 360, 216
Colless M., et al., 2001, MNRAS, 328, 1039
Eisenstein D.J., Hu W., 1998, ApJ, 511, 5
Gregory S. A., Thompson L. A., 1978, ApJ, 222, 784
Gottl[ö]{}ber S., [Ł]{}okas E. L., Klypin A., Hoffman Y., 2003, MNRAS, 344, 715
Hoyle F., Vogeley M. S., Pan D., 2012, MNRAS, 426, 3041
Kirshner R. P., Oemler A., Jr., Schechter P. L., Shectman S. A., 1981, ApJ, 248, L57
Krause E., Chang T.-C., Dor[é]{} O., Umetsu K., 2013, ApJ, 762, L20
Lavaux G., Wandelt B. D., 2010, MNRAS, 403, 1392
Lavaux G., Wandelt B. D., 2012, ApJ, 754, 109
Laureijs R., et al., 2011, arXiv, arXiv:1110.3193
Melchior P., Sutter P. M., Sheldon E. S., Krause E., Wandelt B. D., 2013, arXiv, arXiv:1309.2045
Nadathur S., Hotchkiss S., 2013, arXiv, arXiv:1310.2791
Neyrinck M. C., 2008, MNRAS, 386, 2101
Pan D. C., Vogeley M. S., Hoyle F., Choi Y.-Y., Park C., 2012, MNRAS, 421, 926
Park D., Lee J., 2007, PhRvL, 98, 081301
Patiri S. G., Prada F., Holtzman J., Klypin A., Betancort-Rijo J., 2006b, MNRAS, 372, 1710
Platen E., van de Weygaert R., Jones B. J. T., 2007, MNRAS, 380, 551
Plionis M., Basilakos S., 2002, MNRAS, 330, 399
Quilis V., 2004, MNRAS, 352, 1426
Ricciardelli E., Quilis V., Planelles S., 2013, MNRAS, 434, 1192
Sheth R. K., van de Weygaert R., 2004, MNRAS, 350, 517
Sutter P. M., Lavaux G., Wandelt B. D., Weinberg D. H., 2012a, ApJ, 761, 187
Sutter P. M., Lavaux G., Wandelt B. D., Weinberg D. H., 2012b, ApJ, 761, 44
Sutter P. M., Lavaux G., Wandelt B. D., Hamaus N., Weinberg D. H., Warren M. S., 2013, arXiv, arXiv:1309.5087
York D. G., et al., 2000, AJ, 120, 1579
van de Weygaert R., Platen E., 2011, IJMPS, 1, 41
Varela J., Betancort-Rijo J., Trujillo I., Ricciardelli E., 2012, ApJ, 744, 82
[^1]: E-mail: [email protected]
[^2]: http://sdss.physics.nyu.edu/vagc/
[^3]: http://cas.sdss.org/astrodr7/en
[^4]: The effective radius is defined as the radius of the sphere having the same volume of the void.
[^5]: in the computation of $\rho(<R_{ei})$ the particle masses are weighted according to Eq. \[w1\] and \[w2\].
[^6]: This weighting scheme is the same adopted in the bi-weight estimator
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We theoretically investigate the dependence of exciton transition energies on dielectric constant of surrounding materials. We make a simple model for the relation between dielectric constant of environment and a static dielectric constant describing the effects of electrons in core states, $\sigma$ bonds and surrounding materials. Although the model is very simple, calculated results well reproduce experimental transition energy dependence on dielectric constant of various surrounding materials.'
author:
- 'Y. Miyauchi$^{a}$, R. Saito$^{b}$, K. Sato$^{b}$, Y. Ohno$^{c}$, S. Iwasaki$^{c}$, T. Mizutani$^{c}$, J. Jiang$^{d}$, S. Maruyama$^{a}$[^1]'
title: 'Dependence of exciton transition energy of single-walled carbon nanotubes on surrounding dielectric materials '
---
Introduction
============
Photoluminescence (PL) of single-walled carbon nanotubes (SWNTs) has been intensively studied for elucidating their unusual optical and electronic properties due to one dimensionality [@oconnell02; @weisman02sm; @lebedkin03; @resasco03; @hartschuh03; @MaruyamaCPL; @lefebvre04b; @fwang05; @maultzsch05j; @dukovic05; @c1095; @htoon05; @plentz05; @Ohno06-PRB-s; @miyauchi06prb1; @miyauchi06prb2]. Since both of electron-electron repulsion and electron-hole binding energies for SWNTs are considerably large compared with those for conventional three-dimensional materials, the Coulomb interactions between electron-electron and electron-hole play an important role in optical transition of SWNTs [@ando97a; @zhao04prl; @perebeinos04prl; @spataru04; @jiang07-ex; @jiang07-opphmx]. Optical transition energies of SWNTs are strongly affected by the change of environment around SWNTs such as bundling [@a1119j], surfactant suspension [@lefebvre04b; @Ohno06-PRB-s; @moore03-surf] and DNA wrapping [@strano06-DNA]. Lefebvre [*et al.*]{} [@lefebvre04b] reported that the transition energies for suspended SWNTs between two pillars fabricated by the MEMS technique are blue-shifted relative to the transition energies for micelle-suspended SWNTs. Ohno [*et al.*]{} [@Ohno06-PRB-s] have compared the PL of suspended SWNTs directly grown on a grated quartz substrate using alcohol CVD technique [@MaruyamaCPL] with SDS-wrapped SWNTs [@weisman02sm]. The energy differences between air-suspended and SDS-wrapped SWNTs depend on $(n, m)$ and type of SWNTs \[type I ($(2n+m)$ mod 3 = 1) or type II ($(2n+m)$ mod 3 = 2)[@saitobook; @a1093; @w1089]\].
Recently, Ohno [*et al.*]{} studied $E_{11}$ transition energies of SWNTs in various surrounding materials with different dielectric constant, $\kappa_{\rm env}$ [@Ohno07unp]. Observed dependence of $E_{11}$ on $\kappa_{\rm env}$ for a $(n,m)$ nanotube showed a tendency that can be roughly expressed as $$E_{11}=E_{11}^{\infty}+A_{nm}^{\rm exp}\kappa_{\rm env}^{-\alpha}
\label{eq:E11alpha}$$ where $E_{11}^{\infty}$ denotes a transition energy when $\kappa_{\rm
env}$ is infinity, $A_{nm}^{\rm exp}$ is the maximum value of an energy change of $E_{11}$ by $\kappa_{\rm env}$, and $\alpha$ is a fitting coefficient in the order of 1, respectively. At this stage, the reason why the experimental curve follows Eq.(\[eq:E11alpha\]) is not clear.
In the previous theoretical studies of excitonic transition energies for SWNTs [@ando97a; @perebeinos04prl; @jiang07-ex; @jiang07-opphmx], a screening effect of a surrounding material is mainly described using a static dielectric constant $\kappa$. However, since $\kappa$ consists of both $\kappa_{\rm env}$ and screening effect by nanotube itself, $\kappa_{\rm tube}$, experimental dependence of transition energies on dielectric constants of environment can not directly compared with calculations [@ando97a; @perebeinos04prl; @jiang07-ex; @jiang07-opphmx] using the static dielectric constant $\kappa$. In this study, we make a simple model for the relation between $\kappa_{\rm env}$ and $\kappa$. The calculated results of excitons for different $\kappa_{\rm env}$ reproduced well the experimental transition energy dependence on dielectric constant of various surrounding materials.
Theoretical method
==================
Exciton transition energy
-------------------------
Within the extended tight-binding model [@jiang07-ex; @jiang07-opphmx; @w1089], we calculated transition energies from the ground state to the first bright exciton state by solving the Bethe-Salpeter equation, $$\begin{array}{llll}
& {\displaystyle \left\{[E({\bf k}_{\rm c})-E({\bf k}_{\rm
v})]\delta({\bf k}^{'}_{\rm c}, {\bf k}_{\rm c}) \delta({\bf
k}^{'}_{\rm v},{\bf k}_{\rm v})\right. }\\ &{\displaystyle \left.
+K({\bf k}^{'}_{\rm c} {\bf k}^{'}_{\rm v},{\bf k}_{\rm c} {\bf
k}_{\rm v}) \right\} \Psi^{n}({\bf k}_{\rm c} {\bf k}_{\rm v})
=\Omega_{n}\Psi^{n}({\bf k}^{'}_{\rm c}{\bf k}^{'}_{\rm v}), }
\end{array}
\label{eq:BS}$$ where ${\bf k}_{\rm c}$ and ${\bf k}_{\rm v}$ denote wave vectors of the conduction and valence energy bands and $E({\bf k}_{\rm c})$ and $E({\bf k}_{\rm v})$ are the quasi-electron and quasi-hole energies, respectively. $\Omega_n$ is the energy of the $n$-th excitation of the exciton $(n=0,1,2,\cdots)$, and $\Psi^{n}({\bf k}_{\rm c}{\bf k}_{\rm
v})$ are the excitonic wavefunctions. The kernel $K({\bf k}^{'}_{\rm
c} {\bf k}^{'}_{\rm v},{\bf k}_{\rm c} {\bf k}_{\rm v})$ describes the Coulomb interaction between an electron and a hole. Details of the exciton calculation procedure is the same as presented in Refs [@jiang07-ex; @jiang07-opphmx; @w1089].
The exciton wavefunction $|\Psi^{n}_{{\bf q}}>$ with a center-of-mass momentum ${\bf q}(={\bf k}_c-{\bf k}_v)$ can be expressed as $$|\Psi^{n}_{\bf q}>=\sum_{\bf k} Z^{n}_{{\bf k}c,({\bf k}-{\bf q})v}
c^{+}_{{\bf k}c}c_{({\bf k}-{\bf q})v}|0>
\label{exwf},$$ where $Z^{n}_{{\bf k}c,({\bf k}-{\bf q})v}$ is the eigenvector of the $n$-th $(n=0,1,2,\cdots)$ state of the Bethe-Salpeter equation, and $|0\rangle$ is the ground state. Due to momentum conservation, the photon-excited exciton is an exciton with ${\bf q}\approx0$ for parallel excitations to the nanotube axis. In this Letter, we calculate the $n =
0$ state of ${\bf q} = 0$ exciton for each $(n,m)$ SWNT.
Dielectric screening effect
---------------------------
In our calculation, the unscreened Coulomb potential $V$ between carbon $\pi$ orbitals is modeled by the Ohno potential [@perebeinos04prl]. We consider the dielectric screening effect within the random phase approximation (RPA). In the RPA, the static screened Coulomb interaction $W$ is expressed as [@ando97a] $$\begin{array}{llll}
W=V/\kappa\epsilon({\bf q}),
\end{array}
\label{cbi}$$ where $\epsilon({\bf q})$ is the dielectric function describing effects of the polarization of the $\pi$ bands. $\kappa$ is a static dielectric constant describing the effects of electrons in core states, $\sigma$ bonds, and surrounding materials. In the calculation, we directly calculate only the polarization for the $\pi$ band, and the effects of electrons in core states, $\sigma$ bands, and surrounding materials are represented by a single constant $\kappa$. In the most accurate expression, the inhomogeneous and nonlocal dielectric response of the nanotube itself and the surrounding materials should be considered. However, it is not easy within extended tight binding method. In this study, instead of treating the complicated dielectric response including surrounding materials, we make a simple model for a relation between the static dielectric constant $\kappa$ and $\kappa_{\rm env}$ to obtain the $E_{11}$ dependence on $\kappa_{\rm env}$.
{width="6cm"}
. \[fig:1\]
Relationship between $\kappa$ and $\kappa_{\rm env}$
----------------------------------------------------
Figure \[fig:1\] shows a schematic view for the model relationship between $\kappa$ and $\kappa_{\rm env}$. Here we consider the screening effect related to $\kappa$ as a linear combination of the screening of nanotube itself and the surrounding material $$\frac{1}{\kappa}=\frac{C_{\rm tube}}{\kappa_{\rm tube}}+\frac{C_{\rm env}}{\kappa_{\rm env}},
\label{eq:kapp-rel}$$ where $\kappa_{\rm tube}$ is the dielectric constant within a nanotube except for the $\pi$ bands, and $C_{\rm tube}$ and $C_{\rm env}$ are coefficients for the inside and outside of a nanotube, respectively. As shown in Eq.(\[eq:E11alpha\]), the transition energies observed in the experiment [@Ohno07unp] indicate that there is a limit value [@perebeinos04prl] when $\kappa_{\rm env}$ $\rightarrow$ $\infty$. Hence, when $\kappa_{\rm env}$ $\rightarrow$ $\infty$, $C_{\rm env}/\kappa_{\rm env}$ can be removed from Eq.(\[eq:kapp-rel\]), and $1/\kappa$ is expressed by the limit value as $$\frac{1}{\kappa} = \frac{C_{\rm tube}}{\kappa_{\rm tube}} \equiv \frac{1}{\kappa^{\infty}_{\rm tube}}, (\kappa_{\rm env} \rightarrow \infty )$$ where $\kappa^{\infty}_{\rm tube}$ is the limit value of the net dielectric constant $\kappa$ when $\kappa_{\rm env}$ is infinity. Since electric flux lines through inside of the nanotube remain even when $\kappa_{\rm env}$ $\rightarrow$ $\infty$, we assume there is a certain value of $\kappa$ ($\kappa^{\infty}_{\rm tube}$) that corresponds to the situation when dielectric screening by surrounding material is perfect and only dielectric response of the nanotube itself contributes to the net screening effect.
Replacing $C_{\rm tube}/\kappa_{\rm tube}$ by $\kappa^{\infty}_{\rm
tube}$, Eq.(\[eq:kapp-rel\]) is modified as $$\frac{1}{\kappa}=\frac{1}{\kappa^{\infty}_{\rm tube}}+\frac{C_{\rm env}}{\kappa_{\rm env}}.
\label{kapp-rel-mod}$$ Next, we imagine that the SWNT is placed in the vacuum, which corresponds to $\kappa=\kappa^{\rm vac}$ and $\kappa_{\rm env}=1$, and then $C_{\rm env}$ can be expressed as $$C_{\rm env}=\frac{1}{\kappa^{\rm vac}}-\frac{1}{\kappa^{\infty}_{\rm tube}},$$ where $\kappa^{\rm vac}$ is the static dielectric constant [*not*]{} for the vacuum, but for the situation that the nanotube is placed in the vacuum. We now express $\kappa$ as a function of $\kappa_{\rm
env}$ through two parameters $\kappa^{\infty}_{\rm tube}$ and $\kappa^{\rm vac}$, whose values can be estimated from the following discussions. In the previous papers [@ando97a; @perebeinos04prl; @jiang07-ex; @jiang07-opphmx], $\kappa$ value is put around 2 to obtain a good fit with experiments for SWNTs with surrounding materials. Jiang [*et al.*]{} [@jiang07-ex] have compared the calculated results with the results for the two photon absorption experiments [@dukovic05], and obtained the best fit using $\kappa=2.22$ for SWNTs in a polymer matrix. Here, since $\kappa^{\rm vac}$ is for nanotubes without surrounding materials, $\kappa^{\rm vac}$ should be less than about 2 and close to 1 due to vacancy of inside of the tubes. With regard to $\kappa^{\infty}_{\rm tube}$, according to the experimental results [@lefebvre04b; @Ohno06-PRB-s; @Ohno07unp], transition energy change due to change of surrounding materials is at most 30-100meV. Fig.\[fig:2\](a) shows the calculated $E_{11}$ energy dependence on $\kappa$ for a (9,8) SWNT in a small $\kappa$ region, while the inset shows the $E_{11}$ dependence up to $\kappa_{\rm env}=100$. As shown in Fig.\[fig:2\](a), variation of $\kappa$ that yields the transition energy change of 30 to 100 meV is about 1 to 3 when $\kappa$ is around 2. Therefore, the value of $\kappa^{\infty}_{\rm tube}$ should be around 2 to 3 and that of $\kappa^{\rm vac}$ should be around 1 to 2.
![(a) The $E_{11}$ energy for a $(9,8)$ SWNT as a function of $\kappa$. (b) $\delta E_{11}$ dependence on $\kappa_{\rm env}$. Inset in (a) shows the $E_{11}$ dependence up to $\kappa=100$. In (b), circles denote the experimental data and solid curves denote the calculated results of Eq.(\[eq:kenvdep\]) for $\kappa^{\rm vac}=1.0$ (black), $1.5$ (red) and $2.0$ (blue).[]{data-label="fig:2"}](\HeadDir/fig2.eps){width="6cm"}
Dependence of excitation energy on $\kappa_{\rm env}$
-----------------------------------------------------
As shown in Fig.\[fig:2\](a), the calculated $E_{11}$ energies decrease with increasing $\kappa$. This is mainly due to the fact that the self energy (e-e repulsion) always exceeds to the e-h binding energy and that the both interactions (e-e and e-h) decrease with increasing $\kappa$. The $E_{11}$ almost linearly depends on $\kappa$ around the small $\kappa$ region. We checked that the linear dependence is universal for all $(n,m)$’s for diameters more than 0.7 nm. Assuming the linear dependence, variation of the excitation energy $\delta E_{11} \equiv E_{11}-E_{11}(\kappa_{\rm env}=1)$ for the small $\kappa$ region is approximated by $$\delta E_{\rm 11}=-A_{nm}(\kappa-\kappa^{\rm vac}),
\label{eq:keffdep}$$ where $A_{\rm nm}$ is the gradient of $\delta E_{\rm 11}$ near the small $\kappa$ region for each $(n,m)$ type. After we transform $\kappa$ using the relationship of Eq.(\[eq:kapp-rel\]), Eq.(\[eq:keffdep\]) is modified as $$\begin{array}{llll}
{\displaystyle
\delta E_{\rm 11}=-A_{nm}(\kappa^{\infty}_{\rm tube}-\kappa^{\rm
vac})(\frac{\kappa_{\rm env}-1}{\kappa_{\rm
env}+(\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac})/\kappa^{\rm vac}}).
}
\label{eq:kenvdep}
\end{array}$$
$A_{nm}(\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac})$ corresponds to the maximum value of $\delta E_{\rm 11}$ when $\kappa_{\rm env}$ $\rightarrow$ $\infty$, which corresponds to the value of coefficient $A_{nm}^{\rm exp}$ in the fitting curve of Eq.(\[eq:E11alpha\]). For $(9,8)$ SWNT, the fitted value to the calculated results for $A_{nm}$ is 33 meV and $A_{nm}^{\rm exp}$ obtained by the fit to the experiment[@Ohno07unp] using Eq.(\[eq:E11alpha\]) is 36 meV, and $\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac}$ should be around 1. The values for $\kappa^{\infty}_{\rm tube}$ and $\kappa^{\rm vac}$ are consistent with the values conventionally used for SWNTs in dielectric materials [@ando97a; @perebeinos04prl; @jiang07-ex; @jiang07-opphmx].
Results and Discussion
======================
Figure \[fig:2\](b) compares $\delta E_{11}$ for a (9,8) SWNT depending on $\kappa_{\rm env}$ by the experiment (solid circles) and the calculated results (lines) for $\kappa^{\rm vac}=1,1.5,2.0$ using Eq.(\[eq:kenvdep\]). As shown in Fig.\[fig:2\](b), the qualitative shape of theoretical curves are in good agreement with the experiment and not affected so much by the change of $\kappa^{\rm vac}$. Since the exact value of $\kappa_{\rm vac}$ is unknown, we hereafter set $\kappa^{\rm vac}=1.5$ for each $(n,m)$ SWNT. For the $(9,8)$ SWNT in Fig.\[fig:2\](b), $\kappa^{\infty}_{\rm tube}=2.7$ and $\kappa^{\rm
vac}=1.5$ are fitting values. These values are consistent with the discussion in the previous section. After setting $\kappa^{\rm
vac}=1.5$, Eq.(\[eq:kenvdep\]) turns to be $$\begin{array}{llll}
{\displaystyle
\delta E_{\rm 11}=\frac{-A_{nm}(\kappa_{\rm env}-1)}{\kappa_{\rm
env}/(\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac})+1/1.5}.
\label{eq:kenvdep2}
}
\end{array}$$
Thus, we express $\delta E_{\rm 11}$ as a function of $\kappa_{\rm
env}$ with one parameter $(\kappa^{\infty}_{\rm tube}-\kappa^{\rm
vac})$.
![(a) Calculated values of $A_{nm}$ for each $(n,m)$ SWNT. Open (red) and solid (blue) circles correspond to type I and type II SWNTs, respectively. Solid lines denote the fit curve by Eq.(\[eq:Anm\]). (b) $\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac}$ vs $1/d_t^2$. The values of $\kappa^{\infty}_{\rm tube}-\kappa^{\rm
vac}$ are obtained by the fit of Eq.(\[eq:kenvdep\]) to the experimental data for each $(n,m)$.[]{data-label="fig:3"}](\HeadDir/fig3.eps){width="5cm"}
Figure \[fig:3\](a) shows the calculated values of $A_{nm}$ for each $(n,m)$’s. Family pattern of $(2n+m=const.)$ family is drawn with the $2n+m$ values by dotted lines. We found a slight diameter dependence and relatively large chiral angle dependence of $A_{nm}$ for type II SWNTs (blue) compared with type I SWNTs (red). The type II SWNTs with larger chiral angles tend to have larger value of $A_{nm}$. For a convenient use of Eq.(\[eq:kenvdep\]), we give a fitting function of $A_{nm}$ meV as $$A_{nm}=A+Bd_t+(C+D/d_t)\cos{3\theta},
\label{eq:Anm}$$ which gives the average (maximum) error of $\pm 2\rm meV$$(8 \rm meV)$ for type I, and $\pm 2\rm meV$$(5 \rm meV)$ for type II SWNTs. The fit curve is shown in Fig.\[fig:3\](a) by solid lines. Here $d_t$ (nm) is the diameter of nanotube and $\theta$ is the chiral angle [@saitobook]. The values of (A, B, C, D) are (36, -4, 0, 0) and (33, -3, 6, 7) for type I and for type II SWNTs, respectively.
In order to expand our result to many $(n,m)$ SWNTs, we need a function to describe $(\kappa^{\infty}_{\rm tube}-\kappa^{\rm
vac})$. It is important to note that $\kappa^{\infty}_{\rm tube}$ should depend on the diameter. An exact function should be calculated by taking into account the Coulomb interaction considering induced surface charge at the boundary of the nanotube and surrounding materal for an e-e or e-h pair for each $(n,m)$ SWNT. Instead of calculating this complicated function, here we roughly estimate the $(\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac})$ as a simple function of diameter $d_t$, since $(\kappa^{\infty}_{\rm tube}-\kappa^{\rm
vac})$ should depend on the cross section of a SWNT. As shown in Fig.\[fig:3\](b), $(\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac})$ is roughly proportional to $1/d_t^2$, $$(\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac})=\frac{E}{d_t^2},
\label{eq:Bdt2}$$ with the coefficient $E=1.5\pm0.3$ $\rm nm^{2}$. Here $(\kappa^{\infty}_{\rm
tube}-\kappa^{\rm vac})$ is obtained by the fit using Eq.(\[eq:kenvdep\]) and $A_{nm}$ calculated for each chirality. Fig.\[fig:3\](b) clearly shows that our calculated $A_{nm}$ well describes the chiral angle dependence of $\delta E_{11}$ and that the remaining diameter dependence is understood by $(\kappa^{\infty}_{\rm tube}-\kappa^{\rm vac})$ through $1/d_t^2$. This $1/d_t^2$ dependence implies that $\kappa^{\infty}_{\rm tube}$ depends on the volume of inner space of the nanotube. Although the number of experimental data available for the fit is small and selection of this function is arbitrary to some extent, it is reasonable that $1/\kappa^{\infty}_{\rm tube}$ increase with the increase of the diameter, since $1/\kappa^{\infty}_{\rm tube}$ corresponds to the Coulomb interaction through the inner space of the nanotube. In order to find an accurate form of the function, future experiments and theoretical studies are definitely needed.
![The transition energy dependence plotted as a function of $\kappa_{\rm env}$. (a) experiment and (b) calculated results are indicated by (a) symbols and (b) solid curves. In (b), $(n,m)$ for each curve is indicated by a symbol on the curve. (c) Comparison of $\delta E_{11}$ for the experiment ($\delta E_{11}$(experiment)) and calculation ($\delta E_{11}$(theory)). A dotted line indicates the line of $\delta E_{11}\rm (experiment)=\delta E_{11}\rm (theory)$. Open (red) and solid (blue) symbols correspond to type I and type II SWNTs, respectively. The data in the dotted circle are the data for $\kappa_{\rm env}=1.9$ [@Ohno07unp] (see text).[]{data-label="fig:4"}](\HeadDir/fig4.eps){width="9cm"}
Figure \[fig:4\] shows $\delta E_{11}$ as a function of $\kappa_{\rm
env}$ for (a) the experiment and (b) the calculation using Eq.(\[eq:kenvdep2\]) and (\[eq:Bdt2\]). Fig.\[fig:4\](c) compares $\delta E_{11}$ for the experiment and that for the calculation with the same $\kappa_{\rm env}$ values. The same symbols for an $(n,m)$ are used in three figures of Fig.\[fig:4\]. Details of experimental data will be published elsewhere [@Ohno07unp]. Although our treatment is very simple, the calculated curves for various $(n,m)$ SWNTs well reproduce the experimentally observed tendency for each $(n,m)$ SWNT, and the degree of difference between each $(n,m)$ type is also in good agreement with the experiment. As shown in Fig.\[fig:4\](c), $\delta E_{11}\rm
(theory)$ is in a good agreement with $\delta E_{11}\rm (experiment)$ except for several points indicated by a dotted circle in the figure, which correspond to a case for the smallest $\kappa_{\rm env}=1.9$ (hexane) except for $\kappa_{\rm env}=1$ (air) in the experimental data [@Ohno07unp]. The value of $\kappa_{\rm env}=1.9$ for hexane is adopted as the dielectric constant for the material, in which the dipole moments of liquid hexane are not aligned perfectly even in the presence of the electric field. Since $\kappa_{\rm env}=1.9$ is a macroscopic value, a local dielectric response might be different from the averaged macroscopic response. If the local dielectric constant near SWNTs becomes large (for example, $\kappa_{\rm env}\approx 3$), the fitting of Fig.\[fig:4\](c) becomes better. We expect that the dipole moments of a dielectric material might be aligned locally for a strong electric field near an exciton, which makes the local dielectric constant relatively large. This will be an interesting subject for exciton PL physics. Since the difference of $A_{nm}$ between each $(n,m)$ type decreases with increasing the diameter, it is predicted that the amount of variation due to the change of $\kappa_{\rm env}$ mostly depend on diameter in the larger diameter range. Thus a PL experiment for nanotubes with large diameters would be desirable for a further comparison.
Summary
=======
In summary, the dependence of exciton transition energies on dielectric constant of surrounding materials are investigated. We proposed a model for the relation between dielectric constant of the environment and a static dielectric constant $\kappa$ in the calculation. Although the model is quite simple, calculated results well reproduce the feature of experimentally observed transition energy dependence on dielectric constant of various surrounding materials, and various $d_t$ and $\theta$.
Acknowledgments {#acknowledgments .unnumbered}
===============
Y.M. is supported by JSPS Research Fellowships for Young Scientists (No. 16-11409). R.S. acknowledges a Grant-in-Aid (No. 16076201) from the Ministry of Education, Japan.
[10]{}
M. J. O’Connell, S. M. Bachilo, X. B. Huffman, V. C. Moore, M. S. Strano, E. H. Haroz, K. L. Rialon, P. J. Boul, W. H. Noon, C. Kittrell, J. Ma, R. H. Hauge, R. B. Weisman, R. E. Smalley, Science 297 (2002) 593.
S. M. Bachilo, M. S. Strano, C. Kittrell, R. H. Hauge, R. E. Smalley, R. B. Weisman, Science 298 (2002) 2361.
S. Lebedkin, K. Arnold, F. Hennrich, R. Krupke, B. Renker, M. M. Kappes, New J. Phys. 5 (2003) 140.
S. M. Bachilo, L. Balzano, J. E. Herrera, F. Pompeo, D. E. Resasco, R. B. Weisman, J. Am. Chem. Soc. 125 (2003) 11186.
A. Hartschuh, H. N. Pedrosa, L. Novotny, T. D. Krauss, Science 301 (2003) 1354.
Y. Miyauchi, S. Chiashi, Y. Murakami, Y. Hayashida, S. Maruyama, Chem. Phys. Lett. 387 (2004) 198.
J. Lefebvre, J. Fraser, Y. Homma, P. Finnie, Appl. Phys. A 78 (2004) 1107.
F. Wang, G. Dukovic, L. E. Brus, T. F. Heinz, Science 308 (2005) 838.
J. Maultzsch, R. Pomraenke, S. Reich, E. Chang, D. Prezzi, A. Ruini, E. Molinari, M. S. Strano, C. Thomsen, C. Lienau, Phys. Rev. B 72 (2005) 241402.
G. Dukovic, F. Wang, D. Song, M. Y. Sfeir, T. F. Heinz, L. E. Brus, Nano Lett. 5 (2005) 2314.
S. G. Chou, F. Plentz Filho, J. Jiang, R. Saito, D. Nezich, H. B. Ribeiro, A. Jorio, M. A. Pimenta, G. G. Samsonidze, A. P. Santos, M. Zheng, G. B. Onoa, E. D. Semke, G. Dresselhaus, M. S. Dresselhaus, Phys. Rev. Lett. 94 (2005) 127402.
H. Htoon, M. J. O’Connell, S. K. Doorn, V. I. Klimov, Phys. Rev. Lett. 94 (2005) 127403.
F. Plentz, H. B. Ribeiro, A. Jorio, M. S. Strano, M. A. Pimenta, Phys. Rev. Lett. 95 (2005) 7401.
Y. Ohno, S. Iwasaki, Y. Murakami, S. Kishimoto, S. Maruyama, T. Mizutani, Phys. Rev. B 73 (2006) 235427.
Y. Miyauchi, S. Maruyama, Phys. Rev. B 74 (2006) 35415.
Y. Miyauchi, M. Oba, S. Maruyama, Phys. Rev. B 74 (2006) 205440.
T. Ando, J. Phys. Soc. Jpn. 66 (1997) 1066.
H. Zhao, S. Mazumdar, Phys. Rev. Lett. 93 (2004) 157402.
V. Perebeinos, J. Tersoff, P. Avouris, Phys. Rev. Lett. 92 (2004) 257402.
C. D. Spataru, S. Ismail-Beigi, L. X. Benedict, S. G. Louie, Phys. Rev. Lett. 92 (2004) 077402.
J. Jiang, R. Saito, G. G. Samsonidze, A. Jorio, S. G. Chou, G. Dresselhaus, M. S. Dresselhaus, Phys. Rev. B 75 (2007) 035407.
J. Jiang, R. Saito, K. Sato, J. S. Park, G. G. Samsonidze, A. Jorio, G. Dresselhaus, M. S. Dresselhaus, Phys. Rev. B 75 (2007) 035405.
C. Fantini, A. Jorio, M. Souza, M. S. Strano, M. S. Dresselhaus, M. A. Pimenta1, Phys. Rev. Lett. 93 (2005) 147406.
V. C. Moore, M. S. Strano, E. H. Haroz, R. H. Hauge, R. E. Smalley, Nano Lett. 3 (2003) 1379.
D. A. Heller, E. S. Jeng, T.-K. Yeung, B. M. Martinez, A. E. Moll, J. B. Gastala, M. S. Strano, Science 311 (2006) 508.
R. Saito, G. Dresselhaus, M. S. Dresselhaus, Physical Properties of Carbon Nanotubes, Imperial College Press, London, 1998.
A. Jorio, C. Fantini, M. A. Pimenta, R. B. Capaz, G. G. Samsonidze, G. Dresselhaus, M. S. Dresselhaus, J. Jiang, N. Kobayashi, A. Grüneis, R. Saito, Phys. Rev. B 71 (2005) 075401.
G. G. Samsonidze, R. Saito, N. Kobayashi, A. Grüneis, J. Jiang, A. Jorio, S. G. Chou, G. Dresselhaus, M. S. Dresselhaus, Appl. Phys. Lett. 85 (2004) 5703.
Y. Ohno, S. Iwasaki, Y. Murakami, S. Kishimoto, S. Maruyama, T. Mizutani, arXiv:0704.1018v1 \[cond-mat.mtrl-sci\] (2007).
[^1]: Corresoponding author. FAX: +81-3-5841-6421.\
E-mail: [email protected] (S. Maruyama)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present optical photometric and low-resolution spectroscopic observations of the Type II plateau supernova (SN) 2008in, which occurred in the outskirts of the nearly face-on spiral galaxy M 61. Photometric data in the X-rays, ultraviolet and near-infrared bands have been used to characterize this event. The SN field was imaged with the ROTSE-IIIb optical telescope about seven days before the explosion. This allowed us to constrain the epoch of the shock breakout to JD = 2454825.6. The duration of the plateau phase, as derived from the photometric monitoring, was $\sim$ 98 days. The spectra of show a striking resemblance to those of the archetypal low-luminosity IIP SNe 1997D and 1999br. A comparison of ejecta kinematics of with the hydrodynamical simulations of Type IIP SNe by @dessart10 indicates that it is a less energetic event ($\sim 5\times10^{50}$ erg). However, the light curve indicates that the production of radioactive $^{56}$Ni is significantly higher than that in the low-luminosity SNe. Adopting an interstellar absorption along the SN direction of $A_V \sim$ 0.3 mag and a distance of 13.2 Mpc, we estimated a synthesized mass of $\sim0.015 M_{\sun}$. Employing semi-analytical formulae [@litvinova85], we derived a pre-SN radius of $\sim 126$, an explosion energy of $\sim 5.4\times10^{50}$ erg and a total ejected mass of $\sim 16.7$. The latter indicates that the zero age main-sequence mass of the progenitor did not exceed 20. Considering the above properties of and its occurrence in a region of sub-solar metallicity (\[O/H\] $\sim$ 8.44 dex), it is unlikely that fall-back of the ejecta onto a newly formed black hole occurred in SN 2008in. We therefore favor a low-energy explosion scenario of a relatively compact, moderate-mass progenitor star that generates a neutron star.'
author:
- 'Rupak Roy$^\star$, Brijesh Kumar, Stefano Benetti, Andrea Pastorello, Fang Yuan, Peter J. Brown, Stefan Immler, Timur A. Fatkhullin, Alexander S. Moskvitin, Justyn Maund, Carl W. Akerlof, J. Craig Wheeler, Vladimir V. Sokolov, Rorbert M. Quimby, Filomena Bufano, Brajesh Kumar, Kuntal Misra, S. B. Pandey, Nancy Elias-Rosa, Peter W. A. Roming and Ram Sagar'
title: ' $-$ Bridging the gap between normal and faint supernovae of type IIP'
---
introduction {#intro}
============
Core-collapse Type II supernovae (SNe) mark the violent death of stars with main-sequence masses greater than 8 and as is indicated by the presence of hydrogen lines in their optical spectra, they originate from a progenitor star with a significant amount of hydrogen still intact [@eldridge04]. Of special interest are the Type II plateau SNe which are characterized by a ‘plateau’ in their optical light curve and are more common, constituting about 75% of all Type II SNe [@smith10]. The IIP SNe show a wide range of plateau luminosities, plateau durations, expansion velocities and nickel masses [@hamuy03] and these observational properties are connected with the explosion mechanism as well as the physical properties of the progenitor star such as ejected mass, explosion energy and pre-SN radius [@nadyozhin03; @smartt09]. The IIP SNe are thought to result from progenitor masses in the range $8-25$ [@heger03] with an extended hydrogen envelope necessary to maintain the plateau phase. A detailed study of optical light curves and spectra of only a few nearby IIP SNe has been done so far and there exists a discrepancy in estimating the mass of their progenitors, e.g, for the three well studied events (namely 1999em, 2005cs and 2004et), the determination of progenitor mass from the hydrodynamical modeling of their light curve is found to be higher than that estimated from pre-SN imaging [@utrobin10; @bersten11].
Recently, a number of ‘low luminosity’ Type IIP events have been discovered viz. SNe 1999br, 1999eu, 1994N, 2001dc, 2005cs [@pastorello04; @pastorello09], 2008bk [@vandyk11] and 2009md [@fraser10]. These events have explosion energy ($\sim 10^{50}$ erg) and ejected $^{56}$Ni-mass ($2-7\times10^{-3}$ ), both lower by a factor of 10 than normal, and low expansion velocity $\sim$1000 (@pastorello09 and references therein). The low-luminosity IIP SNe are debated because of the unknown nature of their progenitors. The first reported faint SN was SN 1997D [@turatto98; @benetti01], and the observed properties of its light curves and spectra were explained in terms of significant fallback of ejected material on a newly formed black hole (BH), created through the core collapse of a massive progenitor ($M \ga 20M_\odot$, @zampieri98 [@zampieri03]). Alternatively, SN 1997D was interpreted as the explosion of a less massive progenitor ($8-12M_\odot$, @chugai00), close in mass to the lower limit for stars that can undergo core-collapse. @heger03 suggested that low-luminosity Type IIP events are electron capture SNe produced by low-mass progenitors giving rise to ONeMg cores. This is further supported through investigations of pre-explosion images [@maund05a; @maund05b], though @eldridge07 have ruled out the possibility of such a mechanism for the low-luminosity SN 2005cs. According to the formalism of @heger03 and @eldridge04, no star having initial mass less than 22 can form a BH, which can quench the ejected material and produce a low luminosity SN. Stars with masses above 25$M_\odot$, formed in metal-poor or slightly sub-solar metallicity regions can produce low-luminosity, BH-forming Type IIP SNe. Type IIL/b events can be produced through this process from stars having masses $\textgreater~ 25$ and generated in regions with solar (or super-solar) metallicity. So, the metallicity information at the SN location and the estimation of the initial mass are essential to constrain the triggering mechanisms of these explosions.
SN 2008in was discovered in the nearby galaxy M 61 (NGC 4303). The first unfiltered CCD images of were taken by Koichi Itagaki on 2008 December 26.79 (all times in UT hereafter) and 27.69 at a magnitude of 14.9. Independent observations of this event by K. Kadota showed the transient at an unfiltered mag of 15.1, In addition W. Wells recorded the SN on 2008 December 28.46 at $V$ and $R$ band magnitudes of 14.3 and 13.2, respectively [@nakano08]. Low and mid-resolution spectroscopic observations indicate an early discovery for SN 2008in (within 1$-$2 weeks after core-collapse). The spectra showed highly blueshifted H$_\alpha$ and H$_\beta$ absorptions (by $\sim$ 9000 ) with weaker emission components [@chakraborti08; @foley08; @stritzinger08]. The presence of prominent P-Cygni profiles of Balmer lines leads to its classification as a Type II SN.
The broadband light curve and the initial spectral evolution of SN 2008in were similar to those of normal Type IIP SNe. However, from mid-plateau, the SN started to show a few spectral features (like ) which are similar to under-luminous events. was also observed in the radio with the Very Large Array (VLA) on 2008 December 31.40 UT in two frequency bands at 8.4601 and 22.4601 GHz [@Stockdale08]. Observation for the second epoch was further reported by @Stockdale09 on 2009 January 27. Interacting Type II SNe, like Type IIn events and a few Type IIP events (e.g. SN 2004et) are supposed to be strong sources of radio emission (for review see @weiler02). However, both VLA observations produced null results for this proximate event.
In this paper we present optical and near-infrared photometric and optical spectroscopic observations of . The photometric data cover a time span of about 410 days since the discovery. The [*Swift*]{}/XRT [@burrows05] and [*Swift*]{}/UVOT [@roming05] data covering a time span of 60 days are also presented. In Sections \[res:broadband\] and \[res:spec\] we study the the photometric and spectroscopic evolutions respectively. In Sections \[res:DistExt\] and \[res:ColBol\] we cover the estimates of distance, reddening, intrinsic color and bolometric fluxes. The main physical parameters of the explosion and the mass of the progenitor are derived in Section \[res:parameter\] while a comparison of its properties with other SNe is given in Section \[res:comp\]. The conclusion of the paper is given in the last section. The epoch of explosion JD = 2454825.6 (§\[res:broadband\]) is considered throughout the paper and the times of pre-/post-explosion are referred with $-/+$ signs respectively.
observation and data reduction {#obs}
==============================
Photometric Observation {#obs:phot}
-----------------------
The prompt follow-up of the event was carried out by the ground-based ROTSE-IIIb telescope[^1] having sensitivity in the wavelength region from 0.35 to 1.0 with a peak around 0.6 [@quimby07]. The SN was first detected in the ROTSE-IIIb images on 2008 December 24.45 and it was monitored at 58 phases until +115d. The initial detections of the SN evaded the automated pipeline identification due to poor image quality and low signal-to-noise ratio (S/N). The data reduction was performed afresh on all the available ROTSE data. In order to remove the contamination of the true SN-flux from the underlying galaxy, a pre-SN galaxy template was constructed from images taken in early 2008 and each SN frame was reduced using the galaxy-template subtraction scheme developed by @alard00. The point spread function (PSF) photometry was performed at the SN location in the galaxy-template subtracted images. The unfiltered instrumental magnitudes were calibrated using the USNO B1.0 $R$-band magnitudes of about 15 isolated stars. The light curve thus produced was found to be 0.15 mag off from the Cousins $R$ band light curve produced by multi-band observation carried out at ARIES (described below) and the ROTSE magnitudes were scaled accordingly.
The SN 2008in was also monitored with the Ultraviolet Optical Telescope (UVOT) on board [*Swift*]{} from +5d to +60d. The UVOT filters $uvw2, uvm2, uvw1, u, b$, and $v$ have their effective wavelength at 2030, 2231, 2634, 3501, 4329, and 5402 Å, respectively (Poole et al. 2008). The UVOT data reduction was performed following the prescriptions of @brown09a. A 5 aperture is used to measure the counts for the coincidence loss correction whereas a 3 aperture was used for the photometry. For the filters $uvw2, uvm2$, and $uvw1$, the last epoch (obtained about 60 days after explosion when the UV flux was very weak) data were used to subtract the galaxy light, while in the optical the SN remains much brighter than the underlying light so contamination was negligible. An aperture correction (based on an average PSF in [*Swift*]{} CALDB) as well as zeropoints from @poole08 was applied to put the magnitudes on the UVOT photometric system. The UVOT magnitudes are listed in Table \[tab:uvotsn\].
From +5d to +416d, the follow-up of SN 2008in in optical broadband Johnson $UBV$ and Cousins $RI$ was performed with the 1-m Sampurnanand Telescope (ST) at ARIES, Nainital[^2]. An identification chart showing the field of the galaxy M61 along with the locations of as well as the local standards is presented in Figure \[fig:snid\]. The photometry is performed using standard tasks of IRAF [^3] and [*DAOPHOT*]{} [^4] as described in the paper by @roy11. Bias subtraction and flat fielding were performed on the raw frames and the cosmic removal was done using LACOSMIC routine [@dokkum01]. As the SN lies in the outskirts of the galaxy on a relatively faint and smooth background, the photometry at the initial phases (mostly during the plateau phase when the SN is bright) is estimated using the profile fitting method. During nebular phases, when the SN becomes faint, the true SN flux is estimated using the galaxy template subtraction method following the procedures of @roy11. As a galaxy template, we used post-explosion (+600d) images observed on 2011 January 04 under good seeing conditions. Figure \[fig:subimg\] shows an example of this procedure applied to a late-time $V$-band image of . In the template image of 2011 January 04, we can see a clear flux enhancement located approximately 5away (a linear distance of $\sim$ 343pc) from the SN position. This knot is also present in a deep image from the Sloan Digital Sky Survey (SDSS) and it is identified as a region in the NED catalog. The progenitor of may have a possible association with this star-forming region. The field of was calibrated using @landolt09 standard stars of the field SA98 observed on the same night as the SN. A sample of 10 bright and isolated non-variable stars in the field of was used as local standards to derive the zero points for the SN at each epoch. The location and magnitudes of these local standards are listed in Table \[tab:photstar\]. The entire time span of photometric observation is about 410 days and in Table \[tab:photsn\], we report the $UBVRI$ photometry of the event.
From +4d to +116d, the was also observed in $JHK$ near-infrared (NIR) bands with the 0.6-m REM/REMIR at La Silla [@zerbi04]. The object was clearly visible in the $J$ and $H$ passbands, while it remained undetected in $K$. The instrumental magnitudes were standardized using the 2MASS standards available in the field of . The calibrated $JH$ magnitudes are listed in Table \[tab:remsn\].
X-ray Observations {#obs:xrt}
------------------
The [*Swift*]{} XRT observations were obtained simultaneously with the UVOT observations. To search for X-ray emission from , we extracted X-ray counts from a circular region with a 10 pixel radius ($23\farcs7$, corresponding to the XRT on-axis 90% encircled energy radius) centered on the optical position of the SN. The background was extracted locally from a source-free region of $40\arcsec$ radius to account for the detector, sky background, and the diffuse emission from the host galaxy.
No X-ray source is detected in the merged 27.1 ks XRT data obtained in photon-counting mode. The $3\sigma$ upper limit to the XRT net count rate is $7.2 \times 10^{-4}~{\rm cts~s}^{-1}$, corresponding to an unabsorbed (0.2–10 keV band) X-ray flux of $f_{0.2-10} < 3.4 \times 10^{-14}~{\rm erg~cm}^{-2}~{\rm s}^{-1}$ and a luminosity of $L_{0.2-10} < 7.0 \times 10^{38}~{\rm erg~s}^{-1}$ for an adopted thermal plasma spectrum with a temperature of $kT = 10~{\rm keV}$ (see @fransson96 and references therein), a Galactic foreground column density of $N_{\rm H} = 1.67 \times 10^{20}~{\rm cm}^{-2}$ [@dickey90] and a distance of 13.19 Mpc (§\[res:DistExt\]).
{width="18cm"}
{width="5cm"}{width="5cm"}{width="5cm"}
Optical Spectroscopic Observation {#obs:spec}
---------------------------------
Long-slit low-resolution spectra ($\sim$ 6 to 14 Å) in the optical range (0.33 $-$ 1.0 ) were collected at eleven phases from +7d to +143d, including five phases from the 2-m IGO, three phases from the 9.2-m HET, two phases from the 6-m BTA and one epoch from the 3.6-m NTT [^5]. A journal of spectroscopic observations is given in Table \[tab:speclog\]. The spectroscopic data acquired from IGO, NTT and HET were reduced under the IRAF environment. Bias and flat-fielding were performed on all the frames. Cosmic-ray rejection was done using the Laplacian kernel detection method [@dokkum01]. All the data obtained from the BTA were reduced using programs in the IDL software environment.
The instrumental FWHM resolution of 2-m IGO spectra as measured from the [\[\]]{}$\lambda$5577 emission skyline was found to lie between 6 and 10Å ($\sim$ 322 - 510 ). Flux calibration was done using standard spectrophotometric fluxes from @hamuy94, assuming a mean extinction for the site. For HET, BTA and NTT, the resolution near 6000Å is about 10Å, 14Å, and 12Å respectively.
Light Curve Evolution {#res:broadband}
=====================
According to theoretical interpretation, the entire broadband evolution of Type II SNe can be segmented into three phases: the rising phase, the plateau phase and the nebular phase.
{width="13cm"}
The Rising Phase {#res:rise}
----------------
The rising phase of the light curve is associated with the shock breakout phenomenon having been theoretically predicted for an SN explosion, in which, the radiation-dominated shock wave, generated through the reversal of iron core-collapse, starts to propagate outward through the onion shell-like layers of the progenitor; when the shock reaches regions with an optical depth of a few tens to unity, the radiation behind the shock escapes the outer surface giving rise to a hot (T $>10^5$ K) fireball which emits quasi-black body radiation in UV and soft-Xrays. This phenomenon is called shock breakout and depending on envelope mass, density structure and wind properties of the progenitor, the breakout light curve may last from a few hours to few a days (@grassberg71 [@chevalier76; @falkarnett77]). Due to the short timescale, the detection of shock breakout is rare, and only recently were, the UV light curves of the entire shock breakout phase lasting several hours was reported for two Type IIP SNe $-$ SNLS-04D2dc and SNLS-06D1jd (@gezari08 [@schawinski08]). Observation of the earliest UV and optical light curves of IIP SNe is crucial to model the shock breakout light curves and constrain the properties of SN progenitors [@tominaga09]. The earliest SNe IIP optical light curves have been studied in the past for SN 2005cs by @pastorello09 and for SN 2006bp by @quimby07 and the optical rising phase is rarely observed.
{width="15cm"}
We model the early optical (ROTSE-IIIb $R$-band) light curves of using a simple model and put an observational constraint on the epoch and duration of shock breakout. Following the formulation of @waxman07, it can be shown that just after the shock breakout, the intensity of the SN light at a fixed wavelength is proportional to the intensity of blackbody radiation at that wavelength (see Equation 1 of @cowen10), while for the plateau phase @arnett80 derived an analytical expression (Equation 39 of that paper) and hence by combining these two equations one can approximate the overall intensity profile of a Type II SN during the rising phase through the plateau with the following expression :
I\_[SN]{}(t) = (t-t\_0)\^[1.6]{}
$$+ C\times exp[-\{(t/D) + (t/E)^2\}]$$
Here, the first term represents the phase associated with the shock breakout while the second term represents the plateau phase. Here ‘$t$’ is the time measured in days since 2008 December 23.95 (JD = 2454823.5), the epoch when ROTSE had a non-detection of the SN with limiting magnitude 16.16 in the $R$ band. The time of shock breakout, ‘$t_0$’ along with the constants $A, B, C, D$ and $E$ are free parameters of the fit. The values of these constants depend on the nature of the progenitor and the kind of explosion. The above model is fitted to the ROTSE data using the $\chi$ minimization technique [@press92] and the fit is shown in Figure \[fig:rotse\]. We obtained a value of the shock breakout time as $t_0 = 2.13\pm0.83$ day. This value of $t_0$ is consistent with the first data point observed from ROTSE. We shall adopt $t_0$, corresponding to JD = 2454825.6 as the explosion epoch for all the phases of . It is noted that at $V$ and $R$ bands, the SN was detected by @nakano08, just 2.4 d after the shock breakout. We note, however, that depending on the extent of the envelope the true core reversal marking the SN explosion would have occurred a few hours earlier (e.g., SN 1987A) or a few days earlier for a red supergiant envelope $\sim 50000{\mbox{$R_{\odot}$}}$ [@quimby07].
In Figure \[fig:rotse\], ‘$t_{p}$’ (+5.3 d) corresponds to the peak of the ROTSE $R$-band light curve. A similar peak in the UV light curves of the post-shock breakout phase of the Type IIP supernova SNLS-04D2dc lasting several days was identified as a secondary peak and this peak was explained as the shift of the spectral energy distribution toward longer wavelengths due to the rapid fall in temperature during free adiabatic expansion of optically thick plasma lead by a collisionless shock (@schawinski08 [@tominaga09]). For we identify the peak in the ROTSE $R$-band curve as a secondary peak and also speculate that for a few other type IIP SNe such as 1999em, 1999gi and 2005cs, where very tiny peaks before the plateau light curve were observed, such features are basically the footprints of the secondary peaks which could be a consequence of shock breakout.
The Plateau and Nebular Phase {#res:PlatNeb}
-----------------------------
The entire UV, optical and NIR light curves of are shown in Figure \[fig:lightcur\] and for comparison, the light curves of the archetypal low-luminosity SN 2005cs [@brown09b; @pastorello09] are also plotted by scaling them in magnitude to match it with the observed plateau of . In the early plateau phase, a rapid drop in the UV flux and a slowly declining or constant optical and NIR flux are clearly apparent. The $uvm2$ magnitude declined from about 14.7 to 18.7 mag within 12 days. The decline rates of the flux in UVOT, $U$, $B$, and $V$ bands are almost identical with that of SN 2005cs. Starting from the $uvw2$ until the $R$ band, the measured decline rates from shock breakout to the plateau phase are approximately 0.32, 0.34, 0.25, 0.13, 0.04, 0.01 and 0.01 mag d$^{-1}$. The $IJH$ light curves are almost flat even at early phases of the plateau. For this SN, the plateau is well sampled in the $BVRI$ bands, so we can accurately determine the plateau duration. As discussed in §\[res:rise\], the secondary peak ($t_{p} \sim +5.3$ d) in the prompt light curve is generated by the gradually cooling shock heated SN atmosphere and after that the plateau mechanism starts to dominate. On the other hand, between the plateau and the nebular phase there is another transitional state, when an inflection in the light curve can be seen (§\[res:nick\]). Hence the plateau duration is precisely the time span between the secondary peak and the inflection. The inflection ($t_{i}$) in the $V$ band light curve is observed to occur nearly 103.2 days after the shock breakout. Hence the time interval between the $t_{i}$ and $t_{p}$ is assumed as the duration of the plateau i.e. $(103.2-5.3) \sim 98$ days.
Once the hydrogen envelope is fully recombined and the ejecta becomes optically thin, the light curve enters into the nebular phase and it is sustained mainly by the energy output from the radioactive decays of the iron-group elements. During the plateau to nebular transition phase, the $V$-band magnitude drops from $\sim$ 16.0 mag at around +90d (still in the plateau) to 18.1 mag at +122d (in the exponential light curve tail), i.e. $\sim$ 2 mag in about one month. This drop is remarkably smaller than that of SN 2005cs, but consistent with the 2-3 mag drop observed in normal Type IIP SNe [@olivares10]. A linear fit to the tail from +120d to +400d gives the following decline rates \[in mag (100d)$^{-1}$\]: $\gamma_B~\sim~0.44$, $\gamma_V~\sim~0.84$, $\gamma_R~\sim~0.91$, $\gamma_I~\sim~1.03$ at $B, V, R,$ and $I$ bands which are similar to the values found in normal IIP SNe and comparable with the decay slope of $^{56}$Co to $^{56}$Fe, i.e. 0.98 mag (100d)$^{-1}$.
Spectroscopic Evolution {#res:spec}
=======================
The spectra of at 10 phases from +7d to +143d are shown in Figure \[fig:spec\_all\]. The spectra are corrected for the recessional velocity of the host galaxy ($\sim$ $1567\pm3$ )[^6]. Spectral features are mainly identified as per previously published line identifications for IIP SNe [@leonard02; @pastorello04]. The two earliest spectra (+7d and +14d) show the blue continuum of blackbody emission and have broad P-Cygni profiles of , and $\lambda$5876. The next two spectra (+54d and +60d) represent the mid-plateau phase and are marked by a decrease in the continuum and the appearance of more number of P-Cygni profiles for the permitted metallic (, ,, , , ), $\lambda$7773, D, and IR triplet lines similar to other normal Type IIP SNe (see @roy11 and references therein). The spectra at phases +87d, +90d and +99d represent the end stages of the plateau and these are marked by a redder continuum and decreasing line widths of hydrogen lines. The +118d, +119d and +143d spectra represent the nebular phase having negligible continuum and are marked by pronounced emissions of forbidden lines [\[\]]{} $\lambda\lambda$7291, 7323, and fading of the absorption features of hydrogen and metallic lines.
In Figure \[fig:spec\_ccsne\], we compare the +99d spectrum of with the spectra of other Type IIP SNe observed at similar phases that is at roughly the time when the light curve changes from the photospheric plateau to the nebular phase. The phases quoted for each event are with reference to the moment of inflection ($ti$; see Figure \[fig:rotse\]). The narrow P-Cygni profile and the presence of strong lines of at $\lambda$6142 and $\lambda$6497 of show striking resemblance to the low-luminosity SNe 1997D, 1999br, 1999eu, 2001dc and 2005cs. On the other hand, the normal Type IIP SNe 2004et, 2008gz show broader profiles of and weaker lines of . It is noted that in the blue wing of more metallic lines get resolved than that in the normal IIP SNe and this arises due to smaller line widths of hydrogen lines seen in low-luminosity IIP SNe (@fraser10 and references therein). In addition, the line in the low-luminosity SNe show stronger absorption components due to lower ejecta temperature than that in the normal luminosity Type IIP SNe 2004et and 2008gz [@turatto98]. In Figure \[fig:line\_id\], we have identified the spectral features in a late-plateau phase (+99d) spectrum which covers the full wavelength range from 0.4 to 0.95 . For line identification we have followed @pastorello04, where different spectral lines were identified for the +102d spectrum of a low-luminosity SN 1999br. We are able to identify all the features and the spectral profiles of all the elements are similar to the archetypal low-luminosity IIP SNe 1997D, 2005cs and 1999br (@pastorello04 and references therein).
{width="15cm"}
{width="15cm"}
Figure \[fig:vel\_line\] shows the spectral evolution of , , D, , , and lines. In Figure \[fig:vel\_bulk\] (left panel), we show the expansion velocities of the ejecta derived from Balmer and lines. The later is a good indicator of photospheric velocity. The expansion velocities of the H-envelope are estimated using absorption minima of the P-Cygni profiles and at the two earliest phases (+7d and +14d) and show a broad P-Cygni profile which, with time, becomes narrower at later phases keeping the position of the emission peak centered near zero velocity. The line velocity starts at about 7000 at +10d, reaches 4000 at +50d and flattens at a level of 1200 in the nebular phases. It can be seen from the right panel of Figure \[fig:vel\_bulk\] that in the comparable phases, the line velocities of low-luminosity SNe are less than half those of normal IIP SNe 1999em and 2004et, whereas the velocities are more like low-luminosity SN 2005cs. To estimate the photospheric velocity of the transient we have computed the velocities of different lines $\lambda$4924, $\lambda$5018, and $\lambda$5169 at different phases. The first marginal detection of these lines is in the +14d spectrum and they became prominent in later stages of evolution. For , the average velocity of these lines (and hence roughly the photospheric velocity) at +14d is about 4450 , which is comparable with that of low luminosity SNe and much less than ordinary Type IIP events (see Figure 12 of @pastorello09).
{width="9cm"}{width="9cm"}
{width="15cm"}
The emergence of absorption dips in the blue wings of is clearly seen in the post +60d spectra [^7] and the absorption dips have been marked with A, B and C in Figure 8. We distinguish ‘A’ as an absorption dip due to and and its equivalent width (EW) increases from 12.85 Å at +60d to 19.45 Å at +143d. This progressively stronger absorption dip requires a steep deceleration of the emitting zone along with its depletion by inner metal shells (like ). The broader absorption dip ‘B’ is identified as a blend of $\lambda$6497, , and . An exceptionally strong lines is quite typical for cooler ejecta of less energetic fain SNe. The segment ‘C’ is speculated to be a footprint of multiplets as the evolution of its linewidths seems to be correlated with the other lines. It is noted, however, that in a few normal Type IIP SNe 1999em, 2004et and 2008gz, the spectral feature near ’C’ at early plateau phases was also identified as a signature of high velocity component, and in the present spectra its presence cannot be completely ruled out.
The evolution of is similar to the evolution of . The P-Cygni profile of is clearly visible in the +7d spectrum, and later on, the emission component is blanketed by various metal lines, mainly due to $\lambda$4924, $\lambda$5018 and $\lambda$5169. Traces of and absorption dips indicate that from the beginning of the nebular phase H shells of ejecta and the regions containing ions move with a comparable velocity, reaching an asymptotic value $\sim$ 1000 nearly 140 days after the explosion. In the +14d spectrum, lines are marginally detected while in the spectra between +60d and +99d they are prominent. We investigate the temporal change of EW of a relatively less blended $\lambda$4924 and from the spectra normalized to the peak value, we found that at an EW $\sim$ 0.67 Å at +14d, increased to 17.45 Å at +54d and to 26.53 Å in the +99d spectrum. Later on, the EW decreased to 19.07 Å in the +118d spectrum and finally reaches to 18.27 Å in the +143d spectrum. This rapid fall in the EW of inner metal-rich shells plausibly indicates a decrease in the opacity of the SN inner ejecta.
Traces of the $\lambda$5876 line are also seen in the +7d spectrum (Figure \[fig:vel\_line\]). The ratio of EWs between and for this spectrum is EW()/EW() $\sim 1.99$. In the +14d spectrum, this ratio has increased to 5.36. The steep decrement in EW of seems to be due to a quick recombination of ions as a result of the rapid fall in temperature of the constantly rarefying ejecta. From +54d, the emerging P-Cygni feature of D becomes prominent. This feature seems to be a perfect P-Cygni throughout our entire spectral sequence. It simply indicates an uniform spherical distribution of in the SN ejecta. In the three high S/N spectra labeled with +54d, +60d and +87d, a tiny absorption dip overlaid on the emission component of the D P-Cygni is seen and since these spectra are redshift corrected, these dips are impressions of intervening interstellar matter present in the host galaxy. The absence of any similar absorption dip in the blue wing of the D profile, expected from the Galactic interstellar matter, confirms that there is little absorption from the Milky Way in the line of sight of .
The spectral evolution of $\lambda$6248 and the s-process element $\lambda$6142 is clearly seen in . In the +7d and +14d spectra, there is no clear evidence for the presence of these two elements; however in the +54d spectrum they are prominent. The absorption components become stronger with time and persist until the +143d spectrum. In contrast, in the normal Type IIP SNe 2004et [@sahu06], 1999em [@elmhamdi03a] and 2008gz [@roy11] these features disappeared at $\sim$ 170 day after the explosion. The ejecta of underluminous Type IIP SNe expand at a lower velocity than those of normal ones. So, the Ba lines in low luminosity events persist for a longer time just because the ejecta takes more time to cool-down. As a result it takes longer time to become optically thin.
In the nebular phase, the forbidden lines: \[\] $\lambda\lambda$7291, 7323, \[\] $\lambda\lambda$6300, 6364 and \[\] $\lambda$7155 are among the strongest features visible in the spectra. Also, the line ratios of emission lines of and the doublet in late nebular phases is a good indicator of the progenitor mass [@fransson87]. In Figure \[fig:spec\_oi\], we compare the +139d and +283d spectra of SN 19997D (Benetti et al. 2001) with the nebular phase spectra of . The \[\] doublet is seen in the +99d, +119d and +143d spectra of , whereas a small footprints of the doublet can be seen in the +143d spectrum. Presence of \[\] line is not seen in our spectra. Considering that the explosion epoch of SN 1997D was quite uncertain, it is likely that the spectral evolution of and forbidden lines of is quite similar to that of SN 1997D.
extinction and distance toward {#res:DistExt}
===============================
In order to determine the bolometric light curve and other physical parameters, a correct estimate of the extinction and distance toward the SN is essential. We adopt the Galactic reddening along the line-of-sight of as derived from the 100 all sky dust extinction map [@schlegel98], i.e. = $0.0224\pm0.0003$ mag. In order to estimate reddening due to the host galaxy M61, we used the spectrum of taken on 2009 February 17 from the 2-m IGO telescope having good S/N ($\sim$ 40) and corrected for mean heliocentric radial velocity of the host ($cz \approx 1567\pm3$ ). Near the zero velocity, the spectrum showed a tiny absorption feature overlaid on the emission component of the P-Cygni profile of D (§ \[res:spec\]). The EW of this absorption feature was computed as 0.535$\pm$0.713 Å. The error in EW is estimated using Equation 6 of @vollmann06. It is known that the EW of the interstellar absorption bands is well correlated with the reddening estimated from the tail of SN Ia color curve [@barbon90; @richmond94; @turatto03] and hence using the empirical relation $=$ $-0.01 + 0.16$EW (EW in Å), given by @turatto03, we obtain the host contribution as $\approx 0.076\pm0.104$, which is considerably higher than the Galactic contribution. This is consistent with the absence of any absorption feature due to Galactic interstellar matter in the blue wing of the emission component of D profile. Consequently, we adopt (estimated as a sum of Galactic and host galaxy reddening) of 0.0984$\pm$0.104 mag for . This corresponds to a visual extinction ($A_V$) of $0.305\pm0.322$. Considering the uncertainty in the estimated using D lines, a lower value of of 0.0448$\pm$0.0006 mag (twice the Galactic reddening) and the corresponding $A_V$ of $0.139\pm0.002$ cannot be ruled out. We will discuss its implication for the derived properties of the SN.
The Hubble flow distance of the host galaxy M 61, after correction for the Virgo infall, is estimated as 13.7$\pm$1.1 Mpc [^8]. The distance estimated through the Tully-Fisher method is $12.1\pm2.7$ Mpc[^9]. Additionally, we also calculated the distance following the Standard Candle Method for Type II SNe [@hamuy02; @hamuy05; @hendry05]. It is found that there is a strong correlation between the distance of Type IIP SNe along with their mid-plateau apparent $V$ and $I$ band magnitudes and the mid-plateau photospheric velocity, for a given cosmological model. For we estimate the mid-plateau ($\sim +50$d) apparent magnitudes $\sim 15.56\pm0.05$ mag for $V$ and $\sim 14.77\pm0.06$ mag for the $I$ band. Considering the $V$ and $I$ band extinctions toward the SN, $0.305\pm0.322$ mag and $0.183\pm0.193$ mag respectively, we derive a distance of $\sim 12.23\pm1.87$ Mpc. The adopted value of mid-plateau photospheric velocity is $2694.67\pm70$ (see §\[res:proge\]). Combining the above three measurements, we adopt the weighted mean distance of $13.19\pm1.09$ Mpc, which corresponds to a distance modulus of $30.6\pm0.2$.
{width="15cm"}
Color Evolution and Bolometric Flux {#res:ColBol}
===================================
Figure \[fig:colorcur\] shows the temporal evolution of the reddening-corrected broadband colors of . For comparison, we overplot the color curves for well studied SNe 1987A [@suntzeff90], 1999em [@elmhamdi03a], 2004et [@sahu06; @maguire10], 2005cs [@pastorello09] and SN 2008gz [@roy11]. After small differences during the initial phases, the plateau phase color evolution of all Type II SNe is more or less similar. The $(U-B)_{0}$ and $(B-V)_{0}$ colors are blue during early photospheric phases and they become redder by about 1-2 mag in the plateau phase while the $(V-R)_{0}$ and $(V-I)_{0}$ colors evolves slowly and become red only by about 0.5 mag. The $(J-H)_{0}$ color remains constant at $\sim$ 0.25 mag. During the transition phase from plateau to nebular, the low luminosity SN 2005cs showed a striking red peak in the $(B-V)_0$, $(V-R)_0$ and $(V-I)_0$ colors. For , this red peak is not present and its color evolution is found to be consistent with the normal Type IIP SNe.
The quasi-bolometric light curve estimated from the UV, optical and IR broadband ($UVOIR$) magnitudes of is shown in Figure \[fig:bolcur\] along with those of other Type II SNe. The extinction-corrected magnitudes are first converted into fluxes using zero-points given by @bessell98 and then the total flux in $UVOIR$ bands is obtained after a linear interpolation and integration between 0.203 and 1.67 ${\mbox{$\mu{\rm m}$}}$. The fluxes were calculated on those nights for which we had complete observations in $UBVRI$ bands. For the initial two weeks the SN was detected by [*Swift*]{}/UVOT in near ultraviolet bands with a significant flux density. This contribution has been accounted for while measuring the net $UVOIR$ flux. It has been assumed that contribution in the $U$ band is mainly important during the plateau and decreases rapidly to about 5% in the nebular phase (see @misra07 [@roy11] and references therein). From the light curve it is noticeable that the near ultraviolet flux is almost negligible beyond 20 days after explosion. The $JH$ contribution during the plateau is calculated from our data, whereas for the nebular phase, due to lack of data, we are not able to make any direct measurement. For most low luminosity SNe IIP, the NIR flux contribution in the nebular phase is about 50% of the total flux [@pastorello09]. We have therefore increased the net flux by 55% to account for the maximal contributions from the $U$ and NIR bands at phases later than +140d. Figure \[fig:spec\_ccsne\] and \[fig:bolcur\] clearly demonstrate that spectroscopically appears to be like low-luminosity SNe IIP, but photometrically it appears quite normal. In the following section, the $UVOIR$ curve is used to estimate the amount of radioactive $^{56}$Ni and other physical parameters that characterize the explosion and the progenitor star.
{width="15cm"}
Physical Parameters {#res:parameter}
===================
The physical entities that seem to govern the entire scenario are mainly associated with the nature of the progenitor and the radioactive elements (mainly $^{56}$Ni), generated inside the inner portion of the ejecta during the explosion. $^{56}$Ni is synthesized by the explosive burning of Si and O during the shock breakout [@arnett80; @arnett96]. Over time this material is eventually converted to $^{56}$Co and then to $^{56}$Fe by means of radioactive transitions having decay times of 8.77 and 111.3 days respectively. The $\gamma$-rays and neutrinos emitted during this process sustain the nebular phase light curve and consequently, the observed tail luminosity becomes a good tracer for the ejected synthesized $^{56}$Ni.
Produced Radioactive Nickel {#res:nick}
---------------------------
We use different methods to estimate mass of $^{56}$Ni. @hamuy03 proposed a relation between bolometric tail luminosity and the synthesized $^{56}$Ni during core collapse SNe, by considering the underlying assumption that all $\gamma$-rays emitted during the radioactive decay make the ejecta thermalised. For , the time interval spanned by the observations is about 416 days, where the first $\sim$ 100 days are reserved for photospheric evolutionary processes. The average $V$ band magnitude during the nebular phase calculated using the data between +114d and +416d is $\sim$ 18.91, which corresponds to the $V$ magnitude at $\sim$ +222d. Taking the extinction correction as ($A_V = 0.305\pm0.322$ mag; §\[res:DistExt\]), a bolometric correction of $0.26 \pm 0.06$ mag [@hamuy01] and a distance modulus $30.6\pm0.2$, the derived tail luminosity at this fiducial time is $(3.02\pm1.95)\times10^{40}\,{\rm erg\,s^{-1}}$. Within the errors, this value is consistent with the bolometric flux at a comparable epoch, determined in §\[res:ColBol\]. This implies that the amount of $^{56}$Ni produced is this process is $M_{\rm Ni} =0.0157\pm0.0102$.
All Type IIP SNe show an inflection in the light curve during the transition from the plateau to the nebular phase. Statistically it has been shown that the steepness of the $V$-band light curve slope (defined as S=d$m_{\rm V}$/dt) at the inflection time ($t_{i}$) is anti-correlated with mass [@elmhamdi03b]. For , the $V$ band light curve with its well-sampled transition phase, shows a value of steepness ${\rm S}=0.151\pm0.044$ mag d$^{-1}$ (Figure \[fig:steep\]) and the epoch of inflection is $t_{i} \approx$ +101.5d with respect to date of discovery. This corresponds to $M_{\rm Ni}=0.0175\pm0.002$. This result is consistent with the value measured using the @hamuy03 scheme. According to @elmhamdi03b, the amount of $^{56}$Ni can also be derived from the plateau absolute $V$-band magnitude using the relation ${\rm log} M_{\rm Ni} = -0.438 M_{V}(t_{i}-35)-8.46$. Here $M_{V}(t_{i}-35)$ is the absolute $V$ magnitude 35 days prior to the day of inflection. For , $M_{V}(t_{i}-35) \approx -15.23$, which again corresponds to a $^{56}$Ni mass around 0.016.
Comparison of the tail luminosity with that of SN 1987A $-$ a well studied proximate event, is also used for estimation of the $^{56}$Ni mass. A linear least square fit on the nebular light curve tail shows that at +222d the luminosity of is about 1.54$\times10^{40}\,{\rm erg\,s^{-1}}$, while SN 1987A had a luminosity $\sim 1.10\times10^{41}\,{\rm erg\,s^{-1}}$ (Figure \[fig:bolcur\]). Since the $^{56}$Ni mass produced by SN 1987A is about 0.075, the amount of $^{56}$Ni in the case of is $[(1.54/1.10)\times10^{-1}]\times0.075 \approx 0.0105$.
The above estimates are consistent with each other and hence we adopt a mean value for $^{56}$Ni mass as $0.015\pm0.003$
Explosion Energy and Mass of Progenitor Star {#res:proge}
--------------------------------------------
We use the radiation-hydrodynamical simulations of core-collapse IIP SNe by @dessart10 to infer the explosion energy ($E_{0}$ - kinetic plus thermal, expressed hereafter in units of $10^{51}$ erg or foe). These simulations suggest that in a given progenitor larger explosion energies yield larger ejecta velocities and this implies that the ejecta kinematics can be used to put constraint on $E_{0}$. For , the expansion rate of the H-rich progenitor envelope as derived from absorption minima in at 15d after shock breakout is about 6300 (see Figure \[fig:vel\_line\]), which is higher than 4700 (for SN 2005cs; @pastorello09) and lower than 8800 (for SN 1999em; @elmhamdi03a). For the SNe 2005cs and 1999em, the simulation results (for explosion of a non-rotating solar metallicity pre-SN stars) predicted $E_{0} \sim$ 0.3 foe and $\ga$ 1 foe respectively and these values are found to be consistent with that determined from the actual hydrodynamical modeling of the SN light curves, e.g. 0.4 foe [@utrobin08] and 0.3 foe [@pastorello09] for SN 2005cs; 1.3 foe [@utrobin07a] and 1.25 foe [@bersten11] for SN 1999em. For , the simulations suggest $E_{0}$ $\sim$ 0.5 foe.
Accurate determinations of explosion parameters such as $E_{0}$, the ejected mass ($M_{\rm ej}$) and the pre-SN radius ($R_{0}$) of the progenitor require detailed hydrodynamical modeling of the light curves and spectra and this being a non-trivial task (beyond the scope of this paper) has only been attempted for a few SNe. In order to have an estimate of explosion parameters of here, we employ the analytical relations derived by @litvinova85 correlating the observed parameters ($M_{Vmp}$ - the mid-plateau absolute magnitude at $V$, $v_{mp}$ - the mid-plateau photospheric velocity and $\Delta t_{p}$ -the plateau duration) with the physical parameters ($E_{0}$, $M_{\rm ej}$ and $R_{0}$) based on a grid of hydrodynamical models for different values of physical parameter for Type IIP SNe. We note, however, that these approximate formulae have limitations owing to the poorly-measured observables and the simplified physical conditions such as non-inclusion of the effect of nickel heating, use of old opacity tables, neglecting the effect of line opacity and using outdated pre-SN models (@smartt09 [@bersten11] and references therein). As a result, only approximate values of the physical parameters can be derived using these relations. Fortunately, the observed parameters are derived very accurately for . The $\Delta t_{p}$ is $\sim 98$ days (§\[res:PlatNeb\]), the $v_{\rm mp}$ is $2694\pm70$[^10] and the $M_{Vmp}$ is estimated as $-15.32\pm0.38$ mag[^11] Now, employing analytical relations, we estimate $E_{0}$ $\sim$ 0.54 foe, $M_{\rm ej} \sim 16.7$ and $R_{0}$ $\sim 127 R_{\odot}$. The explosion energy derived in this way is consistent with that predicted from the hydrodynamical simulations of the ejecta kinematics.
The explosion energy of indicates that the event was less energetic than the standard IIP SNe 1999em, 2004et and more energetic than SN 2005cs. Assuming a net mass loss $\sim 0.5$M$_\odot$ due to stellar wind [^12] and accounting for a compact remnant with mass $\sim 1.5-2$M$_\odot$ [@sahu06], we find that the initial mass of the progenitor to be $\leq 20$M$_\odot$.
![Determination of steepness parameter from the apparent $V$-band light curve of . See text for details.[]{data-label="fig:steep"}](elmhamdi_fit.ps){height="13cm"}
{width="15cm"}
Discussion in context of other Type IIP SNe {#res:comp}
===========================================
In Figure \[fig:abs\_lc\], we compare the light curve of in absolute $V$-band magnitude with a sample of 13 other well studied Type II SNe taken from the literature (see @misra07 and @roy11 for the references). The sample includes two low-luminosity Type IIP SNe 1997D, 2005cs; nine normal Type IIP SNe 1990E, 1992H, 1999gi, 1999em, 2003gd, 2004A, 2004dj, 2004et, 2008gz and two peculiar Type II SNe 1987A, 1998A. It is seen that the Type II SNe show a wide range of mid-plateau luminosity, i.e. from $-14$ to $-17$ mag. The data on low luminosity SNe (see also Table 4 in @pastorello06) indicate that their $M_{Vmp}$ lie between $-14$ and $-15$ mag, with the only exception of SN 2002gd, whereas the other SNe are seen to lie between $-16$ and $-17$ mag. With the $M_{Vmp}$ of $-15.32$ mag[^13], the present another case which occupies the gap between a low and normal luminosity Type IIP events.
The tail luminosity is an important indication of ejected nickel mass. The average tail luminosity of (Mv $\sim -12.16$) is nearly 2.3 mag brighter than those of the low luminosity SNe 1997D and 2005cs ($\sim -9.6$ mag), and about 1.3 mag fainter than the normal Type IIP SNe 1992H, 2004et and 2008gz ($\sim -13.5$ mag). A close inspection shows that the average tail luminosity of is roughly similar to Type IIP SNe 2004dj in NGC 2403 and 2003gd in M74, whereas their plateau is brighter ($\sim 0.5-0.8$ mag at mid-plateau) than . These comparison of nebular phase luminosities with other SNe are consistent with a quite modest radioactive $^{56}$Ni production for (§\[res:nick\]).
The luminosity and shape of the tail depend on the $^{56}$Ni mass and the radiant energy per unit ejected mass ($E/M_{\rm ej}$). The first parameter determines the absolute magnitude, while the second determines its decay rate (see @turatto98 and references therein). The measured value of ejected $^{56}$Ni mass for SNe 2003gd, 2004dj and 2008in are respectively $\sim 0.016$ [@hendry05], 0.017 [@venko06] and 0.015. These are more than twice the $^{56}$Ni amount synthesized by the low-luminosity SNe 1997D ($\sim 0.002$, @turatto98) and 2005cs ($0.003-0.004$, @pastorello09), although at least three times less than the $^{56}$Ni produced by normal Type IIP SNe such as 1992H ($\sim 0.075$, @clocchiatti96), 2004et ($\sim 0.06$, @sahu06 [@maguire10]) and 2008gz ($\sim 0.05$, @roy11). Similarly, the measured value of $E/M_{\rm ej}$ ratio for is larger than that for the low-luminosity SN 1997D and lower than other Type IIP events [^14].
Though we have an approximate value of the pre-SN radius (128 ), it can be seen that the progenitor radius of is smaller than or comparable to that of low-luminosity SNe 1997D ($\sim 300$, @turatto98), 2005cs ($100-600$, @pastorello09, @utrobin08), 1999br ($114$, @zampieri03) and much smaller than that of normal Type IIP SNe like 1992H ($\textgreater 600$, @clocchiatti96), 2004et ($\sim 530$, @misra07). We also note that progenitor is only slightly larger than the blue supergiant progenitors of SNe 1987A and 1998A [@pastorello05].
Information about the metallicity of the SNe location is essential to constrain the triggering mechanism of the SN explosion [@heger03]. The oxygen abundance \[O/H\] of is about 8.44 dex [^15] which is marginally lower than the solar value of 8.65 dex [@asplund09]. A comparison [^16] of \[O/H\] of all the events in our sample indicate that there is no clear trend in the type of event and the metallicity.
The photometric and spectroscopic comparisons of with some Type II SNe covering a wide range of physical parameters puts observational constraints on the nature of the progenitor of this event, pointing toward a star that was more compact than a typical M-Type red supergiants, and closer to a blue supergiant, may be an yellow supergiant. However, the direct detection of the progenitors of a few faint SNe IIP in pre-explosion archive [*Hubble Space Telescope*]{} images seems to contradict this conclusion (see @smartt09 for a review). Our spectroscopic study suggests that in the hydrogen envelope ejected by the explosion is smaller than in most Type II events but larger than that ejected in low luminosity events (such as SN 1997D, SN 1999br and SN 1999eu). This may be due to a significant mass loss of the parent star in the latest stages of its life. This conclusion is also partially supported by the upper limit mass loss rate of the progenitor revealed from the X-ray study (§\[res:proge\]). In view of the upper mass limit of 20 M$_\odot$ progenitor, occurrence of the event in a sub-solar metallicity region and following the evolution of a single massive star as a function of metallicity [@heger03], we rule out the possibility of “fall back of ejecta to BH” scenario in case of , supporting the scenario of a weak explosion of a relatively compact progenitor.
conclusion {#concl}
==========
Low-luminosity Type IIP SNe belong to a poorly known class of events due to the unknown nature of their progenitors as well as the explosion mechanism. Spectroscopic as well as photometric characteristics of these events are significantly different from the normal and peculiar Type IIP SNe. The number of such underluminous events discovered so far is relatively small. This could plausibly be a selection effect. If we confine the search of core-collapse SNe to a small volume, the majority of SNe discovered in the local universe are of Type IIP ($\sim 48.2\%$, @smith10) and probably a large fraction of them would turn out to be underluminous. Incidentally SN 1999br, SN 2005cs and SN 2009md are the only events of this group to have been discovered soon after the core-collapse and whose data are publicly available.
In this study we have reported the results of an extensive photometric and spectroscopic follow-up campaign for in the ultraviolet, optical and near-infrared domains. The SN was also observed (although not detected) at X-ray and radio wavelengths. The SN was observed soon after its explosion, during the fast rise of the light curve to the optical maximum. Evidence for a shock breakout in was primarily derived through the analysis of the ROTSE-IIIb $R$-band light curve. An upper limit of 16.16 mag in the $R$-band was estimated about two days before the discovery.
Modeling the $R$-band light curve allowed us to estimate a reliable epoch of the shock breakout, with an uncertainty of about two days. The plateau phase in lasts about 98 days which is marginally shorter than in normal Type IIP SNe. The evolution of the bolometric light curve indicate that the event is somewhat in between the normal and faint Type IIP event. The luminosity of the nebular phase light curve indicates an ejected $^{56}$Ni mass of $\sim0.015{\mbox{$M_{\odot}$}}\,$, a factor two higher than that derived for low-luminosity IIP SNe. The spectroscopic evolution of is similar to those of low-luminosity IIP SNe (1997D, 1999br and 1999eu), indicated particularly by the strong presence of lines, the narrow line widths of lines, and the expansion velocities of SN during the photospheric phases ($\sim 3000$ ) and the nebular phases ($\sim$ 1200 ). The ejecta kinematics of are consistent with less-energetic ($\sim 5\times10^{50}$ erg) Type IIP SNe.
Spectroscopically appears to be like low-luminosity SNe IIP, but photometrically it appears close to a normal type IIP event.
From the light curve and spectra of , we could determine quite accurate values of the observed properties such as plateau duration, mid-plateau luminosity as well as the photospheric velocity and this has helped us to comment on the properties of the explosion and the progenitor star. Using semi-analytical formulae by @litvinova85, we could estimate approximate values of the explosion energy $\sim5.4\times10^{50}$ erg, the ejected mass $\sim17{\mbox{$M_{\odot}$}}$ and the pre-SN radius $\sim127{\mbox{$R_{\odot}$}}$. The explosion energy of is smaller than the normal ($\ge 10^{51}$ erg) Type IIP events, although higher than that estimated in underenergetic ($\sim 10^{50}$ erg) SNe IIP. We could provide an upper limit to the mass loss rate of the progenitor as $5\times10^{-6}~M_{\odot}~{\rm yr}^{-1}~(v_{\rm w}/10~{\rm km~s}^{-1})$ whereas the upper limit for the main-sequence mass of the progenitor star is estimated as 20 .
We thank all the observers at Aryabhatta Research Institute of Observational Sciences (ARIES) who provided their valuable time and support for the observations of this event. We are thankful to the observing staffs of ROTSE, REM, 2-m IGO, 3.6-m NTT, 6-m BTA and 9.2-m HET for their kind cooperation in observation of . This work was supported by the grant RNP 2.1.1.3483 of the Federal Agency of Education of Russia. Timur A. Fatkhullin and Alexander S. Moskvitin were supported by the grant of the President of the Russian Federation (MK-405.2010.2). This is also partially based on observations collected at the European Southern Observatory, Chile under the program 083.D-0970(A). Stefano Benetti and Filomena Bufano are partially supported by the PRIN-INAF 2009 with the project “Supernovae Variety and Nucleosynthesis Yields”. The research of J. Craig Wheeler is supported in part by NSF Grant AST-0707669 and by the Texas Advanced Research Program grant ASTRO-ARP-0094. This research is supported by NASA grant NNX08AV63G and NSF grant PHY-0801007. This work is partially based on observations made with the REM Telescope, INAF Chile. This research has made use of data obtained through the High Energy Astrophysics Science Archive Research Center Online Service, provided by the NASA/Goddard Space Flight Center. We are indebted to the Indo-Russian (DST-RFBR) project No. RUSP-836 (RFBR-08-02:91314) for the completion of this research work.
, C. W., et al. 2003, , 115, 132
, C. 2000, , 144, 363
, W. D. 1996, [Supernovae and nucleosynthesis. an investigation of the history of matter, from the Big Bang to the present]{}, ed. [Arnett, D.]{}
, W. D. 1980, , 237, 541
, M., [Grevesse]{}, N., [Sauval]{}, A. J., & [Scott]{}, P. 2009, , 47, 481
, R., [Benetti]{}, S., [Rosino]{}, L., [Cappellaro]{}, E., & [Turatto]{}, M. 1990, , 237, 79
, S., et al. 2001, , 322, 361
, M. C., [Benvenuto]{}, O., & [Hamuy]{}, M. 2011, , 729, 61
, M. S., [Castelli]{}, F., & [Plez]{}, B. 1998, , 333, 231
, P. J. 2009, Ph.D. thesis, The Pennsylvania State University
, P. J., et al. 2009, , 137, 4517
, D. N., et al. 2005, , 120, 165
, S., [Prabhu]{}, T., [Anupama]{}, G. C., [Kaur]{}, A., [Uday Kumar]{}, G., & [Ray]{}, A. 2008, Central Bureau Electronic Telegrams, 1638, 1
, R. A. 1976, , 207, 872
, N. N., & [Utrobin]{}, V. P. 2000, , 354, 557
, A., et al. 1996, , 111, 1286
, D. F., [Franckowiak]{}, A., & [Kowalski]{}, M. 2010, Astroparticle Physics, 33, 19
, L., [Livne]{}, E., & [Waldman]{}, R. 2010, , 408, 827
, J. M., & [Lockman]{}, F. J. 1990, , 28, 215
, J. J., [Mattila]{}, S., & [Smartt]{}, S. J. 2007, , 376, L52
, J. J., & [Tout]{}, C. A. 2004, , 353, 87
, A., [Chugai]{}, N. N., & [Danziger]{}, I. J. 2003, , 404, 1077
, A., et al. 2003, , 338, 939
, S. W., & [Arnett]{}, W. D. 1977, , 33, 515
, R. J. 2008, Central Bureau Electronic Telegrams, 1638, 2
, C., & [Chevalier]{}, R. A. 1987, , 322, L15
, C., [Lundqvist]{}, P., & [Chevalier]{}, R. A. 1996, , 461, 993
, M., et al. 2010, ArXiv e-prints, 1011.6558
, S., et al. 2008, , 683, L131
, E. K., [Imshennik]{}, V. S., & [Nadyozhin]{}, D. K. 1971, , 10, 28
, M. 2003, , 582, 905
, M. 2005, in IAU Colloq. 192: Cosmic Explosions, On the 10th Anniversary of SN1993J, ed. [J.-M. Marcaide & K. W. Weiler]{}, 535
, M., & [Pinto]{}, P. A. 2002, , 566, L63
, M., [Suntzeff]{}, N. B., [Heathcote]{}, S. R., [Walker]{}, A. R., [Gigoux]{}, P., & [Phillips]{}, M. M. 1994, , 106, 566
, M. A. 2001, Ph.D. thesis, The University of Arizona
, A., [Fryer]{}, C. L., [Woosley]{}, S. E., [Langer]{}, N., & [Hartmann]{}, D. H. 2003, , 591, 288
, M. A., et al. 2005, , 359, 906
, S., et al. 2007, , 664, 435
, A. U. 2009, , 137, 4186
, D. C., et al. 2002, , 114, 35
, I. Y., & [Nadezhin]{}, D. K. 1985, Soviet Astronomy Letters, 11, 145
, K., et al. 2010, , 404, 981
, J. R., & [Smartt]{}, S. J. 2005, , 360, 288
, J. R., [Smartt]{}, S. J., & [Danziger]{}, I. J. 2005, , 364, L33
, K., [Pooley]{}, D., [Chandra]{}, P., [Bhattacharya]{}, D., [Ray]{}, A. K., [Sagar]{}, R., & [Lewin]{}, W. H. G. 2007, , 381, 280
, D. K. 2003, , 346, 97
, S., [Kadota]{}, K., & [Wells]{}, W. 2008, Central Bureau Electronic Telegrams, 1636, 1
, F., et al. 2010, , 715, 833
, A., et al. 2005, , 360, 950
, A., et al. 2006, , 370, 1752
, A., et al. 2009, , 394, 2266
, A., et al. 2004, , 347, 74
, L. S., [V[í]{}lchez]{}, J. M., & [Contini]{}, T. 2004, , 425, 849
, T. S., et al. 2008, , 383, 627
Press, W. H., Teukolsky, S. A., Vetterling, W. T. and Flannery, B. P. 1992, [Numerical recipes in C. The art of scientific computing]{}, ed. [Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. ]{}
, R. M., [Wheeler]{}, J. C., [H[ö]{}flich]{}, P., [Akerlof]{}, C. W., [Brown]{}, P. J., & [Rykoff]{}, E. S. 2007, , 666, 1093
, M. W., [Treffers]{}, R. R., [Filippenko]{}, A. V., [Paik]{}, Y., [Leibundgut]{}, B., [Schulman]{}, E., & [Cox]{}, C. V. 1994, , 107, 1022
, P. W. A., et al. 2005, , 120, 95
, R., et al. 2011, , 414, 167
, D. K., [Anupama]{}, G. C., [Srividya]{}, S., & [Muneer]{}, S. 2006, , 372, 1315
, K., et al. 2008, Science, 321, 223
, D. J., [Finkbeiner]{}, D. P., & [Davis]{}, M. 1998, , 500, 525
, S. J., [Eldridge]{}, J. J., [Crockett]{}, R. M., & [Maund]{}, J. R. 2009, , 395, 1409
, N., [Li]{}, W., [Filippenko]{}, A. V., & [Chornock]{}, R. 2011, , 412, 1522
, P. B. 1987, , 99, 191
, C. J., et al. 2008, The Astronomer’s Telegram, 1883, 1
, C. J., et al. 2009, The Astronomer’s Telegram, 1912, 1
, M. 2008, Central Bureau Electronic Telegrams, 1638, 3
, N. B., & [Bouchet]{}, P. 1990, , 99, 650
, J. N., [Paturel]{}, G., & [Ekholm]{}, T. 2002, , 393, 57
, N., [Blinnikov]{}, S., [Baklanov]{}, P., [Morokuma]{}, T., [Nomoto]{}, K., & [Suzuki]{}, T. 2009, , 705, L10
, M., [Benetti]{}, S., & [Cappellaro]{}, E. 2003, in From Twilight to Highlight: The Physics of Supernovae, ed. [W. Hillebrandt & B. Leibundgut]{}, 200
, M., et al. 1998, , 498, L129
, V. P. 2007, , 461, 233
, V. P., & [Chugai]{}, N. N. 2008, , 491, 507
, V. P., [Chugai]{}, N. N., & [Botticella]{}, M. T. 2010, , 723, L89
, P. G. 2001, , 113, 1420
, S. D., et al. 2010, ArXiv e-prints, 1011.5873
, J., et al. 2006, , 369, 1780
, K., & [Eversberg]{}, T. 2006, Astronomische Nachrichten, 327, 862
, E., [M[é]{}sz[á]{}ros]{}, P., & [Campana]{}, S. 2007, , 667, 351
, K. W., [Panagia]{}, N., [Montes]{}, M. J., & [Sramek]{}, R. A. 2002, , 40, 387
, F. 2010, PhD thesis, University of Michigan, USA
, L., [Pastorello]{}, A., [Turatto]{}, M., [Cappellaro]{}, E., [Benetti]{}, S., [Altavilla]{}, G., [Mazzali]{}, P., & [Hamuy]{}, M. 2003, , 338, 711
, L., [Shapiro]{}, S. L., & [Colpi]{}, M. 1998, , 502, L149
, F. M., et al. 2004, in Presented at the Society of Photo-Optical Instrumentation Engineers (SPIE) Conference, Vol. 5492, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. [A. F. M. Moorwood & M. Iye]{}, 1590
[^1]: The Robotic Optical Transient Search Experiment (ROTSE-III) is a set of four 45 cm fully robotic optical telescopes, installed at Siding Spring Observatory, Australia (ROTSE-IIIa), McDonald Observatory, Texas (ROTSE-IIIb), H.E.S.S. site, Namibia (ROTSE-IIIc) and TUBITAK National Observatory, Turkey (ROTSE-IIId; @akmrs03 [@yuan_thesis]).
[^2]: A 2048 $\times$ 2048 CCD camera mounted at the f/13 Cassegrain focus of the telescope, has a square pixel of 24 micron on a side, and with a plate scale of 0.38 arcsec per pixel, it covers a square area of 13$\arcmin$ on a side in the sky. The gain and readout noise of the CCD camera are 10 electrons per analog-to-digital unit and 5.3 electrons, respectively. To improve the S/N ratio and optimize the sampling, the observations were carried out in a binned mode of 2$\times$2 pixel.
[^3]: IRAF stands for Image Reduction and Analysis Facility distributed by the National Optical Astronomy Observatory which is operated by the Association of Universities for research in Astronomy, Inc. under co-operative agreement with the National Science Foundation
[^4]: [*DAOPHOT*]{} stands for Dominion Astrophysical Observatory Photometry [@stetson87]
[^5]: The full names of the telescope are provided in Table \[tab:speclog\]
[^6]: http://leda.univ-lyon1.fr/
[^7]: This is the first time that the rapidly changing evolution of profiles during the late plateau to nebular phase has been so densely covered in a low-luminosity SN
[^8]: The cosmological model with $H_0$ = 70 Mpc$^{-1}$,$\Omega_{m}$ = 0.3 and $\Omega_{\Lambda}$ = 0.7 is assumed throughout the paper and the uncertainty corresponds to a local cosmic thermal velocity of 208 [@terry02].
[^9]: http://nedwww.ipac.caltech.edu/
[^10]: This is the mean value of the velocities computed from the lines of $\lambda$4924, $\lambda$5018 and $\lambda$5169 in the +54d spectrum.
[^11]: These observed values of $M_{Vmp}$ and $v_{mp}$ can be compared with the estimates derived empirically between the mass of $^{56}$Ni, and the $v_{mp}$ and $M_{Vmp}$ for a larger sample of IIP SNe (see Section 8.3 of @roy11). Using the $^{56}$Ni mass of $0.015\pm0.003$ we find $v_{mp} =2916\pm 220{\mbox{$\rm{\,km\,s^{-1}}$}}\,$ and $M_{Vmp} = -15.37\pm0.23$ mag which are consistent with those measured observationally.
[^12]: was not detected in X-ray. The lack of X-ray emission can be used to constrain the mass-loss rate of the progenitor system that could be heated by the outgoing shock to X-ray emitting temperatures. Following the discussion in @immler07 and references therein, an upper limit to the pre-SN mass-loss rate of $5 \times 10^{-6}~M_{\odot}~{\rm yr}^{-1}~(v_{\rm w}/10~{\rm km~s}^{-1})$ with an uncertainty of a factor of two to three is obtained.
[^13]: Considering a lower value of $A_V = 0.139\pm0.002$, we get $M_{Vmp} = -15.1\pm0.2$
[^14]: The measured value of the $E/M$ ratio for is $3.5\times\,10^{49}$erg $^{-1}$, which is twice the E/M ratio measured for SNe 1997D ($1.7\times\,10^{49}$erg $^{-1}$, @turatto98), but it is half those of SN 2008gz ($7.3\times\,10^{49}$erg $^{-1}$, @roy11), SN 2004dj ($7.3\times\,10^{49}$erg $^{-1}$, @venko06) and SN 2004et ($7.5\times\,10^{49}$erg $^{-1}$, @sahu06). Finally, it is barely a third the E/M ratios of SNe 1987A ($11.3\times\,10^{49}$erg $^{-1}$, @hamuy03) and 2003gd ($13.0\times\,10^{49}$erg $^{-1}$, @hendry05), and about seven times smaller than the luminous event like SN 1998A ($25.4\times\,10^{49}$erg $^{-1}$, @pastorello05).
[^15]: Derived from the \[O/H\]($= 12 + log(N_O/N_H)$)-$M_{B}$ relation given by @pilyugin04 for a given deprojected radius and galaxy type. The host of is a spiral galaxy of SBbc Type and the SN location is $\sim 1.8\arcmin$ away from the center of the host. This corresponds to a deprojected distance $\sim 7.03$ kpc.
[^16]: We have taken the \[O/H\] value from @smartt09, though for a few cases the values are estimated using the \[O/H\]$-M_B$ relation mentioned in @pilyugin04. The \[O/H\] for normal Type IIP SNe 2008gz, 1999em and low-luminosity SNe 1999eu, 2005cs is about 8.6. For the low-luminosity SNe 1994N, 1999br, 2001dc, 2003gd and 2004dj, 1997D and the normal SN 2004et, it is about 8.4. In many cases, we have adopted the \[O/H\] value from @smartt09. For some cases we have estimated the value by adopting the \[O/H\]$-M_B$ relation mentioned in @pilyugin04.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Exploring low-loss two-dimensional plasmon modes is considered central for achieving light manipulation at the nanoscale and applications in plasmonic science and technology. In this context, pump-probe spectroscopy is a powerful tool for investigating these collective modes and the corresponding energy transfer processes. Here, I present a first-principles study on non-equilibrium Dirac plasmon in graphene, wherein damping channels under ultrafast conditions are still not fully explored. The laser-induced blueshift of plasmon energy is explained in terms of thermal increase of the electron-hole pair concentration in the intraband channel. Interestingly, while damping pathways of the equilibrium graphene plasmon are entirely ruled by scatterings with acoustic phonons, the photoinduced plasmon predominantly transfers its energy to the strongly coupled hot optical phonons, which explains the experimentally-observed tenfold increase of the plasmon linewidth. The present study paves the way for an in-depth theoretical comprehension of plasmon temporal dynamics in novel two-dimensional systems and heterostructures.'
author:
- Dino Novko
bibliography:
- 'tdpl.bib'
title: Ultrafast Dynamics of Dirac Plasmon in Graphene
---
Understanding, and thus mastering, the temporal dynamics of charge carriers in graphene and related quasi-two-dimensional materials is pivotal, but highly challenging task in material science. Many recent studies were devoted to explore the time evolution of the laser-excited electrons by means of time-resolved photoemission[@bib:johannsen13; @bib:gierz13; @bib:stange15; @bib:tan17; @bib:rohde18] and pump-probe optical absorption spectroscopies[@bib:kampfrath05; @bib:sun08; @bib:winnerl11; @bib:jnawali13; @bib:jensen14; @bib:frenzel14; @bib:mics15; @bib:tomadin18] in graphene and graphite in order to reach the aforesaid goal. Precise time scales of ultrafast electron interactions were extracted, in particular, electron-electron scattering was shown to rule the dynamics below, while the coupling with the optical phonons (OP) above $\sim50$fs[@bib:kampfrath05; @bib:johannsen13; @bib:stange15; @bib:tan17; @bib:rohde18]. However, underlying microscopic processes still remain largely unexplored, mostly due to a lack of accompanying first-principles methodology that can quantitatively capture these features.
Photoinduced plasmon excitation, i.e., collective electron oscillations under highly nonequilibrium condition,[@bib:macdonald08; @bib:huber16] is one such ultrafast phenomena that requires further insights. In graphene and graphene-based heterostructures, two-dimensional plasmons show quite exceptional features, e.g., electrical tunability[@bib:ju11; @bib:fei12] and low losses[@bib:yan13; @bib:woessner15; @bib:ni18], making these materials promising building blocks for optoelectronic and plasmonic devices. Recently, the relaxation dynamics of laser-induced graphene plasmon was monitored with unprecedented temporal and spatial resolution by using antenna-based near-field nanoscopy[@bib:wagner14; @bib:ni16]. High electron temperatures achieved in these experiments are increasing the energy of graphene plasmon while concurrently increasing (decreasing) its linewidth (lifetime)[@bib:ni16]. The former was explained in terms of increase of the Drude weight, or equivalently increase of thermally excited electron-hole pairs, with elevated electron temperature, while the origin of the latter remains unresolved. Since the relaxation of equilibrium graphene plasmon was shown to be governed by the electron-phonon coupling[@bib:principi14; @bib:woessner15; @bib:ni18; @bib:novko17], mainly coupling with graphene acoustic phonons (AP)[@bib:principi14; @bib:woessner15; @bib:ni18], it was speculated that the enhanced plasmon decay under non-equilibrium condition has the same origin[@bib:ni16]. However, projecting conclusions from the equilibrium situation might be premature, considering highly disparate thermal conditions in the two cases, but also having in mind the results extracted from time-resolved photoemission and optical absorption experiments where OP were proven to play a key role in relaxation processes[@bib:kampfrath05; @bib:johannsen13; @bib:stange15].
In this Letter, I investigate the dynamics of laser-excited plasmon in lightly-doped graphene under nonequilibrium conditions by means of the robust *ab initio* methodology. The work conjoins the electron-phonon coupling theory and the Coulomb screening in random phase approximation to capture the temperature-dependent plasmon decay due to phonons[@bib:shulga91; @bib:novko16; @bib:novko17], while the nonequilibrium electron and phonon temperatures are simulated within the effective temperature model[@bib:allen87; @bib:lin08; @bib:perfetti07; @bib:caruso19; @bib:novko20]. I show, in agreement with previous reports[@bib:principi14; @bib:woessner15; @bib:ni18], that graphene plasmon under equilibrium conditions (i.e., when electrons and phonons are thermalized) is predominantly decaying due to scattering with the AP. In particular, the obtained temperature dependence of the plasmon decay rate due to coupling with AP shows very good agreement with recent measurements done on high-mobility graphene[@bib:ni18]. However, the situation is drastically different for nonequilibrium conditions, where the majority of the laser-induced excess electron energy is transferred to the strongly coupled OP, creating hot phonon bath[@bib:kampfrath05]. In this case, the results show that the low-energy plasmons (i.e., at $\sim 0.1$eV), usually explored in the experiments[@bib:wagner14; @bib:ni16; @bib:ni18], are mostly coupled to the hot OP, which is in contrast to the current belief[@bib:ni16]. The latter interaction is consequently responsible for the large time-dependent modifications of plasmon broadening. In addition, the laser excitation increases the phase space for the interband transitions (i.e., Landau damping), which in turn enhance the plasmon decay rate at higher energies (i.e., around $0.2$eV). Finally, the experimentally-observed ultrafast blueshift of plasmon energy is shown to be induced by the transient increase of the electron-hole pair exctations in the intraband channel (i.e., increase of the Drude weight). All in all, I believe that the *ab initio* methodology and conclusions outlined here will be useful not only for comprehending the interplay of plasmon and phonon dynamics in graphene, but as well in other similar systems hosting a two-dimensional plasmon[@bib:sim15; @bib:in18].
 Electron and (b) phonon band structures of lightly hole-doped graphene. Dashed red line in (a) is the Fermi level, while the shaded brown area schematically represents the skewed electron distribution for finite electron temperature. Green and black dashed lines in panel (b) are Eliashberg function and phonon density of states, respectively. Red and blue colors depict the Bose-Einstein distribution of phonon modes at finite phonon temperature. (c) Plasmon dispersion in graphene for Fermi energy $\varepsilon_F=-250$meV (blue circles). Blue and light blue regions are the intraband and interband excitations regions, respectively. The phase space where plasmon couples only to acoustic (light orange) as well as both acoustic and optical (orange) phonon modes is shown as well. ](fig1.pdf){width="48.00000%"}
Ultrafast electron dynamics is explored here for the hole-doped graphene where the Fermi energy is $\varepsilon_F=-250$meV, as it is the case, e.g., for graphene adsorbed on SiC surface[@bib:johannsen13] (note that the conclusions of the paper would be the same for lightly electron-doped graphene as in the graphene/SiO$_2$ system[@bib:ni16] due to electron-hole symmetry in the low-energy region of band structure). The corresponding electron and phonon band structures are shown in Figs.\[fig:fig1\](a) and \[fig:fig1\](b) (see SI for computational details). Additionally, Fig.\[fig:fig1\](b) shows the results for the Eliashberg function that measures the degree of the electron-phonon coupling with energy resolution. As is well know[@bib:piscanec04], electrons are strongly coupled to OP at K and $\Gamma$ point of the Brillouin zone, i.e., $\omega\gtrsim 0.16$eV where Eliashberg function shows two prominent peaks, while only weakly coupled to the rest of the modes (mostly low-energy AP). Accordingly, the plasmon excitations below $0.16$eV are weakly coupled to AP[@bib:principi14], while at higher energies plasmon decay is mostly due to the strong coupling with OP[@bib:novko17]. The graphene plasmon dispersion $\omega_{\rm pl}$ and the corresponding decay regions are shown in Fig.\[fig:fig1\](c).
 Energy renormalization $1+\lambda_{\rm ep}(\omega)$ and (b) decay rate $\gamma_{\rm ep}(\omega)$ of the electron-hole pairs due to coupling with phonon as a function of frequency and temperature. The energy windows of acoustic and optical phonons are highlighted with light orange and orange areas, respectively. (c) Temperature dependence of $\gamma_{\rm ep}(\omega_{\rm pl})$ when plasmon energy is $\omega_{\rm pl}=110$meV (blue line). The contributions coming from acoustic and optical phonons are shown with light orange and orange lines. The experimentally determined plasmon decay rate for graphene encapsulated in hexagonal boron nitride (green circles) is shown for comparison[@bib:ni18]. The corresponding decay rate without the influence of the boron nitride is shown with blue circles. ](fig2.pdf){width="44.00000%"}
Figures \[fig:fig2\](a) and \[fig:fig2\](b) further depict the energy renormalization $1+\lambda_{\rm ep}(\omega)$ and decay rate $\gamma_{\rm ep}(\omega)$, respectively, of the electron-hole pairs due to coupling with phonons as a function of temperature[@bib:allen71; @bib:allen74; @bib:shulga91; @bib:kupcic15]. These quantities are computed by using density functional and denstiy functional perturbation theories[@bib:baroni01; @bib:giustino17] as in Refs.[@bib:novko17; @bib:novko18; @bib:novko20] (see SI for computational details). The Fermi energy is here chosen to be $\varepsilon_F=300$meV for the sake of comparison with the experiment[@bib:ni18] (the rest of the results are for $\varepsilon_F=-250$meV as mentioned earlier). For plasmon energies, i.e., $\omega=\omega_{\rm pl}$, the quantities $1+\lambda_{\rm ep}(\omega_{\rm pl})$ and $\gamma_{\rm ep}(\omega_{\rm pl})$ are equivalent to the plasmon energy renormalization and plasmon decay rate due to coupling with phonons[@bib:novko17]. The results show that the plasmon energy is insignificantly renormalized due to electron-phonon coupling, with very small temperature modifications. The corresponding plasmon broadening is as well small (especially for $\omega_{\rm pl}<0.16$eV, where electrons mostly couple to AP) but notable. For the temperature range presented here, the temperature-induced change in plasmon broadening is more pronounced for the lower energies since AP are more easily excited with temperature than the high-energy OP (i.e., $\omega_{\rm OP}\gg k_B T$). Figure \[fig:fig2\](c) shows the temperature dependence of the plasmon decay rate when $\omega_{\rm pl}= 110$meV. The results are in very good agreement with the decay rate extracted from the recent experiment done on high-mobility graphene[@bib:ni18]. Further analysis demonstrates that the decay rate of the equilibrium graphene plasmon and its temperature dependence predominantly comes from coupling with AP, in agreement with the previous studies[@bib:principi14; @bib:woessner15; @bib:ni18].
I turn now to the study of the nonequilibrium condition, i.e., of the ultrafast electron dynamics in graphene. In order to simulate the laser-induced electron dynamics I utilize the three temperature model[@bib:perfetti07] with *ab initio* input parameters (see Ref.[@bib:caruso19] and SI for more details). Within this model, the temperature of graphene is divided among three subsystems, i.e., electron temperature ($T_e$), temperature of the strongly coupled OP ($T_{\rm OP}$), and the remnant temperature that mostly belongs to AP ($T_{\rm AP}$). The energy flow between electron and phonon degrees of freedom is then dictated by the electron-phonon coupling[@bib:allen87], while the thermalization between two phonon subsystems goes via anharmonic coupling. The resultant time evolution of these temperatures is shown in Fig.\[fig:fig3\](a), where the laser with fluence of $F=8$J/m$^2$ excites the system at time delay $t_d=0$. Right after the laser excitation, the electron temperature $T_e$ abruptly increases up to almost 3000K and subsequently decays due to coupling with OP, in good agreement with the experiment[@bib:gierz13]. Consequently, the temperature of the strongly coupled OP elevates above 1000K, while the AP remain almost at the same temperature. Since the energy exchange rate between phonon and electron baths is proportional to $\lambda\left\langle \omega^2 \right\rangle / C_{\rm ph}$ (where $\lambda$ is the electron-phonon coupling strength, $\left\langle \omega^2 \right\rangle$ is the second moment of the phonon spectrum, and $C_{\rm ph}$ is the heat capacity of the relevant phonon subsystem)[@bib:allen87], the obtained dramatic difference between the OP and AP temperatures comes not only because the OP are coupled more strongly to the electrons than the AP ($\lambda_{\rm OP}\left\langle \omega^2 \right\rangle_{\rm OP}>\lambda_{\rm AP}\left\langle \omega^2 \right\rangle_{\rm AP}$), but also because the strongly coupled OP subsystem consist of only very few modes around $\Gamma$ and K points of the Brillouin zone \[see Fig.\[fig:fig1\](b)\] and thus $C_{\rm OP}\ll C_{\rm AP}$. Such laser-induced hot OP scenario was already discussed in various spectroscopy studies[@bib:kampfrath05; @bib:johannsen13; @bib:stange15].
 Time evolution of electron temperature $T_e$, temperature of the strongly coupled optical phonons $T_{\rm OP}$, and temperature of the remnant phonon, mostly acoustic, modes $T_{\rm AP}$. The laser with the fluence of $8$J/m$^2$ excites the system at the zero delay time. The extracted experimental results for $T_e$[@bib:gierz13] are shown with blue circles. (b) The corresponding photoinduced charge density modifications for intraband and interband channels. Time dependence of (c) electron-phonon decay rate changes $\delta\gamma_{\rm ep}$ and (d) energy renormalization parameter $1+\lambda_{\rm ep}$ for three different values of excitation energy $\omega$. (e) Transient optical absorption $\sigma_1(\omega)$ for three different values of time delay. Inset: The corresponding low-energy part (Drude peak). (f) Time evolution of photoconductivity $\Delta\sigma_1(\omega)$ for excitation energies below, around, and above the interband threshold $2|\varepsilon_F|$. ](fig3.pdf){width="48.00000%"}
Figure \[fig:fig3\](b) shows the ensuing charge density modifications $\delta n$ as a function of time delay. As laser excites the system, electron-hole pair concentration increases both in intraband (electron and hole are in the same band) and interband (electron and hole are in different bands) channels. In fact, for the lightly-doped graphene with $\varepsilon_F=-250$meV it turns out that $\delta n_{\rm intra} \gg \delta n_{\rm inter}$. The laser-induced changes of the electron-phonon coupling are shown in Figs.\[fig:fig3\](c) and \[fig:fig3\](d). In particular, time-dependant modifications of the electron-phonon decay rate $\delta\gamma_{\rm ep}$ and energy renormalization parameter $1+\lambda_{\rm ep}$ are depicted for three different excitation energies $\omega$. Since $1+\lambda_{\rm ep}$ is already small for the equilibrium situation, it is not surprising that $1+\lambda_{\rm ep}$ experiences only minor changes as a function of pump-probe time delay and excitation energy. One can also note that laser excitation reduces the energy renormalization parameter as a function of time. On the other hand, for the same laser conditions, the photoinduced decay rate modifications $\delta\gamma_{\rm ep}$ are relatively high \[i.e., $\delta\gamma_{\rm ep}\gg \gamma_{\rm ep}(T=300\,{\rm K})$\] and actually follow the variations of both $T_e$ and $T_{\rm OP}$ (see the discussion below). Also, the overall intensity of $\delta\gamma_{\rm ep}$ over time is bigger for smaller excitation energies (i.e., for $\omega=0.1$eV and 0.2eV, compared to $\omega=0.4$eV). This is because for $\omega\lesssim 0.2$eV the equilibrium value of $\gamma_{\rm ep}$ is small and includes mostly the contributions from the weakly-coupled AP modes, while when both $T_e$ and $T_{\rm OP}$ are elevated the probability of scattering on the OP, which are strongly coupled to electrons, increases significantly. A more rapid increase of electron-OP scattering probability when $T\gtrsim 300$K can, for example, be seen in Fig.\[fig:fig2\](c). However, when $\omega\gg 0.2$eV the probability of scattering on the OP is less altered for the present laser conditions, since $\omega> T_e, T_{\rm OP}$. Therefore, the relative modification of damping rate is less pronounced for $\omega\gg 0.2$eV compared to $\omega\lesssim 0.2$eV.
The time dynamics of charge density and decay rate are relevant for comprehending the transient optical absorption, i.e., photoconductivity, as well as the relaxation mechanisms of nonequilibrium plasmons. Figure \[fig:fig3\](e) depicts the time evolution of optical absorption $\sigma_1(\omega)$ (i.e., the real part of optical conductivity $\sigma$) up to 0.7eV (see SI for the corresponding methodology). The observed modifications are due to photoinduced variations of $T_e$, and also due to changes in decay rate $\delta\gamma_{\rm ep}$ (note that $\sigma_1(0)\propto n/\gamma_{\rm ep}$ and $\sigma_1(\omega>0)\propto n\gamma_{\rm ep}/(\omega^2+\gamma_{\rm ep}^2)$[@bib:allen71; @bib:kupcic15; @bib:novko17]). In particular, the low-energy part of $\sigma_1(\omega)$ decreases significantly around $t_d=0$ due to changes in $\delta\gamma_{\rm ep}$ and then it starts to increase back to its equilibrium value \[see the inset in Fig\[fig:fig3\](e)\]. Contrary, for higher excitation energies up to around interband threshold $2|\varepsilon_F|$ the photoinduced modifications of $\sigma_1(\omega)$ are first increasing and then decreasing. The time evolution of photoconductivity $\Delta\sigma_1$ below, around, and above interband threshold $2|\varepsilon_F|$ is also depicted in Fig.\[fig:fig3\](f). In all these three energy regimes, the photoconductivity $\Delta\sigma_1$ shows different time behaviour, which is simply due to modifications of the interband onset when $T_e$ is altered.
Note that the experimentally observed negative and positive values of graphene photoconductivity over different values of $\omega$ are widely discussed and analyzed in literature[@bib:kampfrath05; @bib:sun08; @bib:winnerl11; @bib:jnawali13; @bib:jensen14; @bib:frenzel14; @bib:mics15; @bib:tomadin18]. Here the focus is more on the ultrafast plasmons and the corresponding dynamics in different energy regimes. The time modulations of graphene plasmon properties (e.g., energy loss and Drude weight) under optical pumping were in fact discussed recently[@bib:page16; @bib:hamm16; @bib:sun16; @bib:petersen17; @bib:wilson18], however, the detailed *ab-initio* study on the corresponding plasmon loss channels is still lacking.
![\[fig:fig4\] Modifications of plasmon (a) energy $\delta\omega_{\rm pl}$ and (b) linewidth $\delta\gamma_{\rm pl}$ as well as (c) the corresponding spectral function $S(q,\omega)$ as a function of pump-probe time delay $t_d$ when the plasmon wavevector is $q=3.4\cdot10^{-4}$a.u. Different contributions to $\delta\omega_{\rm pl}$ and $\delta\gamma_{\rm pl}$ are shown: intraband + interband excitations (blue), intraband excitations with the full electron-phonon coupling (orange), intraband excitations without changes in Drude weight $\delta n=0$ (green), intraband excitations when $\delta n=0$ and without electron-acoustic phonon (AP) scattering (purple), and intraband excitations when $\delta n=0$, without AP, and without electron temperature $T_e$ contribution to electron-optical phonon (OP) scattering rate (black). (d)-(f) Same as in (a)-(c) but for larger plasmon wavevector, i.e., $q=6.3\cdot10^{-3}$a.u. Note the different spectral-intensity and energy scales in (c) and (f) panels](fig4.pdf){width="48.00000%"}
Figures \[fig:fig4\](a)-(c) and \[fig:fig4\](d)-(f) show the variations in the plasmon energy $\delta\omega_{\rm pl}$, plasmon linewidth $\delta\gamma_{\rm pl}$, and the spectral function $S(q,\omega)$ as a function of time delay $t_d$ for two different values of wavevector $q$, i.e., for two different energy regimes (note that $\omega_{\rm pl}\propto \sqrt{q}$). For the corresponding computational methods see SI. The obtained time dynamics of the photoinduced graphene plasmon is in good agreement with the experimental observations[@bib:wagner14; @bib:ni16]. Namely, the results show that the graphene plasmon is blueshifted and broadened upon the laser excitation. Also, $\delta\omega_{\rm pl}$ and $\delta\gamma_{\rm pl}$ are notably more pronounced for larger plasmon energies $\omega_{\rm pl}$ (larger wavectors $q$). In order to understand these transient features of graphene plasmon, it is necessary to dissect different contributions to $\omega_{\rm pl}$ and $\gamma_{\rm pl}$. First of all, note that in general both intraband and interband excitations determine the total value of plasmon energy. Here, the laser-induced modifications of plasmon energy predominantly come from the increase of electron-hole pair concentrations in the intraband channel $n_{\rm intra}$ \[see Figs.\[fig:fig4\](a) and \[fig:fig4\](d)\]. In other words, laser elevates $T_e$, i.e., increases the number of thermally excited electrons in the intraband channel \[see also Fig.\[fig:fig3\](b)\], which in turn increases the Drude weight and thus the plasmon energy[@bib:wagner14; @bib:ni16]. As already discussed above, the renormalization of the plasmon energy due to electron-phonon coupling is insignificant \[see also Fig.\[fig:fig3\](d)\].
The processes underlying the plasmon decay rate $\delta\gamma_{\rm pl}$ as a function of time delay are a bit more complex than the processes ruling $\delta\omega_{\rm pl}$. For plasmon energies $\omega_{\rm pl}\approx 0.1$eV, it turns out that the increase of the plasmon broadening $\delta\gamma_{\rm pl}$ is entirely ruled by electron-phonon coupling. What is intriguing and actually at odds with the previous assumptions[@bib:ni16], is that the photoinduced plasmon decays mostly due to scatterings with the OP modes \[cf. green and purple dashed lines in Fig.\[fig:fig4\](b)\]. This is also in contrast with the decay mechanisms of the equilibrium plasmon, for which the scattering with the AP modes is the main loss channel \[see Fig.\[fig:fig2\](c)\]. The present analysis also shows that part of the electron-OP scattering contribution to the $\delta\gamma_{\rm pl}$ is induced by the elevated $T_e$ and part by the elevated $T_{\rm OP}$ \[cf. purple and black dashed lines in Fig.\[fig:fig4\](b)\]. Namely, the laser heats the electrons (i.e., elevates $T_e$), which in turn increases the electron phase space for the electron-OP scattering, especially when $T_e > \omega_{\rm OP}$. In addition, hot electrons transfer the excess energy to the strongly-coupled OP, i.e., $T_{\rm OP}$ rises substantially, which increases the number of thermally excited OP and therefore increases $\delta\gamma_{\rm pl}$. All in all, the results show that the plasmon-OP scattering is behind the experimentally-observed plasmon broadening under non-equilibrium conditions. The coupling with the OP modes is generally much more stronger than with the AP modes. However, for equilibrium case the energy conservation condition $\omega_{\rm pl} \gtrsim \omega_{\rm OP}$ must be met in order to activate this strong coupling. On the other hand, under the strong laser excitations, this energy conservation condition loosens up considerably and graphene plasmons with energy $\omega_{\rm pl} < \omega_{\rm OP}$ couple strongly with the hot OP modes. When the plasmon energy is more closer to the interband threshold $2|\varepsilon_F|$, i.e., $\omega_{\rm pl}\approx 0.2$eV \[see Fig.\[fig:fig4\](f)\], the decay is mostly due to the Landau damping (electron-hole pair interband excitations)[@bib:jensen91], which increases with $T_e$ \[see Figs.\[fig:fig3\](e) and \[fig:fig3\](f)\].
Temporal dynamics of non-equilibrium plasmon under intense laser excitation was explored in lightly-doped graphene by means of density functional and density functional perturbation theories. By considering plasmon-phonon coupling, decay rates of graphene Dirac plasmon were studied under both equilibrium (i.e., electron and nuclear degrees of freedom are thermalized) and ultrafast (i.e., electrons and phonons are thermally excited and have disparate energies) conditions. Due to available phase space and energy constraints, equilibrium plasmon with energy $\sim 0.1$eV and at ambient temperature is mostly damped due to scatterings with the acoustic phonon modes. Interestingly, the photoinduced non-equilibrium graphene plasmon is, on the other hand, underlain by completely different damping mechanism. Namely, the pump laser pulse increases the population of hot electrons, which in turn transfers the large portion of energy to the strongly coupled optical phonons, creating hot optical-phonon bath. Under such out-of-equilibrium condition phase space for electron scatterings and population of optical phonon modes are increased, as well as the energy constraints are loosened. Consequently, the broadening of the non-equlibrium plasmon is increased immediately after the laser excitation, mostly due to coupling with the hot optical phonons. The corresponding strong plasmon energy renormalization is explained in terms of photoinduced Drude weight increase. For energies $\sim 0.2$eV damping pathways of non-equilibrium plasmon are again different and are due to photoinduced interband excitations (Landau damping). Present study might also help elucidate energy-transfer mechanisms under optical pumping in novel quasi-two-dimensional materials that support collective plasmon modes, like metallic transition metal dichalcogenides[@bib:andersen13; @bib:dajornada20], topological insulators[@bib:sim15; @bib:in18], borophene[@bib:huang17], or layered electrides[@bib:druffel16; @bib:wang19].
The author acknowledges financial support from the Croatian Science Foundation (Grant no. UIP-2019-04-6869) and from the European Regional Development Fund for the “Center of Excellence for Advanced Materials and Sensing Devices” (Grant No. KK.01.1.1.01.0001). Financial support by Donostia International Physics Center (DIPC) during various stages of this work is also acknowledged. Computational resources were provided by the DIPC computing center.
| {
"pile_set_name": "ArXiv"
} |
[P.S. Howe and V. Stojevic]{}
[*Department of Mathematics, King’s College, London, UK*]{}
[**Abstract**]{}
In addition to superconformal symmetry, $(1,1)$ supersymmetric two-dimensional sigma models on special holonomy manifolds have extra symmetries that are in one-to-one correspondence with the covariantly constant forms on these manifolds. The superconformal algebras extended by these symmetries close as W-algebras, i.e. they have field-dependent structure functions. It is shown that it is not possible to write down cohomological equations for potential quantum anomalies when the structure functions are field-dependent. In order to do this it is necessary to linearise the algebras by treating composite currents as generators of additional symmetries. It is shown that all cases can be linearised in a finite number of steps, except for $G_2$ and $SU(3)$. Additional problems in the quantisation procedure are briefly discussed.
Introduction
============
There has been a long history of interplay between differential geometry and supersymmetric non-linear sigma models starting with the observation that $N=2$ supersymmetry in two dimensions requires the sigma model target space to be a Kähler manifold [@Zumino:1979et]. It was first pointed out in [@Delius:1989nc] that one could construct conserved currents in $(1,1)$ sigma models given a covariantly constant form on the target space, and in [@Odake:1988bh] it was shown that the $(1,1)$ model on a Calabi-Yau three-fold has an extended superconformal algebra involving precisely such a current constructed from the holomorphic three-form. In [@Howe:1991ic] symmetries of this type were studied systematically in the classical sigma model setting; each manifold on Berger’s list of irreducible non-symmetric Riemannian manifolds has one or more covariantly constant forms which give rise to conserved currents and the corresponding Poisson bracket algebras are non-linear, i.e. they are of W-symmetry type. Subsequently the properties of these algebras were studied more abstractly in a conformal field theory framework [@Shatashvili:1994zw; @Figueroa-O'Farrill:1996hm] and more recently in topological models [@deBoer:2005pt].
In this paper we revisit the symmetries of classical (1,1) supersymmetric non-linear sigma models with target spaces which admit torsion-free connections with special holonomy groups. The structure of the classical Poisson bracket algebra of currents associated with the covariantly constant forms and the supercurrent is investigated and it is shown that, in most cases, it can be linearised by the inclusion of a finite number of composite currents. The exceptional cases are $SU(3)$ and $G_2$, possibly the two cases of most interest in string theory. In these cases derivatives of the original currents are generated and the presence of these suggests that finite linearisations may not be possible.
The main motivation for studying these symmetries is as a preparation for trying to gain a better understanding of them at the quantum level. Such an understanding may be of use in the study of higher-order corrections in string theory in the sigma model context [@Gross:1986iv; @Grisaru:1986px; @Candelas:1986tz], a topic which has recently received renewed attention from the point of view of spacetime supersymmetry [@Lu:2004ng].
Since the symmetry transformations associated with covariantly constant forms of degree greater or equal to three are non-linear, even for a flat target space, one might anticipate that BRST techniques would be necessary in their analysis, and since the algebras only close in a field-dependent way one would also expect that the BV version might be helpful. The idea would be to use these techniques in the context of the algebraic renormalisation programme [@Piguet:1995er] in order to study possible anomalies in a cohomological framework [@Howe:1990pz]. However, it turns out that this formalism is not sufficient to deal with the problems we are mainly interested in, namely the sigma model either by itself or in the presence of external gauge fields. This motivates the search for linearised extensions of the classical special holonomy W-algebras.
The symmetry transformations associated with covariantly constant forms have the property that their parameters are chiral in the sense that they depend on half of the worldsheet superspace coordinates (see (\[2.7\])). The BRST transformation of the matter field $X^i$ is[^1],
$$\label{eq:BRST_matter} s X^i = c^A R_A{}^i(X)$$
where $c^A$ are the parameter ghosts. In order that the BRST operator $s$ is nilpotent on $X^i$ we would like the transformation of the ghosts to take the standard form
$$\label{eq:BRST_ghost} \it{s} c^A = \half c^C c^B f_{BC}{}^A(X) \ ,$$
where $f_{BC}{}^A(X)$ are the field-dependent structure functions defined by $[R_B,R_C]=f_{BC}{}^A
R_A$, the $R_A$s being regarded as vector fields on the space of sigma model fields. However, this is not consistent because one is transforming chiral objects into non-chiral ones. The structure functions depend on the currents and are only chiral on-shell. This then prevents one from writing down a classical master equation in the chiral, or superconformal, theory.
One can circumvent this problem by gauging the algebra in which case the ghosts are no longer taken to be chiral. In the context of a $d=2$ superconformally invariant action $S_0(X)$, for a single chiral sector,[^2] this involves the modification of the original action $S_0$ by
$$\label{eq:gauging} S_0(X) \rightarrow S_0(X) + h^A j_A(X) \ ,$$
where $j_A$ are the conserved currents, and the gauge fields $h^A$ are required to transform as
$$\label{eq:BRST_gauge}
s h^A = D_{-} c^A + h^C c^B f_{BC}{}^A(X) \ .$$
Two further problems now present themselves. The first is that the algebra of transformations may fail to close on the gauge fields and the ghosts due to relations between the currents which show up in the Jacobi identities. This is discussed in more detail in section 4. Even when this is not a problem the BRST transformations (\[eq:BRST\_matter\], \[eq:BRST\_ghost\], \[eq:BRST\_gauge\]) involve the background fields, $c^A$ and $h^A$, transforming into expressions containing quantum fields. It is possible to construct a solution to the classical master equation, but the naive Ward identities involve insertions of composite operators [@Bastianelli:1991yk]. It is not known how to analyse anomalies using cohomological techniques in this situation and it is therefore difficult, if not impossible, to carry through the algebraic renormalisation programme order by order in perturbation theory. On the other hand, a perturbative evaluation of the explicit non-local expressions involving composite operators is hopelessly difficult in the context of a generic special holonomy sigma model.
As far as we have been able to ascertain there seems to be no way out of this apart from linearisation. For the cases where we can establish finite linearisations at the classical level it should be possible to analyse the renormalisation of the symmetries of the effective action itself in perturbation theory. We discuss this further in the final section. However, we note that in order to analyse anomalies in the current algebra, or in the operator product expansion, sources for the currents need to be introduced which implies that the gauging (\[eq:gauging\]) is necessary. So even though writing down OPE expressions involving composite currents is commonplace in the abstract CFT, for special holonomy sigma models it seems that the OPE can really only be understood explicitly in pertubative renormalisation theory if the algebra can be linearised. Moreover, as we shall discuss later, some of the finite linearisations turn out to be unstable in the quantum theory due to the fact that operators which do not appear in the classical Poisson bracket algebra can be generated by the OPE.
Review of basics
================
The action for a $(1,1)$-supersymmetric sigma model without boundary is
$$S=\int\, dz\, g_{ij} D_+ X^i D_- X^j\ , \label{2.1}$$
where $g_{ij}$ is a Riemannian metric on the $n$-dimensional target space $M$. $X^i, i=1,\ldots n$, is the sigma model field represented in some local chart for $M$ and $z$ denotes the coordinates of $(1,1)$ superspace $\S$. We shall use a light-cone basis so that $z=(x^{++},x^{--},\th^{+},\th^{-})$, with $x^{++}=x^0+x^1, x^{--}=x^0- x^1$. $D_+$ and $D_-$ are the usual flat superspace covariant derivatives which obey the relations
$$D_+^2 = i\del_{++};\qquad D_-^2 = i\del_{--};\qquad \{D_+,D_-\}= 0 \ . \label{2.3}$$
We use the convention that $\del_{++} x^{++}=1$. We shall take the superspace measure to be
$$dz:= d^2 x\,D_+ D_- \label{2.4}$$
with the understanding that the superfield obtained after integrating over the odd variables (i.e after applying $D_+ D_-$ to the integrand) is to be evaluated at $\th=0$.
The action [(\[2.1\])]{} is invariant under superconformal transformations which act independently on the left (+) and right (-) light-cone sectors. In the left sector, a superconformal transformation takes the form
$$\d X^i= 2a_{--}\del_{++} X^i -i D_+ a_{--} D_+ X^i\ , \label{2.7}$$
where the parameter $a_{--}$ is chiral, $D_-a_{--}=0$. The corresponding supercurrent is the energy-momentum tensor
$$T_{+3}:=g_{ij} \del_{++}X^i D_+X^j\ , \label{2.8}$$
The current is conserved in the sense that $D_-T_{+3}=0$ on-shell. Similarly, there is a conserved energy-momentum tensor $T_{-3}$ in the right sector.
We shall say that the target space has special holonomy if there are one or more covariantly constant forms which reduce the corresponding holonomy groups from $SO(n)$ to two groups $G$ on Berger’s list. These are: $U(m)$ and $SU(m)$ for $n=2m$; $Sp(k)$ and $Sp(k)\cdot Sp(1)$ for $n=4k$; $G_2$ and $Spin(7)$.
Let $L$ be a vector-valued $l$-form such that the $l+1$-form obtained by lowering the vector index (taken to be in the first slot) is covariantly constant; this form will also be denoted $L$. (It should be clear from the context which is meant). The symmetry transformation associated with $L$ is
$$\d_L X^i=a_L L^i{}_L D_+ X^L \label{2.10}$$
where the parameter $a_L$ has Lorentz weight $-l$ and Grassmann parity $(-1)^l$ and the multi-index $L$ denotes $l$ antisymmetrised indices, $L:=[l_1\ldots l_l]$. We shall use the notation $L_2$ to denote antisymmetrisation over the $l-1$ indices beginning with $l_2$, and so on. Under a general variation of the field $X$ the change in the action is
$$\begin{aligned}
\d S&=&\int\,dz\, 2g_{ij}\d X^i \nab_- D_+ X^j \nn {\\[0.1cm]} &=&-\int\,dz\, 2g_{ij}\d X^i g_{ij} \nab_+
D_- X^j \ . \label{2.11}\end{aligned}$$
If we substitute [(\[2.10\])]{} into the top line of [(\[2.11\])]{} we see that $\d S=0$ provided that the parameter is chiral, $D_-a_L=0$. The corresponding conserved current will also be denoted by $L$; it satisfies $D_- L=0$ on-shell and is given by
$$L=\frac{1}{l+1} L_{iL} D_+ X^{iL}\ . \label{2.12}$$
In order to evaluate the commutator, of two such transformations one needs some algebraic relations which can be proved for any special holonomy forms. If we set
$$\label{2.12.1}
(L\cdot M)_{iL_2,jM_2}:=L_{ki L_2} M^k{}_{j M_2 }\ ,$$
then one can verify that
$$\begin{aligned}
(L\cdot M)_{i[L_2,jM_2]} &=& (-1)^{l+1}P_{ijL_2 M_2} +\frac{m}{2}g_{i[j}Q_{L_2
M_2]}\ ,\nn{\\[0.1cm]}
(L\cdot M)_{[jL_2,|i|M_2]} &=& (-1)^{l}P_{ijL_2 M_2} +\frac{l}{2}g_{i[j}Q_{L_2
M_2]}\ ,\nn{\\[0.1cm]}
(L\cdot M)_{i[L_2,|j| M_2]} + (i\leftrightarrow j)&=& g_{ij} Q_{L_2
M_2}-(l+m-2) g_{(i[l_2} Q_{j)L_3 M_2]}\ .
\la{3.11}\end{aligned}$$
The tensors $P$ and $Q$ are totally antisymmetric and covariantly constant; in particular cases they can vanish. Both of them can be used to define $L$-type symmetry transformations, but in the commutator of two special holonomy transformations, $[\d_L,\d_M]$, $Q$ is combined with the energy-momentum tensor. After some algebra one finds that
$$[\d_L,\d_M]X^i =\d_P X^i + \d_K X^i\ , \label{3.12}$$
where each term is now a symmetry by itself. The $P$ transformation, which is of standard $L$-type has parameter $a_P$ given by
a\_P=(-1)\^[l+1]{} m a\_M D a\_L -(-1)\^m l D a\_M a\_L .
The $K$ transformation is defined as follows. If we set
K\_[i,K]{}:=g\_[i\[k\_1]{} Q\_[K\_2\]]{} ,
where the multi-index $K$ takes on $l+m-1$ values, then it is not difficult to show (for any covariantly constant antisymmetric tensor $Q$) that
\_K X\^i=( a\_K K\^[i]{}\_[jK\_2]{} \_[++]{} X\^j D\_+ X\^[K\_2]{}+ K\^i\_K \_+ (a\_K D\_+ X\^K) )
is a symmetry of the action [(\[2.1\])]{}. In fact, the corresponding conserved quantity is the composite current $T Q$. For the case in hand the parameter $a_K$ is
a\_K=i(-1)\^[l+1]{}a\_M a\_L .
Poisson bracket algebras
========================
In this section we re-examine the algebra of symmetry transformations for the torsion-free model. The idea is to try to linearise the W-type algebraic structure by treating any composite currents as new independent generators. We shall see that for $SU(m), m\geq 4$, $Sp(k)\cdot Sp(1)$ and for $Spin(7)$ it turns out to be rather simple to do this by including a small number of extra generators. Since the $Sp(k)$ case is linear anyway ($N=4$ superconformal symmetry), this only leaves two cases which cannot be linearised straightforwardly, namely $SU(3)$ and $G_2$. The problem here is that derivatives of the original currents turn up and this interferes with the finiteness which is otherwise due to the fact that differential forms only have finite degree.
The subject is best studied using Poisson brackets; these were introduced in [@Howe:1991ic]. These brackets are based on the observation that $(1,1)$ superspace factorises, $\S=\S^\pl\xz\S^\mi$, so that we can view $z^\mi:=(x^{--},\th^-)\in \S^\mi$ as the super-time, while the other coordinates $z^\pl:=(x^{++},\th^+)\in\S^\pl$ are spatial coordinates on which the fields depend. On-shell the currents depend only on the latter as they are conserved in super-time. In the following discussion the minus coordinates are irrelevant, so that we can drop the pluses from the formulae without loss of clarity. In this section, therefore, $D$ will denote $D_+$ while $\del$ denotes $\del_{++}$, with $D^2=i\del$.
The basic Poisson bracket (PB) is
(DX\^i(1),DX\^j(2)) = g\^[ij]{} \_1 \_[12]{}
where $(1,2)$ refer to two different points in $\S^\pl$, $\nab=\nab_+$ and $\d_{12}$ is the delta-function in $\S^\pl$ which, as there is only one odd coordinate, is Grassmann odd. As all of the tensors appearing in the currents are covariantly constant, the covariant derivative in the basic PB can be replaced by the ordinary derivative, and the tensors can be regarded as constants. With this being understood one can write [(\[4.1\])]{} and its corollaries as
$$\begin{aligned}
(DX^i(1),DX^j(2)) &=& g^{ij} D_1 \d_{12} \nn {\\[0.1cm]}
(\del X^i(1),DX^j(2) &=& g^{ij} \del_1 \d_{12} \nn {\\[0.1cm]}
(DX^i(1),\del X^j(2) &=& -g^{ij}\del_1\d_{12} \nn{\\[0.1cm]}
(\del X^i(1),\del X^j(2)) &=& ig^{ij}\del_1 D_1\d_{12} \ .
\la{4.2}\end{aligned}$$
In the following we shall write $j(a)$ to mean a smeared current. For each current the parameter $a$ has the opposite Grassmann parity, so we have
j(a)=dz j(z) a(z)=dz a(z)j(z) ,
where $z$ now denotes $z^\pl$.
For any three currents $A,B,C$ and parameters $f,g$ we have
$$\begin{aligned}
(A(f),BC(g)) &=& (A(f),B(Cg))+(-1)^{BC}(A(f),C(Bg)) \nn{\\[0.1cm]}
(AB(f),C(g)) &=& (B(fA),C(g))+(-1)^{AB}(A(fB),C(g))\ .
\la{4.4}\end{aligned}$$
The three basic PBs are the superconformal algebra,
(T(a),T(b))=T(2(a b-ab) +iDa Db) ,
the PB of the supercurrent with an $L$-current $L$,
(T(a),L(b))=L(la b-2ab +iDa Db) ,
and the PB of two currents, $L,\ M $,
(L(a\_L),M(a\_M))=-P(a\_P) - T Q(a\_K) ,
where $a_P$ and $a_K$ are defined in [(\[3.13\])]{}and [(\[3.20\])]{} respectively. The programme now is to compute the PBs of the composite $T Q$ with all of the other currents including itself. It is easier to do this explicitly case by case (recall that $dim\, M=n$).
$G=U(m);\ n=2m$ {#gum-n2m .unnumbered}
---------------
When $G=U(m)$ there is one extra current associated with the complex structure $J$ and the algebra is just the $N=2$ superconformal algebra. In the present notation this is, in addition to the $N=1$ PB $(T,T)$,
(T(a),J(b))&=& J(a b-2ab + iDa Db) ,1 (J(a),J(b))&=& -iT(ab) ,
where the current is $J=\frac{1}{2}J_{ij} DX^{ij}$. The pair $(J,T)$ together make up the $N=2$ supercurrent which can be viewed as a real $N=2$ superfield. There are further $N=2$ multiplets given by pairs of the form $(J^p,TJ^{p-1})$. The PB algebra generated by these currents closes,
$$\begin{aligned}
(J^p(a),J^q(b)) &=& -ipq T J^{p+q-2}(ab)\nn {\\[0.1cm]}
(TJ^p(a),J^q(b)) &=& \frac{J^{p+q}}{p+q}(q(2q-1)\del a b-2q(p+1)a\del b +iq Da Db)\nn
{\\[0.1cm]}
(TJ^p(a),T J^q(b)) &=& TJ^{p+q}((2q+2)\del a b-(2p+2) a \del b +
iDa Db)\ .
\la{4.8.1}\end{aligned}$$
$G=Sp(k);\ n=4k$ {#gspk-n4k .unnumbered}
----------------
When $G=Sp(k)$, so that $M$ is a hyperKähler manifold, we have three complex structures $\{J_r\},\ r=1,2,3$ giving rise to the $N=4$ superconformal algebra with
(J\_r(a),J\_s(b))=-i\_[rs]{} T(ab) +\_[rst]{} J\_t(Da b + a Db) .
$G=Sp(k)\cdot Sp(1);\ n=4k$ {#gspkcdot-sp1-n4k .unnumbered}
---------------------------
The holonomy groups $Sp(k)\cdot Sp(1)$, which correspond to quaternionic Kähler geometries in $4k$ dimensions, give rise to W-type algebras which admit finite linearisations. There is a set of three complex structures $\{J_r\}$ but they are not globally defined on the target space. This means that one cannot define three additional supercurrents. However, there is a covariantly constant four-form $\o_L=\o_r\wedge \o_r$, where $\o_r$ is the local two-form corresponding to $J_r$. This gives rise to an $L$-type symmetry and hence we have an $N=1$ superconformal algebra extended by this current. The full set of currents is given by $\{L^p,TL^q;\ p=1,\ldots k;\
q=1\ldots k-1\}$.
The PB of two L-currents is
(L(a),L(b))=-4i TL(ab) .
Using this result and the PB of $T$ with $L$ one can verify straightforwardly that
$$\begin{aligned}
(L^p(a),L^q(b)) &=& -4ipq T L^{p+q-1}(ab)\nn {\\[0.1cm]}
(TL^p(a),L^q(b)) &=& \frac{L^{p+q}}{p+q}(q(4q-1)\del a b-2q(2p+1)a\del b +iq Da Db)\nn
{\\[0.1cm]}
(TL(a),T L(b)) &=& TL^{p+q}((4q+2)\del a b-(4p+2 a \del b +
iDa Db)\ .
\la{4.10.1}\end{aligned}$$
$G=SU(m);\ n=2m$ {#gsum-n2m .unnumbered}
----------------
When the holonomy group is $SU(m),m\geq 3$ the target space is a Calabi-Yau manifold. As well as a complex structure there is a covariantly constant $(m,0)$-form $\O$. We shall work with $L$ and $\hat L$ which are respectively the real and imaginary parts of $\O$. So the generating set of currents is $\{T,J,L,{\hat L}\}$, and we shall set $m=l+1$ to be in line with our previous conventions. $T$ and $J$ generate an $N=2$ superconformal algebra and the pair $\{L,\hL\}$ transform as a chiral $N=2$ superconformal field,
$$\begin{aligned}
(J(a),L(b))&=& \hL (lDa b + a Db) \nn{\\[0.1cm]}
(J(a),\hL(b))&=& -L (lDa b + a Db)\ .
\la{4.10.2}\end{aligned}$$
For $m$ even the PBs for the Ls are
$$\begin{aligned}
(L(a),L(b)) &=& -i l l! T J^{l-1} (ab) \nn{\\[0.1cm]}
(\hL(a),\hL(b)) &=& -i l l! T J^{l-1} (ab) \nn \\
(\hL(a),L(b))&=& l! J^{l-1}(Da b + a Db)\ ,
\la{4.10.3}\end{aligned}$$
while for $m$ odd they are
$$\begin{aligned}
(L(a),L(b)) &=& l! J^{l-1}(Da b- aDb)\nn{\\[0.1cm]}
(\hL(a),\hL(b)) &=& l! J^{l-1}(Da b- aDb)\nn{\\[0.1cm]}
(\hL(a),L(b))&=&-i l l!TJ^{l-1}(ab)\ .
\la{4.10.4}\end{aligned}$$
In both cases the new currents at this level are $K:=TJ^{m-2}$ and $M:=J^{m-1}$. As noted above this pair forms an $N=2$ supermultiplet of lowest spin $l$. From [(\[4.8.1\])]{} we can see that these currents have vanishing PBs with themselves unless $m=3$ in which case $(TJ,J^2)\sim J^3$. Moreover, the commutators of powers of $J$ and their products with $T$ with $L$ and $\hL$ are mostly zero. One has
(J\^p(a),L(b))= -2i\_[p2]{}TL(ab) ,
which comes about using the fact that $$\label{4.11.1} JL=0 \ ,$$ as $\o$ is a (1,1) form while $\o_L$ is the sum of $(m,0)$ and $(0,m)$ parts, and
(TJ\^p(a),L(b))=\_[p1]{}(J L + DT )(lab) -T(a Db) .
In order to show this one has to use the identity
DJL=iT .
We therefore see that only $TJ$ and $J^2$ have non-trivial PBs with $L$ and $\hL$, and $J^2$ only produces $J^3$. At the classical level, this pair is only generated for $m=3$ and so we conclude that $SU(m)$ holonomy algebras have finite linearisations for $m\geq 4$. For $m=3$, however, there is a new operator which involves a derivative.
$Spin(7)$ {#spin7 .unnumbered}
---------
The $Spin(7)$ case is similar to $Sp(k)\cdot Sp(1)$. This is an $N=1$ superconformal algebra extended by a superfield current $L$ of weight $2$. The invariant form is the self-dual four-form $\F$. The Poisson bracket of $L$ with itself gives rise to the composite current $K=TL$, and the PB of this with $L$ gives $L^2$ which is simply the current associated with the volume form. The new composites are superconformal fields and have vanishing PBs with $L$ and each other. Explicitly we have
(T(a), L(b))&=& L(3a b -2a b + iDa Db) 1 (L(a),L(b))&=& 9i TL(Da b)1 (L(a),T L(b))&=& L\^2(3 a bab -D aDb) 1 (L(a),L\^2(b))&=& 01 (TL(a),L\^2(b))&=&0 .
This result depends only on the algebraic relations $T^2=L^3=TL^2=0$. In fact, this algebra differs from the $Sp(2)\cdot Sp(1)$ algebra only in the coefficients.
$G_2$ {#g_2 .unnumbered}
-----
The other exceptional special holonomy group is $G_2$ in seven dimensions. This has an invariant three-form $\vf$ and its dual is an invariant four-form; they can be combined to give the $Spin(7)$ four-form $\F$. However, in this case the PB algebra generated by $T$, $L$ (corresponding to $\vf$) and $M$ (corresponding to $*\vf$) leads to derivatives of the original currents and we are unable to conclude that there is a finite linearisation. In more detail, the basic PBs of the weight $3/2$ and $2$ currents $L$ and $M$ are
(L(a),L(b)&=& 2M(Da b-a Db) 1 (L(a),M(b)&=& -18i TL(ab)1 (M(a),M(b)&=&-24i TM(ab) .
The PBs of $L,M$ with $TL,TM$ then give
(L(a),TL(b))&=& TM(Da b-2a Db) + 2(L L+DT M)(ab) 1 (L(a),TM(b))&=& LM(6ab-2ab+ iDa Db) 1 (M(a),TL(b))&=& 0 1 (M(a),TM(b))&=& 0 .
To show these it is necessary to make use of the obvious algebraic identities, such as $L^2=0$, as well as the less obvious ones
DLL&=& TM 1 LDM&=&-D(LM);DL M=D(LM)1 LM&=&(LM); L M=(LM) \[eq:G2\_relations\]
The PBs of the bilnears are
(TL(a),TL(b))&=&-T(7L L +DT M)(ab)1 (TL,TM)&=&(TM,TM)=(TL,LM)=(TM,LM)=(LM,LM)=0 .
After a little algebra one can show that the derivative operator in $(L,TL)$ can be replaced by $A$ which is defined to be the primary part of $DLL+\frac{2}{3}DTM$, i.e. it transforms as a primary of weight $\frac{7}{2}$ under superconformal transformations and that the right-hand side of $(TL,TL)$ is proportional to $TA$. To this level, we therefore find that the only non-algebraic operator that occurs is $A$ together with $TA$. Unfortunately, the algebra does not close on this set and higher derivative operators are generated. It seems highly unlikely that there is a finite linearisation in this case, or for CY3 which is similar in structure.
Jacobi identities for W-algebras
================================
In the BRST/BV language the Jacobi identities (JIs) for the commutator are written as
$$\label{eq:jacobi_brst_problem}
Z^A R_A{}^i := c^D c^C c^B\left(f_{BC}{}^E f_{E D}{}^A-R_B{}^j \del_j f_{CD}{}^A\right) R_A{}^i
= 0 \ .$$
If $Z^A$ does not vanish, the JIs imply relations between the generators. This can also be seen in the JIs for the Poisson bracket. Given a particular field-dependent algebra one can think of these as abstract relations which hold independently of a particular representation.
In this section we investigate the reducibility relations that occur for special holonomy W-algebras. We will show that when $Z^A$ is not zero for the gauged chiral algebras (see (\[eq:gauging\])), it is not possible to solve the classical master equation without introducing further ghosts.
As noted in the introduction, linearisation is necessary to analyse current algebras in an interacting CFT. For linearised algebras $Z^A$ vanishes, but this new obstruction to be discussed here may be of interest in special circumstances when the composite operator expressions can be evaluated more easily, or in the context of W-strings, when all the currents are imposed as constraints and the $h^A$s are treated as quantum fields.
The part of the master equation linear in the ghost antifields, $c^*_A$, contains the term $c^*_A
Z^A$, so if $Z^A$ does not vanish there is a potential obstruction to solving it. It turns out that for the gauged chiral systems $Z^A$ is a function of the currents, and terms proportional to $h^*
c^*$ can be added to the BV action so that this part of the master equation is satisfied. Alternatively, the field dependent closure functions can be set to zero using appropriate terms proportional to $X^* h^*$.
The closure of the transformations acting on the gauge fields (\[eq:BRST\_gauge\]) involves the JIs, and when $Z^A \neq 0$ the algebra closes only modulo certain symmetries which act only on the gauge fields and which reflect the relations between the currents [@Lu:1994sc; @Thielemans:1995hn]. These will be called null symmetries. For example, for the $SU(3)$ case the null (BRST) symmetry, ($c^0$ is a parameter ghost),
$$\label{eq:null_symmetry} s h^J = -D(c^0 L) \ \ \ \ \ s h^{\widehat{L}} = i c^0 T \ \ \ \ \ \ s
X^i = 0 \ ,$$
reflects the relation (\[4.12.1\]):
$$\label{eq:example_null_relation} DJ L - iT \widehat{L} \equiv 0 \ .$$
The gauged chiral action, (\[eq:gauging\]), is
$$S_0 + h^L L+ h^{\hL}\hL +h^T T+ h^J J\ .$$
It is clear that (\[eq:null\_symmetry\]) is not the unique symmetry implied by (\[eq:example\_null\_relation\]). There are many possibilities, all related by transformations which are graded antisymmetric in the equations of motion of the gauge fields. The null symmetries, modulo transformations of this type, are graded symmetric. Therefore they cannot be absorbed by adding terms quadratic in $h^*$ to the BV action, and must be introduced as extra symmetries.
Null symmetries are present in any gauged theory with fermions (due to relations such as $T^2=0$), and normally they should be ignored. In conventional gauge theories the properness condition[^3] ensures the existence of a propagator. In the context of string theory or W-strings the gauge fields are non-propagating, and the propagator for the matter fields exists even if the solution is not proper. Nevertheless, it makes sense to impose the properness condition modulo null symmetries. The reason is that, if the null symmetries are incorporated into the theory, we face the problem that they are infinitely reducible. This fact is easy to demonstrate for null symmetries proportional to the gauge field equations of motion, but it is also true when they are not of this form [@Thielemans:1995hn; @Vandoren:1996ku]. Even more seriously, it is not clear which null symmetries should be included and which should be ignored. If we include one, one might suppose that we should include them all, but then the theory becomes difficult to manage.
When $Z^A \neq 0$ closure forces the introduction generators for a subset of the possible null symmetries. In this case it makes sense to relax the properness condition modulo null symmetries to include this particular subset, but the infinite reducibility still poses a serious obstruction to understanding the theory [@Thielemans:1995hn]. When gauge fields are treated as background fields a proper solution is not required, and reducibility ghosts need not be introduced.[^4] In the examples we encounter it is possible to close the algebra after introducing a finite number of null symmetries. However, even for the non-proper solutions terms increasingly non-linear in the antifields need to be added to the extended action to solve the master equation at higher orders, and it is not clear whether a finite number of terms is sufficient.
Null symmetries arise in this manner for many of the special holonomy W-algebras. In the case of $SU(m)$, using the expressions (\[2.8\]) and (\[2.12\]) for the $T$ and $L$ currents, one finds the basic relations (\[4.11.1\]) and (\[4.12.1\]). These are implied abstractly by the JIs only in the case of $SU(3)$. For $SU(4)$ one obtains many relations which follow from (\[4.11.1\]) and (\[4.12.1\]), but which involve higher powers of currents. For example, the $(L, (L,L))$ Poisson bracket JI implies:
$$\begin{aligned}
& J \partial J L \equiv 0 \ \ , \ \ T D_+J \hL \equiv 0 \ \ , \ \ D_+T J \hL \equiv 0 \ ,
\\ \nonumber & J(T \hL - D_+J L) \equiv 0 \ .\end{aligned}$$
For $SU(m)$, $m \geq 5$, the situation changes. The Poisson bracket JIs now involve high enough powers of currents so that it becomes possible to absorb the null symmetries by terms quadratic in $h^*$. For example, the $(L,(L,L))$ Poisson bracket JI in $SU(5)$ implies
$$J^3 \hL \equiv 0 \ .$$
The $s h^J$ part of the null symmetry vanishes identically, and therefore the $s h^L$ part can be absorbed by adding a term proportional to $h^*_J h^*_L$ to the BV action. In these cases a proper solution (modulo null symmetries) to the master equation can be found.
The $G_2$ case is like $SU(3)$, in that the fundamental relations (\[eq:G2\_relations\]) are implied by the Poisson bracket JIs, so there is a correlation between this point and the problems with linearisation. For the rest of the special holonomy cases the Jacobi identities do not imply any relations between the generators. That is to say, for $Spin(7)$ and $Sp(k)\cdot Sp(1)$ there are no problems with the JIs, and one can define the classical W-string BRST operator, but in the context of the OPE one still has the problem that the background fields transform into the quantum fields.
Discussion
==========
We have seen that all the classical special holonomy algebras admit finite linearisations, except for $G_2$ and $SU(3)$. Thus in all cases except for these we can set up and analyse potential anomalies using cohomological methods. The simplest case to consider is the chiral symmetry algebra for flat target space models. The action is
S=S\_0+ X\_i\^\* s X\^i + c\^\*\_A s c\^A ,
where the ghosts and their antifields are chiral. The ghost term in the action gets no quantum corrections and is only introduced to tidy up the algebra. Since the theory defined by $S_0$ is free it is not affected by anomalies, but there could be anomalies in the antifield sector related to those of the current algebra.
The next step is to gauge models of this type. The action is
S=S\_0+ h\^A j\_A + X\_i\^\* s X\^i + c\^\*\_A s c\^A +h\^\*\_A s h\^A ,
where the ghosts are now no longer chiral. Again the last two terms do not receive quantum corrections. Differentiation of the quantum action twice with respect to the gauge fields gives the two-point function of the currents. Since the OPE, which is straightforward to compute in the free theory, introduces the missing $J^p$ and $TJ^p$ currents in $SU(m), m\geq 4$, it follows that this model will be anomalous in these cases. However, for $Spin(7)$ and $Sp(k)\cdot Sp(1)$, one would expect that these models would be anomaly-free.
In principle, one could apply the same ideas to the gauged and ungauged interacting models. However, in practice one has to specify a quantisation procedure. The best is the background field method which allows one to keep track of the geometrical nature of the theory [@Friedan:1980jm] . This can be accomplished by introducing a family of fields $\bbX(s)$ which interpolate between the background field $X=\bbX(0)$ and the total field $X_t=\bbX(1)$. The field $\bbX(s)$ can be taken to satisfy the geodesic equation, and the quantum field $Y$ is taken to be the tangent vector to the geodesic at $s=0$. The background-quantum split involves a non-linear shift symmetry [@Howe:1986vm] which can be shown to be non-anomalous [@Blasi:1988sh]. This symmetry controls the field dependence of the counterterms and can be understood geometrically in terms of the first jet bundle of the tangent bundle. Symmetries of the sigma model can give rise to linear transformations of the quantum field if the symmetry variation commutes with differentiation with respect to $s$. The condition for this to be the case is that the symmetry variation $\d \bbX(s)$ should satisfy the equation of geodesic deviation. Unfortunately, this is not the case for any of the special holonomy symmetry transformations. This implies that the quantum field has to transform non-linearly. In principle, therefore, in order to undertake a complete analysis of the anomalies of the interacting special holonomy sigma models in the background field method one should analyse these non-linear symmetries for graphs with both background and quantum external lines since the latter can contribute as subgraphs in the effective action for background fields.
Although a full analysis would involve the above steps it is nevertheless not unreasonable to look at the potential anomalies of the background field effective action with no external quantum lines. If one makes a special holonomy transformation of the background field accompanied by the appropriate transformation of the quantum field the local action in the path integral is invariant and the change in the quantum field can be absorbed by a field redefinition in the path integral. Therefore the effective action, defined by
e\^[i]{}= DY e\^[iS\[X\_t\]]{} ,
should satisfy a potentially anomalous Ward Identity of the form
W(a\_L) = (a\_L)
where $a_L$ is the chiral parameter for an $L$-type symmetry, $\D(a_L)$ is the anomaly, and
W(a\_L):=a\_L \_L X\^i .
In this situation one has a Wess-Zumino consistency condition of the form
W(a\_L) (a\_M)-W(a\_M) (a\_L)=((L(a\_L),M(a\_M))) .
Note that for this approach to be valid the algebra in question must be of the linearised type. It would be of interest to investigate this consistency condition. One might expect that special holonomy anomalies would be related to the superconformal anomaly. The above approach can be extended to the gauged case where one would again expect there to be problems for the $SU(m)$ case.
In conclusion, we have seen that it is extremely difficult to analyse the anomalies of special holonomy symmetry W-algebras in the BV framework. It is essential to consider the (chirally) gauged models since the ghosts cannot be consistently taken to be chiral themselves. In these models, however, even in the most favourable cases, such as $Spin(7)$ or $Sp(k)\cdot Sp(1)$, where there are no difficulties due to problems with the Jacobi identity, one is still faced with the problem that the background gauge fields and ghosts transform into the quantum fields. It seems that the only way to avoid this problem in the W-framework is to quantise the gauge fields. Even here, in many cases, one is faced with the problem of infinite reducibility.
These problems can all be avoided to some extent if one utilises the fact that most of the classical special holonomy algebras admit finite linearisations. In the background field method one can then analyse possible anomalies in the effective action with no external quantum lines using the naive classical Ward identities and consistency conditions. On the other hand the inclusion of (background) gauge fields can cause problems due to the fact that the OPE of two currents generates operators which are not in the original set. The only models free from this problem are the $Spin(7)$ and $Sp(k)\cdot Sp(1)$ models.
There are only two models, $SU(3)$ and $G_2$, which do not admit finite linearisations. Technically this is because of the presence of currents in the algebra which then generate others with more and more derivatives. An interesting observation is that CY algebras admits closed linear subalgebras generated by $\{T,J,\O\}$ or $\{T,J,\bar\O\}$ where
L=Ø+|Ø=i(Ø-|Ø) .
These can be analysed in all cases including CY3. It is the complex currents $\O$ which are related to the squares of the spectral flow operator so it may be that commuting the spectral flows in two directions is the source of the problem. The analogous subalgebra in the $G_2$ case is the tri-critical Ising model [@Shatashvili:1994zw].
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported in part by EU grant (Superstring theory) MRTN-2004-512194, PPARC grant number PPA/G/O/2002/00475 and VR grant 621-2003-3454.
[99]{}
B. Zumino, “Supersymmetry And Kahler Manifolds,” Phys. Lett. B [**87**]{} (1979) 203. G. W. Delius, M. Rocek, A. Sevrin and P. van Nieuwenhuizen, “Supersymmetric Sigma Models With Nonvanishing Nijenhuis Tensor And Their Operator Product Expansion,” Nucl. Phys. B [**324**]{} (1989) 523. S. Odake, “Extension Of N=2 Superconformal Algebra And Calabi-Yau Compactification,” Mod. Phys.Lett. A [**4**]{} (1989) 557. P. S. Howe and G. Papadopoulos, “Holonomy groups and W symmetries,” Commun. Math. Phys. [**151**]{}, 467 (1993) \[arXiv:hep-th/9202036\]. S. L. Shatashvili and C. Vafa, “Superstrings and manifold of exceptional holonomy,” Selecta Math. [**1**]{} (1995) 347 \[arXiv:hep-th/9407025\].
J. M. Figueroa-O’Farrill, “A note on the extended superconformal algebras associated with manifolds of exceptional holonomy,” Phys. Lett. B [**392**]{} (1997) 77 \[arXiv:hep-th/9609113\]. J. de Boer, A. Naqvi and A. Shomer, “The topological G(2) string,” arXiv:hep-th/0506211. D. J. Gross and E. Witten, “Superstring Modifications Of Einstein’s Equations,” Nucl. Phys. B [**277**]{} (1986) 1. M. T. Grisaru, A. E. M. van de Ven and D. Zanon, “Four Loop Beta Function For The N=1 And N=2 Supersymmetric Nonlinear Sigma Model In Two-Dimensions,” Phys. Lett. B [**173**]{} (1986) 423. P. Candelas, M. D. Freeman, C. N. Pope, M. F. Sohnius and K. S. Stelle, “Higher Order Corrections To Supersymmetry And Compactifications Of The Heterotic String,” Phys. Lett. B [**177**]{} (1986) 341. H. Lu, C. N. Pope, K. S. Stelle and P. K. Townsend, “String and M-theory deformations of manifolds with special holonomy,” JHEP [**0507**]{} (2005) 075 \[arXiv:hep-th/0410176\]. O. Piguet and S. P. Sorella, “Algebraic renormalization: Perturbative renormalization, symmetries and anomalies,” Lect. Notes Phys. [**M28**]{} (1995) 1. P. S. Howe, U. Lindstrom and P. White, “Anomalies And Renormalization In The Brst - Bv Framework,” Phys. Lett. B [**246**]{} (1990) 430. J. Gomis, J. Paris and S. Samuel, “Antibracket, antifields and gauge theory quantization,” Phys. Rept. [**259**]{} (1995) 1 \[arXiv:hep-th/9412228\]. K. Schoutens, A. Sevrin and P. van Nieuwenhuizen, “A New Gauge Theory For W Type Algebras,” Phys. Lett. B [**243**]{} (1990) 245. C. M. Hull, “Lectures on W gravity, W geometry and W strings,” arXiv:hep-th/9302110. F. Bastianelli, “Ward identities for quantum field theories with external fields,” Phys. Lett. B [**263**]{}, 411 (1991). H. Lu, C. N. Pope, K. Thielemans, X. J. Wang and K. W. Xu, “Quantizing higher spin string theories,” Int. J. Mod. Phys. A [**10**]{}, 2123 (1995) \[arXiv:hep-th/9410005\]. K. Thielemans and S. Vandoren, “Gauging Conformal Algebras with Relations between the Generators,” arXiv:hep-th/9511200. S. Vandoren, “Covariant Quantisation in the Antifield Formalism,” arXiv:hep-th/9601013. D. H. Friedan, “Nonlinear Models In Two + Epsilon Dimensions,” Annals Phys. [**163**]{} (1985) 318. P. S. Howe, G. Papadopoulos and K. S. Stelle, “The Background Field Method And The Nonlinear Sigma Model,” Nucl. Phys. B [**296**]{} (1988) 26. A. Blasi, F. Delduc and S. P. Sorella, “The Background Quantum Split Symmetry In Two-Dimensional Sigma Models: A Regularization Independent Proof Of Its Renormalizability,” Nucl. Phys. B [**314**]{} (1989) 409.
[^1]: The deWitt notation is being used temporarily, with repeated indices implying integration as well as summation over labels. See for example [@Gomis:1994he].
[^2]: For both sectors the gauging procedure is more complicated - see [@Schoutens:1990ja; @Hull:1993kf].
[^3]: See, for example, section 4.3 in [@Gomis:1994he] for the definition of properness.
[^4]: Because the $h^A$s are non-propagating fields, there is also the option that a non-proper solution would make some sense even when the $h^A$ are treated as quantum fields. This has not been investigated in the literature.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a Shared Mobility on Demand Service (SMoDS), dynamic pricing plays an important role in the form of an incentive for the decision of the empowered passenger on the ride offer. Strategies for determining the dynamic tariff should be suitably designed so that the incurred demand and supply are balanced and therefore economic efficiency is achieved. In this manuscript, we formulate a discrete time Markov Decision Process (MDP) to determine the probability desired by the SMoDS platform corresponding to the acceptance rate of each empowered passenger at each state of the system. We use Estimated Waiting Time (EWT) as the metric for the balance between demand and supply, with the goal that EWT be regulated around a target value. We then develop a Dynamic Programming (DP) algorithm to derive the optimal policy of the MDP that regulates EWT around the target value. Computational experiments are conducted that demonstrate the regulation of EWT is effective, through various scenarios. The overall demonstration is carried out offline. The MDP formulation together with the DP algorithm can be utilized to an online determination of the dynamic tariff by integrating with our earlier works on Cumulative Prospect Theory based passenger behavioral modeling and the AltMin dynamic routing algorithm, and form the subject of future works.'
author:
- 'Yue Guan[^1]'
- 'Anuradha M. Annaswamy'
- 'H. Eric Tseng'
bibliography:
- 'references.bib'
title: '**Towards Dynamic Pricing for Shared Mobility on Demand using Markov Decision Processes and Dynamic Programming** '
---
#### Index terms[:]{.nodecor}
Shared Mobility on Demand, Dynamic Pricing, Estimated Waiting Time, Markov Decision Process, Dynamic Programming, Forward Search, Smart Cities.
Introduction {#intro}
============
Shared Mobility on Demand Service (SMoDS) has transformed urban mobility and introduced a continuum of solutions between the traditionally binary modes of private individual vehicles and public mass transit, so as to lead to a range of services with different degrees of cost, flexibility, and carbon footprint. This manuscript pertains to an SMoDS solution that consists of customized dynamic routing and dynamic pricing. We build on our earlier works in [@guan2019dynamicrouting; @guan2019cpt; @annaswamy2018transactive] and propose an Markov Decision Process (MDP) formulation and a Dynamic Programming (DP) planning algorithm towards dynamic pricing for the SMoDS platform.
Dynamic pricing has achieved remarkable successes in emerging ride sharing platforms such as Uber, Lyft, and Didi Chuxing, where passengers are empowered to have the option to decide whether to accept or decline the ride offers. Dynamic tariffs therefore provide the incentive signals that affect the decisions of the passengers, more specifically, tune the probability with which the passengers accept the SMoDS ride offers. The objective of dynamic pricing is to balance the demand and supply of the SMoDS platform, and hence to further achieve economic efficiency. The balance could be measured via several Key Performance Indicators, of which we use Estimated Waiting Time (EWT) in our formulation [@cohen2016using]. EWT approximates the average time that an upcoming passenger would wait until being picked up. A large $\text{EWT}(t)$ at time $t$ indicates that demand exceeds supply and vice versa. The goal is to regulate $\text{EWT}(t)$ around a target value $\text{EWT}^*$ by suitably designing the dynamic pricing strategies.
The main contributions of this manuscript include an MDP formulation that serves as the underlying tool to determine the desired probability of acceptance $p^*$ for each passenger and a DP planning algorithm to derive the optimal policy of the MDP in an offline context. Computational experiments are conducted to demonstrate effective regulation of $\text{EWT}(t)$ around $\text{EWT}^*$ for various $\text{EWT}^*$ values and for a time-varying $\text{EWT}^*$. The extension of the DP algorithm to online scenarios is also discussed. With the desired probability of acceptance addressed in this manuscript, and by integrating with the passenger behavioral model from [@guan2019cpt], the dynamic tariff that nudges the passenger towards $p^*$ can be derived.
Dynamic pricing for the SMoDS has been quite a popular research topic during recently years. [@ma2018spatio] proposes the spatio-temporal pricing mechanism that has prices be smooth in space and time hence drivers will not decline the dispatched rides to seek ones with higher returns nearby. [@banerjee2015pricing] develops a queueing-theoretic economic model for dynamic pricing and proves that dynamic pricing is more robust than static pricing. [@korolko2018dynamic] utilizes the internal data from Uber and calibrates a steady-state model that verifies increased total welfare via dynamic pricing. However, to the best of our knowledge, no prior work has been reported related to the applications of MDP and DP in dynamic pricing for the SMoDS platform.
Preliminaries - Dynamic Routing and Dynamic Pricing {#background}
===================================================
The SMoDS solution we are developing consists of dynamic routing and dynamic pricing, which delivers a customized dynamic route and dynamic tariff to each passenger [@guan2019dynamicrouting; @annaswamy2018transactive; @guan2019cpt]. The overall schematic is illustrated in Fig. \[fig:schematic\], with three building blocks functioning in the following manner. When receiving a ride request, the first block derives an optimized dynamic route. The second and third blocks ensure dynamic pricing that accommodates the possibility that the empowered passenger may either accept or reject the ride offer and still generates an overall performance desired by the SMoDS platform. Of these, the second concerns a passenger behavioral model that derives the actual probability of acceptance for a specified dynamic tariff. The third and final block is the construction of a desired probability of acceptance from each passenger that will ensure the desired performances by the SMODS platform. Using this third block, one can then design the actual dynamic tariff by solving the inverse problem. We focus on determining desired probability of acceptance in this manuscript while only briefly introduce dynamic routing and passenger behavioral modeling in this section, and refer the readers to [@guan2019dynamicrouting] and [@guan2019cpt] for mores details.
![Overall schematic of the SMoDS design.[]{data-label="fig:schematic"}](schematic.pdf){width="90.00000%"}
Dynamic Routing {#dynamicrouting}
---------------
An Alternating Minimization (AltMin) based optimization algorithm is developed for dynamic routing with added spatial flexibility enabled by space window [@guan2019dynamicrouting]. AltMin derives dynamic routes for passengers given their requested pickup, drop-off locations, willingness to walk and other service requirements that minimize a weighted sum of various travel time cost terms. AltMin has been demonstrated via various computational experiments to outperform classic constrained optimization formulations that are solved by standard solvers in terms of both computational complexity and optimality.
In the current SMoDS design, AltMin on the one hand derives dynamic routes that are provided in the ride offer, on the other hand contributes to the MDP formulation towards dynamic pricing, by means of the functions $F_{\text{DR}}(\cdot)$, $F_{\text{EWT}}(\cdot)$, and $F_{\text{wait}}(\cdot, \cdot)$ which we will formally introduce in Section \[MDP\].
Dynamic Pricing {#dynamicpricing}
---------------
Dynamic pricing is divided into two phases as follows.
### Passenger Behavioral Modeling {#behavioral}
The first phase is passenger behavioral modeling. A passenger behavioral model is one that inputs the specifications of the SMoDS ride offer and alternative transportation options, and outputs the actual probability with which the passenger takes the SMoDS. According to discrete choice model [@ben1985discrete], the actual probability of acceptance of the $\ell^{\text{th}}$ transportation option given $Q \in \mathbb{Z}_{>0}$ options to choose from is given by $$\label{discrete_choice_model}
p^{\ell} = \frac{e^{U_{\ell}}}{\sum_{q=1}^Q e^{U_q}}, \quad \forall \ell \in \{1, \cdots, Q\}$$ Here we apply Cumulative Prospect Theory (CPT) in passenger behavioral modeling to capture the subjective decision making of the passengers when facing risk or uncertainty. This is because the SMoDS is exposed to uncertainty since the vehicles need to accommodate new passengers at anytime during the route therefore the service quality is to some extent stochastic. $U_{\ell}$ denotes the subjective utility of taking the $\ell^{\text{th}}$ option perceived by the passenger, which can be computed via the framework we have developed in [@guan2019cpt].
### Desired Probability of Acceptance {#desired}
Given the ride specifications, the passenger is empowered to take the SMoDS with certain probability. The probability of acceptance impacts the expected performances of the SMoDS platform. With the dynamic routes derived via AltMin, dynamic pricing serves as an incentive to tune this probability. As discussed in Section \[intro\], the goal is to regulate $\text{EWT}(t)$ around the target value $\text{EWT}^*$. Therefore, the probability of acceptance desired by the SMoDS platform should minimize the following quantity which essentially measures the average deviation of the incurred $\text{EWT}(t)$ from $\text{EWT}^*$ $$\label{objective}
\frac{1}{T}\int_{0}^{T}|\text{EWT}(t) - \text{EXT}^*| \; dt$$ Here $T > 0$ denotes the evaluation horizon. Note that how to choose an appropriate $\text{EWT}^*$ value is beyond the scope of this manuscript, which requires the knowledge of the demand and supply of the SMoDS platform, i.e., request pattern and fleet portfolio, and explicitly stated objectives, e.g., some combination of revenue and ridership. This is one of our main future directions. We assume that $\text{EWT}^*$ is explicitly given for the rest of the manuscript if not otherwise clarified.
From (\[discrete\_choice\_model\]), we obtain $p$ the actual probability of acceptance from each passenger. And the desired value $p^*$ is to be derived via the methodology developed in this manuscript. We then input $p$ and $p^*$ to the transactive controller to derive the dynamic tariff that nudges the passenger towards $p^*$ by simply setting $p = p^*$ and solving the inverse problem.
A Markov Decision Process Formulation {#MDP}
=====================================
In this section, we propose a discrete time MDP formulation for determining the desired probability of acceptance.
In the MDP formulation, the agent is the SMoDS server, while the fleet as well as the passengers including both existing and future ones make up the environment. The goal is to derive the optimal policy that maximizes the expected total discounted rewards. The MDP is defined using the tuple $\langle \mathcal{S}, \mathcal{A}, \mathcal{P}, \mathcal{R}, \gamma \rangle$ [@sutton2018reinforcement], which denote the state space, action space, state transition function, discount factor, and reward function, respectively. Several clarifications are made as follows before elaborating each element in the tuple.
- The agent receives ride requests in discrete time which are $\Delta t_r \in \mathbb{R}_{>0}$ apart, and $t_r = k \Delta t_r, k \in \mathbb{Z}_{>0}$ is denoted as a general notion of the discrete timestamps. Denotes $f_{\Omega_{t_r}}(\omega_{t_r})$ or just $f(\omega_{t_r})$ as the distribution of $\Omega_{t_r}$, which we assume is independent on the state.
- The action of the agent is to design the desired probability of acceptance for each new passenger at $t_r$, given the state. We assume that by suitably designing the SMoDS ride offer, the passenger will accept the offer with the desired chance. Hence we do not consider the passenger behavioral model in this manuscript, the current MDP formulation is therefore fully observable.
- A policy $\pi(\cdot)$ is a distribution over actions given the state, and an optimal policy $\pi^*(\cdot)$ is one that maximizes the expected total discounted rewards.
- The actions are taken sequentially, one at a time, and are dependent on the decisions from previous passengers in response to the previous actions.
- The SMoDS fleet portfolio is assumed to be fixed.
- We adopt most of the notation in Sutton [@sutton2018reinforcement] and Silver [@silver2015reinforcement]. We use uppercases as general terms to represent states, actions, rewards, ride requests, and decisions from passengers, and lowercases to represent the corresponding realizations. In addition, calligraphic fonts are utilized to represent the state and action spaces as well as the reward and state transition functions.
State Space $\mathcal{S}$ {#state}
-------------------------
We denote $S_t \in \mathcal{S}$ as the state at $t \in \mathbb{R}_{\geq 0}$. In addition, at the timestamps $t_r$ of any new ride request received by the SMoDS server, $S_{{t_r}^-}$ and $S_{{t_r}^+}$ denote the states right before and after the server processes the request and the passenger decides whether to take the ride offer or not at $t_r$, respectively. Although the proposed MDP formulation operates in discrete time since actions can only be taken at discrete $t_r$ when receiving new requests, the states can be defined in continuous time.
$S_t$ describes the status of the SMoDS platform at $t$, which consists of the statuses of both the fleet and passengers, as well as unprocessed requests if any. Denote $m_t$ and $n_t$ as the numbers of active vehicles and passengers at $t$, respectively, $n_t, m_t \in \mathbb{Z}_{\geq 0}$. $S_t$ simply concatenates the states of each vehicle, passenger, and any unprocessed request as follows $$\label{define_S_t}
S_t = \Big[\big\{S_t^{v_i}\big\}_{i \in [m_t]}, \big\{S_t^{p_j}\big\}_{j \in [n_t]}, \Omega_{t} \Big]$$ where $S_t^{v_i}$ and $S_t^{p_j}$ denote the state of the $i^{\text{th}}$ vehicle and $j^{\text{th}}$ passenger, respectively. We denote $[m_t] = \{1,2,\cdots,m_t\}$ and $[n_t]$ similarly for ease of notation. $\Omega_t = \emptyset$ if $t \neq t_r$. $S_t^{v_i}$ and $S_t^{p_j}$ are defined in (\[state\_vehicle\]) and (\[state\_passenger\]), respectively. $$\label{state_vehicle}
S_t^{v_i} = [L_t^{v_i}, G_t^{v_i}, O_t^{v_i}]$$ $S_t^{v_i}$ contains i) $L_t^{v_i}$ the location, ii) $G_t^{v_i}$ a set of consecutive routing points representing the route, and iii) $O_t^{v_i}$ other status indicators of the $i^{\text{th}}$ vehicle at $t$. $O_t^{v_i}$ may contain any status indicator of that is required by the dynamic routing algorithm, for example, the number of passengers onboard should be included if a capacity constraint is imposed. $$\label{state_passenger}
S_t^{p_j} = [L_t^{p_j}, G_t^{p_i}, O_t^{p_i}]$$ $S_t^{p_j}$ contains i) $L_t^{p_j}$ the location, ii) $G_t^{p_j}$ the route, including the requested and negotiated pickup and drop-off locations, and iii) $O_t^{p_j}$ other status indicators of passenger $j$ at $t$. Similarly, $O_t^{p_j}$ may contain any status indicator that is required for dynamic routing, for example, the positions in the request, pickup, and drop-off queues should be included if the maximum position shift constraints are imposed [@guan2019dynamicrouting].
Action Space $\mathcal{A}$ {#action}
--------------------------
We denote $A_{t_r} \in \mathcal{A}$ as the action taken at $t_r$, apparently $A_{t_r} \in [0, 1]$. More specifically, since the desired probability of acceptance is realized through the design of the SMoDS ride offer and ultimately equates the actual probability of acceptance defined in (\[discrete\_choice\_model\]), and the dynamic tariff should be reasonably charged, for example, it should not be negative nor unrealistically high compared with that of the alternative transportation options, hence $A_{t_r}$ does actually lie in a strict subset of $[0, 1]$, i.e., $$\label{action_range}
A_{t_r} \in \Big [\underline{A_{t_r}}, \overline{A_{t_r}} \Big ], \, 0 < \underline{A_{t_r}} < \overline{A_{t_r}} < 1$$ where $\underline{A_{t_r}}$ and $\overline{A_{t_r}}$ denote the lower and upper bounds of $A_{t_r}$ respectively, both of which are functions of the state, and the specifications of the alternatives, denoted as $$\label{action_bounds}
\Big \{ \underline {A_{t_r}}, \overline{A_{t_r}} \Big \} = F_b(S_{t_r}, \Lambda_{t_r})$$ $S_{t_r}$ is utilized to derive the specifications of the SMoDS ride offer via the dynamic routing algorithm, which together with $\Lambda_{t_r}$ the specifications of the alternatives determine the space of action $A_{t_r}$. In fact, the optimal action $a^*_{t_r}$ only takes the value at either end point of the corresponding action space $\Big [ \underline {a_{t_r}}, \overline{a_{t_r}} \Big ]$, which we formally state in Theorem \[thm:optimal\_action\].
\[thm:optimal\_action\] $\forall t_r$, we have the optimal action $$\label{optimal_action}
a^*_{t_r} \in \Big \{ \underline {a_{t_r}}, \overline{a_{t_r}} \Big \}$$
The proof is provided in Appendix.\
Once the action is taken by the agent and delivered to the passenger, the passenger would respond with the decision on the ride offer. We denote $D_{t_r}$ as the decision of the passenger at $t_r$, where $d_{t_r} = 1$ denotes acceptance and $d_{t_r} = 0$ denotes rejection. Apparently $D_{t_r}$ obeys the Bernoulli distribution as $$\label{d_distribution}
D_{t_r} \sim \text{B}(1, a_{t_r})$$
State Transition Function $\mathcal{P}$ {#transition}
----------------------------------------
Consider $S_t, t \in \mathbb{Z}_{\geq 0}$ defined in the continuous time domain, there are three distinct scenarios where the state might transit from one to another as follows.
### State Transition due to Internal Dynamics {#transition_one}
When there is no new ride request received nor existing request processed, the transition is purely due to the internal dynamics of the state, i.e., movements of the vehicles and passengers following the routes. This type of state transition is deterministic and can be described as follows $$\label{transition_no}
\mathcal{P}_{s_t s_{t + \tau}} =
\begin{cases}
1, \, & \text{if} \, s_{t + \tau} = F_{\text{D}}(s_t, \tau) \\
0, \, & \text{otherwise}
\end{cases}$$ (\[transition\_no\]) is valid $\forall t, \tau \geq 0$ and when no ride request is received during $[t, t + \tau]$. $F_{\text{D}}(\cdot, \cdot)$ essentially captures the internal dynamics when there is no external disturbance, i.e., request received or processed. We omit the superscript that represents the action since there is no action taken here.
### State Transition due to Receiving New Requests {#transition_two}
When there is new request received at $t_r$, state transition occurs via augmenting the current state by the request specifications $$\label{transition_yes_receive}
\mathcal{P}_{s_{t_r^-} s_{t_r}} =
\begin{cases}
f(\omega_{t_r}), \; & \text{if} \; s_{t_r} =[s_{t_r^-} , \omega_{t_r}] \\
0, \; & \text{otherwise}
\end{cases}$$ recall that $f(\omega_{t_r})$ denotes the distribution of $\Omega_{t_r}$. Similar as in (\[transition\_no\]), we omit the representation of actions in the superscripts due to absence of actions in this scenario.
### State Transition due to Processing Existing Requests {#transition_three}
When the agent takes an action to process the new request, and the passenger responds with the decision on the offer, the state transits by possibly updating the status of the vehicles and passengers according to the decision from the passenger, as well as eliminating the existing request. $$\label{transition_yes}
\mathcal{P}_{s_{t_r} s_{t_r^+}} ^ {a_{t_r}} =
\begin{cases}
a_{t_r}, \; & \text{if} \; s_{t_r^+} = F_{\text{DR}}(s_{t_r}) \\
1 - a_{t_r}, \; & \text{if} \; s_{t_r^+} = s_{t_r^-} \\
0, \; & \text{otherwise}
\end{cases}$$ (\[transition\_yes\]) essentially states that the passenger accepts the SMoDS ride offer with the probability of $a_{t_r}$ and as a result the state $s_{t_r}$ transits to $s_{t_r^+} = F_{\text{DR}}(s_{t_r})$. $F_{\text{DR}}(\cdot)$ is essentially the dynamic routing algorithm, i.e., the AltMin algorithm in the SMoDS solution we are developing [@guan2019dynamicrouting], which is a function of $s_{t_r}$ the state when the request is received at $t_r$. While the passenger might decline the offer with the probability of $(1 - a_{t_r})$ and hence the state transits back to that right before receiving the request by eliminating the exiting request $\omega_{t_r}$, i.e., $s_{t_r^+} = s_{t_r^-}$. The superscripts of $a_{t_r}$ indicate that this type of state transition is dependent on the actions.
Combining (\[transition\_no\]) through (\[transition\_yes\]), we derive the state transition function for $S_{t_r}$ the states defined in the discrete time domain when actions are taken as follows $$\label{transition_combine}
\mathcal{P}_{s_k s_{k+1}} ^ {a_k} =
\begin{cases}
a_kf(\omega_{k+1}), \; & \text{if} \; s_{k+1} = [F_{\text{D}} ( F_{\text{DR}}(s_k), \Delta t_r ), \omega_{k+1} ] \\
(1 - a_k)f(\omega_{k+1}), \; & \text{if} \; s_{k+1} = [F_{\text{D}}(s_{k^-}, \Delta t_r), \omega_{k+1}] \\
0, \; & \text{otherwise}
\end{cases}$$ where we omit $\Delta t_r$ in the subscripts and use $k \in \mathbb{Z}_{> 0}$ to represent $k \Delta t_r$ for ease of notation.
\[thm:markov\] The states $S_{t_r}$ defined in (\[define\_S\_t\]) through (\[state\_passenger\]) with the transition function derived in (\[transition\_combine\]) are Markov.
The proof is provided in Appendix.
Reward Function $\mathcal{R}$ {#reward}
-----------------------------
As has been discussed in Section \[background\], the objective of the MDP formulation is to derive the optimal policy $\pi^*(\cdot)$ such that $\text{EWT}(t)$ is well regulated around $\text{EWT}^*$. We break down the average deviation defined in (\[objective\]) into segments by timestamps of requests, and define the reward as the direct contribution to (\[objective\]) of taking action $a_{t_r}$ at state $s_{t_r}$ as follows $$\label{reward_function}
\begin{split}
\mathcal{R}_{s_{t_r}}^{a_{t_r}} = - \mathbb{E}_{D_{t_r}} \bigg[ \frac{1}{\Delta t_r} \int_{t_r^+}^{{(t_r + \Delta t_r)}^-} |\text{EWT}(\tau) - \text{EWT}^*| \; d\tau \bigg]
\end{split}$$ where the expectation is taken over the decision from the passenger in response to $a_{t_r}$. (\[reward\_function\]) essentially measures the average closeness, i.e., the opposite of average deviation, of $\text{EWT}(t)$ from $\text{EWT}^*$ in response to $a_{t_r}$, from $t_r^+$ till ${(t_r + \Delta t_r)}^-$ right before when the next action is about to take place. Following the discussions in Section \[background\], $\text{EWT}(t)$ is a panel parameter of the SMoDS platform measuring the expected waiting time if an upcoming request is received, where the expectation is taken over $f_{\Omega_{t_r}}(\omega_{t_r})$. Hence $\text{EWT}(t)$ is a function of the state $s_t$ following $$\label{calculate_EWT}
\text{EWT}(t) = \mathbb{E}_{\Omega_{t_r}} [ F_{\text{wait}}(s_t , \omega_{t_r}) ] = F_{\text{EWT}}(s_t)$$ where $F_{\text{wait}}(\cdot , \cdot)$ denotes the waiting time derived via the dynamic routing algorithm given $s_t$ and $\omega_{t_r}$. However, (\[calculate\_EWT\]) might be hard to derive even if $f_{\Omega_{t_r}}(\omega_{t_r})$ is known. One could therefore choose appropriate approaches to approximate $\text{EWT}(t)$ instead of deriving it exactly if necessary.
Discount Factor $\gamma$ {#discount}
------------------------
The discount factor $\gamma \in [0, 1]$. Typically $\gamma < 1$, while if the episode is guaranteed to terminate, we let $\gamma = 1$.
Value Function {#value}
--------------
With the discrete time MDP formulation elaborated in Sections \[state\] through \[discount\], we rewrite the objective function from (\[objective\]) in the form of expected total discounted rewards $$\label{objective_rewrite}
\mathbb{E} \Bigg[ \sum_{k = 0} ^ {\infty} \gamma^k \mathcal{R}_{S_{k \Delta t_r}} ^ {A_{k \Delta t_r}} \Bigg]$$ where we let $A_0 = \emptyset$ for ease of notation. $\pi^*(\cdot)$ can be derived either directly via policy based approaches or indirectly via value based ones, or both. For value based approaches, the optimal value function of any state $S_{t_r} = s_{t_r}$ is defined as $$\label{value_state}
v^*(s_{t_r}) = \max_{\pi(\cdot)} \mathbb{E} \Bigg[ \sum_{k = 0} ^ {\infty} \gamma^k \mathcal{R}_{s_{t_r + k \Delta t_r}} ^ {\pi(s_{t_r + k \Delta t_r})} \; \Bigg | \; s_{t_r}\Bigg]$$ where the expectation is taken over the upcoming requests, and the decisions from the passengers in response to the actions (including $D_{t_r}$). Using Bellman Optimality Equation, we rewrite (\[value\_state\]) as $$\label{value_state_rewite}
v^*(s_{t_r}) = \max_{a_{t_r}} \mathcal{R}_{s_{t_r}}^{a_{t_r}} + \gamma \mathbb{E} \Big[v^*(s_{t_r + \Delta t_r}) \; \Big| \; s_{t_r}\Big]$$ where the expectation is taken over $D_{t_r}$ and $\Omega_{t_r + \Delta t_r}$.
When the numbers of active vehicles and passengers are large, and considering that the state space is continuous, deriving $\pi^*(\cdot)$ through $v^*(\cdot)$ exactly may be computationally impractical. Alternatively, one could apply function approximation [@bertsekas1995neuro] to derive a parametrized suboptimal policy as $$\label{value_state_rewite_approximation}
\tilde{\pi}(s_{t_r};f_v) = \text{arg}\min_{\pi(\cdot)} \mathbb{E} \Big[ \mathcal{R}_{s_{t_r}}^{\pi(s_{t_r})} + \gamma \tilde{v}(s_{t_r + \Delta t_r};f_v) \; \Big| \; s_{t_r} \Big]$$ where $\tilde{v}(\cdot;f_v)$ is an approximation of $v^*(\cdot)$ parametrized by the feature vector $f_v$, whose dimension is much smaller than that of the state and therefore results in more efficient computation. The state-action value function and the corresponding function approximation can be defined similarly.
Moreover, the action space at any $t_r$ is a function of the state $S_{t_r}$, which can be appropriately parametrized and approximated as well. Denote $f_b$ as the feature vector that parametrizes $S_{t_r}$. Similarly, the dimension of $f_b$ should be relatively small compared with that of the state to ensure efficient computation. The approximation is denoted as $$\label{action_bounds_approximation}
\Big [ \underline {A_{t_r}}, \overline{A_{t_r}} ; f_b\Big ] = \tilde{f}(S_{t_r}, \Lambda_{t_r} ; f_b)$$
In addition, policy or Actor Critic based approaches can also be approximated with appropriate parametrization. With these, one can apply a brand range of learning or/and planning algorithms to derive $\pi(\cdot)$ or $\tilde{\pi}(\cdot; \cdot)$.
A Dynamic Programming Algorithm to Determine $\pi^*(\cdot)$ {#experiments}
===========================================================
In this section, we start with a specific offline scenario of the MDP formulation, develop a DP planning algorithm and demonstrate how the desired probability of acceptance could be designed to regulate $\text{EWT}(t)$ around $\text{EWT}^*$. We then proceed to discuss the implications of the DP algorithm for the general online case.
Problem Setup {#setup}
-------------
In this section, we adopt an offline setup. 12 synthetic ride requests are generated with origins and destinations uniformly distributed in a square of one by one mile, and are served by one single vehicle. The first 4 requests are scheduled at $t = 0$ and these passengers are assumed to accept the ride offers, which initialize the simulation episode. The following $N=8$ requests are scheduled one by one and arrive 4 minutes apart over an interval of 28 minutes. Each request is assumed to have one passenger, and the vehicle has a capacity of 6. The offline setup indicates that the agent has the information of all 12 requests beforehand and takes subsequent requests into consideration when designing the policy. However, the requests are still processed sequentially, meaning that the offers will be provided to the passengers sequentially and the policy for the upcoming passengers may adapt depending on the decisions from previous ones. Dynamic routing are conducted using the AltMin algorithm developed in [@guan2019dynamicrouting]. Since the simulation episodes are guaranteed to terminate in the offline setup, we set $\gamma = 1$.
As the first attempt, and without loss of generality, here we adopt simplified approximations of $\Big \{ \underline {a_{t_r}}, \overline{a_{t_r}} \Big \}$ given $s_{t_r}, \lambda_{t_r}$, and $\text{EWT}(t)$ given $s_t$, respectively. The alternative is considered as the MoD service without sharing, where the waiting time for each request is set as $\frac{2}{3}\text{EWT}^*$, and the riding time is the direct travel time from the origin to the destination. Given $s_{t_r}$, the specifications of the SMoDS ride offer are derived using AltMin, if the total travel time, i.e., waiting plus riding times of the SMoDS does not exceed $\frac{3}{2}$ of that of the alternative, the action space is set as $\{0.5, 0.9\}$, otherwise $\{0.2, 0.6\}$. Given $s_t$, $\text{EWT}(t)$ is approximated as the average time that four representative passengers who request pickup locations at the four corners of the one by one mile square would wait. As has been discussed in Section \[action\] and \[reward\], accurate derivations of $\Big \{ \underline {a_{t_r}}, \overline{a_{t_r}} \Big \}$ and $\text{EWT}(t)$ require knowledge of the alternative transportation options, passenger behavioral model and distribution of requests, and are our future directions. Here in this section, we adopt approximation approaches that are sufficiently reasonable to illustrate the central idea of regulating $\text{EWT}(t)$ via $\pi^*(\cdot)$.
Algorithmic Design {#algorithms}
------------------
Since we deploy an offline setup, all ride requests are known to the SMoDS server beforehand and deterministic. This reduces the cardinality of the space of $S_{t_r}$ from continuous to finite. In addition, according to Theorem \[thm:optimal\_action\], the action space of each $S_{t_r}$ is finite as well, and has a cardinality of $2$. These enable the development of the value based DP planning algorithm that derives $\pi^*(\cdot)$ exactly to be feasible.
![Illustration of the DP planning algorithm to derive $\pi^*(\cdot, \cdot)$.[]{data-label="fig:tree"}](tree.pdf){width="80.00000%"}
Fig. \[fig:tree\] illustrates a directed and rooted tree structure that explains the DP algorithm. Except for the root and leaves, i.e., terminal vertices represented by black squares, each vertex represents a circumstance when the agent takes an action, and has two children representing two distinct scenarios of subsequent states resulted from the decision of the passenger in response to the action. If the passenger decides to accept the ride offer, the child on the top is chosen and a rejection leads to the one on the bottom. The action is labeled inside each nonterminal vertex, and we let that of the root be $\emptyset$ since no action is taken therein. Each edge represents the transition after the decision of the previous passenger till when the new request is received and a subsequent action is to be taken, and is associated with a realization of the reward. The height of the tree is $H = N + 1$ and each longest path with length equal to $H$ represents a complete scenario as a result of consecutive decisions from all passengers. For ease of notation, we omit $\Delta t_r$ in the subscripts of states, actions and rewards for the rest of the manuscript if not otherwise clarified, for example, $s_{k}^{d_{[k-1]}}$ represents $s_{k \Delta t_r}^{d_{[k-1]}}$. In addition, $d_{[k]} = [d_1, \cdots, d_k], \forall k \in [N]$ represents the decisions of the first $k$ passengers. We have $d_{[k]}$ in the superscripts explicitly since the decisions of previous passengers determine which vertex and edge that the DP algorithm traverses. Moreover, we denote $\omega_{k}, \forall k \in[N]$ as the $k^{\text{th}}$ ride request, and $\omega_{N+1} = \emptyset$ for ease of notation.
The algorithm consists of two steps. The first step is to conduct dynamic routing to process the requests consecutively which essentially traverses the tree from the root to the leaves, therefore to derive the reward using functions $F_\text{D}(\cdot, \cdot), F_\text{DR}(\cdot), F_\text{EWT}(\cdot)$ and (\[reward\_function\]) on each edge. The second step is to conduct backward recursion from the leaves to the root, therefore to derive the optimal value function at each state and the associated optimal action at each vertex, via Bellman Optimality Equation delineated in (\[value\_state\_rewite\]). The details of the DP exact algorithm are outlined in Algorithm \[DP\_exact\].
Initialize the route via AltMin and derive $s_{0^+}$ Initialize $s_1^{\emptyset} \leftarrow [F_{\text{D}}(s_{0 ^ +}, \Delta t_r),\omega_1], d_{[]}, d_{[0]} \leftarrow \emptyset$ Initialize $v^*(s_{k}^{d_{[k-1]}}) \leftarrow 0, \forall k \in [N + 1], d_{[k-1]}$ \[algo\_DPE\_start\] ()[$k = 1:N$]{}
\[algo\_DPE\_mark\]
()[$k = N:-1:1$]{}[ ]{} \[algo\_DPE\_end\]
Similar idea can be exploited to develop a DP heuristic algorithm that improves computational efficiency at the cost of optimality. Instead of conducting forward search till the end of the future, i.e., number of lookahead steps being $N$, the DP heuristic algorithm conducts forward search with steps being $\tilde{N} < N$ , which poses as a hyper parameter to tune the trade-off between efficiency and optimality. For example, to derive $a_k^{d_{[k-1]}}$, the DP-H($\tilde{N}$) algorithm essentially carries out DP-E($\tilde{N}$) on the sub tree rooted at the vertex $a_k^{d_{[k-1]}}$ with a height of $\tilde{N} + 1$, instead of traversing the entire tree in Fig. \[fig:tree\]. The pseudo codes are provided in Algorithm \[DP\_heuristics\].
Note that in line \[algo\_DPE\_mark\_mark\], if $1 < p < N - \tilde{N} + 1$, only execute when $k=\tilde{N}$ in line \[algo\_DPE\_mark\] in Algorithm \[DP\_exact\] since $r_q^{d_{[q-1]}}$ with $q<p + \tilde{N} - 1$ have been evaluated in previous iterations.
Results of Computational Experiments {#results}
------------------------------------
With the problem setup described in Section \[setup\], we exploit the DP exact algorithm delineated in Algorithm \[DP\_exact\] to derive $\pi^*(\cdot)$ under various $\text{EWT}^*$ values. The results are summarized in Fig. \[fig:basic\]. The subplots on the left, middle and right demonstrate the results with $\text{EWT}^*$ being 4, 5, and 6 minutes, respectively. In the subplots on the top, the blue curves illustrate how the expected $\text{EWT}(t)$ varies with $t$ under the optimal policy $\pi^*(\cdot)$ derived using DP-E($N$), the orange curves illustrate $\text{EWT}(t)$ in the benchmark scenario where passengers are not empowered and hence all get onboard, and the black curves represent $\text{EWT}^*$. The benchmark curve is identical in the three subplots since it does not depend on $\text{EWT}^*$. The subplots on the bottom summarize expected acceptance rate under $\pi^*(\cdot)$ for each passenger and the mean across all passengers. Fig. \[fig:basic\] demonstrate that the blue curves could be brought closer to the target dashed lines compared with the orange ones, indicating effective regulation of $\text{EWT}(t)$ around $\text{EWT}^*$ through $\pi^*(\cdot)$ derived via DP-E($N$). The regulation can be achieved with different $\text{EWT}^*$ values that are within a reasonable range. Though how to choose an appropriate $\text{EWT}^*$ value is beyond the scope of this manuscript, we could roughly argue the impacts of the $\text{EWT}^*$ value on the performances of the SMoDS platform. According to Fig. \[fig:basic\], the average acceptance rate, i.e., the ridership, increases with $\text{EWT}^*$ because the tolerance of the agent to having passengers wait gets higher therefore it is more likely to get the passengers onboard. However, on the other hand, the downgrade of the service quality might occur due to the increased actual waiting times as a response to increased $\text{EWT}^*$. In order to maintain relatively high acceptance rates, the dynamic tariff should decrease. Therefore, the overall impacts of increased $\text{EWT}^*$ on the total revenue depend on two competing contributions, one is increased ridership and the other is decreased revenue per ride. Hence $\text{EWT}^*$ should be suitably determined to tune the trade-off in order to achieve desired overall performances.
In addition to Fig. \[fig:basic\] where $\text{EWT}^*$ is held constant within episodes, we also experiment with time-varying $\text{EWT}^*$. The results are demonstrated in Fig. \[fig:jump\]. $\text{EWT}^*$ is set as 4 minutes during the first half of the episode, while when the SMoDS platform gets fairly crowded and $\text{EWT}(t)$ is about to increase substantially, we relax the regulation by increasing $\text{EWT}^*$ up to 6 minutes at $t = 20 $ minutes, and the regulation is adapted instantaneously. These results support our argument above that $\text{EWT}^*$ could be actively tuned to adapt to the actual conditions of the system during practical operations.
![Regulation of $\text{EWT}(t)$ around changing $\text{EWT}^*$.[]{data-label="fig:jump"}](jump.pdf){width="45.00000%"}
Moreover, the DP-H($\tilde{N}$) algorithm is implemented with various $\tilde{N}$ values for the case where $\text{EWT}^* = 5$ minutes. Fig. \[fig:heuristics\] plots the average deviation of $\text{EWT}(t)$ from $\text{EWT}^*$, i.e., the negative of the average rewards defined in (\[reward\_function\]), with respect to $\tilde{N}$, under the suboptimal policy derived using DP-H($\tilde{N}$).[^2] Fig. \[fig:heuristics\] indicate that the optimality of DP-H($\tilde{N}$) is fairly comparable to that of DP-E($N$), for example, with just one lookahead step, comparable regulation in terms of the average deviation can be achieved. Therefore, it is promising to apply DP-H($\tilde{N}$) in more general cases with sufficient computational efficiency and preserved optimality.
![Average deviation of $\text{EWT}(t)$ from $\text{EWT}^*$ with respect to the number of lookahead steps.[]{data-label="fig:heuristics"}](heuristics.pdf){width="45.00000%"}
Remarks
-------
In this section, we construct a special case of the MDP formulated in Section \[MDP\], where the offline setup is deployed. Being offline reduces the state space from continuous to finite, together with relatively small number of requests and vehicles, enabling the development of the DP algorithm. The policies derived via both DP-E($N$) and DP-H($\tilde{N}$) demonstrate effective regulation of $\text{EWT}(t)$ around $\text{ETW}^*$, under various $\text{ETW}^*$ values and time-varying $\text{ETW}^*$ as well.
When the state space gets more complicated, either due to fairly large number of requests and vehicles, or the adoption of an online setup, one can exploit function approximation briefly discussed in Section \[value\] and develop a broad range of planning and/or learning algorithms. Notably, DP-H($\tilde{N}$) would be quite useful in either scenario. For example, when dealing with online requests in real time, one can integrate learning and planning. A preliminary value network is learned offline via historical or self-play data, and lookahead search is conducted via DP-H($\tilde{N}$) to refine the value function on the fly and then take actions accordingly. This is very similar to the algorithmic architecture of AlphaGo [@silver2016mastering].
Concluding Remarks {#conclusions}
==================
In this manuscript, we propose a discrete time MDP formulation to determine desired probabilities of acceptance for empowered passengers towards dynamic pricing in the context of SMoDS, and develop a DP algorithm to derive the optimal policy that regulates $\text{EWT}(t)$ around $\text{EWT}^*$ for a specific scenario of the MDP with an offline setup. Computational experiments are carried out that demonstrate effective regulation of $\text{EWT}(t)$ around $\text{EWT}^*$, for various $\text{EWT}^*$ values and for a time-varying $\text{EWT}^*$ as well. The heuristic version of the DP algorithm, DP-H($\tilde{N}$), could be exploited as the lookahead search algorithm when large state space or online setups are encountered. The MDP formulation together with our previous works of the AltMin dynamic routing algorithm in [@guan2019dynamicrouting] and CPT based passenger behavioral modeling in [@guan2019cpt], provide a complete solution to the SMoDS.
Future works include developing integrated learning and planning algorithms for large state space or online setups, and investigating disciplines that guide the choice of appropriate $\text{EWT}^*$ values leading to the desired combination of revenue and ridership for the SMoDS platform. Moreover, the integration of passenger behavioral model with the MDP hence directly designing dynamic tariffs is of interest as well.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by the Ford-MIT Alliance.
Appendix: Proofs of Theorems {#appendix .unnumbered}
============================
\[proof:optimal\_action\] Expand (\[value\_state\_rewite\]), we have $$\label{backup_value}
\begin{split}
v^*( s_{t_r}) & = \max_{a_{t_r} \in \Big [\underline{a_{t_r}}, \overline{a_{t_r}} \Big ]} \mathcal{R}_{s_{t_r}}^{a_{t_r}} + \gamma \mathbb{E} \Big[ v^*(s_{t_r + \Delta t_r}) \; \Big| \; s_{t_r}, a_{t_r} \Big] \\
& = \max_{a_{t_r} \in \Big [\underline{a_{t_r}}, \overline{a_{t_r}} \Big ]} a_{t_r} \Big\{r_{t_r}^{[1]} + \gamma \mathbb{E} \Big[ v^* \Big( s_{t_r+\Delta t_r} ^ {[1]} \Big) \Big ] \Big\} + (1 - a_{t_r}) \Big\{r_{t_r}^{[0]} + \gamma \mathbb{E} \Big[ v^* \Big( s_{t_r+\Delta t_r}^{[0]}\Big) \Big ] \Big\} \\
& = \max_{a_{t_r} \in \Big \{ \underline{a_{t_r}}, \overline{a_{t_r}} \Big \}} a_{t_r} \Big\{r_{t_r}^{[1]} + \gamma \mathbb{E} \Big[ v^* \Big( s_{t_r+\Delta t_r} ^ {[1]} \Big) \Big ] \Big\} + (1 - a_{t_r}) \Big\{r_{t_r}^{[0]} + \gamma \mathbb{E} \Big[ v^* \Big( s_{t_r+\Delta t_r}^{[0]}\Big) \Big ] \Big\}
\end{split}$$ The first equality is essentially Bellman Optimality Equation and holds by definition. The second equality holds due to the expansion of the expectation term over $D_{t_r}$. The third equality holds because the right-hand side of the second equality is linear in $a_{t_r}$, since the remaining expectations are taken over the distribution of $\Omega_{t_r + \Delta t_r}$ hence $r_{t_r}^{[1]}$, $v^* ( s_{t_r+\Delta t_r} ^ {[1]} )$, $r_{t_r}^{[0]}$, or $v^* ( s_{t_r+\Delta t_r}^{[0]})$ do not depend on $a_{t_r}$, therefore the maximum must be reached at either end point. Here $r_{t_r}^{[1]} = - \frac{1}{\Delta t_r} \int_{0^+}^{{(\Delta t_r)}^-} |F_{\text{EWT}} [F_{\text{D}}(F_{\text{DR}}(s_{t_r}), \tau)] - \text{EWT}^*| \; d\tau$, $s_{t_r+\Delta t_r} ^ {[1]} = [ F_{\text{D}} ( F_{\text{DR}}(s_{t_r}), \Delta t_r ) , \omega_{t_r+ \Delta t_r} ]$, $r_{t_r}^{[0]} = - \frac{1}{\Delta t_r} \int_{0^+}^{{(\Delta t_r)}^-} |F_{\text{EWT}}[F_{\text{D}}(s_{t_r^-}, \tau)] - \text{EWT}^*| \; d\tau$, and $s_{t_r+\Delta t_r}^{[0]} = [ F_{\text{D}} ( s_{t_r^-} , \Delta t_r ), \omega_{t_r + \Delta t_r} ]$[^3].
\[proof:markov\] According to (\[transition\_combine\]), we have $$\begin{gathered}
\label{markov_property}
\mathbb{P} \Big [S_{k + 1} = s_{k + 1} \; \Big | \; S_k = s_k, \cdots, S_1 = s_1 \Big ] = \mathbb{P} \Big [ S_{k+1} = s_{k+1}\; \Big | \; S_k = s_k \Big ]
\end{gathered}$$ $\forall k \in \mathbb{Z}_{> 0}$ and $ \{ s_1, \cdots, s_{k+1} \} \subset \mathcal{S}$. Obviously, (\[markov\_property\]) holds for the first and third subcases in (\[transition\_combine\]). (\[markov\_property\]) holds for the second subcase in (\[transition\_combine\]) because $\mathbb{P} \Big [ S_{k+1} = s_{k+1}\; \Big | \; S_k = s_k \Big ] = \mathbb{P} \Big [ S_{k+1} = s_{k+1}\; \Big | \; S_k = s_k, S_{k^-} = s_{k^-} \Big ] $. Hence $\forall k \in \mathbb{Z}_{>0}$, $S_k$ are Markov.
[^1]: Corresponding author. Email: [email protected].
[^2]: When $\tilde{N} = 0$, actions are determined by comparing $F_{\text{EWT}}(s_{t_r^+})$ values, instead of rewards and value functions in lines \[algo\_compare\_start\] through \[algo\_compare\_end\] in Algorithm \[DP\_exact\] for cases when $\tilde{N} > 0$.
[^3]: We omit the decisions from previous passengers in the superscripts as the formulas of $r_{t_r}^{[1]}$, $r_{t_r}^{[0]}$, $s_{t_r+\Delta t_r} ^ {[1]}$ and $s_{t_r+\Delta t_r} ^ {[0]}$ hold for any scenario.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present $H$-band VLT/NACO polarized light images of the Herbig Ae/Be star HD169142 probing its protoplanetary disk as close as $\sim$0.1$''''$ to the star. Our images trace the face-on disk out to $\sim$1.7$''''$ ($\sim$250 AU) and reveal distinct sub-structures for the first time: 1) the inner disk ($\lesssim$20 AU) appears to be depleted in scattering dust grains; 2) an unresolved disk rim is imaged at $\sim$25 AU; 3) an annular gap extends from $\sim$40 – 70 AU; 4) local brightness asymmetries are found on opposite sides of the annular gap. We discuss different explanations for the observed morphology among which ongoing planet formation is a tempting – but yet to be proven – one. Outside of $\sim$85 AU the surface brightness drops off roughly $\propto r^{-3.3}$, but describing the disk regions between 85–120 AU / 120–250 AU separately with power-laws $\propto r^{-2.6}$/$\propto r^{-3.9}$ provides a better fit hinting towards another discontinuity in the disk surface. The flux ratio between the disk integrated polarized light and the central star is $\sim 4.1\cdot 10^{-3}$. Finally, combining our results with those from the literature, $\sim$40% of the scattered light in the $H$-band appears to be polarized. Our results emphasize that HD169142 is an interesting system for future planet formation or disk evolution studies.'
author:
- 'Sascha P. Quanz$^{1,2}$, Henning Avenhaus$^2$, Esther Buenzli$^3$, Antonio Garufi$^2$, Hans Martin Schmid$^2$, and Sebastian Wolf$^4$'
title: 'Gaps in the HD169142 protoplanetary disk revealed by polarimetric imaging: Signs of ongoing planet formation?'
---
Introduction
============
To study the physical and chemical conditions for planet formation and to search for morphological evidence of forming planets, the inner few tens of AU of protoplanetary disks have to be investigated. In particular, annular gaps in those disks are considered to be possible signatures of ongoing planet formation. In this letter we report the detection of such gaps in the disk around the Herbig Ae/Be star HD169142 using polarimetric differential imaging (PDI). PDI is a powerful high-contrast technique to study the dusty surface layer of protoplanetary disks very close to the star [e.g., @quanz2011b; @hashimoto2011; @muto2012; @quanz2012a; @hashimoto2012; @tanii2012; @kusakabe2012; @mayama2012; @grady2013].
[llc]{} RA (J2000) & 18$^h$24$^m$29$^s$.79 & (1)\
DEC (J2000) & -29$^\circ$46$'$49$''$.22 & (1)\
$J$ \[mag\] & 7.31$\pm 0.02$ & (1)\
$H$ \[mag\]& 6.91$\pm 0.04$ &(1)\
$K_s$ \[mag\]& 6.41$\pm 0.02$ & (1)\
Sp. Type & A9III/IVe / A7V& (2),(3)\
$v\,{\rm sin}\,i$ \[km$\,$s$^{-1}$\] & 55$\pm 5$ & (2)\
Age \[Myr\]& 1–5 / 12 / 3–12 & (2),(3),(4)\
$[{\rm Fe/H}]$ & $ -0.5\pm 0.1$ / -0.25 – -0.5 & (2),(5)\
log $g$ & 3.7$\pm$0.1 / 4.0–4.1 & (2),(5)\
$T_{\rm eff}$ \[K\] & 7500$\pm$200 / 6500 / 7650$\pm$150 & (2),(3),(5)\
Mass \[M$_\sun$\] & $\sim$1.65 & (3)\
$R_*$ \[R$_\sun$\]& $\sim$1.6 & (3),(5)\
$L_*$ \[L$_\sun$\] & $\sim$8.6 & (3)\
$\dot{M}$ \[$10^{-9}$ ${\rm M}_\sun{\rm yr}^{-1}$\]& $\sim$3.1 / $\leq$1.25$\pm$0.55 & (3),(4)\
Distance \[pc\] & 145 / 151& (6),(3)\
Our new $H$-band PDI observations of HD169142 resolve for the first time distinct structures in the inner disk regions with high signal-to-noise and allow an in-depth analysis of the disk profile and morphology. The basic stellar parameters of HD169142 are summarized in Table \[parameters\].
[lccccc]{} $NB1.64$ & 1 s $\times$ 15 & 3 & $\sim$1.02 & $\sim 1.57''$ & $\sim 14 $ ms\
$H$ & 1 s $\times$ 45 & 2$\times$6 & $\sim$1.06 & $\sim 1.04''$ &$\sim 22 $ ms\
$NB1.64$ & 1 s $\times$ 20 & 3 & $\sim$1.11 & $\sim 1.07''$ & $\sim 20 $ ms\
HD169142 was classified as a Herbig group Ib member [@meeus2001], indicating that its spectral energy distribution (SED) rises in the mid-infrared (MIR), but does not show a 10 $\mu$m silicate emission feature [e.g., @boekel2005]. However, PAH emission features have been detected in MIR spectroscopy [@meeus2001; @boekel2005] and were spatially resolved at 3.3 $\mu$m [@habart2006]. The disk was previously resolved in polarized light in the NIR from the ground [@kuhn2001; @hales2006], but those data did not allow an in-depth analysis of the disk structure and profile. @grady2007 and @fukagawa2010 detected the disk in scattered light at 1.1 $\mu$m with HST/NICMOS coronagraphy and in $H$-band from the ground, respectively, and they derived azimuthally averaged surface brightness profiles. In all direct imaging studies the disk was consistent with being seen face-on and no sub-structures were detected. In the MIR the disk was resolved at 12, 18, 18.8 and 24.5 $\mu$m [@honda2012; @marinas2011] and most of this emission is thought to arise from disk regions between $\sim$30 – 60 AU. From dust continuum and molecular line observations the mass and outer radius of the disk are estimated to be $\sim$0.005 - 0.04 M$_\sun$ and $\sim$235 AU [@raman2006; @dent2006; @panic2008; @meeus2010; @sandell2011]. The kinematic pattern and molecular line profiles are best fitted with a disk inclination of $\sim$13$^\circ$ [@raman2006; @panic2008]. The lack of a silicate emission feature at 10 $\mu$m can be explained with a lack of small ($<$3–5 $\mu$m) dust grains in the inner $\sim$10 – 20 AU of the disk possibly due to grain growth [@boekel2005]. From SED modeling a disk hole was inferred for the inner disk regions [@grady2007; @meeus2010; @honda2012].
Observations and data reduction {#observations_section}
===============================
The observations were carried out with VLT/NACO [@lenzen2003; @rousset2003] in the $H$ and $NB1.64$ filter. A description of the camera properties and the PDI observing mode is given in @quanz2011b. In short, a Wollaston prism splits the light in an ordinary and extraordinary beam with orthogonal, linear polarization directions. Both beams are imaged simultaneously on the detector. The polarization direction can be changed by rotating a half-wave plate (HWP). One full polarization cycle includes exposures at four HWP positions (0$^\circ$ and $-$45$^\circ$ for Stokes $Q$, $-$22.5$^\circ$ and $-$67.5$^\circ$ for Stokes $U$). We summarize the observations for HD169142 in Table \[observations\]. The central few pixels of the point spread function (PSF) were saturated in the $H$ filter. The exposures in the $NB1.64$ filter were unsaturated and used for photometric calibration (see below).
NACO suffers from instrumental polarization and cross-talk between the different Stokes components [@witzel2011; @quanz2011b]. To obtain reliable results in PDI observations these effects need to be corrected for. Having obtained data sets for five more Herbig Ae/Be stars allowed us to do a systematic analysis and we refer to an upcoming publication for a detailed description of our data reduction and calibration approach (Avenhaus et al., in prep.). The key steps are:
All exposures were dark current corrected and flat-fielded. Bad pixels were replaced with the mean value of surrounding pixels. Each quadrant of the detector showed a noise pattern affecting every other row, which was eliminated by subtracting the mean value of each row from each pixel in that row. Detector regions close to the star were excluded to compute the mean values. The ordinary and extraordinary images in each exposure were extracted and the stellar position in all individual images determined. We rebinned the images to a factor three higher resolution (bicubic interpolation) and shifted all images to a common center (bilinear interpolation). Assuming that the central star is unpolarized in $H$-band we corrected instrumental polarization and polarization cross-talk effects[^1]. For each dither position the fractional Stokes parameters $p_Q$ and $p_U$ were computed using the double ratio approach [see e.g., @tinbergen1996; @schmid2006]: $$p_Q=\frac{{\rm R}_Q-1}{{\rm R}_Q+1}\quad;\quad p_U=\frac{{\rm R}_U-1}{{\rm R}_U+1}$$ with $${\rm R}_Q=\sqrt{\frac{{I^{0^{\circ}}_{\rm ord}}/{I^{0^{\circ}}_{\rm extra}}}{I^{-45^{\circ}}_{\rm ord}/I^{-45^{\circ}}_{\rm extra}}}\quad ; \quad {\rm R}_U=\sqrt{\frac{{I^{-22.5^{\circ}}_{\rm ord}}/{I^{-22.5^{\circ}}_{\rm extra}}}{I^{-67.5^{\circ}}_{\rm ord}/I^{-67.5^{\circ}}_{\rm extra}}}$$ Here, the subscripts refer to either the ordinary or extraordinary beam and the superscripts refer to the angular position of the HWP. The final $p_Q$ and $p_U$ images were obtained by averaging over all dither positions. From $p_Q$ and $p_U$ and the intensity images $I$ we obtained the Stokes $Q$ and $U$ images, from which we computed the radial Stokes parameters $Q_{\rm r}$ and $U_{\rm r}$: $$Q_{\rm r}=+Q\,{\rm cos}\,2\phi+U\,{\rm sin}\,2\phi$$ $$U_{\rm r}=-Q\,{\rm sin}\,2\phi+U\,{\rm cos}\,2\phi$$ with $$\phi ={\rm arctan} \frac{x-x_0}{y-y_0}+\theta$$ being the polar angle of a given position $(x, y)$ on the detector, $(x_0,y_0)$ denoting the central position of the star [see, e.g., @schmid2006] and $\theta$ being an offset due to polarization cross-talk effects (Avenhaus et al., in prep.), which amounts to 7$^\circ$ in our case. $Q_{\rm r}$ is equivalent to the polarized flux $P$ under the assumption that the polarized flux has only a tangential component, but, in contrast to $P$, $Q_{\rm r}$ is free of any biases introduced from computing the squares of the $Q$ and $U$ components [@schmid2006]. Under the same assumption, and assuming the correction for instrumental effects was done properly, $U_{\rm r}$ only contains noise and can be used to estimate the error in $Q_{\rm r}$ via $$\Delta Q_{\rm r} = \sqrt{\sigma^2_{U_{\rm r}}} / \sqrt{n_{\rm res}}\quad.$$ Here, $\Delta Q_{\rm r}$ refers to the 1$\sigma$ uncertainty in $Q_{\rm r}$, $\sigma^2_{U_{\rm r}}$ is the variance in the final $U_{\rm r}$ image and $n_{\rm res}$ the number of resolution elements in the region of interest.
The calibration of the surface brightness level in the final $Q_{\rm r}$ image was done using the non-saturated images in the $NB1.64$ filter following the description in @quanz2011b. We estimate that the absolute flux calibration is good to $\sim$30%.
Results and analysis {#results}
====================
Figure \[images\] shows the final $Q_{\rm r}$ and $U_{\rm r}$ images, an intensity image, and a polar coordinate mapping of the $Q_{\rm r}$ image. While the $Q_{\rm r}$ image shows extended polarized flux that we interpret as the disk around HD169142, the $U_{\rm r}$ image does not reveal any significant signal as expected for scattering of dust particles without any preferred alignment.
Figure \[images\] a) reveals a protoplanetary disk with a complex radial morphology. Very close to the star, but outside the innermost, saturated pixels, the detected polarized flux is low and it increases going outwards to reach a ring-like maximum at $\sim$0.17$''$ ($\sim$25 AU). Outside of the narrow ring the surface brightness decreases and an annular gap stretches from $\sim$0.28–0.48$''$ ($\sim$40 – 70 AU) with a local minimum at $\sim$0.38$''$ ($\sim$55 AU). Outside of $\sim$0.52$''$ ($\sim$75 AU) the disk surface brightness drops off smoothly.
The bright inner ring does not appear fully circular (Figure \[images\] panels a and d) and the radius of the peak brightness varies between $\sim$0.16$''$ and $\sim$0.19$''$ with a mean value of $\sim$0.17$''$. A dip in the brightness of the ring seems to be present at position angle (PA) $\approx$80$^\circ$ (east of north). On the opposite side of the annular gap at PA$\approx$90$^\circ$, next to an AO artifact, a region of enhanced brightness is seen stretching a bit into the gap. The full-width-half-maximum of the ring varies slightly with PA but the ring is not spatially resolved. Using the ring we estimated an upper limit in disk inclination of $i<20^\circ$, so that $i$ is still in agreement with the fluctuations in ring radius. However, our image is consistent with a face-on disk. Analyzing the mean surface brightness as a function of azimuth for different radii also supports a small inclination angle as no significant variation is seen (@muto2012 used this method for the SAO206462 disk).
In Figure \[profile\] we show the azimuthally averaged surface brightness profile and the signal-to-noise profile for the $Q_{\rm r}$ and $U_{\rm r}$ images. Fitting a power-law to the brightness profile between 85–250 AU yields an exponent of $-3.31\pm0.11$, where the error denotes the 95% confidence level (black dashed line in top panel of Figure \[profile\]). However, fitting the regions between 85–120 AU and 120–250 AU separately with power-laws yields a mathematically better fit in a $\chi^2$ sense (black solid line). In this case the power-law exponents are $-2.64^{+0.15}_{-0.17}$ and $-3.90^{+0.10}_{-0.11}$ for the inner and outer region, respectively. The power-law fit to the inner region connects well with the bright inner ring at $\sim$25 AU. A 0.15$''\times$0.1$''$ region centered on the AO feature seen in Figure \[images\] has been masked out before the power-law fits were computed.
The maximum surface brightness in polarized flux of $S \approx$ 11.3 mag/arcsec$^2$ is reached at the location of the bright ring. Close to the inner edge of the outer disk ($\sim$75 AU) we find $\approx$14.3 mag/arcsec$^2$ and at $\sim$220 AU $\approx$18.1 mag/arcsec$^2$. Inside the gap ($\sim$55 AU) the surface brightness is reduced by a factor of $\sim$3 compared to the corresponding value of the power-law fit. Integrating the polarized surface brightness between $\sim$14.5–250 AU and comparing it to the $H$-band flux of the central star yields a flux ratio of $F_{\rm Disk}^{\rm polarized}/F_{\rm Star}\approx4.1\cdot10^{-3}$. Here, the regions between $\sim$14.5–40 AU (the ring), $\sim$40–70 AU (the gap), and $\sim$70–250 AU (the outer disk) contribute $\sim$49%, $\sim$14%, and $\sim$37% to the integrated disk flux, respectively.
Discussion
==========
The face-on orientation and general extent of the disk have been reported in earlier scattered or polarized light studies [@kuhn2001; @hales2006; @grady2007; @fukagawa2010]. However, no sub-structures were revealed so far.
The bright inner ring
---------------------
We interpret the bright ring as the inner rim of the main disk. Its location fits well to recent model predictions [@meeus2010; @honda2012] but is not in agreement with a rim at 44 AU as suggested by @grady2007. The drop in polarized light inside of the rim is significant and supports the idea that a hole might be present in the inner $\lesssim$20 AU. The brightness dip in the rim at PA$\approx$80$^\circ$ hints to additional substructure or local change in dust grain properties. Although local minima in polarized flux can result from the scattering function of the dust grains even for small disk inclinations [see, e.g., @perrin2009; @muto2012] also the outer disk regions should show a minimum at the same PA, which is not the case.
The annular gap
---------------
To our knowledge this is the first time that such a symmetric and finite gap-like structure has been detected on the surface layer of a protoplanetary disk. This gap does not appear to be completely void of scattering material as the remaining surface brightness is above our detection limits. Seeing a gap-like feature in scattered light could indicate a deficiency in surface density possibly penetrating deeper into the disk mid-plane. If so, one (or several) forming planet(s) could be responsible for the observed morphology. Theory and simulations predict that planets open annular gaps in disks and the gap width and depth depend on the planet mass and on local disk properties [e.g., @lin1986; @bryden1999; @ruge2013]. @jang-condell2012 simulated how planet-induced gaps would alter scattered light images of disks. They showed that depending on the mass of the planet the gap would not necessarily appear empty [cf. @dong2012]. In addition, SPH simulations suggest that localized streams through the disk gap connect the outer and inner disk with the forming planet [e.g., @ruge2013]. One could speculate whether the observed local brightness asymmetries on opposite sides of the gap (see section 3) could be signposts of such features. Finally, according to the simulations by @jang-condell2012, the impact of gaps and planets on the SEDs of star+disk systems is limited. This could explain why previous SED modeling attempts for HD169142 did not reveal this feature.
One could in principle infer mass constraints for the companion from the geometry of the gap [@lin1986; @bryden1999]. However, this exercise requires some specific assumptions about the protoplanetary disk (e.g., viscosity, scale height) which are poorly or not at all constrained empirically. Additionally, not knowing how empty the gap is in the disk mid-plane complicates any attempt to derive a useful companion mass estimate analytically and we leave it out of the scope of this letter. However, a direct search for companions could be attempted with high-contrast imaging instruments even though this might be challenging if the planet is still embedded [@wolf2005; @wolf2007].
We strongly emphasize that there are alternative explanations for the observed disk morphology without invoking an embedded planet. For instance, the disk rim could cast some shadow onto the disk surface. Puffed-up disk rims have been predicted at least for the inner edges of ’typical’ protoplanetary disks even if the exact geometry and the question whether these rims can cause disk shadows is still debated [see, @dullemond2010 for a review]. Concerning HD169142, the disk model by @honda2012 does *not* predict a shadowing disk rim. However, a combination of radially varying dust grain properties with dust settling might possibly be able to create a shadow-like feature without the need of a local deficiency in the surface density. Dust continuum observations with sufficient spatial resolution can shed light on this open question.
The disk surface brightness
---------------------------
The single power-law index we derive from fitting the surface brightness profile agrees with the results from @fukagawa2010, who found $S \propto r^{-3.4\pm0.4}$ between 120 – 200 AU in $H$ band scattered light images. @grady2007 found a slightly shallower profile with $S \propto r^{-3.0\pm0.1}$ between $\sim$70 – 200 AU using HST/NICMOS at 1.1$\mu$m.
The errors in our profile fits justify the use of a broken power law. Broken power-law surface brightness profiles are observed in resolved scattered light images of debris disks, where the inner disk regions show a shallower profile due to the presence of larger grains while the outer disk profile is steeper and dominated by blown-out smaller grains [e.g., @augereau2001b; @liu2004]. The disk around HD169142 is certainly still gas rich, but the observed brightness profile could be indicative of different dust grain populations. However, also changes in the dust density distribution could lead to the observed effect. Until now, all studies focusing on the gas content or the dust continuum of the HD169142 disk did not have enough spatial resolution to securely detect gaps or discontinuities in the disk structure. Of particular interest would be to see whether the broken power law discovered here is accompanied with a radial change in the dust-to-gas ratio.
Comparing the disk surface brightness derived by @fukagawa2010 with our values between 120 and 200 AU suggests that roughly $\sim$40% of the scattered light of HD169142 seems to be polarized in the $H$ band. However, as the errors bars in both studies are large, this value for the polarization fraction should be taken as a rough indication. At this level the polarization fraction appears to be higher than that for the HD100546 disk [$\sim$14%; @quanz2011b] but comparable to the value for the transition disk around AB Aur [@perrin2009] for scattering angles around 90$^\circ$.
Conclusions
===========
Our PDI images of the protoplanetary disk around HD169142 reveal previously undetected features and emphasize the power of this observing technique. With PDI on ground-based 8-m class telescopes the inner few tens of AU of the dusty surface layer of protoplanetary disks can be reliably resolved opening up a new era of disk imaging. Several PDI studies in the last few years have shown that sub-structures in disks (e.g., holes, gaps, spiral arms) appear to be rather the rule than the exception [e.g., @quanz2011b; @hashimoto2011; @muto2012; @hashimoto2012; @mayama2012; @grady2013].
In the case of HD169142, our images support previous model results suggesting that the inner $\lesssim$20 AU are largely devoid of scattering dust particles. In addition, our images reveal an annular gap roughy between $\sim$40–70 AU, localized brightness asymmetries on opposite sides of this gap and a discontinuity in the radial surface brightness profile at $\sim$120 AU. Our data alone do not allow us to unambiguously reveal the physical nature of the annular gap, i.e, whether it is a gap in surface density possible induced by forming planet, a shadowing effect, or combination of radially changing dust properties and dust settling. However, our results make HD169142 an excellent target for follow-up studies aiming at a better understanding of disk physics and potentially ongoing planet formation.
This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. We thank the staff at VLT for their excellent support during the observations and M.R. Meyer, F. Meru and O. Pani[ć]{} for useful discussions. This work is supported by the Swiss National Science Foundation.
[*Facilities:*]{}
[43]{} natexlab\#1[\#1]{}
, J. C., [Nelson]{}, R. P., [Lagrange]{}, A. M., [Papaloizou]{}, J. C. B., & [Mouillet]{}, D. 2001, , 370, 447
, P. F. C. & [Djie]{}, H. R. E. T. A. 2006, , 456, 1045
, G., [Chen]{}, X., [Lin]{}, D. N. C., [Nelson]{}, R. P., & [Papaloizou]{}, J. C. B. 1999, , 514, 344
, R. M., [Skrutskie]{}, M. F., [van Dyk]{}, S., [Beichman]{}, C. A., [Carpenter]{}, J. M., [Chester]{}, T., [Cambresy]{}, L., [Evans]{}, T., [Fowler]{}, J., [Gizis]{}, J., [Howard]{}, E., [Huchra]{}, J., [Jarrett]{}, T., [Kopan]{}, E. L., [Kirkpatrick]{}, J. D., [Light]{}, R. M., [Marsh]{}, K. A., [McCallon]{}, H., [Schneider]{}, S., [Stiening]{}, R., [Sykes]{}, M., [Weinberg]{}, M., [Wheaton]{}, W. A., [Wheelock]{}, S., & [Zacarias]{}, N. 2003, [2MASS All Sky Catalog of point sources.]{} (The IRSA 2MASS All-Sky Point Source Catalog, NASA/IPAC Infrared Science Archive. http://irsa.ipac.caltech.edu/applications/Gator/)
, W. R. F., [Torrelles]{}, J. M., [Osorio]{}, M., [Calvet]{}, N., & [Anglada]{}, G. 2006, , 365, 1283
, R., [Rafikov]{}, R., [Zhu]{}, Z., [Hartmann]{}, L., [Whitney]{}, B., [Brandt]{}, T., [Muto]{}, T., [Hashimoto]{}, J., [Grady]{}, C., [Follette]{}, K., [Kuzuhara]{}, M., [Tanii]{}, R., [Itoh]{}, Y., [Thalmann]{}, C., [Wisniewski]{}, J., [Mayama]{}, S., [Janson]{}, M., [Abe]{}, L., [Brandner]{}, W., [Carson]{}, J., [Egner]{}, S., [Feldt]{}, M., [Goto]{}, M., [Guyon]{}, O., [Hayano]{}, Y., [Hayashi]{}, M., [Hayashi]{}, S., [Henning]{}, T., [Hodapp]{}, K. W., [Honda]{}, M., [Inutsuka]{}, S., [Ishii]{}, M., [Iye]{}, M., [Kandori]{}, R., [Knapp]{}, G. R., [Kudo]{}, T., [Kusakabe]{}, N., [Matsuo]{}, T., [McElwain]{}, M. W., [Miyama]{}, S., [Morino]{}, J.-I., [Moro-Martin]{}, A., [Nishimura]{}, T., [Pyo]{}, T.-S., [Suto]{}, H., [Suzuki]{}, R., [Takami]{}, M., [Takato]{}, N., [Terada]{}, H., [Tomono]{}, D., [Turner]{}, E. L., [Watanabe]{}, M., [Yamada]{}, T., [Takami]{}, H., [Usuda]{}, T., & [Tamura]{}, M. 2012, , 750, 161
, C. P. & [Monnier]{}, J. D. 2010, , 48, 205
, M., [Tamura]{}, M., [Itoh]{}, Y., [Oasa]{}, Y., [Kudo]{}, T., [Hayashi]{}, S. S., [Kato]{}, E., [Ootsubo]{}, T., [Itoh]{}, Y., [Shibai]{}, H., & [Hayashi]{}, M. 2010, , 62, 347
, C. A., [Muto]{}, T., [Hashimoto]{}, J., [Fukagawa]{}, M., [Currie]{}, T., [Biller]{}, B., [Thalmann]{}, C., [Sitko]{}, M. L., [Russell]{}, R., [Wisniewski]{}, J., [Dong]{}, R., [Kwon]{}, J., [Sai]{}, S., [Hornbeck]{}, J., [Schneider]{}, G., [Hines]{}, D., [Moro Mart[í]{}n]{}, A., [Feldt]{}, M., [Henning]{}, T., [Pott]{}, J.-U., [Bonnefoy]{}, M., [Bouwman]{}, J., [Lacour]{}, S., [Mueller]{}, A., [Juh[á]{}sz]{}, A., [Crida]{}, A., [Chauvin]{}, G., [Andrews]{}, S., [Wilner]{}, D., [Kraus]{}, A., [Dahm]{}, S., [Robitaille]{}, T., [Jang-Condell]{}, H., [Abe]{}, L., [Akiyama]{}, E., [Brandner]{}, W., [Brandt]{}, T., [Carson]{}, J., [Egner]{}, S., [Follette]{}, K. B., [Goto]{}, M., [Guyon]{}, O., [Hayano]{}, Y., [Hayashi]{}, M., [Hayashi]{}, S., [Hodapp]{}, K., [Ishii]{}, M., [Iye]{}, M., [Janson]{}, M., [Kandori]{}, R., [Knapp]{}, G., [Kudo]{}, T., [Kusakabe]{}, N., [Kuzuhara]{}, M., [Mayama]{}, S., [McElwain]{}, M., [Matsuo]{}, T., [Miyama]{}, S., [Morino]{}, J.-I., [Nishimura]{}, T., [Pyo]{}, T.-S., [Serabyn]{}, G., [Suto]{}, H., [Suzuki]{}, R., [Takami]{}, M., [Takato]{}, N., [Terada]{}, H., [Tomono]{}, D., [Turner]{}, E., [Watanabe]{}, M., [Yamada]{}, T., [Takami]{}, H., [Usuda]{}, T., & [Tamura]{}, M. 2013, , 762, 48
, C. A., [Schneider]{}, G., [Hamaguchi]{}, K., [Sitko]{}, M. L., [Carpenter]{}, W. J., [Hines]{}, D., [Collins]{}, K. A., [Williger]{}, G. M., [Woodgate]{}, B. E., [Henning]{}, T., [M[é]{}nard]{}, F., [Wilner]{}, D., [Petre]{}, R., [Palunas]{}, P., [Quirrenbach]{}, A., [Nuth]{}, III, J. A., [Silverstone]{}, M. D., & [Kim]{}, J. S. 2007, , 665, 1391
, M. M., [Alencar]{}, S. H. P., [Corradi]{}, W. J. B., & [Vieira]{}, S. L. A. 2006, , 457, 581
, E., [Natta]{}, A., [Testi]{}, L., & [Carbillet]{}, M. 2006, , 449, 1067
, A. S., [Gledhill]{}, T. M., [Barlow]{}, M. J., & [Lowe]{}, K. T. E. 2006, , 365, 1348
, J., [Dong]{}, R., [Kudo]{}, T., [Honda]{}, M., [McClure]{}, M. K., [Zhu]{}, Z., [Muto]{}, T., [Wisniewski]{}, J., [Abe]{}, L., [Brandner]{}, W., [Brandt]{}, T., [Carson]{}, J., [Egner]{}, S., [Feldt]{}, M., [Fukagawa]{}, M., [Goto]{}, M., [Grady]{}, C. A., [Guyon]{}, O., [Hayano]{}, Y., [Hayashi]{}, M., [Hayashi]{}, S., [Henning]{}, T., [Hodapp]{}, K., [Ishii]{}, M., [Iye]{}, M., [Janson]{}, M., [Kandori]{}, R., [Knapp]{}, G., [Kusakabe]{}, N., [Kuzuhara]{}, M., [Kwon]{}, J., [Matsuo]{}, T., [Mayama]{}, S., [McElwain]{}, M. W., [Miyama]{}, S., [Morino]{}, J.-I., [Moro-Martin]{}, A., [Nishimura]{}, T., [Pyo]{}, T.-S., [Serabyn]{}, G., [Suenaga]{}, T., [Suto]{}, H., [Suzuki]{}, R., [Takahashi]{}, Y., [Takami]{}, M., [Takato]{}, N., [Terada]{}, H., [Thalmann]{}, C., [Tomono]{}, D., [Turner]{}, E. L., [Watanabe]{}, M., [Yamada]{}, T., [Takami]{}, H., [Usuda]{}, T., & [Tamura]{}, M. 2012, , 758, L19
, J., [Tamura]{}, M., [Muto]{}, T., [Kudo]{}, T., [Fukagawa]{}, M., [Fukue]{}, T., [Goto]{}, M., [Grady]{}, C. A., [Henning]{}, T., [Hodapp]{}, K., [Honda]{}, M., [Inutsuka]{}, S., [Kokubo]{}, E., [Knapp]{}, G., [McElwain]{}, M. W., [Momose]{}, M., [Ohashi]{}, N., [Okamoto]{}, Y. K., [Takami]{}, M., [Turner]{}, E. L., [Wisniewski]{}, J., [Janson]{}, M., [Abe]{}, L., [Brandner]{}, W., [Carson]{}, J., [Egner]{}, S., [Feldt]{}, M., [Golota]{}, T., [Guyon]{}, O., [Hayano]{}, Y., [Hayashi]{}, M., [Hayashi]{}, S., [Ishii]{}, M., [Kandori]{}, R., [Kusakabe]{}, N., [Matsuo]{}, T., [Mayama]{}, S., [Miyama]{}, S., [Morino]{}, J.-I., [Moro-Martin]{}, A., [Nishimura]{}, T., [Pyo]{}, T.-S., [Suto]{}, H., [Suzuki]{}, R., [Takato]{}, N., [Terada]{}, H., [Thalmann]{}, C., [Tomono]{}, D., [Watanabe]{}, M., [Yamada]{}, T., [Takami]{}, H., & [Usuda]{}, T. 2011, , 729, L17
, M., [Maaskant]{}, K., [Okamoto]{}, Y. K., [Kataza]{}, H., [Fukagawa]{}, M., [Waters]{}, L. B. F. M., [Dominik]{}, C., [Tielens]{}, A. G. G. M., [Mulders]{}, G. D., [Min]{}, M., [Yamashita]{}, T., [Fujiyoshi]{}, T., [Miyata]{}, T., [Sako]{}, S., [Sakon]{}, I., [Fujiwara]{}, H., & [Onaka]{}, T. 2012, , 752, 143
, H. & [Turner]{}, N. J. 2012, , 749, 153
, J. R., [Potter]{}, D., & [Parise]{}, B. 2001, , 553, L189
, N., [Grady]{}, C. A., [Sitko]{}, M. L., [Hashimoto]{}, J., [Kudo]{}, T., [Fukagawa]{}, M., [Muto]{}, T., [Wisniewski]{}, J. P., [Min]{}, M., [Mayama]{}, S., [Werren]{}, C., [Day]{}, A. N., [Beerman]{}, L. C., [Lynch]{}, D. K., [Russell]{}, R. W., [Brafford]{}, S. M., [Kuzuhara]{}, M., [Brandt]{}, T. D., [Abe]{}, L., [Brandner]{}, W., [Carson]{}, J., [Egner]{}, S., [Feldt]{}, M., [Goto]{}, M., [Guyon]{}, O., [Hayano]{}, Y., [Hayashi]{}, M., [Hayashi]{}, S. S., [Henning]{}, T., [Hodapp]{}, K. W., [Ishii]{}, M., [Iye]{}, M., [Janson]{}, M., [Kandori]{}, R., [Knapp]{}, G. R., [Matsuo]{}, T., [McElwain]{}, M. W., [Miyama]{}, S., [Morino]{}, J.-I., [Moro-Martin]{}, A., [Nishimura]{}, T., [Pyo]{}, T.-S., [Suto]{}, H., [Suzuki]{}, R., [Takami]{}, M., [Takato]{}, N., [Terada]{}, H., [Thalmann]{}, C., [Tomono]{}, D., [Turner]{}, E. L., [Watanabe]{}, M., [Yamada]{}, T., [Takami]{}, H., [Usuda]{}, T., & [Tamura]{}, M. 2012, , 753, 153
, R., [Hartung]{}, M., [Brandner]{}, W., [Finger]{}, G., [Hubin]{}, N. N., [Lacombe]{}, F., [Lagrange]{}, A., [Lehnert]{}, M. D., [Moorwood]{}, A. F. M., & [Mouillet]{}, D. 2003, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4841, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. [M. Iye & A. F. M. Moorwood]{}, 944–952
, D. N. C. & [Papaloizou]{}, J. 1986, , 309, 846
, M. C. 2004, Science, 305, 1442
, N., [Telesco]{}, C. M., [Fisher]{}, R. S., & [Packham]{}, C. 2011, , 737, 57
, S., [Hashimoto]{}, J., [Muto]{}, T., [Tsukagoshi]{}, T., [Kusakabe]{}, N., [Kuzuhara]{}, M., [Takahashi]{}, Y., [Kudo]{}, T., [Dong]{}, R., [Fukagawa]{}, M., [Takami]{}, M., [Momose]{}, M., [Wisniewski]{}, J. P., [Follette]{}, K., [Abe]{}, L., [Akiyama]{}, E., [Brandner]{}, W., [Brandt]{}, T., [Carson]{}, J., [Egner]{}, S., [Feldt]{}, M., [Goto]{}, M., [Grady]{}, C. A., [Guyon]{}, O., [Hayano]{}, Y., [Hayashi]{}, M., [Hayashi]{}, S., [Henning]{}, T., [Hodapp]{}, K. W., [Ishii]{}, M., [Iye]{}, M., [Janson]{}, M., [Kandori]{}, R., [Kwon]{}, J., [Knapp]{}, G. R., [Matsuo]{}, T., [McElwain]{}, M. W., [Miyama]{}, S., [Morino]{}, J.-I., [Moro-Martin]{}, A., [Nishimura]{}, T., [Pyo]{}, T.-S., [Serabyn]{}, E., [Suto]{}, H., [Suzuki]{}, R., [Takato]{}, N., [Terada]{}, H., [Thalmann]{}, C., [Tomono]{}, D., [Turner]{}, E. L., [Watanabe]{}, M., [Yamada]{}, T., [Takami]{}, H., [Usuda]{}, T., & [Tamura]{}, M. 2012, , 760, L26
, G., [Pinte]{}, C., [Woitke]{}, P., [Montesinos]{}, B., [Mendigut[í]{}a]{}, I., [Riviere-Marichalar]{}, P., [Eiroa]{}, C., [Mathews]{}, G. S., [Vandenbussche]{}, B., [Howard]{}, C. D., [Roberge]{}, A., [Sandell]{}, G., [Duch[ê]{}ne]{}, G., [M[é]{}nard]{}, F., [Grady]{}, C. A., [Dent]{}, W. R. F., [Kamp]{}, I., [Augereau]{}, J. C., [Thi]{}, W. F., [Tilling]{}, I., [Alacid]{}, J. M., [Andrews]{}, S., [Ardila]{}, D. R., [Aresu]{}, G., [Barrado]{}, D., [Brittain]{}, S., [Ciardi]{}, D. R., [Danchi]{}, W., [Fedele]{}, D., [de Gregorio-Monsalvo]{}, I., [Heras]{}, A., [Huelamo]{}, N., [Krivov]{}, A., [Lebreton]{}, J., [Liseau]{}, R., [Martin-Zaidi]{}, C., [Mora]{}, A., [Morales-Calderon]{}, M., [Nomura]{}, H., [Pantin]{}, E., [Pascucci]{}, I., [Phillips]{}, N., [Podio]{}, L., [Poelman]{}, D. R., [Ramsay]{}, S., [Riaz]{}, B., [Rice]{}, K., [Solano]{}, E., [Walker]{}, H., [White]{}, G. J., [Williams]{}, J. P., & [Wright]{}, G. 2010, , 518, L124
, G., [Waters]{}, L. B. F. M., [Bouwman]{}, J., [van den Ancker]{}, M. E., [Waelkens]{}, C., & [Malfait]{}, K. 2001, , 365, 476
, T., [Grady]{}, C. A., [Hashimoto]{}, J., [Fukagawa]{}, M., [Hornbeck]{}, J. B., [Sitko]{}, M., [Russell]{}, R., [Werren]{}, C., [Cur[é]{}]{}, M., [Currie]{}, T., [Ohashi]{}, N., [Okamoto]{}, Y., [Momose]{}, M., [Honda]{}, M., [Inutsuka]{}, S., [Takeuchi]{}, T., [Dong]{}, R., [Abe]{}, L., [Brandner]{}, W., [Brandt]{}, T., [Carson]{}, J., [Egner]{}, S., [Feldt]{}, M., [Fukue]{}, T., [Goto]{}, M., [Guyon]{}, O., [Hayano]{}, Y., [Hayashi]{}, M., [Hayashi]{}, S., [Henning]{}, T., [Hodapp]{}, K. W., [Ishii]{}, M., [Iye]{}, M., [Janson]{}, M., [Kandori]{}, R., [Knapp]{}, G. R., [Kudo]{}, T., [Kusakabe]{}, N., [Kuzuhara]{}, M., [Matsuo]{}, T., [Mayama]{}, S., [McElwain]{}, M. W., [Miyama]{}, S., [Morino]{}, J.-I., [Moro-Martin]{}, A., [Nishimura]{}, T., [Pyo]{}, T.-S., [Serabyn]{}, E., [Suto]{}, H., [Suzuki]{}, R., [Takami]{}, M., [Takato]{}, N., [Terada]{}, H., [Thalmann]{}, C., [Tomono]{}, D., [Turner]{}, E. L., [Watanabe]{}, M., [Wisniewski]{}, J. P., [Yamada]{}, T., [Takami]{}, H., [Usuda]{}, T., & [Tamura]{}, M. 2012, , 748, L22
, O., [Hogerheijde]{}, M. R., [Wilner]{}, D., & [Qi]{}, C. 2008, , 491, 219
, M. D., [Schneider]{}, G., [Duchene]{}, G., [Pinte]{}, C., [Grady]{}, C. A., [Wisniewski]{}, J. P., & [Hines]{}, D. C. 2009, , 707, L132
, S. P., [Birkmann]{}, S. M., [Apai]{}, D., [Wolf]{}, S., & [Henning]{}, T. 2012, , 538, A92
, S. P., [Schmid]{}, H. M., [Geissler]{}, K., [Meyer]{}, M. R., [Henning]{}, T., [Brandner]{}, W., & [Wolf]{}, S. 2011, , 738, 23
, A., [Lisanti]{}, M., [Wilner]{}, D. J., [Qi]{}, C., & [Hogerheijde]{}, M. 2006, , 131, 2290
, G., [Lacombe]{}, F., [Puget]{}, P., [Hubin]{}, N. N., [Gendron]{}, E., [Fusco]{}, T., [Arsenault]{}, R., [Charton]{}, J., [Feautrier]{}, P., [Gigan]{}, P., [Kern]{}, P. Y., [Lagrange]{}, A., [Madec]{}, P., [Mouillet]{}, D., [Rabaud]{}, D., [Rabou]{}, P., [Stadler]{}, E., & [Zins]{}, G. 2003, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4839, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. [P. L. Wizinowich & D. Bonaccini]{}, 140–149
, J. P., [Wolf]{}, S., [Uribe]{}, A. L., & [Klahr]{}, H. H. 2013, , 549, A97
, G., [Weintraub]{}, D. A., & [Hamidouche]{}, M. 2011, , 727, 26
, H. M., [Joos]{}, F., & [Tschan]{}, D. 2006, , 452, 657
, R. J., [Skinner]{}, C. J., [Barlow]{}, M. J., & [Mannings]{}, V. 1996, , 279, 915
, R., [Itoh]{}, Y., [Kudo]{}, T., [Hioki]{}, T., [Oasa]{}, Y., [Gupta]{}, R., [Sen]{}, A. K., [Wisniewski]{}, J. P., [Muto]{}, T., [Grady]{}, C. A., [Hashimoto]{}, J., [Fukagawa]{}, M., [Mayama]{}, S., [Hornbeck]{}, J., [Sitko]{}, M., [Russell]{}, R., [Werren]{}, C., [Cure]{}, M., [Currie]{}, T., [Ohashi]{}, N., [Okamoto]{}, Y., [Momose]{}, M., [Honda]{}, M., [Inutsuka]{}, S.-I., [Takeuchi]{}, T., [Dong]{}, R., [Abe]{}, L., [Brandner]{}, W., [Brandt]{}, T., [Carson]{}, J., [Egner]{}, S., [Feldt]{}, M., [Fukue]{}, T., [Goto]{}, M., [Guyon]{}, O., [Hayano]{}, Y., [Hayashi]{}, M., [Hayashi]{}, S. S., [Henning]{}, T., [Hodapp]{}, K. W., [Ishii]{}, M., [Iye]{}, M., [Janson]{}, M., [Kandori]{}, R., [Knapp]{}, G. P., [Kusakabe]{}, N., [Kuzuhara]{}, M., [Matsuo]{}, T., [McElwain]{}, M. W., [Miyama]{}, S., [Morino]{}, J.-I., [Moro-Martin]{}, A., [Nishimura]{}, T., [Pyo]{}, T.-S., [Serabyn]{}, G., [Suto]{}, H., [Suzuki]{}, R., [Takami]{}, M., [Takato]{}, N., [Terada]{}, H., [Thalmann]{}, C., [Tomono]{}, D., [Turner]{}, E. L., [Watanabe]{}, M., [Yamada]{}, T., [Takami]{}, H., [Usuda]{}, T., & [Tamura]{}, M. 2012, ArXiv e-prints
, J. 1996, [Astronomical Polarimetry]{}
, R., [Min]{}, M., [Waters]{}, L. B. F. M., [de Koter]{}, A., [Dominik]{}, C., [van den Ancker]{}, M. E., & [Bouwman]{}, J. 2005, , 437, 189
, G., [Eckart]{}, A., [Buchholz]{}, R. M., [Zamaninasab]{}, M., [Lenzen]{}, R., [Sch[ö]{}del]{}, R., [Araujo]{}, C., [Sabha]{}, N., [Bremer]{}, M., [Karas]{}, V., [Straubmeier]{}, C., & [Muzic]{}, K. 2011, , 525, A130
, S. & [D’Angelo]{}, G. 2005, , 619, 1114
, S., [Moro-Mart[í]{}n]{}, A., & [D’Angelo]{}, G. 2007, , 55, 569
[^1]: The results were consistent with those using the approach described in @quanz2011b.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Saul Cohen\
Department of Physics, Boston University, Boston, MA 02215
- |
John M. Bulava, Justin Foley, Colin Morningstar and Ricky Wong\
Department of Physics, Carnegie Mellon University, Pittsburgh, PA 15213
- |
Robert G. Edwards, Bálint Joó, David G. Richards\
Thomas Jefferson National Accelerator Facility, Newport News, VA 23606
- |
Eric Engelson and Stephen J. Wallace\
Department of Physics, University of Maryland, College Park, MD 20742, USA
- |
K. Jimmy Juge\
Department of Physics, University of the Pacific, Stockton, CA 95211, USA
- |
\
Department of Physics, University of Washington, Seattle, WA 98195-1560\
E-mail:
- |
Nilmani Mathur\
Department of Theoretical Physics, Tata Institute of Fundamental Research, India
- |
Michael J. Peardon and Sinéad M. Ryan\
School of Mathematics, Trinity College, Dublin 2, Ireland
title: 'Excited-Nucleon Spectroscopy with 2+1 Fermion Flavors'
---
Introduction
============
A new generation of experiments devoted to hadron spectroscopy, including GlueX at Jefferson Lab, PANDA at GSI/FAIR and BES III intend to make measurements with unprecedented precision and in previously unexplored mass ranges and quantum numbers. Both in meson and baryon spectroscopy, there are many experimentally observed excited states whose physical properties are poorly understood and could use theoretical input from lattice QCD to solidify their identification. Aside from masses, other excited-state quantities that could be computed on the lattice, such as form factors and coupling constants, would be useful to groups such as the Excited Baryon Analysis Center (EBAC) at Jefferson Lab, where dynamical reaction models have been developed to interpret experimentally observed properties of excited nucleons in terms of QCD [@ebac]. In certain cases, input from the lattice may be helpful in determining the composition of controversial states, which may be interpreted as ordinary hadrons, tetra- or pentaquarks, hadronic molecules or unbound resonances.
Among the excited nucleon states, the nature of the Roper resonance, $N(1440)\ P_{11}$, has been the subject of interest since its discovery in the 1960s. It is quite surprising that the rest energy of the first excited state of the nucleon is less than the ground-state energy of the nucleon’s negative-parity partner, the $N(1535)\ S_{11}$ [@Yao:2006px], a phenomenon never observed in meson systems. There are several interpretations of the Roper state, for example, as the hybrid state that couples predominantly to QCD currents with some gluonic contribution or as a five-quark (meson-baryon) state [@others]
Due to the greater impact of systematic errors, such as finite-volume effects and discretization errors, on excited states, there is an essential need to examine calculations of excited masses more carefully during analysis. It is common in such calculations to use the ground-state masses, such as nucleon mass, as a starting point for analysis but without further checks of the consistency of the approach. The nucleon mass has been demonstrated with good consistency (a great demonstration of the universality of lattice QCD using different fermion actions by various groups with independent analyses) in both quenched and dynamical $N_f=2+1$ cases. For dynamical ensembles, there are more variables in terms of algorithm, scale setting, etc., but an approximate universality is achieved among different groups (see middle panel of Fig. \[fig:Resonance-FV\]). However, beyond the ground state, there exists a diverse distribution of excited-state masses as functions of $m_\pi$. There are big discrepancies in the calculated nucleon first-excited mass, creating an apparently chaotic atmosphere, as shown in the left panel of Fig. \[fig:Resonance-FV\]. Note that the errorbars here are just statistical, none of the systematic errors (such as quenching, finite-volume effect, etc.) are estimated. We re-address the same issue by modifying the axes to be in terms of the dimensionless quantity $ML$ (as shown in right panel of Fig. \[fig:Resonance-FV\]); we find that the Roper masses are roughly inversely proportional to lattice size. Now we see a better agreement (or universality) among the lattice QCD Roper-mass calculations; most of the Roper masses agree within 2 standard deviations of the numbers in Ref. [@Lasscock:2007ce]. This suggests that finite-volume effects can be more severe for excited states than ground states and that careful examination of such systematic errors is crucial. Similarly, we cannot ignore other systematic errors that may arise even if we did not observe the effect in the ground state. Excited-state analyses should proceed with greater caution.
To tackle the challenge of extracting reliable excited-state energies, the Hadron Spectrum Collaboration (HSC) has been devoting resources and effort into resolving the mystery from the fundamental point of view. It is difficult to extract the excited state reliably without sufficient information carried in the nucleon correlators. We need correlators which clearly contain the masses of specific quantum numbers and allow us access to data points that reach higher excited states before the signal decays exponentially away. To successfully and reliably extract these excited-state energies, we need better resolution in temporal direction (large superfine isotropic lattices or anisotropic lattices), separation of the signals for individual states (variational method), and most importantly, operators that have good overlap with various desired quantum numbers (cubic-group irreducible representations and some way to put in many operators that contribute unique signals).
![\[fig:Resonance-FV\] Summary of published quenched ($N_f=0$) lattice QCD calculations of the nucleon and Roper masses in GeV (left) and in terms of the dimensionless product of the Roper mass and lattice size $L$ (right). The middle panel shows the dimensionless products of lattice nucleon mass and $r_0$ as a function of $m_\pi r_0$ from $N_f=2+1$ ensemble. ](nf0NR.eps "fig:"){width="32.00000%"} ![\[fig:Resonance-FV\] Summary of published quenched ($N_f=0$) lattice QCD calculations of the nucleon and Roper masses in GeV (left) and in terms of the dimensionless product of the Roper mass and lattice size $L$ (right). The middle panel shows the dimensionless products of lattice nucleon mass and $r_0$ as a function of $m_\pi r_0$ from $N_f=2+1$ ensemble. ](r0MNplotHSC.eps "fig:"){width="32.00000%"} ![\[fig:Resonance-FV\] Summary of published quenched ($N_f=0$) lattice QCD calculations of the nucleon and Roper masses in GeV (left) and in terms of the dimensionless product of the Roper mass and lattice size $L$ (right). The middle panel shows the dimensionless products of lattice nucleon mass and $r_0$ as a function of $m_\pi r_0$ from $N_f=2+1$ ensemble. ](MLFVplot.eps "fig:"){width="32.00000%"}
$N_f=2+1$ Clover Anisotropic Lattices and Updates
==================================================
To improve our ability to extract higher excited energies (even with the aid of the variational method), we need to use finer temporal lattice spacings. This allows us to use as many time slices of information as possible to reconstruct more accurately the signal due to a particular state, since the signal in the Euclidean space declines exponentially as $E t$. In principle, we could use very fine lattices while keeping the volume big enough to avoid the “squeezing” systematic effect; however, the computational cost scales significantly (power of 5–6) with the lattice discretization. We adopt the approach of anisotropic clover lattices to keep the spatial lattice spacing coarse, avoiding finite-volume systematic error, and to make the temporal lattice spacing fine enough to extract towers of nucleon excited states. However, tuning the $O(a)$-improved parameters for fermion actions in the dynamical gauge generation correctly is a much more difficult task for anisotropic lattices. We use Symanzik-improved gauge action and clover fermion action with 3-dimensional stout-link smeared gauge fields; the gauge ensemble is generated using the (R)HMC algorithm. Our spatial lattice spacing is $a_s=0.1227(8)$ fm (determined using $m_\Omega$), and the renormalized anisotropy $\xi_R=a_s/a_t$ is $3.5$. Ref. [@Edwards:2008ja] shows a detailed study of the dynamical anisotropic lattice parameter settings, and Ref. [@Lin:2008pr] reports basic lattice properties along with the ground-state hadron spectrum. Since then, we have also moved on to generate $m_\pi=230$ MeV $24^3$ and $32^3$ lattices. A stream of $m_\pi=180$ MeV ensembles is slowly progressing (for both $24^3$ and $32^3$ volumes). In Ref. [@Lin:2008pr], we demonstrated and compared various strange-quark mass-setting approaches. We tuned the parameter $s_\Omega=\frac{9(2m_K^2-m_\pi^2)}{4m_\Omega^2}$ as close as possible to the corresponding experimental value at the $SU(3)_f$-symmetric point (in our case, it is the 875-MeV–pion ensemble). When we reduce the sea-quark mass, the $s_\Omega$ parameter remains roughly constant as low as the 383 MeV pion ensemble. (Furthermore, the $s_\Omega$ parameter is more sensitive to the strange-quark mass than other methods, such as $J$-parameter. Since it is a ratio, there is no need to worry about estimating the shift in the lattice spacing when decreasing the sea-quark mass or generating more statistics.) We measure the same quantity on the 230-MeV $24^3$ ensemble and $s_\Omega$ does not deviate from the expected value. The $s_\Omega$ parameter is a stable and useful observable for setting the strange-quark mass; we have not observed any notable deviation for a wide range of sea-quark masses.
In Ref. [@Lin:2008pr], we reported a naive extrapolation in mass ratios (using $m_\Omega$ or $m_\Xi$ as the reference mass) to obtain the meson and baryon masses. Note that at each sea-strange mass, we only use the larger-volume set for the lightest pion mass to avoid large finite-volume effects. The advantage of using the mass ratios instead of the masses is that we gain smaller statistical (and possibly systematic) errors than the mass itself due to cancellation of systematics and removal of the ambiguity associated with setting the lattice spacing. Using the same data, we update the mass-ratio chiral extrapolations by modifying the next-to-leading-order heavy-baryon chiral perturbation theory. We find the finite-volume corrections are negligible according to ChPT estimates and the extrapolated baryon masses agree with experimental values significantly better than with a linear extrapolation. Preliminary results are shown in Fig. \[fig:ChPTMassRatio\].
![\[fig:ChPTMassRatio\] (Left panel) The location of the dynamical ensembles used in this work in the $s_\Omega$-$l_\Omega$ plane. The leftmost circle (black) indicates the physical point {$l_\Omega^{\rm phys}$, $s_\Omega^{\rm phys}$}, while the red, green and blue points are the lattices with $a_tm_s=-0.0540, -0.0618, -0.0743$ respectively in Ref. [@Lin:2008pr]. The horizontal dashed (pink) line indicates constant $s_\Omega$ at the physical point, and the diagonal line indicates three-flavor degenerate theories. The purple points are from recent measurements of the 230 MeV ensemble, showing that it does not deviate away from the $N_f=3$ value. Mass-ratio chiral extrapolations as functions of the $l_\Xi$ and $s_\Xi$ for octets (center panel) and $l_\Omega$ and $s_\Omega$ for decuplets (right panel). The lines indicate the “projected” leading chiral extrapolation fits in $l_\Xi$ and $s_\Xi$ ($l_\Omega$ and $s_\Omega$) while keeping the other one fixed. The black (circular) point is the extrapolated point at physical $l_\Xi$ and $s_\Xi$ ($l_\Omega$ and $s_\Omega$). The masses for $N$, $\Sigma$ and $\Lambda$ are 0.962(29), 1.203(11), 1.122(16) GeV, respectively, and the masses for $\Delta$, $\Sigma^*$ and $\Xi^*$ are 1.275(38), 1.426(25) and 1.580(13) GeV, respectively. ](nnplot2.eps "fig:"){width="32.00000%"} ![\[fig:ChPTMassRatio\] (Left panel) The location of the dynamical ensembles used in this work in the $s_\Omega$-$l_\Omega$ plane. The leftmost circle (black) indicates the physical point {$l_\Omega^{\rm phys}$, $s_\Omega^{\rm phys}$}, while the red, green and blue points are the lattices with $a_tm_s=-0.0540, -0.0618, -0.0743$ respectively in Ref. [@Lin:2008pr]. The horizontal dashed (pink) line indicates constant $s_\Omega$ at the physical point, and the diagonal line indicates three-flavor degenerate theories. The purple points are from recent measurements of the 230 MeV ensemble, showing that it does not deviate away from the $N_f=3$ value. Mass-ratio chiral extrapolations as functions of the $l_\Xi$ and $s_\Xi$ for octets (center panel) and $l_\Omega$ and $s_\Omega$ for decuplets (right panel). The lines indicate the “projected” leading chiral extrapolation fits in $l_\Xi$ and $s_\Xi$ ($l_\Omega$ and $s_\Omega$) while keeping the other one fixed. The black (circular) point is the extrapolated point at physical $l_\Xi$ and $s_\Xi$ ($l_\Omega$ and $s_\Omega$). The masses for $N$, $\Sigma$ and $\Lambda$ are 0.962(29), 1.203(11), 1.122(16) GeV, respectively, and the masses for $\Delta$, $\Sigma^*$ and $\Xi^*$ are 1.275(38), 1.426(25) and 1.580(13) GeV, respectively. ](octExp.eps "fig:"){width="32.00000%"} ![\[fig:ChPTMassRatio\] (Left panel) The location of the dynamical ensembles used in this work in the $s_\Omega$-$l_\Omega$ plane. The leftmost circle (black) indicates the physical point {$l_\Omega^{\rm phys}$, $s_\Omega^{\rm phys}$}, while the red, green and blue points are the lattices with $a_tm_s=-0.0540, -0.0618, -0.0743$ respectively in Ref. [@Lin:2008pr]. The horizontal dashed (pink) line indicates constant $s_\Omega$ at the physical point, and the diagonal line indicates three-flavor degenerate theories. The purple points are from recent measurements of the 230 MeV ensemble, showing that it does not deviate away from the $N_f=3$ value. Mass-ratio chiral extrapolations as functions of the $l_\Xi$ and $s_\Xi$ for octets (center panel) and $l_\Omega$ and $s_\Omega$ for decuplets (right panel). The lines indicate the “projected” leading chiral extrapolation fits in $l_\Xi$ and $s_\Xi$ ($l_\Omega$ and $s_\Omega$) while keeping the other one fixed. The black (circular) point is the extrapolated point at physical $l_\Xi$ and $s_\Xi$ ($l_\Omega$ and $s_\Omega$). The masses for $N$, $\Sigma$ and $\Lambda$ are 0.962(29), 1.203(11), 1.122(16) GeV, respectively, and the masses for $\Delta$, $\Sigma^*$ and $\Xi^*$ are 1.275(38), 1.426(25) and 1.580(13) GeV, respectively. ](decExp.eps "fig:"){width="32.00000%"}
Methodology: Cubic-Group Operators and Distillation
===================================================
The Hadron Spectrum Collaboration (HSC) has been investigating interpolating operators projected into irreducible representations (irreps) of the cubic group [@HSC2005] in order to better calculate two-point correlators for nucleon spectroscopy. Baryon correlation functions were evaluated using the displaced-quark operators described in Refs. [@HSC2005] and employed in spectrum studies in Refs. [@HSCnum]. In the cubic group $O_h$, for baryons, there are four two-dimensional irreps $G_{1g}, G_{1u}, G_{2g}$, $G_{2u}$ and two four-dimensional irreps $H_g$ and $H_u$. (The subscripts “$g$” and “$u$” indicate positive and negative parity, respectively.) Each lattice irrep contains parts of many continuum states. The $G_1$ irrep contains $J=\frac{1}{2},\frac{7}{2},\frac{9}{2},\frac{11}{2},\dots$ states, the $H$ irrep contains $J=\frac{3}{2},\frac{5}{2},\frac{7}{2}, \frac{9}{2},\dots$ states, and the $G_2$ irrep contains $J=\frac{5}{2},\frac{7}{2},\frac{11}{2},\dots$ states. The continuum-limit spins $J$ of lattice states must be deduced by examining degeneracy patterns among the different $O_h$ irreps. Using these operators, we construct an $r \times r$ correlator matrix and extract individual excited-state energies by fitting with single- and multiple-exponential functions.
However, the calculations using these cubic-group operators require multiple orientations in order to maximally overlap with a wide range of quantum numbers, which is quite expensive. Furthermore, as we go to lighter and lighter pion masses, there will be more decay modes open, even for the lowest energy at a specific quantum number. We need to extend the matrix to include the multiple-particle operators (so that we can further understand the nature of the “state” in our calculation) and “disconnected” operators. Further, we need to achieve better precision for each state to distinguish among them.
A new way to calculate timeslice-to-all propagators, “distillation”, has been proposed in Ref. [@Peardon:2009gh]. The method is useful for creating complex operators, such as those used in the variational method, allows the operators to be decided after performing the Dirac inversions and reduces the amount of time needed for contractions. Distillation uses color-eigenvector sources to improve on noisy estimators, giving better coverage of relevant degrees of freedom. Increasing the number of “hits” improves statistics faster than 1/$\sqrt{N}$. The method can combine with stochastic methods, which might be desirable if the number of sources needed to cover the volume becomes too large.
The distillation operator on time-slice $t$ can be written as $$\Box(t) = V(t) V^\dagger(t)
\rightarrow
\Box_{xy}(t) = \sum_{k=1}^N v_x^{(k)} (t) v_y^{(k)\dag} (t),
\label{eqn:box}$$ where the $V(t)$ is a matrix containing the first through $k^{\rm th}$ eigenvectors of the lattice spatial Laplacian. The baryon operators involve displacements (${\cal D}_i$) as well as coefficients ($S_{\alpha_1\alpha_2\alpha_3}$) in spin space: $$\chi_B(t) = \epsilon^{abc} S_{\alpha_1\alpha_2\alpha_3}
({\cal D}_1\Box d)^a_{\alpha_1}
({\cal D}_2\Box u)^b_{\alpha_2}
({\cal D}_3\Box u)^c_{\alpha_3}(t),$$ where the color indices of the quark fields acted upon by the displacement operators are contracted with the antisymmetric tensor, and sum over spin indices. Then one can construct the two-point correlator; for example, in the case of proton, $$\begin{aligned}
C^{(2)}_B[\tau_d,\tau_u,\tau_u](t',t) &=
\Phi^{(i,j,k)}(t')
\tau_d^{(i,\bar{i})}(t',t)
\tau_u^{(j,\bar{j})}(t',t)
\tau_u^{(k,\bar{k})}(t',t)
\Phi^{(\bar{i},\bar{j},\bar{k})*}(t)
\nonumber\\
&\quad- \Phi^{(i,j,k)}(t')
\tau_d^{(i,\bar{i})}(t',t)
\tau_u^{(j,\bar{k})}(t',t)
\tau_u^{(k,\bar{j})}(t',t)
\Phi^{(\bar{i},\bar{j},\bar{k})*}(t),\end{aligned}$$ where the “baryon elemental” $$\Phi^{(i,j,k)}_{\alpha_1\alpha_2\alpha_3}(t) = \epsilon^{abc}
\left({\cal D}_1 v^{(i)}\right)^{a}
\left({\cal D}_2 v^{(j)}\right)^{b}
\left({\cal D}_3 v^{(k)}\right)^{c}(t)\;
S_{\alpha_1\alpha_2\alpha_3}
\label{eqn:baryon_op}$$ can be used for all flavors of baryon and quark masses with the same displacements on the same ensemble, and the “perambulator” $$\tau_{\alpha\beta }(t',t) = V^\dagger(t') M^{-1}_{\alpha\beta}(t',t) V(t)
\label{eqn:peram}$$ can be reused for different baryon (and meson) operators after a single inversion of the $M$ matrix. There is a large factor of computational power saved by “factorizing” correlators in terms of elementals and perambulators, and we can use the same elementals and perambulators to contract various different correlators.
The distillation method is demonstrated on four nucleon $G_{1g}$ operators, two of which are local to a single site and the remaining two having a singly displaced quark field. The corresponding $4 \times 4$ matrix of correlators was computed on 316 configurations of the $16^3 \times 128$ ($\approx 380$ MeV) lattice ensemble. Modeling the correlator noise-to-signal ratio with $ a + \frac{b}{N^p}$ as a function of the number of distillation eigenvectors included ($N$ in Eq. \[eqn:box\]) gives a best-fit exponent $p \sim 1.1(2)$; that is, increasing the number of vectors decreases the noise considerably faster than simple statistical scaling. The variational method [@VM] is used to extract the masses of the lowest two states in the $G_{1g}$ spectrum. The extracted mass dependence on the number of eigenvectors is shown in the left panel of Fig. \[fig:spec-840\]. We find consistent masses among the larger numbers of eigenvectors, with an increase in statistical precision as the number of vectors is increased.
A procedure called “pruning” is used to reduce the number of operators included. A practical procedure is to calculate the correlators with the same source and sink operator (i.e. the diagonal elements of the full correlator matrix) and sort them according to their “individuality”. One way to systematically prune is to take the matrix of inner products of the effective masses of each correlator across all time slices and sort them from there. The middle panel of Fig. \[fig:spec-840\] shows a subset of various $G_{1g}$ displacement-operator correlators sorted by the values of their inner products with respect to one another. Any overlap larger than 70% is marked by yellow, while the remaining values are depicted as magenta to blue colors. ($G_{1g}$ has the largest overlap amongst its operators of all the baryon irreps.) Notice that the inner-product matrix has a clear block structure when organized this way. Each yellow block consists of a set of operators that yield nearly identical effective masses as functions of time. We exclude many operators by such a selection process and pick a total of 24 operators among different quark orientations for each irrep. Then we refine to a smaller sub-matrix of $8\times 8$ by filtering the matrix by condition numbers; we perform a variational-method analysis on the resulting $8\times 8$ matrix and retain the lowest 4 eigenstates.
Fig. \[fig:spec-840\] shows preliminary results for nucleon spectroscopy with pion mass around 380 MeV on a $16^3\times 128$ volume, using “distillation” [@Peardon:2009gh] with $N=32$ eigenvectors. We observe a similar distribution of states as the previous study. Further measurements on the larger volume and investigation of decay thresholds and potential two-particle states are underway. (See the proceeding by K. Juge, reporting two-particle results using distillation.)
 The fitted energies for the first-excited (top) and ground state (bottom) as functions of the number of distillation eigenvectors $N$. (Middle) A subset of $G_{1g}$ irrep operator correlators, grouped by their inner product. The yellow blocks indicate overlap $>70$%. $G_{1g}$ is the worst case among all the irreps. (Right) Nucleon excited spectrum sorted according to cubic-group irrep. ](massNev.eps "fig:"){width=".32\textwidth"}  The fitted energies for the first-excited (top) and ground state (bottom) as functions of the number of distillation eigenvectors $N$. (Middle) A subset of $G_{1g}$ irrep operator correlators, grouped by their inner product. The yellow blocks indicate overlap $>70$%. $G_{1g}$ is the worst case among all the irreps. (Right) Nucleon excited spectrum sorted according to cubic-group irrep. ](G1g.eps "fig:"){width=".32\textwidth"}  The fitted energies for the first-excited (top) and ground state (bottom) as functions of the number of distillation eigenvectors $N$. (Middle) A subset of $G_{1g}$ irrep operator correlators, grouped by their inner product. The yellow blocks indicate overlap $>70$%. $G_{1g}$ is the worst case among all the irreps. (Right) Nucleon excited spectrum sorted according to cubic-group irrep. ](nucleonplot.eps "fig:"){width=".34\textwidth"}
Conclusion and Outlook
======================
We report an ongoing effort to solve the mysteries of baryon resonances. The ground-state baryon masses are in reasonable agreement with experiment and consistent among different groups with different actions if systematics are taken into consideration. Our new technique, “distillation”, will greatly improve precision in our future calculations for extracting excited-state masses using cubic group-irrep operators, which provide powerful probes to extract highly excited resonances. A preliminary result for nucleons on $N_f= 2+1$ $m_\pi=380$ MeV is shown in this proceeding. Work on larger volumes (with a modified stochastic distillation) are under development. Meanwhile, parallel work from the HSC for meson spectroscopy with exotic quantum numbers and baryons using derivative operators are also in progress. Multi-particle operators are under investigation to distinguish these from resonances. We are also investigating the application of the distillation method to form factors to help us understand the nature of specific states.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The time interval between successive bursts from SGR 1806-20 and the intensity of these bursts are both consistent with lognormal distributions. Monte Carlo simulations of lognormal burst models with a range of distribution parameters have been investigated. The main conclusions are that while most sources like SGR 1806-20 should be detected in a time interval of 25 years, sources with means about 100 times longer have a probability of about 5% of being detected in the same interval. A new breed of experiments that operate for long periods are required to search for sources with mean recurrence intervals much longer than SGR 1806-20.'
author:
- 'K. J. Hurley, B. McBreen, M. Delaney and A. Britton'
date: 'astro-ph/9508074: presented at 29 ESLAB Symposium, April 1995'
title: 'Lognormal Properties of SGR 1806-20 and Implications for Other SGR Sources'
---
Introduction
============
The lognormal properties of the soft repeater SGR1806-20 have been previously reported by Hurley, K.J. [[*et al.*]{}]{} (1994). In particular, both the time interval between repeater events and the luminosity function of the source were fit with lognormal distributions (see Aitchison and Brown, 1957, for a comprehensive introduction to lognormal statistics). This analysis used the data-base of 111 events detected by the International Cometary Explorer (ICE) mission, as reported by Laros [[*et al.*]{}]{} (1987).
While the present number of events observed from the other two sources (Norris [[*et al.*]{}]{}, 1991, Kouveliotou [[*et al.*]{}]{}, 1993) does not allow any detailed analysis, the intervals between successive events of SGR 0526-66 (Golenetskiǐ [[*et al.*]{}]{}, 1987) is also suggestive of lognormal behaviour. Continued observations by BATSE of these sources may reveal lognormal properties for one or both of the remaining two repeaters if either passes into a phase of activity similar to the behaviour of SGR1806-20 during 1983.
The relationship between the number of active (i.e observable) sources and the true number of SGRs in the galaxy is one which is the subject of some debate (see discussions in Kouveliotou [[*et al.*]{}]{}, 1992, Kouveliotou [[*et al.*]{}]{}, 1994 and Hurley, K. [[*et al.*]{}]{}, 1994). If the time interval between SGR events proves to be lognormal then there may be long quiescent periods where the source could be undetectable, leading to an underestimate of the population.
Simulations
===========
In order to investigate the behaviour of sources with much longer mean recurrence times we generated Monte Carlo simulations with a variety of distribution parameters. The Monte Carlo simulations were performed using the random normal generator with Matlab 4.0 for Windows, which is based on a random number generator algorithm given by Park and Miller (1988) with the transformation to the standard normal variate given by Forsythe, Malcolm and Moler (1977). The normal variates were then transformed to lognormal variates using the relationship $Y=e^{\sigma X+\mu}$ where $Y$ is lognormally distributed (with parameters $\mu$ and ) and $(\sigma \! X+\mu)$ is normally distributed with mean $\mu$ and variance .
=
The parameters of the lognormal density function which were fit to the distribution of recurrence intervals for SGR1806-20 were ${\mbox{$\mu_0$}}=9.64$, ${\mbox{$\sigma_0$}}=3.44$ (Hurley, K.J. [[*et al.*]{}]{}, 1994). Initially we generated 100 year long simulations of SGR1806-20 (Fig. \[sgr-sim\]) using these parameters, to check the algorithm. The samples produced were tested for compatibility with a lognormal population using a $\chi^2$ test (Sachs 1986) and were compatible at the 99% confidence level, indicating that the Monte Carlo simulator was functioning correctly. Two further simulations were then performed to investigate how the source behaviour varied as and varied. The results (illustrated in Fig. \[probplot\]) are discussed below.
Discussion
==========
=
Presented in Figure \[probplot\] are the probabilities for source activity in a 25 year period as calculated from the results of the two simulations described above. Figure \[probplot\](a) shows that as the parameter increases the chance of one or more event in 25 years falls from $\approx 80\%$ at $\mu=9.5 \approx \mu_0$ (geometric mean of 0.15 days) to less than 5% at $\mu=18$ (geometric mean of $\sim \! 500$ days).
The chance of one or more events in 25 years for a source like SGR1806-20 (that is with $\mu\approx\mu_0 = 9.64$ and $\approx\sigma_0 = 3.44$) as predicted by Figure \[probplot\] indicates that the majority of this type of source should be observed in $\sim$ 25 years. For $\mu\!\gg\!\mu_0$ experiments which operate for a long time must be devised and maintain a continuous search over the whole sky for longer periods than any spaceborne experiments designed so far. Such experiments could reveal a larger population of sources with significant gaps of inactivity.
The lognormal distribution arises in statistical processes whose completion depend on a product of probabilities, arising from a combination of independent events (Montroll and Shlesinger, 1982). Lognormal statistics have previously been used in connection with gamma-ray bursts by McBreen [[*et al.*]{}]{} (1994) and Brock [[*et al.*]{}]{} (1994). In the context of this investigation the physical significance of this statistical behaviour may lie in the connection between SGRs and neutron stars. In their paper, Hurley, K.J. [[*et al.*]{}]{} also presented a similar statistical analysis of the behaviour of microglitches from the Vela pulsar (Cordes, Downes and Krause-Polstroff, 1988). The time separation and the intensity of these small ( $\mid \! \Delta\nu/\nu
\! \mid \sim 10^{-9}$ ) frequency adjustments were both compatible with lognormal distributions, and there was no correlation between waiting time and intensity: just as observed with SGR1806-20 (Laros [[*et al.*]{}]{}, 1987). This result, combined with the identification of X-ray point sources (Murakami [[*et al.*]{}]{}, 1994, Rothschild [[*et al.*]{}]{}, 1994) embedded in plerion-powered SNR (Kulkarni [[*et al.*]{}]{}, 1993) as counterparts to the SGR sources, suggests structural adjustments in neutron stars may be the cause of SGRs.
Conclusion
==========
Previously it was shown that the time intervals between successive events from SGR1806-20 and the associated luminosity function were both lognormally distributed. Structural adjustments in neutron stars may be responsible for this behaviour. The activity of sources with longer mean recurrence times was investigated using Monte Carlo simulations. The results of the simulations indicate that there could exist a significant population of SGRs with means longer than SGR1806-20 that remain undetected. A new breed of experiments with very long observation times will be required to search for this type of source.
Aitchison, J. and Brown, J.A.C., 1957, The Lognormal Distribution, Cambridge University Press: Cambridge.
Brock, M. [[*et al.*]{}]{}: 1994, in Fishman, G.J.M, Brainerd, J.J, Hurley, K., ed(s), [*A.I.P. Conf. Proc. 307,*]{} 672.
Cordes, J.M., Downs, G.S. and Krause-Polstorff, J., 1988, Astrophys. J., 330, 847.
Forsythe, G.E., Malcolm, M.A. and Moler C.B., 1977, Computer Methods for Mathematical Computations, Prentice-Hall.
Golenetskiǐ, S.V. [[*et al.*]{}]{}, 1987, Sov. Astron. Lett., 13(3), 166.
Hurley, K. [[*et al.*]{}]{}, 1994, Astrophys. J., 423, 709.
Hurley, K.J., McBreen, B., Rabbette, M. and Steel, S., 1994, Astron. Astrophys., 288, L49.
Kouveliotou, C. [[*et al.*]{}]{}, 1992, Astrophys. J., 392, 179.
Kouveliotou, C. [[*et al.*]{}]{}, 1993, Nature, 362, 728.
Kouveliotou, C. [[*et al.*]{}]{}, 1994, Nature, 368, 125.
Kulkarni, S.R. [[*et al.*]{}]{}, 1994, Nature, 368, 129.
Laros, J.G. [[*et al.*]{}]{}, 1987, Astrophys. J., 320, L111.
McBreen, B., Hurley, K.J., Long, R. and Metcalfe, L., 1994, MNRAS, 271, 662.
Montroll, E.W., Shlesinger, M.F., 1982, Proc Nat Acad Sci USA, 79, 3380.
Murakami, T. [[*et al.*]{}]{}, 1994, Nature, 368, 127.
Norris, J.P., [[*et al.*]{}]{}, 1991, Astrophys. J., 366, 240.
Park, S.K. and Miller, K.W., 1988, Comm ACM, 32(10), 1192.
Rothschild, R.E., Kulkarni, S.R. and Lingenfelter, R.E., 1994, Nature , 368, 432.
Sachs, L., 1986, Applied Statistics, Springer-Verlag: New York.,
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Looking at the automata defined over a group alphabet as a nearring, we see that they are a highly complicated structure. As with ring theory, one method to deal with complexity is to look at semisimplicity modulo radical structures. We find some bounds on the Jacobson 2-radical and show that in certain groups, this radical can be explicitly found and the semisimple image determined.'
author:
- |
Tim Boykett,\
Time’s Up Research, Industriezeile 33b, 4020 Linz, Austria\
and Institute for Algebra, Johannes Kepler University, Linz, Austria\
email:[email protected], [email protected]\
and\
Gerhard Wendt,\
Institute for Algebra, Johannes Kepler University, Linz, Austria\
email:[email protected]
bibliography:
- 'automataradicals.bib'
title: '${\cal J}_2$ radical in automata nearrings'
---
*Dedicated to Professor Günter Pilz on the occasion of his retirement.*
Keywords: Nearrings; Radical; Jacobson; Automata; State Machines.
Introduction
============
Automata defined over a group alphabet can be interpreted as mappings from a group $G$ into itself. A natural algebraic structure to study mappings from a group into itself is a nearring. The link between automata we are studying and nearrings will be further explained in the following. First we fix and introduce some notation and elementary facts concerning nearrings.
A *nearring* $(N,+,*)$ is an algebra with two binary operations such that $(N,+)$ is a group and $(N,*)$ is a semigroup and the right distributive law $(a+b)*c=a*c+b*c$ holds for all $a,b,c \in N$. Usually we will omit the operation symbol $*$ in the following. The notation we use is that of [@pilzbook].
Let $0$ be the neutral element of the group $(N,+)$ of the nearring $N$. A nearring $N$ is called *zero symmetric* if for all $n \in N$, $n*0=0$. Nearrings arise naturally when studying mappings from a group $G$ into itself. Let $M(G):=\{f: G \rightarrow G |\text{f is a function}\}$. Then, $M(G)$ is a nearring with respect to pointwise addition of functions and function composition. In fact, any nearring can be embedded as a subnearring into some $M(G)$ for a suitable group $G$. If the zero $0$ of the group $G$ is preserved by the mappings, then we get the zero symmetric subnearring $M_0(G):=\{f: G \rightarrow G|\text{f is a function and} f(0)=0\}$ of $M(G)$. We will call a function $f: G \rightarrow G$ *zero preserving* if $f(0)=0$.
A *left ideal* $L$ of a nearring $N$ is a normal subgroup of the group $(N, +)$ such that $\forall n,m \in N\, \forall l \in L: n(m + l)-nm \in L$. An *ideal* $I$ of a nearring is a left ideal which is closed under multiplication from the right hand side, so $IN \subseteq I$. Ideals are kernels of nearring homomorphisms. An *$N$-subgroup* $S$ of the nearring $N$ is a subgroup of $(N,+)$ such that $NS \subseteq S$. In case $N$ is a zero symmetric nearring, a left ideal $L$ of $N$ is also an $N$-subgroup of $N$. If $N$ is a nearring with identity element $1$, then an element $n \in N$ is called *quasiregular* if there exists an element $m \in N$ such that $m(1-n)=1$.
In many situations it is interesting to study the action of a nearring on a group $G$. Let $N$ be a nearring and $G$ be a group. A group $G$ is called an $N$-group if there is an action $N \times G \rightarrow G$, written $ng$, such that $n(mg)=(nm) g$ and $(n+m) g=ng + mg$ for all $n,m \in N$ and $g \in G$. A subgroup $S$ of an $N$-group $G$ will be called an $N$-subgroup of $G$ if $NS \subseteq S$. For an $N$-group $G$ we define $(0:G):=\{n \in N| ng=0 \forall g\in G\}$ to be the *annihilator* of this $N$-group. Annihilators of $N$-groups are always ideals in the nearring $N$.
Let $S$ be a set of symbols. As is usual, $S^*$ is the set of finite sequences of symbols from the set $S$, with $\lambda$ being the empty sequence. $S^*$ forms a monoid under concatenation with $\lambda$ being the identity. We will write this operation by juxtaposition.
We write $S^{\mathbb N}$ for the set of infinite sequences over $S$.
Prefix preserving maps
======================
Let $(G,+)$ be a group so $(G^{\mathbb N},+)$ is a group. Thus $M_0(G^{\mathbb N})$ is a nearring.
$n \in M_0(G^{\mathbb N})$ is *prefix preserving* if $$\forall k \in {\mathbb N}, x,y \in G^{\mathbb N}:\: x_i=y_i \forall i<k
\Rightarrow (nx)_i = (ny)_i \forall i<k.$$
Prefix preserving maps form a subnearring of $M_0(G^{\mathbb N})$, we will call it $PP(G)$. Prefix preserving maps in $M(G)$, which are not necessarily zero symmetric, will be called $PP_c(G)$
As an example, take some $f \in M_0(G)$ and $x \in G^{\mathbb N}$. Define $\bar f \in PP(G)$ by $(\bar f x)_i = f(x_i)$. This is a simple prefix preserving map.
A map in $PP(G)$ can be seen as an element of $M_0(G^n)$ for any finite $n$. Let $x \in G^n$, $m\in PP(G)$. Then $mx = (m(x_1,\ldots,x_n,0,0,\ldots))\vert_{\{1,\ldots,n\}}$. We will call this action the *restricted action* of $PP(G)$ on $M_0(G^n)$.
Given a group $(G,+)$, it is possible to define the nearring of state automata or state machines over the alphabet $G$ [@pilzautomata]. If we fix the input-output alphabet $G$, a *state automaton* is defined as $(Q,t,f,s)$, where
- $Q$ is a set of states,
- $t:Q\times G \rightarrow Q$ is a state transition map,
- $f:Q \times G \rightarrow G$ is the output map and
- $s \in Q$ is the start state.
A state automaton or state machine can be seen as a mapping of $G$-sequences to $G$-sequences, which is how we will see them. In order to calculate the mapping, we proceed as follows. Let $a= (Q,t,f,s)$ be a state machine, $x \in G^{\mathbb N}$ be the input sequence. Let $q_1=s$ and $q_{i+1}= t(q_i,x_i)$ for $i \geq 1$, the *state sequence*. Then $y_i = f(q_i,x_i)$ and $y=ax \in G^{\mathbb N}$ is the *output sequence*, the image of $x$ under the state machine mapping. Note that this is defined on finite as well as infinite sequences.
The concatenation of state machines and the addition of state machines are then precisely defined as composition and addition of the maps on $G^{\mathbb N}$. These operations can be defined as operation on the state machines, see [@pilzautomata paragraphs before Prop 2] for details. If two state machines agree as maps on $G^{\mathbb N}$, we regard them as equal. Call the collection of state machines on a group $SM(G)$.
We will have occasion to look at the *state output maps* of a state machine $(Q,t,f,s)$. These are the mappings $f_q:G\rightarrow G$ $g \mapsto f(q,g)$.
Note that we use the formulation of state machines as *Mealy Machines*. If the state output map does not depend upon the input, but only upon the state, i.e. the mappings $f_q: g \mapsto f(q,g)$ are constant maps, then we have a *Moore Machine* (see e.g. [@leeseshia p. 58]).
As an example, let $f \in M_0(G)$, $Q=\{s\}$ a single element set, $\bar f (s,g)=f(g)$ and $t(s,g)=s$. Then $(Q,t,\bar f,s)$ is a state machine with a single state, with state output map $\bar f_s = f$.
Let $G$ be a group. $PP_c(G)$ is isomorphic to $SM(G)$.
Proof: We have defined state machines $SM(G)$ and prefix preserving maps $PP_c(G)$ as subnearrings of $M(G^{\mathbb N})$, so we need only show that they are the same as subsets.
From the definition of the sequence mapping, state machines are prefix preserving, so $SM(G) \subseteq PP_c(G)$.
Let $n\in PP_c(G)$. Define $Q = G^*$. Define $t(q,g) = qg \in G^*$ by concatenation and $f(q,g)= n(qg)_{i+1}$ where $i$ is the length of the string $q$, $n$ acting by the restricted action on $G^{i+1}$. Then the state machine $(Q,t,f,\lambda)$ will induce the same mapping as $n$, so $PP_c(G) \subseteq SM(G)$ and we are done. $\Box$
We will use this equivalence often, in order to define, manipulate and analyse maps in the clearest way possible. We will concentrate upon $PP(G)$ and are thus interested in the zero symmetric state machines in $SM(G)$.
A state $r \in Q$ is called *reachable* if there is an input sequence $x \in G^{\mathbb N}$ such that $r=q_i$ for some $q_i$ in the state sequence $q$. It is clear that unreachable states do not affect the properties of state machines. A state $r\in Q$ is *0-reachable* if for some $i$, $r=q_i$ in the state sequence induced by the zero sequence $(0,0,\ldots)$.
A state machine is zero symmetric if the state output maps $f_q$ are 0-preserving for every state that is 0-reachable.
Proof: Let $a=(Q,t,f,s)$ be a zero symmetric state machine. We know that a state machine is prefix preserving. Let $r\in Q$ be 0-reachable, $r=q_i$ with $q$ being the state sequence from the zero input sequence. Thus $f(q,0)=0$ and we see that every 0-reachable state has a 0-preserving state output map.
Let $a=(Q,t,f,s)$ be a state machine with all 0-reachable states having 0-preserving state output maps. Then $a(0,0,\ldots) = (0,0,\ldots)$ so $a$ is 0-preserving and hence $a$ is a zero symmetric state machine. $\Box$
The following result indicates that the nearring of prefix preserving maps is complex and complicated. Let ${\mathbb V}(G)$ be the variety generated by the group $G$.
Let $K$ be a finite group. Then for all finite groups $G \in {\mathbb V}(K)$, $M_0(G) < PP(K)$.
Proof: We will show that we can encode $G$ into $K^{\mathbb N}$ and mappings in $M_0(G)$ as state automata mappings on $K$.
From [@burrsan Thm 10.16], $G$ is a homomorphic image of a subgroup $S$ of a finite power of $K$, so there exists some natural number $n$ and some $\beta:S \subseteq K^n \rightarrow G$ which is a homomorphism. Let $\alpha:G \rightarrow S$ be some mapping such that $\beta \circ \alpha$ is the identity on $S$.
Let $f \in M_0(G)$ be arbitrary. Define $F:S\rightarrow S$ by $F=\alpha f\beta$.
Define $Q = \cup_{j=0,\ldots, n-1} K^j \times S \times \{1,\ldots,n\} \cup \{z\}$. $Q$ is thus made up of the cartesian product of $K$-sequences up to length $n-1$, a recognized element of $S$ and an index. The state $z$ indicates an *error* state, which will not be reached by correctly encoded inputs. Let the initial state be $s=(\lambda,0,1)$. Define the state transition function $t:Q\times K \rightarrow Q$ as
- $t((q_1,0,1),k) = (q_1k,q_1k,1)$ if $q_1k \in S$.
- $t((q_1,0,1),k) = (q_1k,0,1)$ if $q_1k$ is a prefix of some element of $S$.
- $t((q_1,0,1),k) = z$ otherwise.
- $t((q_1,s,n),k) = (q_1k,s,1)$ if $q_1k =s$.
- $t((q_1,s,i),k) = (q_1k,s,1+1)$ if $q_1k$ is a prefix of $s$.
- $t((q_1,s,i),k) = z$ otherwise.
- $t(z,k)=z$
The state transition function recognises an input sequence as consisting of a sequence of $0$s of length a multiple of $n$, followed by repeats of an element of $S$.
Define the output function $o:Q\times K \rightarrow K$ as
- $o((q_1,0,1),k)=0$
- $o((q_1,s,i),k)=F(s)_i$ for $s\in S$
- $o(z,k)=0$
The output map implements componentwise the mapping $F$ on the recognized element of $S$.
For the following, let $\hat i = ((i-1) \mod n)+1$.
The encoding map $e:G\rightarrow K^{\mathbb N}$ has $e(g)_i = \alpha(g)_{\hat i}$. The decoding map $d:K^{\mathbb N}\rightarrow G$ is more complex. If there exists some $m\in {\mathbb N}$ such that $\forall i \geq mn$, $x_i = x_{i-n}$, i.e. $x$ cycles, then $d(x) = \beta(x_{mn+1},x_{mn+2},\ldots,x_{mn+n})$, otherwise $d(x)=0$.
Then the automaton defined as $a=(Q,t,o,\lambda)$ on $K$ acts on $G$ by $ag := d\circ a \circ e (g)$. We claim that $ag = f(g)$.
Let $x \in K^{\mathbb N}$ be of the form $x=(0,\ldots,0,s_1,\ldots,s_n,s_1,\ldots)$ with $mn$ $0$s in the prefix, $(s_1,\ldots,s_n)=\alpha(g) \in G$. Note that $e(g)$ is of this form with $m=0$. Then $a(x)$ will have the state sequence $q$ with $q_i = (o_i,0,1)$ with $o_i$ being a sequence of $(i \mod n)$ $0$s, for $i \leq mn$, $q_i= ((s_1,\ldots,s_{\hat i}),0,1)$ for $mn < i < (m+1)n$, $q_{(m+1)n} = ((s_1,\ldots,s_n),(s_1,\ldots,s_n),1)$ and $q_i=((s_1,\ldots,s_{\hat i}),(s_1,\ldots,s_n),\hat i)$ for $i > (m+1)n$.
This will give the output sequence $y=ax \in K^{\mathbb N}$ with $y_i = 0$ for $i \leq (m+1)n$ and $y_i= (F((s_1,\ldots,s_n)))_{\hat i}$ for $i > (m+1)n$. Let $g \in G$, $x=e(g)$. Then $d(y) = \beta(F(s_1,\ldots,s_n)) = \beta \circ \alpha \circ f \circ \beta \circ \alpha (g) = f(g)$.
We have shown that the above construction maps $M_0(G)$ injectively into $PP(K)$.
Let $f_1,f_2 \in M_0(G)$, $a_1,a_2 \in PP(K)$ their encodings by the above construction. Then $d (a_1 a_2) e=(d a_1 e)(d a_2 e)$ by the way $d$ ignores initial $0$ sequences, so we see that the mapping of $M_0(G)$ into $PP(K)$ is a multiplicative homomorphism. Similarly it is an additive homomorphism. Thus the construction is an isomorphic embedding and we are done. $\Box$
Note that the complex definition of $d$ was necessary because the automaton delays its output by $n$ places. Thus the composition of two such automata will delay by $2n$ places.
One approach to deal with complex structures in ring and nearring theory in to consider semisimple rings and nearrings. Our goal in this paper is to investigate the Jacobson 2-radical structure of $PP(G)$ in order to see what the semisimple image is. An $N$-group $\Gamma$ is of *type $2$* if there is $\gamma \in \Gamma$ such that $N\gamma = \Gamma$ and there do not exist non-trivial $N$-subgroups in $\Gamma$.
[@pilzbook] Let $N$ be a nearring. Then the *Jacobson 2-radical* of $N$, written ${\cal J}_2(N)$, is the intersection of the annihilators of all $N$-groups of type $2$.
Amnesia and Procrastination
===========================
In this section we introduce two classes of automata and use them to bound ${\cal J}_2(PP(G))$.
The map $\alpha:PP(G)\rightarrow PP(G)$ defined by: $\forall n \in PP(G)$, $\forall x \in G^{\mathbb N}$ $$((\alpha n)x)_i = (n(0,0,\ldots,0,x_i,0,\ldots))_i$$ is called the *amnesiac map*.
The amnesiac map is a nearring homomorphism.
Proof: Let $n,m \in PP(G)$, $x \in G^{\mathbb N}$. Then $$\begin{aligned}
(\alpha (n+m)x)_i &= ((n+m)(0,\ldots,0,x_i,0,\ldots))_i \\
&= (n(0,\ldots,0,x_i,0,\ldots) + m(0,\ldots,0,x_i,0,\ldots))_i \\
&= (n(0,\ldots,0,x_i,0,\ldots))_i + (m(0,\ldots,0,x_i,0,\ldots))_i \\
&= ((\alpha n)x + (\alpha m)x)_i\end{aligned}$$ so $\alpha (n+m) = \alpha n + \alpha m$.
$$\begin{aligned}
((\alpha (nm))x)_i &= (nm(0,\ldots,0,x_i,0,\ldots))_i \\
&= (n(m(0,\ldots,0,x_i,0,\ldots)))_i \\
&= (n(0,\ldots,0,((\alpha m)x)_i,0,\ldots))_i \\
&= ((\alpha n)((\alpha m) x))_i \\
&= ((\alpha n)(\alpha m)x)_i\end{aligned}$$
so $\alpha (nm) = (\alpha n)( \alpha m)$ and we see that $\alpha$ is a nearring homomorphism,. $\Box$
Let $n \in PP(G)$. The following are equivalent:
1. $n \in \ker \alpha$
2. $\forall x\in G^{\mathbb N},\, ((\alpha n)x)_i = (n(0,0,\ldots,0,x_i,0,\ldots))_i =0$
3. $\forall i,\,n(\underbrace{0,0,\ldots ,0}_{i-1},x,\ldots) = (\underbrace{0,0,\ldots,0,0}_{i},\ldots)$
4. $\forall i,\,(n(\underbrace{0,0,\ldots ,0}_{i-1},x))\vert_{\{1,\ldots,i\}} = (0,0,\ldots,0) \in G^i$
This can be seen because 2 is a rewording of the definition in 1, while 3 is simply rewriting 2. The fourth statement uses the restricted action to say the same as 3.
Let $f\in M_0(G)$ and $i,j \in {\mathbb N}$. Define $f^i \in PP(G)$ as $(f^i(x))_i = f(x_i)$ and $(f^i(x))_j = 0$ $\forall j\neq i$. Define $f^{i,j} \in PP(G)$ as $(f^{i,j}(x))_{i+kj} = f(x_{i+kj})$ for all $k\in {\mathbb N}_0$ and $(f^i(x))_l = 0$ otherwise. Let $M_{i,j}(G)= \{f^{i,j}: f \in M_0(G)\}$. Then $M_{i,0}(G)= \{f^i: f \in M_0(G)\}$.
The $M_{i,0}(G)$ automata react only at one time step and are elsewhere zero, ignoring any inputs before or after that time step. In some sense these are the most amnesiac of automata. Note that $\alpha f^i = f^i$ and $\alpha f^{i,j} = f^{i,j}$.
Let $G$ be a group. Then $$J_2(PP(G)) \subseteq \ker \alpha$$
Proof: Suppose $n \not\in \ker \alpha$. $n \not\in \ker \alpha $ iff $\exists \bar g\in G $ such that $(n(0,\ldots,0,\bar g,\ldots))_i \neq 0$ with $\bar g$ in the $i$th place.
Define an action of $PP(G)$ on $G$ as: for $m \in PP(G)$, $g \in G$, $mg = (m(0,0,\ldots,0,g,\ldots))_i $ with $g$ in the $i$th place. We need to show that $mg$ is indeed an action of $PP(G)$ on $G$. The additive property is clear, while the multiplicative property is shown by $$\begin{aligned}
(nm)g &= ((nm)(0,\ldots,0,g,\ldots))_i \\
&= (n(m(,\ldots,0,g,\ldots))_i \\
&= (n(0,\ldots,0,mg,\ldots))_i\\
&= n(mg)
\end{aligned}$$
For all $f\in M_0(G)$, $f^i \in PP(G)$ acts as $f$ under this action, i.e. $f^ig=f(g)$.
So the action is the same as the action of $M_0(G)$ on $G$. Consequently, $G$ is a $PP(G)$ group of type $2$ which means that $J_2(PP(G)) \subseteq (0:G)$. But under this action, $n \not \in (0:G)$ because $n\bar g \neq 0$ , so $n \not \in J_2(PP(G))$. $\Box$
A procrastinating or delaying automaton never reacts immediately to an input. Maybe it should be called a burocrat?
$n \in PP(G)$ is *delaying* if $$\forall k \in {\mathbb N}: x_i=y_i \forall i < k \Rightarrow (nx)_i=(ny)_i \forall i \leq k$$
We write $D(G)$ for this set of state machines.
$D(G)$ is an $N$-subgroup.
Proof: By rudimentary calculations it is clear that $D(G)$ is closed under addition and composition, forming a subnearring.
Now we need to show $PP(G)D(G) \subseteq D(G)$. Let $n\in PP(G)$, $d\in D(G)$, $x,y \in G^{\mathbb N}$ with $x_i=y_i$ for $i<k$. Then $(dx)_i = (dy)_i$ for $i \leq k$ by the delaying property, so $((nd)x)_i=(n(dx))_i = (n(dy))_i=((nd)y)_i$ for $i \leq k$ by the prefix preserving property, so $nd \in D(G)$.
Thus we see that $D(G)$ is an $N$-subgroup. $\Box$
The following result gives us an idea of what the automata in $D(G)$ are like.
$n\in D(G)$ iff $n$ is a Moore machine iff every state output map is constant.
Proof: Being a Moore machine is defined as having every output map being constant. So we need only concern ourselves with the first and last statements.
$(\Rightarrow)$: Let $n = (Q,t,f,s) \in D(G)$, so for all reachable $r \in Q$, there is some input sequence $x \in G^{\mathbb N}$ such that $r=q_{i+1}$ in the state sequence for some $i$. Then by the delaying property, for all $g_1,g_2 \in G$, $f(r,g_1)=n(x_1,\ldots,x_i,g_1,\ldots)_{i+1} = n(x_1,\ldots,x_i,g_2,\ldots)_{i+1}=f(r,g_2)$ so $f_r: g \mapsto f(r,g)$ is constant. Thus every state output map is constant.
$(\Leftarrow)$: Let $n = (Q,t,f,s) \in PP(G)$ with $f_q: g \mapsto f(q,g)$ constant for all $q \in Q$. Then for all $x,y\in G^{\mathbb N}$ with $x_i=y_i$ $\forall i <k$, the state sequences induced by $x$ and $y$ will match up to $q_k = t(q_{k-1},x_{k-1})$. Then $(nx)_k = f(q_k,x_k) = f(q_k,y_k) = (ny)_k$ so $n \in D(G)$ and we are done. $\Box$
Let $G$ be a group. Then $D(G) \leq J_2(PP(G)) \leq \ker \alpha$.
Proof: We know the second inclusion from above. So we need only show the first inclusion.
Let $n\in D(G)$, as a state machine $(Q,t,f,s)$. The automaton $1-n$ can be written as $(Q,t,h,s)$ with $h(q,g) = g-f(q,g)$. As every state output map $f_q:g \mapsto f(q,g)$ is a constant map by the claim above, the automaton $1-n$ has output map $h(q,g)=g-f(q,g)$, state output maps $h_q:g \mapsto g-f_q(g)$ that are permutations. Let $\bar h_q:G \rightarrow G$ be the inverse of this permutation, $\bar h(q,g) = \bar h_q(g)$. Define $\bar t:Q\times G \rightarrow Q$ by $\bar t(q,g) = t(q,\bar h(q,g))$. Then the state machine $m=(Q,\bar t, \bar h, s)$ is the inverse of $1-n$.
The composition state machine $m \circ (1-n) = (Q\times Q, T, F, (s,s))$ for some maps $T$ and $F$. Note that the start state is on the diagonal. Let $q \in Q,\,g\in G$. Then $$T((q,q),g) = (\bar t(q,h_q(g)), t(q,g)) = (t(q,\bar h_q(h_q(g))),t(q,g)) = (t(q,g),t(q,g))$$ so we see that the state remains on the diagonal $\{(q,q):q\in Q\} \subset Q \times Q$. Thus we need only consider the output maps on the diagonal, as no other states are reachable. The output function of the composition can be seen by $$F((q,q),g) = \bar h(q,h(q,g)) = \bar h_q (h_q(g)) = g$$ to be the identity, $m \circ (1-n) = 1$, so by Beidleman [@pilzbook 3.37c], $n$ is quasiregular.
Thus the whole $N$-subgroup $D(G)$ is quasiregular, so by Ramakotaiah [@ramakotaiah67] (also [@pilzbook Theorem 5.44]) $D(G)$ lies within the Jacobson 2-radical ${\cal J}_2(PP(G))$ and we are done. $\Box$
We know that the radical is an ideal, so we are interested in the ideal generated by $D(G)$. Our job is to work out when these nearrings coincide as ideals.
${\cal J}_2$ for certain groups
===============================
In this section we show that there are a class of groups where we can identify the Jacobson 2-radical of the prefix preserving maps on that group.
\[defnX\] A group $(G,+)$ has *property X* if there is an element $k\in G$ and a function $f:G \rightarrow G$ such that $f(x+k)-f(x)=x$ for all $x \in G$.
Finite abelian groups have property X iff they are of odd order.
Proof: First we show that all cyclic groups of odd order have property X. Then we show that odd order abelian groups have property X. Then we will show that a group with an element of even order will not have property X.
Let $n>1$ be an odd integer. Let $k=1 \in {\mathbb Z}_n$ be a generator. Let $x_i=f(i)$. Property X can be written as a collection of linear equations over $Z_n$ with $-x_i + x_{i+1} = i$ for $0\leq i <n$ with suffix addition modulo $n$. The $n$th equation is $-x_{n-1} + x_0 = n-1$. The sum of the first $n-1$ equations is $$-x_0 +x_1 - x_1 + x_2+\ldots +x_{n-1} = 0+1+2+\ldots+{n-2}$$ which reduces to $-x_0 + x_{n-1} = 1$, the additive inverse of the $n$th equation. Thus the $n$th equation is redundant and we are left with $n-1$ equations having a set of solutions with one parameter. We select $x_0$ then for each $0<i<n$, $x_i=x_{i-1}+i-1$.
Let $G$ be an abelian group of odd order. Let $H$ be a maximal cyclic subgroup of $G$, generated by $k$ of order $2n+1$. $H$ has property X by the above argument. Let $c \not\in H$. Then in $c+H$ the same argument holds (that the last equation is redundant) because the order of $c$ divides $2n+1$ and thus $(2n)c = -c$. Thus in each coset of $H$ we can find values for $f$ and we are done.
Suppose $G$ is an abelian group of even order satisfying property X. If the element $k$ has even order, say $n=2m$, then the set of the first $n-1$ equations, when added, gives $x_0 - x_{n-1} = 1+m$. The $n$th equation is $x_{n-1} - x_0 = n-1$ so we add them to get $0=n-1+1+m = n+m$, i.e. $n=m=0$, a contradiction. Thus the element $k$ for property X should be of odd order, let the order be $n$. Let $c\in G$ be of order 2. Then there will be a set of $n$ equations of the form $$f(c+ik) - f(c+(i+1)k) = c+ik$$ Adding all these equations together we get $$0 = nc + \frac{n(n+1)}{2}k = nc$$ but $c$ is of even order and $n$ is odd, a contradiction.
Thus no abelian group of even order can have property X. $\Box$
Infinite cyclic groups have property X.
Proof: Suppose $G$ is an infinite order cyclic group, we will assume it is ${\mathbb Z}$. Defining $f(0)=f(1)=\ldots=f(k-1)=0$ and $f(x+k)=f(x)+x$ otherwise gives a function for property X. $\Box$
Note that without loss of generality, we can insist that $f(0)=0$.
Let $(G,+)$ be a group with property X. Then ${\cal J}_2(PP(G)) = \ker \alpha$.
Proof: Let $d\in \ker \alpha$.
Let $g\in G$ be some arbitrary but fixed nonzero element. Let $c\in D(G)$ be the delaying automaton defined as $(\{a,b\},t,o,a)$ with $t(a,0)=a$, $t(a,x)=t(b,y)=b$ and $o(a,y)=0$, $o(b,y)=g$ for all $x,y \in G, x\neq 0$.
The initial $0$ inputs are mapped to $0$, the first nonzero input as well, then all outputs are $g\neq 0$.
Let $f$ be a one state automaton with the output function being the function defined by property X. Then $f(d+c) - fd = d$ so the left ideal generated by $D(G)$ is all of $\ker \alpha$ which is an ideal. Thus ${\cal J}_2(PP(G)) = \ker \alpha$. $\Box$
Semisimplicity
==============
The radical is, in some sense, the “bad” part of a nearring. So we factor that out to get the “good” part. What is the image of $PP(G)$ under the map $\alpha$?
The following result shows what the amnesiac map does to an automaton.
Let $n=(Q,t,o,s)$ be an automaton. Define $\bar t (q,g) = t(q,0)$ $ \bar n = (Q,\bar t,o,s)$. Then $\bar n = \alpha n$ and $\alpha$ is idempotent.
Proof: For any $i$, $(\alpha n x)_i = (n(0,\ldots,0,x_i,0,\ldots))_i$, the state sequence $q$ induced by $(0,\ldots,0,x_i,0,\ldots)$ in $\alpha n$ agrees, up to the $i$th entry, with the state sequence induced by $x$ in $\bar n$. So $(\alpha n x)_i = o(q_i,x_i) = (\bar n x)_i$. Thus $\alpha n$ and $\bar n$ agree for all $i$, so they are equal.
$\bar \bar t (q,g) \bar t(q,0) = t(q,0)= \bar t(q,g)$, so $\bar \bar n = \bar n$ so the mapping $\alpha$ is idempotent. $\Box$
Let $(G,+)$ be a group such that ${\cal J}_2(PP(G)) = \ker \alpha$. Then, $PP(G) / {\cal J}_2(PP(G)) = \sum M_{i,0}(G)$.
Proof: First we show that the sum is direct. Let $i\neq j$, $f^i \in M_{i,0}(G),\, g^j \in M_{j,0}(G)$. Then $(f^ig^j(x))_i = f(g^j(x)_i) = f(0) = 0$, and for $k \neq i$, $(f^ig^j(x))_k = 0$. So $f^ig^j=0$ and we are done.
Let $n = (Q,t,f,s)$ be an automaton, $\alpha n = n$. Then $t(q,g)=t(q,0)$ for all $q\in Q$. Define $\tau :Q\rightarrow Q$ by $\tau: q \mapsto t(q,0)$. Then the state sequence will be $q_i = \tau^{i-1}s$.
Define $f_i(g) = f(\tau^{i-1}s,g) \in M_0(G)$. Then we claim that $$n = \sum_{i\in {\mathbb N}} f_i^{i,0}$$
Thus $\alpha n = n \Rightarrow n \in \sum M_{i,0} (G)$. We know the converse, so $\alpha PP(G) = \sum M_{i,0}(G)$.
Now $PP(G) / {\cal J}_2(PP(G)) = PP(G) / \ker \alpha = \alpha(PP(G))$, so we are done. $\Box$
We note that a finite state automata will give, in general, an infinite sum in $\sum M_{i,0}(G)$. It will, however, give a finite sum in $\sum M_{i,j}(G)$, by summing over the 0-reachable states.
Summary
=======
We examined the properties of nearrings of state automata and found that the nearrings are highly complex. In order to get a handle on them, we looked at the radicals and the semisimple images modulo these radicals. The Jacobson 2-radical could be constrained and, in the case of noneven order abelian groups, fully determined. It equals the kernel of the amnesiac endomorphism. Thus we can say that the representatives of the 2-semisimple nearring of automata are the amnesiac automata. These are sequences of mappings from $M_0(G)$. These sequences are infinite, but in the case of finite state automata they will cycle after an initial sequence.
In difference to ring theory, several different types of simplicity of $N$-groups exist (see [@pilzbook]). One obvious next stage would be to determine other radicals in $PP(G)$ and to see in which way their semisimple images behave as automata.
In difference to ring theory the radical theory for nearrings is much more complex. Several types of simplicity for $N$-groups of a nearring $N$ exist. Apart from $N$-groups of type $2$ the most common ones are $N$ groups of type $0$ and type $1$. The intersection of the annihilators of the $N$-groups of type $1$, type $0$ respectively, defines the *Jacobson 1-radical* ${\cal J}_1(N)$, *Jacobson 0-radical* respectively. Note that $PP(G)$ contains the identity mapping, so ${\cal J}_2(N)={\cal J}_1(N)$ by [@pilzbook Proposition 5.3]. There are also a range of further radicals, all of which would offer further insights.
On the other hand, it would be of interest to determine the 2-radical for further classes of groups.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors gratefully acknowledge support from the Austrian Science Fund FWF under project number 23689-N18.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A conjecture of Morel asserts that the sheaf $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ of ${{\mathbb A}}^1$-connected components of a simplicial sheaf ${{\mathcal X}}$ is ${{\mathbb A}}^1$-invariant. A conjecture of Asok and Morel asserts that the sheaves of ${{\mathbb A}}^1$-connected components of smooth schemes over a field coincide with the sheaves of their ${{\mathbb A}}^1$-chain-connected components. Another conjecture of Asok and Morel states that the sheaf of ${{\mathbb A}}^1$-connected components is a birational invariant of smooth proper schemes. In this article, we exhibit examples of schemes for which conjectures of Asok-Morel fail to hold and whose $Sing_*$ is not ${{\mathbb A}}^1$-local. We also give equivalent conditions for Morel’s conjecture to hold and obtain an explicit conjectural description of $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$. A method suggested by these results is then used to prove Morel’s conjecture for non-uniruled surfaces over a field.'
address:
- 'School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Bombay 400005, India'
- 'School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Bombay 400005, India'
- 'School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Bombay 400005, India'
author:
- Chetan Balwe
- Amit Hogadi
- Anand Sawant
title: '${{\mathbb A}}^1$-connected components of schemes'
---
[^1]
Introduction
============
${{\mathbb A}}^1$-homotopy theory, developed by Morel and Voevodsky [@Morel-Voevodsky] in the 1990’s, is a *homotopy theory* for schemes in which the affine line ${{\mathbb A}}^1$ plays the role of the unit interval in usual homotopy theory. In the last few years, Morel has developed *${{\mathbb A}}^1$-algebraic topology* over a perfect field [@Morel] and obtained analogues of many important theorems in classical algebraic topology in the ${{\mathbb A}}^1$-homotopic realm. In ${{\mathbb A}}^1$-homotopy theory, one needs to suitably enlarge the category of smooth schemes over a field in order to be able to perform various categorical constructions. For various technical reasons, in ${{\mathbb A}}^1$-homotopy theory, one enlarges the category of smooth schemes over a field to the category of simplicial presheaves/sheaves of sets over the site of smooth schemes over a field with the Nisnevich topology. By inverting *${{\mathbb A}}^1$-weak equivalences*, one obtains the *${{\mathbb A}}^1$-homotopy category*, defined in [@Morel-Voevodsky] by Morel-Voevodsky. Analogous to algebraic topology, given a simplicial presheaf/sheaf ${{\mathcal X}}$, one then studies the ${{\mathbb A}}^1$-homotopy sheaves $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ and $\pi_i^{{{\mathbb A}}^1}({{\mathcal X}}, x)$, where $i\geq 1$ and $x \in {{\mathcal X}}({{\rm Spec \,}}k)$ is a basepoint. Throughout this paper, we will work over a perfect base field $k$ and denote the unpointed ${{\mathbb A}}^1$-homotopy category by ${{\mathcal H}}(k)$.
A sheaf of sets ${{\mathcal F}}$ on the Nisnevich site $Sm/k$ of smooth schemes over $k$ is said to be *${{\mathbb A}}^1$-invariant* if the projection map $U \times {{\mathbb A}}^1 \to U$ induces a bijection ${{\mathcal F}}(U) \to {{\mathcal F}}(U \times {{\mathbb A}}^1)$, for all smooth schemes $U$. Morel has proved the ${{\mathbb A}}^1$-invariance of the higher ${{\mathbb A}}^1$-homotopy sheaves of groups $\pi_i^{{{\mathbb A}}^1}({{\mathcal X}},x)$, for $i \geq 1$, for any simplicial sheaf of sets ${{\mathcal X}}$ over smooth schemes over a perfect field $k$ (see [@Morel Theorem 6.1 and Corollary 6.2]). Morel has conjectured that the sheaf of ${{\mathbb A}}^1$-connected components $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ of a simplicial sheaf ${{\mathcal X}}$ is ${{\mathbb A}}^1$-invariant. Morel’s conjecture has been proved for motivic $H$-groups and homogeneous spaces for motivic $H$-groups by Choudhury in [@Choudhury]. Any scheme $X$ over $k$ can be viewed as an object of the ${{\mathbb A}}^1$-homotopy category by means of the Nisnevich sheaf on $Sm/k$ given by the functor of points of $X$. The ${{\mathbb A}}^1$-invariance of $\pi_0^{{{\mathbb A}}^1}(X)$ is easily seen to hold when $X$ is a scheme of dimension $\leq 1$ or an *${{\mathbb A}}^1$-rigid scheme* (for example, an abelian variety; see Definition \[definition A1 rigid\]) or an *${{\mathbb A}}^1$-connected scheme* (such as the affine line ${{\mathbb A}}^1$). It is also known to hold for smooth toric varieties (see Wendt [@Wendt Lemma 4.2 and Lemma 4.4]). We now briefly describe the main results of this paper.
\[Intro-theorem lim S\^n\] If ${{\mathcal F}}$ is a Nisnevich sheaf of sets on $Sm/k$, considered as a simplicially constant sheaf in $\bigtriangleup^{op}Sh(Sm/k)$. Let $\mathcal S({{\mathcal F}})$ denote the sheaf of ${{\mathbb A}}^1$-chain connected components of ${{\mathcal F}}$ (see Definition \[definition A1 chain connected components\]). Then the sheaf ${{\mathcal L}}({{\mathcal F}}):= \underset{n}{\varinjlim}~ \mathcal S^n({{\mathcal F}})$ is ${{\mathbb A}}^1$-invariant. Furthermore, if Morel’s conjecture (Conjecture \[a1invariance\]) is true, then the canonical map $$\pi_0^{{{\mathbb A}}^1}({{\mathcal F}}) \to {{\mathcal L}}({{\mathcal F}})$$ is an isomorphism.
The sheaf ${{\mathcal S}}(X)$ (referred to as $\pi_0^{ch}(X)$ by Asok and Morel) of ${{\mathbb A}}^1$-chain-connected components of a smooth scheme $X$ over a field $k$ has been studied in the work of Asok and Morel [@Asok-Morel], where they provide explicit connections between the notions of ${{\mathbb A}}^1$-connectedness and chain-connectedness and various rationality and near-rationality properties of the scheme $X$. They have conjectured that the canonical map $\pi_0^{ch}(X) \to \pi_0^{{{\mathbb A}}^1}(X)$ is an isomorphism, for all smooth schemes $X$ (see [@Asok-Morel Conjecture 2.2.8]). In support of this, they have shown that for a proper scheme $X$, the sections of the two sheaves $\pi_0^{ch}(X)$ and $\pi_0^{{{\mathbb A}}^1}(X)$ agree over all finitely generated, separable field extensions of $k$. This is proved in [@Asok-Morel Theorem 2.4.3 and Section 6] by constructing another sheaf $\pi_0^{b{{\mathbb A}}^1}(X)$ of *birational ${{\mathbb A}}^1$-connected components* of $X$ and by showing that its sections over finitely generated separable field extensions of $k$ are the same as those of $\pi_0^{ch}(X)$. Asok and Morel have also conjectured that the natural map $\pi_0^{{{\mathbb A}}^1}(X) \to \pi_0^{b{{\mathbb A}}^1}(X)$ is an isomorphism, for all proper schemes $X$ of finite type over a field (see [@Asok-Morel Conjecture 6.2.7]).
We obtain a refinement of the result of Asok and Morel [@Asok-Morel Theorem 2.4.3] about sections of $\pi_0^{{{\mathbb A}}^1}(X)$ over field extensions of $k$ using a method suggested by Theorem \[Intro-theorem lim S\^n\]. Our proof gives the result of Asok and Morel by a direct geometric argument and completely avoids the use of $\pi_0^{b{{\mathbb A}}^1}(X)$.
\[Intro-theorem field case\] Let $X$ be a proper scheme over $k$ and let $L$ be a finitely generated field extension of $k$. For every positive integer $n$, we have ${{\mathcal S}}(X)({{\rm Spec \,}}L) = {{\mathcal S}}^n(X)({{\rm Spec \,}}L)$. Consequently, the canonical epimorphism ${{\mathcal S}}(X)({{\rm Spec \,}}L) \to \pi_0^{{{\mathbb A}}^1}(X)({{\rm Spec \,}}L)$ is an isomorphism.
Theorem \[Intro-theorem lim S\^n\] suggests that the conjectures of Morel and of Asok-Morel could be proved for a scheme $X$ if we prove that ${{\mathcal S}}(X) = {{\mathcal S}}^2(X)$. This observation allows one to bring in geometric methods to study this question. In this paper, we have taken this approach to prove the conjectures of Morel and of Asok-Morel for (possibly singular) proper non-uniruled surfaces (see Theorem \[theorem non-uniruled\]). However, we also present an example of a smooth, proper variety $X$ (of dimension $>2$, over $\mathbb C$) for which ${{\mathcal S}}(X) \neq {{\mathcal S}}^2(X)$. For this variety $X$, we show that the morphism ${{\mathcal S}}(X) \to \pi_0^{{{\mathbb A}}^1}(X)$ is not an isomorphism (disproving the conjecture of Asok-Morel [@Asok-Morel Conjecture 2.2.8]). We also show that $Sing_*(X)$ is not ${{\mathbb A}}^1$-local, thus answering a question raised in [@Asok-Morel Remark 2.2.9]. We also exhibit an example of a smooth proper scheme $X$ (Example \[example pi\_0 not birational\]) for which the conjecture by Asok-Morel about the birationality of $\pi_0^{{{\mathbb A}}^1}(X)$ fails to hold.
Section \[Section lim S\^n\] begins with some preliminaries and generalities on ${{\mathbb A}}^1$-homotopy theory. We then give a proof of Theorem \[Intro-theorem lim S\^n\] and obtain equivalent characterizations of Morel’s conjecture about ${{\mathbb A}}^1$-invariance of $\pi_0^{{{\mathbb A}}^1}$ of a simplicial sheaf and an explicit conjectural description of $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$. In Section \[Section Asok-Morel\], we study the sheaves of ${{\mathbb A}}^1$-connected and chain-connected components of a scheme and give a proof of Theorem \[Intro-theorem field case\] and prove the conjectures of Morel and of Asok-Morel for non-uniruled surfaces. Finally, in Section \[section counterexamples\], we give an examples of schemes $X$ for which the conjectures of Asok-Morel about birationality of $\pi_0^{{{\mathbb A}}^1}(X)$ and equivalence of ${{\mathbb A}}^1$-connected and chain-connected components fail to hold.
A1-connected components of a simplicial sheaf {#Section lim S^n}
=============================================
Basic definitions and generalities {#subsection generalities}
----------------------------------
We will always work over a base field $k$. Let $Sm/k$ denote the category of smooth finite type schemes over $k$ with Nisnevich topology. We let ${\Delta^{op}(P{{\mathcal S}}h(Sm/{k}))}$ (resp. ${\Delta^{op}({{\mathcal S}}h(Sm/{k}))}$) denote the category of simplicial presheaves (resp. sheaves) of sets on $Sm/k$. All presheaves or sheaves will also be considered as constant simplicial objects. There is an adjunction $${\Delta^{op}(P{{\mathcal S}}h(Sm/{k}))}\overset{a_{Nis}}{\underset{i}{\rightleftarrows}} {\Delta^{op}({{\mathcal S}}h(Sm/{k}))},$$ where $i$ denotes the inclusion and $a_{Nis}$ denotes the Nisnevich sheafification. Both the categories have a model category structure, called the *Nisnevich local model structure* (see [@Morel-Voevodsky] and [@Jardine Appendix B]), where weak equivalences are local (stalkwise) weak equivalences and cofibrations are monomorphisms. The corresponding homotopy category of ${\Delta^{op}({{\mathcal S}}h(Sm/{k}))}$ is called the *simplicial homotopy category* and is denoted by $\mathcal H_s(k)$. Moreover, the above adjunction is a Quillen equivalence.
The left Bousfield localization of this model structure on $\bigtriangleup^{op}P{{{\mathcal S}}h}(Sm/k)$ with respect to the collection of projection maps $\mathcal{X} \times \mathbb{A}^1 \to \mathcal{X}$ is called the *$\mathbb{A}^1$-model structure*. The resulting homotopy category is called the *${{\mathbb A}}^1$-homotopy category* and is denoted by $\mathcal{H}(k)$.
There exists an *${{\mathbb A}}^1$-fibrant replacement functor* $$L_{\mathbb{A}^1} : {\Delta^{op}(P{{\mathcal S}}h(Sm/{k}))}\to {\Delta^{op}(P{{\mathcal S}}h(Sm/{k}))}$$ such that for any space $\mathcal{X}$, the object $L_{{{\mathbb A}}^1}(\mathcal{X})$ is an $\mathbb{A}^1$-fibrant object. Moreover, there exists a canonical morphism $\mathcal{X} \to L_{{{\mathbb A}}^1}(\mathcal{X})$ which is an $\mathbb{A}^1$-weak equivalence. A simplicial sheaf on $Sm/k$ is $\mathbb{A}^1$-fibrant if and only if it is simplicially fibrant (that is, fibrant in the Nisnevich local model structure) and $\mathbb{A}^1$-local ([@Morel-Voevodsky 2, Proposition 3.19]). See [@Morel-Voevodsky 2, Theorem 1.66 and p. 107] for the definition of the ${{\mathbb A}}^1$-fibrant replacement functor and more details.
With the above notation, it is easy to see that $i$ preserves fibrations, ${{\mathbb A}}^1$-fibrations and ${{\mathbb A}}^1$-local objects, and $a_{Nis}$ preserves cofibrations and ${{\mathbb A}}^1$-weak equivalences.
For any simplicial presheaf (or a sheaf) ${{\mathcal X}}$ on $Sm/k$, we define $\pi_0^s({{\mathcal X}})$ to be the Nisnevich sheafification of the presheaf of sets ${\pi_0}({{\mathcal X}})$ given by $${\pi_0}({{\mathcal X}})(U) := {{\rm Hom}}_{\mathcal H_s(k)}(U, {{\mathcal X}}).$$
\[notation1\] If we have two sets $R,S$ with maps $f,g:R \to S$, we denote by $\dfrac{S}{_gR_f}$, the quotient of $S$ by the equivalence relation generated by declaring $f(t)\sim g(t)$ for all $t\in R$. In this notation, for any simplicial presheaf ${{\mathcal X}}$, $${\pi_0}({{\mathcal X}})(U) = \frac{{{\mathcal X}}_0(U)}{_{d_0}{{\mathcal X}}_1(U)_{d_1}},$$ where $d_0,d_1: {{\mathcal X}}_1(U) \to {{\mathcal X}}_0(U)$ are the face maps in the simplicial set ${{\mathcal X}}_{\bullet}(U)$.
\[definition a1 connected components\] The sheaf of *${{\mathbb A}}^1$-connected components* of a simplicial sheaf ${{\mathcal X}}\in {\Delta^{op}({{\mathcal S}}h(Sm/{k}))}$ is defined to be $$\pi_0^{{{\mathbb A}}^1}({{\mathcal X}}) := \pi_0^s(L_{{{\mathbb A}}^1}({{\mathcal X}})).$$
In other words, for ${{\mathcal X}}\in {\Delta^{op}({{\mathcal S}}h(Sm/{k}))}$, $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ is the sheafification of the presheaf $$U \in Sm/k ~ \mapsto ~ {{\rm Hom}}_{\mathcal H(k)}(U, {{\mathcal X}}).$$
\[definition a1 invariance\] A presheaf ${{\mathcal F}}$ of sets on $Sm/k$ is said to be [*${{\mathbb A}}^1$-invariant*]{} if for every $U$, the map $ {{\mathcal F}}(U)\to {{\mathcal F}}(U\times {{\mathbb A}}^1)$ induced by the projection $U\times {{\mathbb A}}^1 \to U$, is a bijection.
It is easy to see that the presheaf $$U \mapsto {\pi_0}(L_{{{\mathbb A}}^1}({{\mathcal X}})(U))$$ is ${{\mathbb A}}^1$-invariant. A conjecture of Morel (see [@Morel 1.12]) asserts that its Nisnevich sheafification $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ remains ${{\mathbb A}}^1$-invariant.
\[a1invariance\] For any simplicial sheaf ${{\mathcal X}}$, $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ is ${{\mathbb A}}^1$-invariant.
The following example shows that sheafification in Nisnevich topology can destroy ${{\mathbb A}}^1$-invariance of a presheaf, in general.
Consider the presheaf ${{\mathcal F}}$ on $Sm/k$ whose sections ${{\mathcal F}}(U)$ are defined to be the set of $k$-morphisms from $U \to {{\mathbb A}}^1_{k}$ which factor through a proper open subset of ${{\mathbb A}}^1_{k}$. One can observe that ${{\mathcal F}}$ is ${{\mathbb A}}^1$-invariant, but its Nisnevich sheafification, which is the sheaf represented by ${{\mathbb A}}^1_{k}$, is clearly not ${{\mathbb A}}^1$-invariant.
In [@Choudhury Theorem 4.18], Morel’s conjecture was proved in the special case when ${{\mathcal X}}$ is a motivic $H$-group or a homogeneous space for a motivic $H$-group. As observed in [@Choudhury], such simplicial sheaves ${{\mathcal X}}$ have a special property - for $U \subset Y$, an inclusion of dense open subscheme, $\pi_0^s({{\mathcal X}})(Y)\to \pi_0^s({{\mathcal X}})(U)$ is injective (see [@Choudhury Corollary 4.17]). Unfortunately, this injectivity property fails for general simplicial sheaves ${{\mathcal X}}$, as shown by the following example.
\[nonseparated\] Let $U$ be a non-empty proper subscheme of an abelian variety $A$ over $k$. Let $X$ be the non-separated scheme obtained by gluing two copies of $A$ along $U$. Let $X$ continue to denote the corresponding constant simplicial sheaf in ${\Delta^{op}({{\mathcal S}}h(Sm/{k}))}$. Clearly, $X(A)\to X(U)$ is not injective. However, one can verify that the sheaf represented by $X$ is ${{\mathbb A}}^1$-invariant and hence, $X=\pi_0^{{{\mathbb A}}^1}(X)$.
For any simplicial sheaf ${{\mathcal X}}$, it is easy to see from the definitions that the sheaf $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ has the following universal property.
\[universal property pi\_0A1\] Let ${{\mathcal X}}\in {\Delta^{op}({{\mathcal S}}h(Sm/{k}))}$ be a simplicial sheaf and ${{\mathcal F}}$ be an ${{\mathbb A}}^1$-invariant sheaf of sets on $Sm/k$. Then any morphism ${{\mathcal X}}\to {{\mathcal F}}$ factors uniquely through the canonical morphism ${{\mathcal X}}\to \pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$.
We have a commutative diagram $$\xymatrix{
{{\mathcal X}}\ar[d] \ar[r] & {{\mathcal F}}\ar[d] \\
L_{{{\mathbb A}}^1}({{\mathcal X}}) \ar[r] & L_{{{\mathbb A}}^1}({{\mathcal F}})
.}$$ Since ${{\mathcal F}}$ is ${{\mathbb A}}^1$-invariant, the map ${{\mathcal F}}\to L_{{{\mathbb A}}^1}({{\mathcal F}})$ is an isomorphism. Hence, the map ${{\mathcal X}}\to {{\mathcal F}}$ factors through $L_{{{\mathbb A}}^1}({{\mathcal X}})$. The map $L_{{{\mathbb A}}^1}({{\mathcal X}}) \to {{\mathcal F}}$ factors through $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}}) = \pi_0^s(L_{{{\mathbb A}}^1}({{\mathcal X}}))$ since ${{\mathcal F}}$ is of simplicial dimension $0$.
We now recall the construction of the functor $Sing_*$ as given in [@Morel-Voevodsky p. 87]. Let $\Delta_{\bullet}$ denote the cosimplicial sheaf where $$\Delta_n = {{\rm Spec \,}}\left(\frac{k[x_0,...,x_n]}{(\sum_ix_i=1)}\right)$$ with obvious coface and codegeneracy maps motivated from those on topological simplices. For any simplicial presheaf (or a sheaf) ${{\mathcal X}}$, define $Sing_*({{\mathcal X}})$ to be the diagonal of the bisimplicial presheaf $\underline{{{\rm Hom}}}(\Delta_{\bullet},{{\mathcal X}})$, where $\underline{{{\rm Hom}}}$ denotes the internal Hom. Concretely, $$Sing_*({{\mathcal X}})_n = \underline{{{\rm Hom}}}(\Delta_n,{{\mathcal X}}_n).$$ Note that if ${{\mathcal X}}$ is a sheaf, then so is $Sing_*({{\mathcal X}})$. There is a functorial morphism ${{\mathcal X}}\to Sing_*({{\mathcal X}})$ induced by ${{\mathcal X}}_n(U) \to {{\mathcal X}}_n(U\times\Delta_n)$, which is an ${{\mathbb A}}^1$-weak equivalence. The functor $Sing_*$ takes ${{\mathbb A}}^1$-fibrant objects to ${{\mathbb A}}^1$-fibrant objects and takes ${{\mathbb A}}^1$-homotopic maps to simplicially homotopic maps.
\[definition A1 chain connected components\] Let ${{\mathcal F}}$ be a Nisnevich sheaf of sets on $Sm/k$, considered as a simplicially constant sheaf in $\bigtriangleup^{op}Sh(Sm/k)$. Define ${{\mathcal S}}({{\mathcal F}})$ to be the sheaf associated to the presheaf ${{\mathcal S}}^{pre}$ given by $${{\mathcal S}}^{pre}(U) := \frac{{{\mathcal F}}(U)}{_{\sigma_0}{{\mathcal F}}(U\times{{\mathbb A}}^1)_{\sigma_1}},$$ for $U \in Sm/k$, where $\sigma_0, \sigma_1 : {{\mathcal F}}(U \times {{\mathbb A}}^1) \to {{\mathcal F}}(U)$ are the maps induced by the $0$- and $1$-sections $U \to U \times {{\mathbb A}}^1$, respectively (see Notation \[notation1\]). Note that, even if ${{\mathcal F}}$ is a sheaf, ${{\mathcal S}}^{pre}({{\mathcal F}})$ need not be a sheaf. For any $n > 1$, we inductively define the sheaves $${{\mathcal S}}^n({{\mathcal F}}) := {{\mathcal S}}({{\mathcal S}}^{n-1}({{\mathcal F}})).$$
For any sheaf ${{\mathcal F}}$, there exists a canonical epimorphism ${{\mathcal F}}\to {{\mathcal S}}({{\mathcal F}})$. This gives a canonical epimorphism $${{\mathcal F}}\to \underset{n}{\varinjlim}~ {{\mathcal S}}^n({{\mathcal F}}).$$
\[remark-S-Sing\] It is clear from the definition of $Sing_*({{\mathcal F}})$ that $${\pi_0}(Sing_*({{\mathcal F}}))(U) = \frac{{{\mathcal F}}(U)}{_{\sigma_0}{{\mathcal F}}(U\times{{\mathbb A}}^1)_{\sigma_1}}.$$ Thus, ${{\mathcal S}}({{\mathcal F}}) = \pi_0^s(Sing_*({{\mathcal F}}))$, for any sheaf ${{\mathcal F}}$.
Let $X$ be a smooth scheme over $k$ and view it as a Nisnevich sheaf of sets over $Sm/k$. The sheaf ${{\mathcal S}}(X)$ defined above (Definition \[definition A1 chain connected components\]) is none other than the sheaf $\pi_0^{ch}(X)$ of *${{\mathbb A}}^1$-chain connected components* of $X$ introduced by Asok and Morel [@Asok-Morel Definition 2.2.4]. We use ${{\mathcal S}}$ instead of $\pi_0^{ch}$ in this paper only for typographical reasons.
For smooth schemes over a field $k$, Asok and Morel have conjectured the following (see [@Asok-Morel Conjecture 2.2.8]):
\[conjecture Asok-Morel\] For any smooth scheme $X$ over a field $k$, the natural epimorphism ${{\mathcal S}}(X) \to \pi_0^{{{\mathbb A}}^1}(X)$ is an isomorphism.
Asok and Morel also mention that this would be true if one proves that $Sing_*(X)$ is ${{\mathbb A}}^1$-local, for any smooth scheme $X$ over $k$ (see [@Asok-Morel Remark 2.2.9]). However, in Section \[section counterexamples\], we give examples of schemes for which this property and Conjecture \[conjecture Asok-Morel\] fail to hold.
Remarks on Morel’s conjecture on A1-invariance of pi0A1(X)
----------------------------------------------------------
In this subsection, we use the sheaf of ${{\mathbb A}}^1$-chain connected components of a (Nisnevich) sheaf ${{\mathcal F}}$ on $Sm/k$ to construct a sheaf ${{\mathcal L}}({{\mathcal F}})$ that is closely related to $\pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$ and prove Theorem \[Intro-theorem lim S\^n\] stated in the introduction.
\[theorem lim S\^n\] For any sheaf of sets ${{\mathcal F}}$ on $Sm/k$, the sheaf $${{\mathcal L}}({{\mathcal F}}) := \underset{n}{\varinjlim}~ S^n({{\mathcal F}})$$ is ${{\mathbb A}}^1$-invariant.
It is enough to show that for every smooth $k$-scheme $U$, the map $${{\mathcal L}}({{\mathcal F}})(U)\to {{\mathcal L}}({{\mathcal F}})(U\times{{\mathbb A}}^1)$$ is surjective, since it is already injective because the projection map $U \times {{\mathbb A}}^1 \to U$ admits a section. Let $t\in {{\mathcal L}}({{\mathcal F}})(U\times {{\mathbb A}}^1)$. Since the sheaf ${{\mathcal L}}({{\mathcal F}})$ is a filtered colimit, we have for each $U \in Sm/k$ $${{\mathcal L}}({{\mathcal F}})(U) = {\mathop{\varinjlim}\limits}_{n \in {{\mathbb N}}} {{\mathcal S}}^n({{\mathcal F}})(U).$$ Thus, $t$ is represented by an element $t_n$ of ${{\mathcal S}}^n({{\mathcal F}})(U\times {{\mathbb A}}^1)$ for some $n$. We will show that $t$ is contained in the image of ${{\mathcal L}}({{\mathcal F}})(U)$ by showing that the image of $t_n$ in ${{\mathcal S}}^{n+1}(U \times {{\mathbb A}}^1)$ is contained in the image of ${{\mathcal S}}^{n+1}({{\mathcal F}})(U)$. Let $m$ denote the *multiplication* map $$U\times {{\mathbb A}}^1 \times {{\mathbb A}}^1 \to U\times {{\mathbb A}}^1 \ \ ; \ \ (u,x,y) \mapsto (u,xy)$$ Consider the element $$m^*(t_n) \in {{\mathcal S}}^n({{\mathcal F}})(U\times {{\mathbb A}}^1\times {{\mathbb A}}^1).$$ Then $$\sigma_1^*\circ m^* (t_n) = t_n, \ \ \text{and} \ \ \sigma_0^*\circ m^*(t_n) = p^*\circ \sigma_0^*(t_n)$$ where $U\times {{\mathbb A}}^1 \stackrel{p}{\to} U$ is the projection. Thus, the image of $t_n$ in ${{\mathcal S}}^{n+1}({{\mathcal F}})(U\times {{\mathbb A}}^1)$ is contained in the image of the map $${{\mathcal S}}^{n+1}({{\mathcal F}})(U) \to {{\mathcal S}}^{n+1}({{\mathcal F}})(U\times {{\mathbb A}}^1).$$ This proves the result.
\[remark factorization through pi\_0-a1\] In view of Theorem \[theorem lim S\^n\], the canonical map ${{\mathcal F}}\to {{\mathcal L}}({{\mathcal F}})$ uniquely factors through the canonical map ${{\mathcal F}}\to \pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$.
\[remark universal property L(F)\] Note that ${{\mathcal L}}({{\mathcal F}})$ satisfies the following universal property: any map from a sheaf ${{\mathcal F}}$ to an ${{\mathbb A}}^1$-invariant sheaf uniquely factors through the canonical map ${{\mathcal F}}\to {{\mathcal L}}({{\mathcal F}})$. Recall that $\pi_0^{{{\mathbb A}}^1}$ also satisfies the same universal property (Lemma \[universal property pi\_0A1\]) but is not known to be ${{\mathbb A}}^1$-invariant in general.
We now see how Theorem \[theorem lim S\^n\] gives us a reformulation of Morel’s conjecture on the ${{\mathbb A}}^1$-invariance of the ${{\mathbb A}}^1$-connected components sheaf $\pi_0^{{{\mathbb A}}^1}$.
\[a1-invariance-equiv-sheaves\] Let ${{\mathcal F}}$ be a sheaf of sets on $Sm/k$, considered as an object of $\bigtriangleup^{op}{{{\mathcal S}}h}(Sm/k)$. The following are equivalent:\
$(1)$ The sheaf $\pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$ is ${{\mathbb A}}^1$-invariant.\
$(2)$ The canonical map $$\pi_0^{{{\mathbb A}}^1}({{\mathcal F}}) \to {{\mathcal L}}({{\mathcal F}})$$ admits a retract.
$(1) \Rightarrow (2)$ is straightforward by Remark \[remark universal property L(F)\] and the universal property of $\pi_0^{{{\mathbb A}}^1}$ (Lemma \[universal property pi\_0A1\]). $(2) \Rightarrow (1)$ follows since ${{\mathcal L}}({{\mathcal F}})$ is ${{\mathbb A}}^1$-invariant (Theorem \[theorem lim S\^n\]) and since a retract of an ${{\mathbb A}}^1$-invariant sheaf is ${{\mathbb A}}^1$-invariant.
\[a1-invariance-equiv-spaces\] The following are equivalent:\
$(1)$ The sheaf $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ is ${{\mathbb A}}^1$-invariant, for all ${{\mathcal X}}\in \bigtriangleup^{op}{{{\mathcal S}}h}(Sm/k)$.\
$(2)$ The canonical map $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}}) \to \pi_0^{{{\mathbb A}}^1}(\pi_0^s({{\mathcal X}}))$ is an isomorphism, for all ${{\mathcal X}}\in \bigtriangleup^{op}{{{\mathcal S}}h}(Sm/k)$.
$(1) \Rightarrow (2):$ Consider the following commutative diagram with the natural morphisms $$\xymatrix{
{{\mathcal X}}\ar[r] & \pi_0^s({{\mathcal X}}) \ar[r] \ar[d] & \pi_0^{{{\mathbb A}}^1}({{\mathcal X}}) \ar[dl] \\
& \pi_0^{{{\mathbb A}}^1}(\pi_0^s({{\mathcal X}})) & }$$ All the morphisms in the above diagram are epimorphisms (the right horizontal map is an epimorphism by the unstable ${{\mathbb A}}^1$-connectivity theorem [@Morel-Voevodsky 2, Corollary 3.22]). Since $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ is ${{\mathbb A}}^1$-invariant by assumption, the map $\pi_0^s({{\mathcal X}}) \to \pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ has to uniquely factor through $\pi_0^{{{\mathbb A}}^1}(\pi_0^s({{\mathcal X}}))$, by the universal property of $\pi_0^{{{\mathbb A}}^1}(\pi_0^s({{\mathcal X}}))$. This gives an inverse to the map $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}}) \to \pi_0^{{{\mathbb A}}^1}(\pi_0^s({{\mathcal X}}))$, by uniqueness.\
$(2) \Rightarrow (1):$ Let ${{\mathcal X}}$ be a space and let ${{\mathcal F}}:= \pi_0^s({{\mathcal X}})$ be the sheaf of (simplicial) connected components of ${{\mathcal X}}$. We are given that $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}}) \to \pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$ is an isomorphism. We will prove that $\pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$ is ${{\mathbb A}}^1$-invariant. We have a natural map $${{\mathcal S}}({{\mathcal F}}) = \pi_0^s(Sing_*({{\mathcal F}})) \to \pi_0^{{{\mathbb A}}^1}(Sing_*({{\mathcal F}})) \simeq \pi_0^{{{\mathbb A}}^1}({{\mathcal F}}).$$ By Remark \[remark-S-Sing\], we have $${{\mathcal S}}^2({{\mathcal F}}) = {{\mathcal S}}(\pi_0^s(Sing_*({{\mathcal F}}))) = \pi_0^s(Sing_* \pi_0^s(Sing_*({{\mathcal F}}))),$$ whence we have a map ${{\mathcal S}}^2({{\mathcal F}}) \to \pi_0^{{{\mathbb A}}^1}(Sing_* \pi_0^s(Sing_*({{\mathcal F}})))$. But hypothesis $(2)$ implies that $$\pi_0^{{{\mathbb A}}^1}(Sing_* \pi_0^s(Sing_*({{\mathcal F}}))) \simeq \pi_0^{{{\mathbb A}}^1}(\pi_0^s(Sing_*({{\mathcal F}}))) \simeq \pi_0^{{{\mathbb A}}^1}(Sing_*({{\mathcal F}})) \simeq \pi_0^{{{\mathbb A}}^1}({{\mathcal F}}).$$ The composition of these maps gives a natural map ${{\mathcal S}}^2({{\mathcal F}}) \to \pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$, which makes the following diagram commute, where the maps are the ones defined above. $$\xymatrix{
{{\mathcal S}}({{\mathcal F}}) \ar[dr] \ar[rr] & & {{\mathcal S}}^2({{\mathcal F}}) \ar[dl]\\
{} & \pi_0^{{{\mathbb A}}^1}({{\mathcal F}}) & {}
}$$ Continuing in this way, we obtain a compatible family of maps ${{\mathcal S}}^i({{{\mathcal F}}}) \to \pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$, for each $i$, giving a retract ${{\mathcal L}}({{\mathcal F}}) \to \pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$. Lemma \[a1-invariance-equiv-sheaves\] now implies that $\pi_0^{{{\mathbb A}}^1}({{\mathcal F}})$ is ${{\mathbb A}}^1$-invariant, proving $(1)$.
These results immediately give a description of the ${{\mathbb A}}^1$-connected components sheaf $\pi_0^{{{\mathbb A}}^1}({{\mathcal X}})$ for a simplicial sheaf of sets ${{\mathcal X}}$ on $Sm/k$, provided Morel’s conjecture is true.
\[a1-invariance-candidate\] Suppose Conjecture \[a1invariance\] is true. Let ${{\mathcal X}}$ be a simplicial sheaf of sets on $Sm/k$ and let ${{\mathcal F}}:= \pi_0^s({{\mathcal X}})$. Then the canonical maps $$\pi_0^{{{\mathbb A}}^1}({{\mathcal X}}) \to \pi_0^{{{\mathbb A}}^1}({{\mathcal F}}) \to {{\mathcal L}}({{\mathcal F}})$$ are isomorphisms.
A1-connected/chain-connected components of schemes {#Section Asok-Morel}
==================================================
Sections of pi0A1(X) over finitely generated field extensions of k {#subsection field case}
------------------------------------------------------------------
The aim of this subsection is to give a proof of Theorem \[Intro-theorem field case\] stated in the introduction. The strategy is the one suggested by Theorem \[theorem lim S\^n\]: we shall prove that for a proper scheme $X$ over $k$, ${{\mathcal S}}(X)({{\rm Spec \,}}L) = {{\mathcal S}}^2(X)({{\rm Spec \,}}L)$, for every field extension $L$ of $k$.
Let ${{\mathcal F}}$ be any sheaf of sets on $Sm/k$. In order to address the question of ${{\mathbb A}}^1$-invariance of ${{\mathcal S}}({{\mathcal F}})$, we need to obtain an explicit description for the elements of the set ${{\mathcal S}}({{\mathcal F}})(U)$ where $U$ is a smooth scheme over $k$. We will then specialize to the case when ${{\mathcal F}}$ is represented by a proper scheme $X$ over $k$.
As before, for any scheme $U$ over $k$, we let $\sigma_0$ and $\sigma_1$ denote the morphisms $U \to U \times {{\mathbb A}}^1$ given by $u \mapsto (u,0)$ and $u \mapsto (u,1)$, respectively.
\[definition homotopy\] Let ${{\mathcal F}}$ be a sheaf of sets over $Sm/k$ and let $U$ be a smooth scheme over $k$.
- An *${{\mathbb A}}^1$-homotopy* of $U$ in ${{\mathcal F}}$ is a morphism an element $h$ of ${{\mathcal F}}(U \times {{\mathbb A}}^1_{k})$. We say that $t_1, t_2 \in {{\mathcal F}}(U)$ are *${{\mathbb A}}^1$-homotopic* if there exists an ${{\mathbb A}}^1$-homotopy $h \in {{\mathcal F}}(U \times {{\mathbb A}}^1)$ such that $\sigma_0^*(h) = t_1$ and $\sigma_1^*(h)= t_2)$.
- An *${{\mathbb A}}^1$-chain homotopy* of $U$ in ${{\mathcal F}}$ is a finite sequence $h=(h_1, \ldots, h_n)$ where each $h_i$ is an ${{\mathbb A}}^1$-homotopy of $U$ in ${{\mathcal F}}$ such that $\sigma_1^*(h_i) = \sigma_0^*(h_{i+1})$ for $1 \leq i \leq n-1$. We say that $t_1, t_2 \in {{\mathcal F}}(U)$ are *${{\mathbb A}}^1$-chain homotopic* if there exists an ${{\mathbb A}}^1$-chain homotopy $h=(h_1, \ldots, h_n)$ such that $\sigma_0^*(h_1) = t_1$ and $\sigma_1^*(h_n) = t_2$.
Clearly, if $t_1$ and $t_2$ are ${{\mathbb A}}^1$-chain homotopic, they map to the same element of ${{\mathcal S}}({{\mathcal F}})(U)$. The converse is partially true - if $t_1, t_2 \in {{\mathcal F}}(U)$ are such that they map to the same element in ${{\mathcal S}}({{\mathcal F}})(U)$, then there exists a Nisnevich cover $V \to U$ such that $t_1|_{V}$ and $t_2|_{V}$ are ${{\mathbb A}}^1$-chain homotopic.
Now we explicitly describe the sections of ${{\mathcal S}}({{\mathcal F}})$. Let $t \in {{\mathcal S}}({{\mathcal F}})(U)$. Since ${{\mathcal F}}\to {{\mathcal S}}({{\mathcal F}})$ is an epimorphism of sheaves, there exists a finite Nisnevich cover $V \to U$ such that $t|_{V}$ can be lifted to $s \in {{\mathcal F}}(V)$. Let $pr_1,pr_2: V \times_{U} V \to V$ be the two projections. Then, since the two elements $pr_i^*(s)$ map to the same element in ${{\mathcal S}}({{\mathcal F}})(V \times_U V)$, there exists a finite Nisnevich cover $W \to V \times_U V$ such that $pr_1^*(s)|_{W}$ and $pr_2^*(s)|_{W}$ are ${{\mathbb A}}^1$-chain homotopic. Conversely, given a finite cover $V \to U$ and an element $s$ of ${{\mathcal F}}(V)$ such that $pr_1^*(s)$ and $pr_2^*(s)$ become ${{\mathbb A}}^1$-chain homotopic after restricting to a Nisnevich cover of $V \times_U V$, we can obtain a unique element $t$ of ${{\mathcal S}}({{\mathcal F}})(U)$.
Applying the same argument to to the sections of ${{\mathcal S}}({{\mathcal F}})$ over ${{\mathbb A}}^1_U$, we are led to define the notion of an *${{\mathbb A}}^1$-ghost homotopy*.
\[definition ghost homotopy\] Let ${{\mathcal F}}$ be a sheaf of sets and let $U$ be a smooth scheme over $k$. An *${{\mathbb A}}^1$-ghost homotopy* of $U$ in ${{\mathcal F}}$ consists of the data $${{\mathcal H}}:=(V \to {{\mathbb A}}^1_U, W \to V \times_{{{\mathbb A}}^1_U} V, h, h^W)$$ which is defined as follows:
- $V \to {{\mathbb A}}^1_U$ is a Nisnevich cover of ${{\mathbb A}}^1_U$.
- $W \to V \times_{{{\mathbb A}}^1_U} V$ is a Nisnevich cover of $V \times_{{{\mathbb A}}^1_U} V$.
- $h$ is a morphism $V \to {{\mathcal F}}$.
- $h^W$ is an ${{\mathbb A}}^1$-chain homotopy connecting the two morphisms $W \to V\times_{{{\mathbb A}}^1_U} V \overset{pr_i}{\to} V \to F$ where $pr_1$ and $pr_2$ are the projections $V \times_{{{\mathbb A}}^1_U} V \to V$. (Thus $h^W$ is a finite sequence $(h^{W}_1, \ldots )$ of ${{\mathbb A}}^1$-homotopies satisfying the appropriate conditions as given in Definition \[definition homotopy\].)
Let $t_1, t_2 \in {{\mathcal F}}(U)$. We say that *${{\mathcal H}}$ connects $t_1$ and $t_2$* (which are then said to be *${{\mathbb A}}^1$-ghost homotopic*) if the morphisms $\sigma_0, \sigma_1: U \to {{\mathbb A}}^1_U$ admit lifts $\~\sigma_0: U \to V$ and $\~\sigma_1: U \to V$ such that $h \circ \~\sigma_0 = t_1$ and $h \circ \~\sigma_1 = t_2$.
By the discussion above, a homotopy in ${{\mathcal S}}({{\mathcal F}})$ gives rise (non-uniquely) to a ghost homotopy in ${{\mathcal F}}$. On the other hand, a ghost homotopy in ${{\mathcal F}}$ gives rise to a unique homotopy in $S({{\mathcal F}})$. Also, if $t_1$, $t_2$ are ${{\mathbb A}}^1$-ghost homotopic elements of ${{\mathcal F}}(U)$, then it is clear that their images in ${{\mathcal S}}({{\mathcal F}})(U)$ are ${{\mathbb A}}^1$-homotopic. The converse is partially true – if $\~{t_1}, \~{t_2} \in {{\mathcal S}}({{\mathcal F}})(U)$ are ${{\mathbb A}}^1$-homotopic, then they have preimages $t_1, t_2 \in {{\mathcal F}}(U)$ which become ${{\mathbb A}}^1$-ghost homotopic over some Nisnevich cover of $U$. In general, if ${{\mathcal H}}$ is an ${{\mathbb A}}^1$-ghost homotopy of $U$, it is possible that there may not exist two elements $t_1, t_2 \in {{\mathcal F}}(U)$ which are connected by ${{\mathcal H}}$. This is because the lifts $\~\sigma_0$ and $\~\sigma_1$ may not exist until we pass to a Nisnevich cover of $U$. However, if $U$ is a smooth Henselian local scheme over $k$, then the morphisms $\~\sigma_0$ and $\~\sigma_1$ do exist and thus there exist preimages of $\~{t_1}$ and $\~{t_2}$ which are ${{\mathbb A}}^1$- ghost homotopic.
From the above discussion, we obtain the following obvious lemma:
\[lemma strategy\] Let ${{\mathcal F}}$ be a sheaf of sets over $Sm/k$. Then, ${{\mathcal S}}({{\mathcal F}}) = {{\mathcal S}}^2({{\mathcal F}})$ if and only if for every smooth Henselian local scheme $U$, if $t_1, t_2 \in {{\mathcal F}}(U)$ are ${{\mathbb A}}^1$-ghost homotopic, then they are also ${{\mathbb A}}^1$-chain homotopic.
The rest of this subsection is devoted to the proof of Theorem \[Intro-theorem field case\]. We introduce the concept of *almost proper sheaves* (see Definition \[definition-almost-proper\] below), which we shall find helpful while handling the iterations of the functor ${{\mathcal S}}$, for the purposes of this paper.
\[notation-agree-on-points\] Let $Z$ be a variety over $k$ and let $x$ be a point of $Z$. Let $k(x) = {{\mathcal O}}_{Z,x}/\mathfrak{m}_x$ be the residue field at $x$. The quotient homomorphism ${{\mathcal O}}_{X,x} \to k(x)$ induces a morphism ${{\rm Spec \,}}(k(x)) \to {{\rm Spec \,}}({{\mathcal O}}_{Z,x}) \to Z$ which will also be denoted by $x: {{\rm Spec \,}}(k(x)) \to Z$.
Let ${{\mathcal F}}$ be a sheaf of sets and let $Z$ be a variety over $k$. Let $s_1, s_2 \in {{\mathcal F}}(Z)$. We will write $s_1 \sim_0 s_2$ if for any point $x$ of $Z$, the morphisms $s_1 \circ x, s_2 \circ x: {{\rm Spec \,}}(k(x)) \to {{\mathcal F}}$ are identical.
In Notation \[notation-agree-on-points\], the extension $k(x)/k$ may have positive transcendence degree. For instance, this may happen when $U$ is a smooth, irreducible scheme of dimension $\geq 1$ over $k$ and $x$ is the generic point of $U$. Then ${{\rm Spec \,}}(k(x))$ is viewed as an essentially smooth scheme over $k$ and the set ${{\mathcal F}}({{\rm Spec \,}}(k(x))$ is the direct limit $${{\mathcal F}}({{\rm Spec \,}}(k(x)) := \varinjlim_{U^{\prime}} {{\mathcal F}}(U^{\prime})$$ where $U^{\prime}$ varies over all the open subschemes of $U$. Thus any morphism ${{\rm Spec \,}}(k(x)) \to {{\mathcal F}}$ can be represented by a morphism $U^{\prime} \to {{\mathcal F}}$ where $U^{\prime}$ is some open subscheme of $U$. In the following discussion, we will also have to consider morphisms of the form ${{\mathbb A}}^1_{k(x)} \to {{\mathcal F}}$. Such a morphism can be represented by a morphism ${{\mathbb A}}^1_{U^{\prime}} \to {{\mathcal F}}$ where $U^{\prime}$ is an open subscheme of $U$.
\[definition-almost-proper\] Let ${{\mathcal F}}$ be a sheaf of sets over $Sm/k$. We say that ${{\mathcal F}}$ is *almost proper* if the following two properties hold:
- Let $U$ be a smooth, irreducible variety of dimension $\leq 2$ and let $s: U \to {{\mathcal F}}$ be any morphism. Then there exists a smooth projective variety $\-U$, a birational map $i: U \dashrightarrow \-U$ and a morphism $\-s: \-U \to {{\mathcal F}}$ which “extends $s$ on points”. By this we mean that if $U^{\prime}$ is the largest open subscheme of $U$ such that $i$ is represented by a morphism $i^{\prime}: U^{\prime} \to \-U$, then we have $\-s \circ i^{\prime} \sim_0 s|_{U^{\prime}}$.
- Let $U$ be a smooth, irreducible curve over $k$ and let $U^{\prime}$ be an open subscheme. Suppose we have two morphisms $s_1, s_2: U \to {{\mathcal F}}$ such that $s_1|_{U^{\prime}} = s_2|_{U^{\prime}}$. Then $s_1 \sim_0 s_2$.
It is clear that a proper scheme over $k$ represents an almost proper sheaf. Thus, in order to prove Theorem \[Intro-theorem field case\] in the introduction, it will suffice to prove that if ${{\mathcal F}}$ is an almost proper sheaf, then ${{\mathcal S}}({{\mathcal F}})(K) = {{\mathcal S}}^n({{\mathcal F}})(K)$ for any finitely generated field extension $K/k$. This will be done in Theorem \[theorem field case\] below.
\[lemma-connect-points\] Let ${{\mathcal F}}$ be an almost proper sheaf of sets on $Sm/k$. Let $U$ be a smooth curve, let $x: {{\rm Spec \,}}(k) \to U$ be a $k$-rational point on $U$ and let $U^{\prime}$ be the open subscheme $U \backslash \{x\}$ of $U$. Let $f,g: U \to {{\mathcal F}}$ be such that the morphisms $f|_{U^{\prime}}$ and $g|_{U^{\prime}}$ are ${{\mathbb A}}^1$-chain homotopic. Then the morphisms $f \circ x, g \circ x: {{\rm Spec \,}}(k) \to {{\mathcal F}}$ are ${{\mathbb A}}^1$-chain homotopic.
First we prove the result in the case when $f|_{U^{\prime}}$ and $g|_{U^{\prime}}$ are simply ${{\mathbb A}}^1$-homotopic, i.e. there exists a morphism $h: U^{\prime} \times {{\mathbb A}}^1 \to {{\mathcal F}}$ such that $h|_{U^{\prime} \times \{0\}} = f|_{U^{\prime}}$ and $h|_{U^{\prime} \times \{1\}} = g|_{U^{\prime}}$.
By (AP1), there exists a smooth proper surface $X$ and a proper, birational morphism $i: U' \times {{\mathbb A}}^1 \dashrightarrow X$ and a morphism $\-h: X \to {{\mathcal F}}$ “extending $h$ on points". As in Definition \[definition-almost-proper\], we elaborate on the precise sense in which $\-h$ extends $h$. Suppose that $W$ is the largest open subscheme of $U^{\prime} \times {{\mathbb A}}^1$ on which $i$ is defined. Then $\-h \circ i|_W \sim_0 h|_W$. Note that $(U^{\prime} \times {{\mathbb A}}^1) \backslash W$ is a closed subscheme of codimension $2$. Thus there exists an open subscheme $U'' \subset U'$ such that $U'' \times {{\mathbb A}}^1 \subset W$. The curves $U'' \times \{0\}$ and $U'' \times \{1\}$, which are contained in $W$, are mapped into $X$ by $i$. The restrictions of $i$ to these curves induce morphisms $t_0, t_1: \-U \to X$, where $\-U$ is a smooth compactification of $U$. Since $\-h \circ i|_W \sim_0 h|_W$, we see that $\-h \circ t_0|_{U''} \sim_0 f|_{U''}$ and $\-h \circ t_1|_{U''} \sim_0 g|_{U''}$. This implies that there exists a Zariski open subset $U'''$ of $U''$ such that $\-h \circ t_0|_{U'''} = f|_{U'''}$ and $\-h \circ t_1|_{U'''} = g|_{U''}$. Thus, by (AP2), $\-h \circ t_0 \circ x = f \circ x$ and $\-h \circ t_1 \circ x = g \circ x$. Thus in order to show that $f \circ x$ and $g \circ x$ are ${{\mathbb A}}^1$-chain homotopic, it suffices to show that $t_0 \circ x$ and $t_1 \circ x$ are ${{\mathbb A}}^1$-chain homotopic in $X$. Since $X$ is birational to $\-U \times {{\mathbb P}}^1$, by resolution of indeterminacy (see, for example, [@Hartshorne Chapter II, 7.17.3]), there exists a smooth projective surface $\~X$ and birational proper morphisms $\~X \to \-U \times {{\mathbb P}}^1$ and $\~X \to X$. $$\xymatrix{
& \~X \ar[dl] \ar[dr] & \\
\-U \times {{\mathbb P}}^1 & & X \\
& W \ar@{_(->}[ul] \ar[ur]^i&
}$$ Let $\-t_0: \-U \to \-U \times {{\mathbb P}}^1$ be the map that identifies $\-U$ with the curve $\-U \times \{(0:1)\}$ and $\-t_1: \-U \to \-U \times {{\mathbb P}}^1$ be the map that identifies $\-U$ with the curve $\-U \times \{(1:1)\}$. It is easy to see that there exist unique morphisms $\~t_0, \~t_1: \-U \to \~X$ such that $\~t_0$ is a lift of both $t_0$ and $\-t_0$ and $\~t_1$ is a lift of both $t_1$ and $\-t_1$. This completes the proof of the result when $f|_{U^{\prime}}$ and $g|_{U^{\prime}}$ are simply ${{\mathbb A}}^1$-homotopic by [@Asok-Morel Lemma 6.2.9], since $\~t_0 \circ x$ and $\~t_1 \circ x$ are ${{\mathbb A}}^1$-chain homotopic.
Now, suppose $f|_{U^{\prime}}$ and $g|_{U^{\prime}}$ are ${{\mathbb A}}^1$-chain homotopic. Thus, we have a sequence $ f_0 = f|_{U^{\prime}}, \ldots, f_n =g|_{U^{\prime}}$ where $f_i$ is ${{\mathbb A}}^1$-homotopic to $f_{i+1}$ for every $i$. By (AP1), for each $i$, there exists a $\~f_i: U \to {{\mathcal F}}$ such that $\~f_i|_{U^{\prime}} = f_i$. Thus $\~f_i \circ x$ and $\~f_{i+1} \circ x$ are ${{\mathbb A}}^1$-chain homotopic for every $i$. By (AP2), $\~f_0 \circ x = f \circ x$ and $\~f_n \circ x = g \circ x$.
\[lemma-almost-proper\] Let ${{\mathcal F}}$ be an almost proper sheaf. Then ${{\mathcal S}}({{\mathcal F}})$ is also an almost proper sheaf.
We first check condition (AP2). Thus, let $U$ be a smooth curve and let $U^{\prime} \subset U$ be an open subscheme. We assume that we have two morphisms $s_1, s_2: U \to {{\mathcal S}}({{\mathcal F}})$ such that $s_1|_{U^{\prime}} = s_2|_{U^{\prime}}$ and prove that $s_1 \sim_0 s_2$. Without any loss of generality, we may assume that $U \backslash U^{\prime}$ consists of a single closed point. By a change of base, if necessary, we may assume that this point is $k$-rational, that is, it is the image of a morphism $x: {{\rm Spec \,}}(k) \to U$. We need to prove that $s_1 \circ x = s_2 \circ x$.
The morphism $\phi: {{\mathcal F}}\to {{\mathcal S}}({{\mathcal F}})$ is an epimorphism of sheaves. Thus, there exists a smooth, irreducible curve $C$ and an étale morphism $p: C \to U$ such that $x$ can be lifted to a morphism $c_0: {{\rm Spec \,}}(k) \to C$, (i.e. $p \circ c_0 = x$) and such that the morphisms $p \circ s_1$ and $p \circ s_2$ can be lifted to ${{\mathcal F}}$. In other words, there exist morphisms $f,g: C \to {{\mathcal F}}$ such that $\phi \circ f =s_1 \circ p$ and $\phi \circ g = s_2 \circ p$. Since $s_1|_{U^{\prime}} = s_2|_{U^{\prime}}$, there exists a Nisnevich cover $q: V \to U^{\prime}$ such that $q \circ f$ and $q \circ g$ are ${{\mathbb A}}^1$-chain homotopic. Let $K$ denote the function field of $C$. Then the canonical morphism $\eta: {{\rm Spec \,}}(K) \to {{\mathcal F}}$ lifts to $V$. Thus the morphisms $f \circ \eta$ and $g \circ \eta$ are ${{\mathbb A}}^1$-chain homotopic. Thus, there exists an open subscheme $C^{\prime} \subset C$ such that if $i: C^{\prime} \hookrightarrow C$ is the inclusion, the morphisms $f \circ i$ and $g \circ i$ are ${{\mathbb A}}^1$-chain homotopic. Applying Lemma \[lemma-connect-points\], we see that $f \circ c_0, g \circ c_0: {{\rm Spec \,}}(k) \to {{\mathcal F}}$ are ${{\mathbb A}}^1$-chain homotopic. Thus $$s_1 \circ x = s_1 \circ p \circ c_0 = \phi \circ f \circ c_0 = \phi \circ g \circ c_0 = s_2 \circ p \circ c_0 = s_2 \circ x$$ as desired. This shows that the sheaf ${{\mathcal S}}({{\mathcal F}})$ satisfies the condition (AP2).
Now we check the condition (AP1). First suppose that $U$ is a smooth, irreducible variety of dimension $\leq 2$ and we have a morphism $s: U \to {{\mathcal S}}({{\mathcal F}})$. Since $\phi: {{\mathcal F}}\to {{\mathcal S}}({{\mathcal F}})$ is surjective, there exists a Nisnevich cover $p: V \to U$ of such that the morphism $s \circ p$ lifts to ${{\mathcal F}}$. Thus there exists an open subscheme $U^{\prime} \subset U$ such that the morphism $s|_{U^{\prime}}$ lifts to a morphism $t: U^{\prime} \to {{\mathcal F}}$. Applying the condition (AP1) for ${{\mathcal F}}$, there exists a smooth projective variety $\-U$, a birational map $i: U^{\prime} \dashrightarrow \-U$ and a morphism $\-t: \-U \to {{\mathcal F}}$ “extending" $t$. Let $U''$ be the largest open subscheme of $U^{\prime}$ such that the rational map $i$ can be represented by a morphism $i^{\prime}: U'' \to \-U$. We know that $\-t \circ i^{\prime} \sim_0 t|_{U''}$.
We claim that $\phi \circ \-t$ is the required morphism. We note that $i^{\prime}$ defines a rational map from $U$ to $\-U$. Let $U''' \subset U$ be the largest open subscheme of $U$ such that this rational map can be represented by a morphism $i'': U''' \to \-U$. So, we clearly have $U'' \subset U'''$. We wish to prove that for any point $x$ in $U'''$, the equality $s \circ x = \phi \circ \-s \circ i'' \circ x$ holds.
When $x \in U^{\prime \prime}$, we already know that $\-s \circ i^{\prime} \circ x = t|_{U^{\prime \prime}} \circ x$. Composing by $\phi$, we get that $\phi \circ \-s \circ i^{\prime} \circ x = \phi \circ t|_{U^{\prime \prime}} \circ x$. But $\phi \circ t|_{U^{\prime \prime}} = s|_{U^{\prime \prime}}$. Thus we see that $\phi \circ t|_{U^{\prime \prime}} \circ x = s \circ x$. Since $i''|_{U''} = i^{\prime}|_{U''}$, we obtain the desired equality in this case.
Thus we now assume that $x \in U''' \backslash U^{\prime \prime}$. Since $U^{\prime \prime}$ is a dense open subscheme of $U$, there exists a smooth curve $D$ over $k$ and a locally closed immersion $j: D \rightarrow U'''$ such that $j(D)$ contains $x$ and also intersects $U''$. (When $\dim(U) = 1$, $D$ will be an open subscheme of $U$.) Let $z$ be the generic point of $D$. Then we note that $j(z)$ is a point of $U''$ with residue field $k(z)$. By what we have proved in the previous paragraph, we know that $s \circ j(z) = \phi \circ \-s \circ i'' \circ j(z)$. Thus there exists an open subscheme $D^{\prime}$ of $D$ such that $j$ maps $D^{\prime}$ into $U''$ and $s \circ j|_{D^{\prime}} = \phi \circ \-s \circ i'' \circ j|_{D^{\prime}}$. Since we have already verified (AP2) for ${{\mathcal S}}({{\mathcal F}})$, we can conclude that $s \circ x = \phi \circ \-s \circ i'' \circ x$. This completes the proof of the fact that ${{\mathcal S}}({{\mathcal F}})$ is almost proper.
\[theorem field case\] Let ${{\mathcal F}}$ be an almost proper sheaf. Then for any field extension $K$ of $k$, ${{\mathcal S}}({{\mathcal F}})(K) = {{\mathcal S}}^n({{\mathcal F}})(K)$ for any integer $n \geq 1$.
By a base change, we may assume that $K = k$. Thus we need to prove that if ${{\mathcal F}}$ is almost proper, we have ${{\mathcal S}}({{\mathcal F}})(k) = {{\mathcal S}}^{n}({{\mathcal F}})(k)$ for all $n \geq 1$. In view of Lemma \[lemma-almost-proper\], it suffices to show that ${{\mathcal S}}({{\mathcal F}})(k) = {{\mathcal S}}^2({{\mathcal F}})(k)$.
Let $\phi$ denote the morphism of sheaves ${{\mathcal F}}\to {{\mathcal S}}({{\mathcal F}})$. Let $x$ and $y$ be elements of ${{\mathcal S}}({{\mathcal F}})(k)$ and suppose there exists a morphism $h: {{\mathbb A}}^1_k \to {{\mathcal S}}({{\mathcal F}})$ such that $h(0) = x$ and $h(1) = y$. As $\phi$ is an epimorphism of Nisnevich sheaves, there exists an open subscheme $U$ of ${{\mathbb A}}^1_k$ such that $h|_U$ can be lifted to ${{\mathcal F}}$, i.e. there exists a morphism $h^{\prime}: U \to {{\mathcal F}}$ such that $\phi \circ h^{\prime} = h|_{U}$. By (AP1), there exists a morphism $\-h: {{\mathbb P}}^1_k \to {{\mathcal F}}$ which “extends $h^{\prime}$ on points”, i.e. if $i: U \rightarrow {{\mathbb P}}^1_k$ is the composition $U \hookrightarrow {{\mathbb A}}^1_k \hookrightarrow {{\mathbb P}}^1_k$, for all points $x$ of $U$, we have $\-h \circ i \circ x = h^{\prime} \circ x$. (Here we identify ${{\mathbb A}}^1_k$ with the open subscheme of ${{\mathbb P}}^1_k = Proj(k[T_0,T_1])$ given by $T_1 \neq 0$.) Applying this to the generic point of $U$, we see that there exists an open subscheme $U^{\prime}$ of $U$ such that $\-h \circ i|_{U^{\prime}} = h^{\prime}|_{U^{\prime}}$. Since $i$ is just the inclusion, we may write this as $\-h|_{U'} = h'
|_{U'}$.
Consider the morphism $\~h:=\phi \circ \-h|_{{{\mathbb A}}^1_k}: {{\mathbb A}}^1_k \to {{\mathcal S}}({{\mathcal F}})$. We have the equalities $\~h|_{U^{\prime}} = \phi \circ h^{\prime}|_{U^{\prime}} = h|_{U^{\prime}}$. Thus, by (AP2), we see that $\~h \sim_0 h$. But this means that $x = (\phi \circ \-h)((0:1))$ and $y = (\phi \circ \-h)((1:1))$. In other words, $x$ and $y$ are images under $\phi$ of two ${{\mathbb A}}^1$-homotopic morphisms from ${{\rm Spec \,}}(k)$ into ${{\mathcal F}}$. But by the definition of the functor ${{\mathcal S}}$, this implies that $x = y$.
\[corollary field case\] If ${{\mathcal F}}$ is an almost proper sheaf of sets, the canonical surjection ${{\mathcal S}}(F)(K) \to \pi_0^{{{\mathbb A}}^1}(K)$ is a bijection, for any finitely generated field extension $K/k$. In particular, this holds for any proper scheme $X$ of finite type over $k$.
Theorem \[theorem field case\] and Theorem \[theorem lim S\^n\] together imply that $${{\mathcal S}}({{\mathcal F}})(K) \to {{\mathcal L}}({{\mathcal F}})(K)$$ is a bijection for all finitely generated field extensions $K$ of $k$. The result now follows by Remark \[remark factorization through pi\_0-a1\].
This completes the proof of Theorem \[Intro-theorem field case\] stated in the introduction.
The conjectures of Morel and of Asok-Morel for non-uniruled surfaces {#subsection non-uniruled}
--------------------------------------------------------------------
In this subsection, we show that Conjectures \[a1invariance\] and \[conjecture Asok-Morel\] hold for a non-uniruled surface over a field $k$. In the case of non-uniruled surfaces, it turns out that the condition in Lemma \[lemma strategy\] is true even when $\dim(U) >0$. The key observation in this case is that the problem can be reduced to $1$-dimensional schemes. We begin by recalling the definition of a ${{\mathbb P}}^1$-bundle from [@Kollar Chapter II, Definition 2.5] which will used in the proof.
\[definition P1-bundle\] Let $X$ be a scheme over $k$. We say that $f:Y \to X$ is a ${{\mathbb P}}^1$-bundle if $f$ is a smooth, proper morphism such that for every point $x \in X$, the fibre $f^{-1}(x)$ is a rational curve.
The following property of ${{\mathbb P}}^1$-bundles seems to be well-known to experts and is mentioned without proof in [@Kollar Chapter II, Definition 2.5]. We include a proof here for the sake of completeness.
\[lemma P1-bundles\] Let $\pi: E \to B$ be a smooth, projective morphism of varieties over $k$ such that for any point $b \in B$, the fibre $E_b$ is isomorphic to ${{\mathbb P}}^1_b$. Then $\pi$ is an étale-locally trivial fibre bundle.
First we consider the case that $\pi$ admits a section $\tau: B \to E$. In this case, $E$ is isomorphic to ${{\mathbb P}}({{\mathcal E}})$ where ${{\mathcal E}}$ is a rank $2$ vector bundle over $B$. For the proof of this statement, we refer to [@Hartshorne Chapter V, Lemma 2.1, Proposition 2.2]. This result is proved there under the assumption that $B$ is a smooth curve, but the proof applies in general.
For the general case, we use the fact that the smooth, surjective morphism $\pi$ has an étale-local section, i.e. there exists an étale cover $V \to B$ and a $B$-morphism $V \to E$. Then $E_V:= E \times_B V \to V$ is also a smooth, projective morphism with fibres isomorphic to ${{\mathbb P}}^1$. However, the $B$-morphism $E \to V$ induces a section $V \to E_V$. Thus, by the special case given above, $E_V \to V$ is a Zariski-locally trivial ${{\mathbb P}}^1$-fibre bundle. It follows that $\pi$ is an étale-locally trivial ${{\mathbb P}}^1$-fibre bundle over $B$.
\[proposition A1-invariance for curves\] Let $X$ be a reduced, separated $1$-dimensional proper scheme (possibly singular) over $k$. Then ${{\mathcal S}}(X) = {{\mathcal S}}^2(X)$.
Let $U$ be a smooth Henselian local scheme over $k$. We will show that ${{\mathcal S}}(X)(U) = {{\mathcal S}}^2(X)(U)$.
If $X$ is smooth, it is a disjoint union of irreducible, smooth, projective curves $\{C_i\}_{i\in I}$ where $I$ is a finite set. We need to verify that ${{\mathcal S}}(C_i)(U) = {{\mathcal S}}^2(C_i)(U)$ for every $i \in I$. If $C_i$ is not geometrically rational (i.e. rational over the algebraic closure $\-k$ of $k$), it is ${{\mathbb A}}^1$-rigid (see Definition \[definition A1 rigid\]) and thus ${{\mathcal S}}(C_i) = C_i$. So, in this case, ${{\mathcal S}}(C_i)(U) = {{\mathcal S}}^2(C_i)(U)$ as desired. If $C_i$ is geometrically rational, it is isomorphic to a smooth plane conic. Any morphism $f: U \to C_i$ factors as $$U \overset{\Gamma_f}{\to} (C_i)_U:= C_i \times_{{{\rm Spec \,}}(k)} U \to C_i$$ where $\Gamma_f$ is the graph of $f$. If the morphism $(C_i)_U \to U$ has no section, then clearly $${{\mathcal S}}(C_i)(U) = {{\mathcal S}}^2(C_i)(U) = \emptyset$$ which verifies our claim in this case. On the other hand, if there exists a section for the morphism $(C_i)_U \to U$, then $(C_i)_U \to U$ is a smooth, projective morphism such that the fibre over any point $u$ of $U$ is isomorphic to ${{\mathbb P}}^1_u$. (Recall that a conic is isomorphic to ${{\mathbb P}}^1$ if and only if it admits a rational point over the base field.) By Lemma \[lemma P1-bundles\], the morphism $(C_i)_U \to U$ is an étale-locally trivial ${{\mathbb P}}^1$-bundle. As $U$ is a Henselian local scheme, it is a trivial ${{\mathbb P}}^1$-fibre bundle. Thus, in this case $${{\mathcal S}}(C_i)(U) = {{\mathcal S}}^2(C_i)(U) = \ast$$ as desired. Thus we have proved the lemma in case $X$ is smooth.
Suppose now that $X$ is not smooth. For any smooth irreducible scheme $U$ over $k$, any ${{\mathbb A}}^1$-homotopy or ${{\mathbb A}}^1$-ghost homotopy of $U$ in $X$ must factor through the normalization of $X$ if its image is dense and the result follows from the smooth case. If the image of an ${{\mathbb A}}^1$-homotopy or ${{\mathbb A}}^1$-ghost homotopy of $U$ is not dense in $X$, then it is a single point and the result follows immediately.
\[theorem non-uniruled\] Let $X$ be a proper, non-uniruled surface over $k$. Then we have ${{\mathcal S}}(X) = {{\mathcal S}}^2(X)$.
Since $X$ is not uniruled, we know (see [@Kollar], Chapter IV, 1.3) that for any variety $Z$ over $k$, if we have a rational map ${{\mathbb P}}^1 \times Z \dashrightarrow X$, then either
- this rational map ${{\mathbb P}}^1 \times Z \dashrightarrow X$ is not dominant, or
- for every $z \in Z$, the induced map ${{\mathbb P}}^1_z \dashrightarrow X$ is constant.
It follows that if, for an essentially smooth scheme $U$ over $k$, we have an ${{\mathbb A}}^1$-homotopy $U \times {{\mathbb A}}^1 \to X$ , then it is either the constant homotopy or it factors through a $1$-dimensional, reduced, closed subscheme of $X$. We will make use of this observation in the following discussion.
Let $U$ be a smooth, Henselian, local scheme over $k$. We will show that ${{\mathcal S}}(X)(U) \to {{\mathcal S}}(X)(U \times {{\mathbb A}}^1)$ is a bijection. If ${{\rm dim}}(U) = 0$, this has already been proved in Theorem \[theorem field case\]. Thus we now assume that ${{\rm dim}}(U) \geq 1$.
We will use the notation of Definition \[definition ghost homotopy\] in the following arguments. So, let $${{\mathcal H}}:=\left(V \to {{\mathbb A}}^1_U, W \to V \times_{{{\mathbb A}}^1_U} V, h, h^W \right)$$ be an ${{\mathbb A}}^1$-ghost homotopy of $U$ in $X$ connecting morphisms $t_1,t_2: U \to X$. As discussed before ${{\mathcal H}}$ determines an ${{\mathbb A}}^1$-homotopy $\~h: {{\mathbb A}}^1_U \to {{\mathcal S}}(X)$. We will show that either $\~h$ lifts to an ${{\mathbb A}}^1$-homotopy of $U$ in $X$ or that the given ${{\mathbb A}}^1$-ghost homotopy factors through a $1$-dimensional, reduced, closed subscheme of $X$.
We write $V = \coprod_{i \in I} V_i$ for some indexing set $I$ where each $V_i$ is irreducible. We will denote the morphism $h|_{V_i}$ by $h_i$. Also, we write $$W = \coprod_{i,j \in I} \left(\coprod_{l \in K_{ij}} W^{ij}_l\right)$$ for indexing sets $K_{ij}$ depending on pairs $i,j \in I$, so that each $W^{ij}_l$ is irreducible and $\left(\coprod_{l \in K_{ij}} W^{ij}_l\right) \to V_i \times_{{{\mathbb A}}^1_U} V_j$ is a Nisnevich cover. For every pair $i,j \in I$ and $l \in K_{ij}$, we have the sequence of morphisms $$W^{ij}_l \to V_i \times_{{{\mathbb A}}^1_U} V_j \to V_i \stackrel{h_i}{\to} X.$$ In the following argument, we will use $h_i|_{W^{ij}_l}$ (resp. $h_i|_{V_i \times_{{{\mathbb A}}^1_U} V_j}$) to denote the composition of the morphisms in this sequence starting from $W^{ij}_l$ (resp. $V_i \times_{{{\mathbb A}}^1_U} V_j$).
As we observed, the fact that $X$ is non-uniruled implies that the restrictions of $h^W$ to $W^{ij}_l$ for various $i,j \in I$ and $l \in K_{ij}$ are either constant ${{\mathbb A}}^1$-chain homotopies or they factor through some $1$-dimensional, reduced closed subscheme of $X$. If the restrictions of $h^W$ to all the $W^{ij}_k$ are constant for all $i,j \in I$ and $l \in K_{ij}$, then we have $h_i|_{W^{ij}_l} = h_j|_{W^{ij}_l}$ for all choices of $i,j$ and $l$. Since the morphism $\left(\coprod_{l \in K_{ij}} W^{ij}_l\right) \to V_i \times_{{{\mathbb A}}^1_U} V_j$ is an epimorphism, this implies that $h_i|_{V_i \times_{{{\mathbb A}}^1_U} V_j} = h_j|_{V_i \times_{{{\mathbb A}}^1_U} V_j}$. But this means that the morphisms $h_i$ give rise to a morphism ${{\mathbb A}}^1_U \to X$ which lifts $\~h: {{\mathbb A}}^1_U \to {{\mathcal S}}(X)$. By the definition of ${{\mathcal S}}(X)$, this proves $t_1 = t_2$.
Thus we now assume that for some fixed $i,j \in I$ and some fixed $l \in K_{ij}$, the restriction of $h^W$ to $W^{ij}_l$ is a non-constant ${{\mathbb A}}^1$-chain homotopy. Hence, (2) above cannot hold. We claim that in this case all the morphisms $h_i$ for $i \in I$ as well as the ${{\mathbb A}}^1$-chain homotopy $h^W$ factor through some $1$-dimensional, reduced, closed subscheme of $D$ of $X$. Once this is proved, the result will follow from Proposition \[proposition A1-invariance for curves\]. Indeed, then we see that the $t_1,t_2: U \to X$ must factor through $D$ and, being connected by an ${{\mathbb A}}^1$-ghost homotopy within $D$, must also be connected by an ${{\mathbb A}}^1$-chain homotopy in $D$. It now remains to prove the existence of $D$.
The restriction of $h^W$ to $W^{ij}_l$ must factor through some $1$-dimensional, closed subscheme of $C^{ij}_l \subset X$, by (1) above. In particular, the morphisms $h_i|_{W^{ij}_{l}}$ and $h_j|_{W^{ij}_{l}}$ also factor through $C^{ij}_l \hookrightarrow X$. As $V_i$ and $V_j$ are assumed to be irreducible and since the morphisms $W^{ij}_l \to V_i$ and $W^{ij}_l \to V_j$ are dominant, both $h_i$ and $h_j$ also factor through $C^{ij}_l \hookrightarrow X$.
Now we consider any other index $m \in K_{ij}$. If the restriction of $h^W$ to $W^{ij}_m$ is a non-constant ${{\mathbb A}}^1$-chain homotopy, the argument in the previous paragraph shows that this ${{\mathbb A}}^1$-chain homotopy factors through some $1$-dimensional, reduced, closed subscheme $C^{ij}_{m}$. On the other hand, if this ${{\mathbb A}}^1$-chain homotopy is constant, it must factor through the variety $C^{ij}_l$ obtained above and thus we may define $C^{ij}_m:= C^{ij}_l$. Taking the union $C^{ij}:= \bigcup_{m \in K_{ij}} C^{ij}_m$, we see that $h_i$, $h_j$ and the restriction of $h^W$ to $\coprod_{m \in K_{ij}} W^{ij}_m$ factor through $C^{ij}$.
Now let $i' \in I$ be any index. We claim that $h_{i'}$ factors through some $1$-dimensional, reduced, closed subscheme $C^i$ of $X$. For $i' = i$ or $j$, we define $C^{i'}:= C^{ij}$. For $i' \neq i,j$, we prove the claim by contradiction. Suppose that $h_{i'}$ does not factor through any $1$-dimensional, reduced, closed subscheme of $X$. First note that the scheme $V_i \times_{{{\mathbb A}}^1_U} V_{i^{\prime}}$ is non-empty since ${{\mathbb A}}^1_U$ is irreducible and $V_i, V_{i^{\prime}}$ are étale over ${{\mathbb A}}^1_U$. Thus the scheme $\coprod_{m \in K_{ii'}} W^{ii'}_m$ is non-empty. The restriction of $h^W$ to $\coprod_{m \in K_{ii'}} W^{ii'}_m$ cannot be the constant ${{\mathbb A}}^1$-chain homotopy since $h_i$ factors through a $1$-dimensional subscheme of $X$ while $h_{i'}$ does not. But then, by the argument in the previous paragraph, the fact that this ${{\mathbb A}}^1$-chain homotopy is nonconstant implies that $h_{i'}$ factors through some $1$-dimensional, reduced, closed subscheme of $C^{ii'}$ of $X$. This gives us a contradiction. Thus for every index $i' \in I$, the morphism $h_{i'}$ factors through some $1$-dimensional, reduced, closed subscheme $C^{i'}$ of $X$.
Finally, we wish to prove that for any two indices $i', j' \in I$, the restriction of $h^W$ to $\coprod_{m \in K_{i'j'}} W^{i'j'}_m$ factors through some $1$-dimensional, reduced, closed subscheme of $X$. If this restriction is a non-constant ${{\mathbb A}}^1$-homotopy, the above argument gives us a $1$-dimensional, reduced, closed subscheme $C^{i'j'}$ of $X$ through which this homotopy must factor. If it is the constant homotopy, we note that both $h_{i'}$ and $h_{j'}$ factor through $C^{i'} \cup C^{j'}$. Thus, the restriction of $h^W$ to $\coprod_{m \in K_{i'j'}} W^{i'j'}_m$ must also factor through this subscheme.
Now we may define $D$ to be the scheme $(\bigcup_{i,j \in I} C^{ij}) \cup (\bigcup_{i \in I} C^i)$. This completes the proof of the theorem.
Conjectures \[a1invariance\] and \[conjecture Asok-Morel\] hold for a proper, non-uniruled surface over a field.
One can prove that ${{\mathcal S}}(X) \simeq {{\mathcal S}}^2(X)$, when $X$ is a rational surface (or in fact, any rational variety) or a ${{\mathbb P}}^1$-bundle over a curve. However, for a ruled surface $X$ (over an algebraically closed base field of characteristic 0), one can only prove that ${{\mathcal S}}(X)(U) \simeq {{\mathcal S}}^2(X)(U)$, where $U$ is the Spec of a Henselian discrete valuation ring. In fact, the conclusion of Theorem \[theorem non-uniruled\] is not true for ruled surfaces (even over an algebraically closed base field of characteristic $0$). Indeed, if $C$ is a curve of genus $>0$ and $X$ is the blowup of ${{\mathbb P}}^1 \times C$ at a single closed point, then it is possible to show that ${{\mathcal S}}(X)(U) \neq {{\mathcal S}}^2(X)(U)$, for a two-dimensional smooth Henselian local scheme $U$. This shows that ${{\mathcal S}}(X) \neq {{\mathcal S}}^2(X)$ in general, even for smooth projective surfaces. Surprisingly, for such ruled surfaces $X$, one can prove that ${{\mathcal S}}^2(X) \simeq {{\mathcal S}}^3(X)$. For details, see [@Sawant-thesis]. However, our proof for these results involves a lot more effort and the techniques used there substantially differ from the ones used in this paper. We do not include it here since we can demonstrate this phenomenon in a much easier manner in higher dimensions (see Section \[section counterexamples\]). The details about ${{\mathbb A}}^1$-connected components of ruled surfaces will appear in a forthcoming paper [@BS].
Although the method of showing ${{\mathcal S}}(X) \simeq {{\mathcal S}}^2(X)$ is insufficient to address the question of the ${{\mathbb A}}^1$-invariance of $\pi_0^{{{\mathbb A}}^1}$ for an arbitrary smooth projective variety $X$, it is interesting to examine the question of whether, for any smooth projective variety $X$, the sequence of sheaves $({{\mathcal S}}^n(X))_{n \geq 1}$ stabilizes at some finite value of $n$. If it does, one may ask whether it stabilizes to $\pi_0^{{{\mathbb A}}^1}(X)$, since that is what is predicted by Morel’s conjecture (see Corollary \[a1-invariance-candidate\]).
Examples of schemes for which the conjectures of Asok-Morel fail to hold {#section counterexamples}
========================================================================
Examples of schemes whose Sing\* is not A1-local
------------------------------------------------
The aim of this subsection is to construct an example of a smooth, projective variety $X$ over ${{\mathbb C}}$ such that:
- ${{\mathcal S}}(X) \neq {{\mathcal S}}^2(X)$,
- ${{\mathcal S}}(X) \to \pi_0^{{{\mathbb A}}^1}(X)$ is not a monomorphism, and
- $Sing_*(X)$ is not ${{\mathbb A}}^1$-local.
We will use a well-known characterization of Nisnevich sheaves, which we will recall here for the sake of convenience.
For any scheme $U$, an *elementary Nisnevich cover* of $U$ consists of two morphisms $p_1: V_1 \to U$ and $p_2: V_2 \to U$ such that:
- $p_1$ is an open immersion.
- $p_2$ is an étale morphism and its restriction to $p_2^{-1}(U \backslash p_1(V_1))$ is an isomorphism onto $U \backslash p_1(V_1)$.
Then a presheaf of sets ${{\mathcal F}}$ on $Sm/k$ is a sheaf in Nisnevich topology if and only if the morphism $${{\mathcal F}}(U) \to {{\mathcal F}}(V_1) \times_{{{\mathcal F}}(V_1 \times_U V_2)} {{\mathcal F}}(V_2)$$ is an isomorphism, for all elementary Nisnevich covers $\{ V_1, V_2\}$ of $U$. If ${{\mathcal F}}$ is a Nisnevich sheaf, the fact that this morphism is an isomorphism is an immediate consequence of the definition of a sheaf. The converse (see [@Morel-Voevodsky 3, Proposition 1.4, p.96] for a proof) is useful since it simplifies the criterion for checking whether sections $s_i \in {{\mathcal F}}(V_i)$ determine a section $s \in {{\mathcal F}}(U)$. Indeed, suppose $pr_1$ and $pr_2$ are the two projections $V \times_U V \to V$. Then we do not need to check that the two elements $pr_1^*(s_1)$ and $pr_2^*(s_2)$ of ${{\mathcal F}}(V \times_U V)$ are identical. We merely need to check that the images of $s_1$ and $s_2$ under the maps ${{\mathcal F}}(V_1) \to {{\mathcal F}}(V_1 \times_U V_2)$ and ${{\mathcal F}}(V_2) \to {{\mathcal F}}(V_1 \times_U V_2)$ respectively are identical.
Now suppose ${{\mathcal F}}$ is a sheaf of sets. Applying the above criterion to the sheaf ${{\mathcal S}}({{\mathcal F}})$, we see that when we work with elementary Nisnevich covers, we can construct morphisms ${{\mathbb A}}^1_U \to {{\mathcal S}}({{\mathcal F}})$ with only part of the data that is required for an ${{\mathbb A}}^1$-ghost homotopy of $U$ in ${{\mathcal F}}$.
\[lemma Sing\_\* not A1-local\] Let ${{\mathcal F}}$ be a sheaf of sets over $Sm/k$. Let $U$ be a smooth scheme over $k$ and let $f,g: U \to {{\mathcal F}}$ be two morphisms. Suppose that we are given data of the form $$(\{p_i: V_i \to {{\mathbb A}}^1_U\}_{i \in \{1,2\}}, \{\sigma_0,\sigma_1\}, \{h_1, h_2\}, h^W)$$ where:
- The two morphisms $\{p_i: V_i \to {{\mathbb A}}^1_U\}_{i=1,2}$ constitute an elementary Nisnevich cover.
- For $i \in \{0,1\}$, $\sigma_i$ is a morphism $U \to V_1 \coprod V_2$ such that $(p_1 \coprod p_2) \circ \sigma_i: U \to U \times {{\mathbb A}}^1$ is the closed embedding $U \times \{i\} \hookrightarrow U \times {{\mathbb A}}^1$,
- For $i \in \{1,2\}$, $h_i$ denotes a morphism $V_i \to {{\mathcal F}}$ such that $(h_1 \coprod h_2) \circ \sigma_0 = f$ and $(h_1 \coprod h_2) \circ \sigma_1 = g$.
- Let $W := V_1 \times_{{{\mathbb A}}^1_U}V_2$ and let $pr_i: W \to V_i$ denote the projection morphisms. Then $h^W = (h_1, \ldots, h_n)$ is a ${{\mathbb A}}^1$-chain homotopy connecting the two morphisms $h_i \circ pr_i: W \to {{\mathcal F}}$.
Then, $f$ and $g$ map to the same element under the map ${{\mathcal F}}(U) \to \pi_0^{{{\mathbb A}}^1}({{\mathcal F}})(U)$. Moreover, $Sing_*({{\mathcal F}})$ is not ${{\mathbb A}}^1$-local.
Choose a simplicially fibrant replacement $Sing_*({{\mathcal F}}) \to {{\mathcal X}}$. We denote by $H_i$ the composition $V_i \overset{h_i}{\to} {{\mathcal F}}\to Sing_*({{\mathcal F}}) \to {{\mathcal X}}$, for $i = 1,2$. According to the discussion preceding the statement of this lemma, the given data induces a morphism $\psi: {{\mathbb A}}^1_U \to {{\mathcal S}}({{\mathcal F}}) = \pi_0^s({{\mathcal X}})$.
The ${{\mathbb A}}^1$-homotopy $h^{W}$ connects $h_i|_W$ for $i = 1,2$. Thus, by the definition of $Sing_*({{\mathcal F}})$, it gives rise to a simplicial homotopy $h^s: \Delta^1 \to {{\mathcal X}}(W)$ connecting $H_i$ for $i=1,2$. Thus we have the following diagram (where $*$ denotes the point sheaf) $$\xymatrix{
{*} \ar[r] \ar[d] & {{\mathcal X}}(V_2) \ar[d] \\
\Delta^1 \ar[r] \ar@{-->}[ur]^{\exists} & {{\mathcal X}}(W)
}$$ where the upper horizontal arrow maps $*$ to $H_2$. Since $W = V_2 \backslash (U \times \pi^{-1}({0})) \to V_2$ is a cofibration, the morphism ${{\mathcal X}}(V_2) \to {{\mathcal X}}(W)$ of simplicial sets is a fibration. Thus the dotted diagonal lift exists in the above diagram.
Thus we see that we can find a morphism $H^{\prime}: V_2 \to {{\mathcal X}}$ such that:
- $h^{\prime}$ and $h_2$ induce the same morphism $V_2 \to \pi_0^s({{\mathcal X}})$ (since they are simplicially homotopic in ${{\mathcal X}}(V_2)$), and
- $H^{\prime}|_W = H_1|_{W}$.
Since ${{\mathcal X}}$ is a sheaf, the following diagram is cartesian: $$\xymatrix{
{{\mathcal X}}({{\mathbb A}}^1_U) \ar[r] \ar[d] & {{\mathcal X}}(V_2) \ar[d] \\
{{\mathcal X}}(V_1) \ar[r] & {{\mathcal X}}(W)\text{.}
}$$ Thus, $H_1$ and $H^{\prime}$ can be glued to give an element of ${{\mathcal X}}({{\mathbb A}}^1_U)$ which we can think of as a morphism $\~\psi: {{\mathbb A}}^1_U \to {{\mathcal X}}$. Clearly, this lifts $\psi$, as can be checked by restriction to the Nisnevich cover $\{V_1, V_2\}$.
It is easily seen that the morphisms that $\~\psi$ connects are simplicially chain homotopic to the morphisms $\~f$ and $\~g$. Thus we see that the morphisms $\~f$ and $\~g$ are simplicially chain homotopic. From the explicit construction of the ${{\mathbb A}}^1$-fibrant replacement functor $L_{{{\mathbb A}}^1}$ given in [@Morel-Voevodsky], we see that the two morphisms $U \to L_{{{\mathbb A}}^1}({{\mathcal F}})$ obtained by composing $f$ and $g$ with the canonical morphism ${{\mathcal F}}\to L_{{{\mathbb A}}^1}({{\mathcal F}})$ are simplicially homotopic. Therefore, $f$ and $g$ map to the same element of $\pi_0^{{{\mathbb A}}^1}({{\mathcal F}})(U)$. Consequently, ${{\mathcal X}}$ (and hence, $Sing_*({{\mathcal F}})$) is not ${{\mathbb A}}^1$-local.
We will refer to the data of the form $$(\{p_i: V_i \to {{\mathbb A}}^1_U\}_{i \in \{1,2\}}, \{\sigma_0,\sigma_1\}, \{h_1, h_2\}, h^W)$$ as in Lemma \[lemma Sing\_\* not A1-local\] by an *elementary ${{\mathbb A}}^1$-ghost homotopy*.
We now proceed to the construction of the schemes satisfying properties (i)-(iii) stated at the beginning of the section.
\[notation example 1\] To clarify the idea behind this construction, we first construct a non-proper, singular variety satisfying (i)-(iii). Later (in Construction \[notation example 2\]), we will modify this example suitably to create an example of a smooth, projective variety over ${{\mathbb C}}$ which will also satisfy (i)-(iii).
- Let $\lambda_i \in {{\mathbb C}}\backslash \{0\}$ for $i = 1,2,3$. Let $E \subset {{\mathbb A}}^2_{{{\mathbb C}}}$ be the affine curve cut out by the polynomial $y^2 - \prod_i (x-\lambda_i)$. Let $\pi: E \to {{\mathbb A}}^1_{{{\mathbb C}}}$ denote the projection onto the $x$-axis.
- Let $S_1$ and $S_2$ be the closed subschemes of ${{\mathbb A}}^3_{{{\mathbb C}}}$ cut out by the polynomials $y^2 = t^2 \prod_i (x - \lambda_i)$ and $y^2 = \prod_i (x-\lambda_i)$ respectively. Let $f: S_2 \to S_1$ be the morphism corresponding to the homomorphism of coordinate rings given by $x \mapsto x$, $y \mapsto yt$ and $t \mapsto t$.
- Let $\alpha_0$ be a square-root of $-\lambda_1\lambda_2\lambda_3$ so that the point $(0,\alpha_0)$ is mapped to the point $x=0$ under the projection $\pi$. Let $p_0$, $p_1$ and $q$ denote the points $(0,0,0)$, $(0,\alpha_0,1)$ and $(1,0,0)$ of $S_1$.
- Consider the morphism ${{\mathbb A}}^1_{{{\mathbb C}}} \to S_2$ given by $s \mapsto (0,\alpha_0,s)$ for $s \in {{\mathbb A}}^1_{{{\mathbb C}}}$. Composing this morphism with $f$, we obtain a morphism ${{\mathbb A}}^1_{{{\mathbb C}}} \to S_1$, the image is a closed subscheme of $S_1$ which we denote by $D$.
Now let $S \subset {{\mathbb A}}^3_{{{\mathbb C}}}$ be the scheme $S_1 \backslash \{p_0\}$. We claim that $S$ satisfies condition (i) above. We claim that any morphism $h:{{\mathbb A}}^1_{{{\mathbb C}}} \to S$ such that $p_1$ lies in the image of $h$ is constant. To see this, first observe that this is obvious if the image of $h$ is contained completely inside $D \backslash \{p_0\}$ (since any morphism from ${{\mathbb A}}^1_{{{\mathbb C}}}$ into ${{\mathbb A}}^1_{{{\mathbb C}}} \backslash \{0\}$ is a constant morphism). If this is not the case, note that the morphism $f: S_2 = E \times {{\mathbb A}}^1_{{{\mathbb C}}} \to S_1$ is an isomorphism outside the locus of $t=0$. Hence, $h$ induces a rational map ${{\mathbb A}}^1_{{{\mathbb C}}} \dashrightarrow E$, which can be completed to a morphism ${{\mathbb P}}^1_{{{\mathbb C}}} \to \-E$, where $\-E$ denotes a projective compactification of $E$. This has to be a constant morphism by the Hurwitz formula, since the genus of $\-E$ is greater than that of ${{\mathbb P}}^1_{{{\mathbb C}}}$. Thus, there is no non-constant morphism ${{\mathbb A}}^1_{{{\mathbb C}}} \to S$ having $p_1$ in its image. Thus we see that $p_1$ and $q$ are not connected by an ${{\mathbb A}}^1$- chain homotopy. However, we claim that they map to the same element in ${{\mathcal S}}^2(S)({{\mathbb C}})$.
To construct an elementary ${{\mathbb A}}^1$-ghost homotopy, we first construct a Nisnevich cover of ${{\mathbb A}}^1_{{{\mathbb C}}}$. We observe that $\pi$ is étale except at the points where $x = \lambda_i$, $i = 1,2,3$. Now we define $\-V_1 = {{\mathbb A}}^1_{{{\mathbb C}}} \backslash \{0\}$ and $\-V_2 = \pi^{-1}({{\mathbb A}}^1_{{{\mathbb C}}} \backslash \{\lambda_1, \lambda_2, \lambda_3\}) \subset E$. The morphism $\-V_1 \to {{\mathbb A}}^1_{{{\mathbb C}}}$ is the inclusion and the morphism $\-V_2 \to {{\mathbb A}}^1_{{{\mathbb C}}}$ is simply $\pi|_{\-V_2}$. As $\lambda_i \neq 0$ for all $i$, we see that this is an elementary Nisnevich cover of ${{\mathbb A}}^1_{{{\mathbb C}}}$. We define $\-W:= \-V_2 \backslash \pi^{-1}(0)$ and observe that $\-V_1 \times_{{{\mathbb A}}^1_{{{\mathbb C}}}} \-V_2 \cong \-W$. In order to give an elementary ${{\mathbb A}}^1$-ghost homotopy in $S$ using this Nisnevich cover, we will now define morphisms $\-h_i: \-V_i \to S$ and a morphism $h^{\-W}: \-W \times {{\mathbb A}}^1 \to S$, which is a homotopy connecting $(\pi \circ \-h_1)|_{\-W}$ and $\pi \circ \-h_2|_{\-W}$.
We define $\-h_1: \-V_1 \to S$ by $s \mapsto (s,0,0)$ for $s \in {{\mathbb A}}^1_{{{\mathbb C}}}$ and $\-h_2: \-V_2 \to S$ by $(\alpha,\beta) \mapsto (\alpha,\beta,1)$ for $(\alpha,\beta) \in \-V_2$. To define $h^{\-W}$, we first identify $S_2$ with $E \times {{\mathbb A}}^1$ and define $h^{\-W}$ to be $f|_{\-W \times {{\mathbb A}}^1}$ (recall that $\-W \subset \-V_2 \subset E$). It is clear that the the ${{\mathbb C}}$-valued points $p_1$ and $q$ are mapped to the same element in ${{\mathcal S}}^2(S)({{\mathbb C}})$. This shows that the surface $S$ satisfies the condition (i) listed above.
We have thus obtained the data $$\left( \-V_1 \coprod \-V_2 \to {{\mathbb A}}^1_U, \-W \to \-V_1 \times_{{{\mathbb A}}^1_U} \-V_2, \-h_1 \coprod \-h_2, \-h^{\-W}\right),$$ which satisfies the hypotheses of Lemma \[lemma Sing\_\* not A1-local\], which proves that $S$ satisfies (ii) and (iii).
The following lemma is the key tool in creating a smooth example that satisfies (i) - (iii) stated at the beginning of the section:
\[lemma rigid embedding\] Let $k$ be a field and let $S$ be an affine scheme over $k$. Then there exists a closed embedding of $S$ into a smooth scheme $T$ over $k$ such that for any field extension $L/k$, if $H: {{\mathbb A}}^1_L \dashrightarrow T$ is a rational map such that the image of $H$ meets $S$, then $H$ factors through $S \hookrightarrow T$.
Choose an embedding of $S$ into ${{\mathbb A}}^n_k$ for some $n$ and suppose that as a subvariety of ${{\mathbb A}}^n_k$, $S$ is given by the ideal $\<f_1, \ldots, f_r\> \subset k[X_1, \ldots, X_n]$. Then the polynomials $f_i$ define a morphism $f: {{\mathbb A}}^n \to {{\mathbb A}}^r$ such that $S$ is the fibre over the origin. Now let $g: C \to {{\mathbb A}}^1_k$ be an étale morphism such that $C$ is a curve of genus $\geq 1$ and the preimage of the origin consists of a single $k$-valued point $c_0$ of $C$. Then this gives us an étale morphism $g^r: C^r \to {{\mathbb A}}^r$ such that the preimage of the origin is the point $(c_0, \ldots, c_0)$. Let $T = {{\mathbb A}}^n_k \times_{f,{{\mathbb A}}^r_k, g^r} C^r$. It is clear that $T \to {{\mathbb A}}^n_k$ induces an isomorphism over $S$. We claim that $T$ is the desired scheme.
Let $L/k$ be a field extension and let $H: {{\mathbb A}}^1_L \dashrightarrow T$ be a rational such that the image of $H$ meets the preimage of $S$ in $T$. Since $C$ has genus $\geq 1$, the Hurwitz formula implies that any rational map ${{\mathbb A}}^1_L \dashrightarrow C$ is constant. Thus, the composition of $H$ with the projection $T \to C^r$ factors through $(c_0, \ldots, c_0)$. It follows from this that $H$ factors through $S$, which is the preimage of $(c_0, \ldots, c_0)$ in $T$.
\[notation example 2\] We now construct a smooth and projective variety $X$ over ${{\mathbb C}}$ satisfying the conditions (i)-(iii) listed at the beginning of this subsection:
- Using Lemma \[lemma rigid embedding\], we construct a scheme $T$ such that $S_1 \subset T$ as a closed subscheme and such that for any field extension $L/{{\mathbb C}}$, if $H: {{\mathbb A}}^1_L \to T$ is a morphism such that the image of $H$ meets $S_1$, then $H$ factors through $S_1$.
- Let $T^c \supset T$ be a smooth, projective compactification of $T$.
- Let $C$ be a smooth, projective curve over ${{\mathbb C}}$ of genus $>0$. Let $c_0$ be a ${{\mathbb C}}$-valued point of $C$. Let $R = {{\mathcal O}}^{h}_{C,c_0}$ and let $U = {{\rm Spec \,}}(R)$. Let $u$ be the closed point of $U$. Let $\gamma: U \to C$ be the obvious morphism.
- Let $p_2$ by any closed point of $D \backslash \{p_0, p_1\}$ (see Construction \[notation example 1\], (4)). Let $X$ be the blowup of $C \times T^c$ at the points $(c_0,p_0)$ (see Construction \[notation example 1\], (3)) and $(c_0,p_2)$.
- For any ${{\mathbb C}}$ valued point $p: {{\rm Spec \,}}({{\mathbb C}}) \to T^c$ of $T^c$, let $\theta_p$ be the induced morphism $\gamma \times p: U = U \times {{\rm Spec \,}}({{\mathbb C}}) \to C \times T^c$. Note that $\theta_p$ is a lift of $\gamma$ to $C \times T^c$ with respect to the projection $C \times T^c \to C$.
- Let $\xi_{p_1}$ and $\xi_q$ denote the lifts of $\theta_{p_1}$ and $\theta_{q}$ to $X$ (which exist and are unique).
We claim that $\xi_{p_1}$ and $\xi_q$ are connected by an elementary ${{\mathbb A}}^1$-ghost homotopy. To prove this, we construct an elementary ${{\mathbb A}}^1$-ghost homotopy connecting $\theta_{p_1}$ and $\theta_{q}$ which lifts to $X$. This is done by simply taking the product of the morphism $\gamma: U \to C$ with the elementary ${{\mathbb A}}^1$-ghost homotopy of ${{\rm Spec \,}}({{\mathbb C}})$ in $S \subset T^c$ which we constructed above. In other words, we define $V_i = U \times \-V_i$ and $W = U \times \-W$. We define $h_i := \gamma \times \-h_i: V_i = U \times \-V_i \to C \times T^c$ and define $h^W:= \gamma \times h^{\-W}: W \times {{\mathbb A}}^1 = U \times \-W \to C \times T^c$. (Here we have abused notation by viewing $\-h_i$ and $h^{\-W}$ as morphisms into $T^c$ rather than as morphisms into $S \subset T^c.$) We observe that $W \cong V_1 \times_{{{\mathbb A}}^1_U} V_2$ and we have constructed an elementary ${{\mathbb A}}^1$-ghost homotopy ${{\mathcal H}}$ of $U$ in $C \times T^c$. Since this elementary ${{\mathbb A}}^1$-ghost homotopy factors through the complement of the the points $(c_0,p_0)$ and $(c_0,p_2)$, it lifts to an elementary ${{\mathbb A}}^1$-ghost homotopy $\~{{{\mathcal H}}}$ of $U$ in $X$ and connects $\xi_{p_1}$ and $\xi_q$.
We now claim that $\xi_{p_1}$ and $\xi_q$ are not ${{\mathbb A}}^1$-chain homotopic. For this it will suffices to show that if $\xi_{p_1}$ is ${{\mathbb A}}^1$-homotopic to any $\xi: U \to X$, then $\xi(u) = \xi_{p_1}(u)$. Projecting this ${{\mathbb A}}^1$-homotopy down to $C \times T^c$, we see that it suffices to prove that if $h$ is an ${{\mathbb A}}^1$-homotopy of $U$ in $C \times T^c$ which lifts to $X$ and if $\sigma_0 \circ h = \theta_{p_1}$, then $h$ maps the closed fibre ${{\mathbb A}}^1_{{{\mathbb C}}} \subset {{\mathbb A}}^1_U$ to $\theta_{p_1}(u)$.
Let $\-D$ be the closure of $D$ (see Construction \[notation example 1\], (4)) in $C \times T^c$. We see by Construction \[notation example 2\], (4) that $\-D$ is the only rational curve through the point $\theta_{p_1}(u)$. We claim that any ${{\mathbb A}}^1$-homotopy $h$ of $U$ in $C \times T^c$ such that $h \circ \sigma_0 = \theta_{p_1}$ must map ${{\mathbb A}}^1_{{{\mathbb C}}} \subset {{\mathbb A}}^1_U$ into $D$. Indeed, this follows immediately from the fact that the composition of this ${{\mathbb A}}^1$-homotopy with the projection $pr_1: C \times T^c \to C$ must be the constant homotopy (since $C$ is ${{\mathbb A}}^1$-rigid). In fact, it is easy to see that $pr_1 \circ h$ must factor through $\gamma$.
Now we claim that the image of $h|_{{{\mathbb A}}^1_{{{\mathbb C}}}}$ does not contain the point $p_0$. Indeed, if it were to contain this point, since $h$ can be lifted to $X$, the preimage of $p_0$ under $h$ would be a non-empty, codimension $1$ subscheme of ${{\mathbb A}}^1_U$ which is contained in ${{\mathbb A}}^1_{{{\mathbb C}}} \subset {{\mathbb A}}^1_U$. Thus it would be equal to ${{\mathbb A}}^1_{{{\mathbb C}}}$. Since $h \circ \sigma (u) = p_1 \neq p_0$, this is impossible. Thus, we see that the image of $h|_{{{\mathbb A}}^1_{{{\mathbb C}}}}$ does not contain $p_0$. By the same argument it does not contain $p_2$.
However, a morphism from ${{\mathbb A}}^1_{{{\mathbb C}}}$ into a rational curve which avoids at least two points on the rational curve must be a constant (since ${{\mathbb G}}_m$ is ${{\mathbb A}}^1$-rigid). Thus $h|_{{{\mathbb A}}^1_{{{\mathbb C}}}}$ is a constant. This completes our proof of the claim that $\xi_{q}$ and $\xi_{p_1}$ are not ${{\mathbb A}}^1$-chain homotopic. Thus, we have now proved that $X$ satisfies the property (i) listed above. Lemma \[lemma Sing\_\* not A1-local\] implies that it also satisfies property (ii). If ${{\rm Sing}}_*(X)$ were ${{\mathbb A}}^1$-local, the morphism ${{\mathcal S}}(X):= \pi_0^s(Sing_*(X)) \to \pi_0^{{{\mathbb A}}^1}(X)$ would be an isomorphism. Thus $X$ also satisfies property (iii).
Note that this method can be used to prove that $Sing_*(X)$ is not ${{\mathbb A}}^1$-local for any scheme $X$ for which the following holds: there exists a smooth Henselian local scheme $U$ and two morphisms $f, g:U \to X$, which are not ${{\mathbb A}}^1$-chain homotopic and are such that there exists an ${{\mathbb A}}^1$-ghost homotopy connecting $f$ to $g$ that is defined on an elementary Nisnevich cover of ${{\mathbb A}}^1_U$. Such a scheme $X$ is obviously a counterexample to ${{\mathcal S}}(X) \simeq {{\mathcal S}}^2(X)$ and hence, also a counterexample to Conjecture \[conjecture Asok-Morel\] of Asok and Morel.
Example showing that pi0A1 is not a birational invariant
--------------------------------------------------------
We end this paper by noting a counterexample to the birationality of $\pi_0^{{{\mathbb A}}^1}$ of smooth proper schemes over a field $k$.
\[definition A1 rigid\] A scheme $X \in {\mathcal Sm}_k$ is said to be *${{\mathbb A}}^1$-rigid*, if for every $U \in {\mathcal Sm}_k$, the map $$X(U) \longrightarrow X(U \times {{\mathbb A}}^1)$$ induced by the projection map $U \times {{\mathbb A}}^1 \to U$ is a bijection.
Since schemes are simplicially fibrant objects, by [@Morel-Voevodsky 2, Lemma 3.19], it follows that ${{\mathbb A}}^1$-rigid schemes are ${{\mathbb A}}^1$-fibrant objects. Thus, if $X \in Sm/k$ is ${{\mathbb A}}^1$-rigid, then for any $U \in Sm/k$ the canonical map $X(U) \to {{\rm Hom}}_{\mathcal H(k)}(U,X)$ is a bijection. Consequently, the canonical map $X \to \pi_0^{{{\mathbb A}}^1}(X)$ is an isomorphism of Nisnevich sheaves of sets.
\[example pi\_0 not birational\] Let $X$ be an abelian variety of dimension at least $2$. Since $X$ is an ${{\mathbb A}}^1$-rigid scheme, $\pi_0^{A^1}(X) \simeq X$, as noted above. Now, blow up $X$ at a point to get another scheme $Y$. Then the blow-up map $Y \to X$ is a birational proper map which induces a map $\pi_0^{{{\mathbb A}}^1}(Y) \to \pi_0^{{{\mathbb A}}^1}(X)$. If $\pi_0^{{{\mathbb A}}^1}$ is a birational invariant, then $\pi_0^{{{\mathbb A}}^1}(Y) \to \pi_0^{{{\mathbb A}}^1}(X)$ must be an isomorphism. However, $Y\to \pi_0^{{{\mathbb A}}^1}(Y)$ is an epimorphism, implying that the map $Y\to X$ is an epimorphism (in Nisnevich topology). But a blow-up cannot be an epimorphism in Nisnevich topology, unless it is an isomorphism. Indeed, if $Y \to X$ is an epimorphism, it would admit a Nisnevich local section, which in turn, would imply that it is an isomorphism, $Y \to X$ being a birational proper morphism. This is a contradiction. Consequently, $\pi_0^{{{\mathbb A}}^1}(X)$ cannot be a birational invariant of smooth proper schemes.
[AAAA]{}
A. Asok, F. Morel, Smooth varieties up to ${{\mathbb A}}^1$-homotopy and algebraic $h$-cobordisms, Adv. Math. 227(2011), no. 5, 1990 – 2058.
C. Balwe, A. Sawant, *${{\mathbb A}}^1$-connected components of ruled surfaces*, in preparation.
U. Choudhury, Connectivity of motivic $H$-spaces, Algebr. Geom. Topol. 14 (2014), no. 1, 37 – 55.
R. Hartshorne: *Algebraic geometry*, Graduate Texts in Mathematics, No. 52. Springer-Verlag, New York-Heidelberg, 1977.
J. F. Jardine, Motivic symmetric spectra, [Doc. Math.]{} [5]{} (2000), 445 – 553.
J. Kollár, Rational curves on algebraic varieties, Ergebnisse der Mathematik und ihrer Grenzgebiete [32]{}, Springer Verlag, Berlin, 1996.
F. Morel, ${{\mathbb A}}^1$-algebraic topology over a field, LNM [2052]{}, Springer, Heidelberg, 2012.
F. Morel, V. Voevodsky, ${{\mathbb A}}^1$-homotopy theory of schemes, [Publ. Math. I.H.E.S.]{} [90]{}(1999), 45 – 143 (2001).
A. Sawant, ${{\mathbb A}}^1$-connected components of schemes, Ph.D. Thesis, Tata Institute of Fundamental Research, 2014.
M. Wendt, On the ${{\mathbb A}}^1$-fundamental groups of smooth toric varieties, Adv. Math. 223(2010), 352 – 378.
[^1]: Anand Sawant was supported by the Council of Scientific and Industrial Research, India under the Shyama Prasad Mukherjee Fellowship SPM-07/858(0096)/2011-EMR-I
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We are concerned with the variation of the supercritical nearest neighbours contact process such that first infection occurs at a lower rate; it is known that the process survives with positive probability. Regarding the rightmost infected of the process started from one site infected and conditioned to survive, we specify a sequence of space-time points at which its behaviour regenerates and thus obtain the corresponding strong law and central limit theorem. We also extend complete convergence in this case.'
author:
- 'Achilleas Tzioufas[^1]'
title: |
On the growth of one-dimensional\
reverse immunization contact processes
---
Introduction and main results {#S0}
=============================
We begin by defining a class of processes that includes the processes we are especially interested in. The *nearest neighbours three state contact process* with parameters $(\lambda,\mu)$ is a continuous time Markov process $\zeta_{t}$ with state space $\{-1,0,1\}^{{\mathbb{Z}}}$, elements of which are called configurations. The evolution of $\zeta_{t}$ is described locally as follows. Think of configurations as functions from ${\mathbb{Z}}$ to $\{-1,0,1\}$, transitions at each site $x$, $\zeta_{t}(x)$, occur according to the rules: $$\label{rates}
\begin{array}{cl}
-1 \rightarrow 1 & \mbox{ at rate } \lambda \hspace{0.5mm} |\{ y = x-1,x+1 :\zeta_{t}(y)= 1\}|, \\
\mbox{ } 0 \rightarrow 1 & \mbox{ at rate } \mu \hspace{0.5mm} |\{ y = x-1,x+1 :\zeta_{t}(y)= 1\}|, \\
\mbox{ } 1 \rightarrow 0 & \mbox{ at rate } 1,
\end{array}$$ for all times $t \geq0$, where $|B|$ denotes the cardinal of $B\subset {\mathbb{Z}}$. Typically, the process started from configuration $\eta$ is denoted as $\zeta_{t}^{\eta}$. For general information about interacting particle systems, such as the fact that the above rates specify a well-defined process, we refer to Liggett [@L]. We note that the cases $\lambda = \mu$ and $\mu=0$ correspond to the extensively studied processes known as the contact process and as the forest fire model respectively, see $\mbox{e.g.}$ $\cite{L99}$, $\cite{D88}$. Furthermore in the literature various survival aspects of the three state contact process on the $d$-dimensional lattice were studied by Durrett and Schinazi [@DS] and by Stacey [@S], the latter also includes results for the process on homogeneous trees.
The process is thought of according to the following epidemiological interpretation. Given a configuration $\zeta$, each site $x$ is regarded as infected if $\zeta(x) = 1$, as susceptible and never infected if $\zeta(x) = -1$ and, as susceptible and previously infected if $\zeta(x) = 0$. The *standard initial configuration* is such that the origin is infected while all other sites are susceptible and never infected. We will use $\zeta_{t}^{O}$ to denote the nearest neighbours three state contact process process started from the standard initial configuration. We say that the three state contact process *survives* if ${\mathbf{P}}(\zeta_{t}^{O} \mbox{ survives})>0$, where the event $\{\forall \hspace{0.5mm}t\geq0, \exists \hspace{0.5mm} x: \zeta_{t}(x) =1\}$ is abbreviated as $\{\zeta_{t} \mbox{ survives}\}$.
For $\zeta_{t}^{O}$ we have that transitions $-1 \rightarrow 1$, $ 0 \rightarrow 1$, and $1 \rightarrow 0$ correspond respectively to initial infections, subsequent infections and recoveries. Accordingly, the initial infection of a site induces a permanent alternation of the parameter proportional to which it will be susceptible; hence the parameter either decreases, corresponding to (partial) immunization, or increases, i.e. the reverse occurs. Our results concern the three state contact process under the constraint that $\mu\geq\lambda$, this explains the title of the article. When modelling an epidemic, the case that $\mu \leq \lambda$ could be a consequence of imperfect inoculation of individuals following their first exposure to the disease, while the case that $\mu \geq \lambda$ could be a consequence of debilitation of individuals caused by their first exposure to the disease. Specifically, tuberculosis and bronchitis are plausible examples of a disease that captures the latter characteristic.
When $(\lambda,\mu)$ are such that $\lambda = \mu$ the process is reduced to the well known contact process. In this case we will identify a configuration with the subset of ${\mathbb{Z}}$ that corresponds to the set of its infected sites, since states $-1$ and $0$ are effectively equivalent. Also, it is well known that the contact process exhibits a phase transition phenomenon, $\mu_{c}$ will denote its (one-dimensional nearest neighbours) critical value, i.e., $0<\mu_{c}<\infty$ and, if $\mu< \mu_{c}$ the process dies out while if $\mu> \mu_{c}$ the process survives, for an account of various related results and proofs see [@L], [@D88] and [@L99].
It is known that the three state contact process with parameters $(\lambda,\mu)$ such that $\mu > \mu_{c}$ and $\lambda>0$ survives, see [@DS]. We are concerned with the behaviour of the process when survival occurs assuming additionally that $\mu\geq\lambda$. The following theorem summarizes the main results of this paper, in words, parts *(i)* and *(ii)* are respectively a law of large numbers and the corresponding central limit theorem for the rightmost infected while parts *(iii)* and *(iv)* are respectively a law of large numbers and complete convergence for the set of infected sites of the process. For demonstrating our results we introduce some notation. The standard normal distribution function is represented by $N(0,\sigma^{2}), \sigma^{2}>0,$ also, weak convergence of random variables and of set valued processes are denoted by “$\overset{w}{\rightarrow}$” and by “$\Rightarrow$” respectively. Further, we denote by $\bar{\nu}_{\mu}$ the upper invariant measure of the contact process with parameter $\mu$, and by $\delta_{\emptyset}$ the probability measure that puts all mass on the empty set. (For general information about the upper invariant measure and weak convergence of set valued processes we refer to Liggett [@L99]).
\[THEthm\] Consider $\zeta_{t}^{O}$ with parameters $(\lambda, \mu)$, and let $I_{t}= \{x: \zeta_{t}^{O}(x)=1\}$ and $r_{t} = \sup I_{t}$. If $(\lambda, \mu)$ are such that $\mu \geq \lambda>0$ and $\mu > \mu_{c}$ then there exists $\alpha>0$ such that conditional on $\{\zeta_{t}^{O} \mbox{\textup{ survives}}\}$,
\(i) $\displaystyle{ \frac{r_{t}}{t} \rightarrow \alpha}$, almost surely;
\(ii) $\displaystyle{ \frac{r_{t} - \alpha t}{\sqrt{t}} \overset{w}{\rightarrow} N(0,\sigma^{2})}$, for some $\sigma^{2}>0$;
\(iii) let $\theta= \theta(\mu)$ be the density of $\bar{\nu}_{\mu}$, then, $\displaystyle{ \frac{|I_{t}|}{t} \rightarrow 2 \alpha \theta}$, almost surely.
\(iv) Let $\beta= {\mathbf{P}}(\zeta_{t}^{O} \mbox{ survives})>0$, then, $\displaystyle{ I_{t} \Rightarrow (1-\beta) \delta_{\emptyset} + \beta \bar{\nu}_{\mu}}$.
We comment on the proof of Theorem \[THEthm\]. The cornerstone for acquiring parts *(i) (ii)* is to ascertain the existence of a sequence of space-time points, termed *break points*, strictly increasing in both space and time, among which the behaviour of $r_{t}$ conditional on $\{\zeta_{t}^{O} \mbox{\textup{ survives}}\}$ stochastically replicates; these type of arguments have been established by Kuczek, see [@K]. We also note that proofs of parts *(iii) (iv)* are based on variations of the arguments for the contact process case due to Durrett and Griffeath, see [@DG] and [@D80], [@G].
In the next section we introduce the graphical construction, we also present monotonicity and give some elementary coupling results. Section \[Sexp\] is intended for the proof of two exponential estimates that we need for latter. Section \[S3\] is devoted to the study of break points and the proof of Theorem \[THEthm\].
Preliminaries {#S2}
=============
The graphical construction {#grrep}
--------------------------
The graphical construction will be used in order to visualize the construction of various processes on the same probability space; we will repeatedly use it throughout this paper.
Consider parameters $(\lambda,\mu)$ and suppose that $\mu \geq \lambda$, the other case is similar. To carry out our construction for all sites $x$ and $y=x-1,x+1$, let $(T_{n}^{x,y})_{n\geq1}$ and $(U_{n}^{x,y})_{n\geq1}$ be the event times of Poisson processes respectively at rates $\lambda$ and $\mu-\lambda$; further, let $(S_{n}^{x})_{n\geq 1}$ be the event times of a Poisson process at rate $1$. (All Poisson processes introduced are independent).
Consider the space ${\mathbb{Z}}\times [0,\infty)$ thought of as giving a time line to each site of ${\mathbb{Z}}$; Cartesian product is denoted by $\times$. Given a realization of the before-mentioned ensemble of Poisson processes, we define the *graphical construction* and $\zeta_{t}^{[\eta,s]}$, $t\geq s$, the nearest neighbours three state contact process with parameters $(\lambda,\mu)$ started from $\eta$ at time $s\geq0$, i.e. $\zeta_{s}^{[\eta,s]} = \eta$, as follows. At each point $x \times T_{n}^{x,y}$ we place a directed $\lambda$-*arrow* to $y \times T_{n}^{x,y}$; this indicates that at all times $t=T_{n}^{x,y}$, $t\geq s$, if $\zeta_{t-}^{[\eta,s]}(x)=1$ and $\zeta_{t-}^{[\eta,s]}(y)=0$ *or* $\zeta_{t-}^{[\eta,s]}(y) = -1$ then we set $\zeta_{t}^{[\eta,s]}(y) = 1$ (where $\zeta_{t-}(x)$ denotes the limit of $\zeta_{t-\epsilon}(x)$ as $\epsilon\rightarrow 0$). At each point $x \times U_{n}^{x,y}$ we place a directed $(\mu-\lambda)$-*arrow* to $y \times U_{n}^{x,y}$; this indicates that at any time $t=U_{n}^{x,y}$, $t\geq s$, if $\zeta_{t-}^{[\eta,s]}(x)=1$ and $\zeta_{t-}^{[\eta,s]}(y)=0$ then we set $\zeta_{t}^{[\eta,s]}(y) = 1$. While at each point $x \times S_{n}^{x}$ we place a *recovery mark*; this indicates that at any time $t= S_{n}^{x}, t\geq s,$ if $\zeta_{t-}^{[\eta,s]}(x)=1$ then we set $\zeta_{t}^{[\eta,s]}(x) = 0$. The reason we introduced the special marks is to make connection with percolation and hence the contact process, we define the contact process $\xi_{t}^{A}$ with parameter $\mu$ started from $A \subset {\mathbb{Z}}$ as follows. We write $A \times 0 \rightarrow B \times t$, $t\geq 0$, if there exists a connected oriented path from $x \times 0$ to $y \times t$, for some $x \in A$ and $y \in B$, that moves along arrows (of either type) in the direction of the arrow and along time lines in increasing time direction without passing through a recovery mark, defining $\xi_{t}^{A} := \{x: A \times 0 \rightarrow x \times t\}$, $t\geq0$, we have that $(\xi_{t}^{A})$ is a set valued version of the contact process with parameter $\mu$ started from $A$ infected.
It is important to emphasize that the graphical construction, for fixed $(\lambda,\mu)$, defines all $\zeta_{t}^{[\eta,s]}$, $t\geq s$, for any configuration $\eta$ and time $s\geq0$, and all $\xi_{t}^{A}$, for any $A \subset {\mathbb{Z}}$, simultaneously on the same probability space, i.e. provides a coupling of all these processes.
\[calI\]
To simplify our notation, consistently to Section \[S0\], $\zeta_{t}^{[\eta,0]}$ is denoted as $\zeta_{t}^{\eta}$, and, letting $\eta_{0}$ be the standard initial configuration, $\zeta_{t}^{[\eta_{0},0]}$ is denoted as $\zeta_{t}^{O}$. Additionally, the event $\{\mathcal{I}(\zeta_{t}^{[\eta,s]}) \not= \emptyset \mbox{ for all } t\geq s\}$ will be abbreviated below as $\{\zeta_{t}^{[\eta,s]} \mbox{ survives}\}$.
Finally, we note that we have produced a version of $\zeta_{t}^{\eta}$ via a countable collection of Poisson processes, this provides well-definedness of the process. Indeed, whenever one assumes that $|\mathcal{I}(\eta)|<\infty$, this is a consequence of standard Markov chains results having an almost surely countable state space; otherwise, this is provided by an argument due to Harris [@H], see Theorem 2.1 in Durrett [@D95].
Monotonicity, coupling results {#coupling}
------------------------------
To introduce monotonicity concepts, we endow the space of configurations $\{-1,0,1\}^{{\mathbb{Z}}}$ with the *component-wise partial order*, $\mbox{i.e.}$, for any two configurations $\eta_{1}, \eta_{2}$ we have that $\eta_{1} \leq \eta_{2}$ whenever $\eta_{1}(x) \leq \eta_{2}(x)$ for all $x \in {\mathbb{Z}}$. The following theorem is a known result, for a proof we refer to section 5 in Stacey [@S].
\[moninit1\] Let $\eta$ and $\eta'$ be any two configurations such that $\eta\leq \eta'$. Consider the respective three state contact processes $\zeta_{t}^{\eta}$ and $\zeta_{t}^{\eta'}$ with the same parameters $(\lambda,\mu)$ coupled by the graphical construction. For all $(\lambda,\mu)$ such that $\mu \geq \lambda>0$, we have that $\zeta_{t}^{\eta} \leq \zeta_{t}^{\eta'}$ holds. We refer to this property as monotonicity in the initial configuration.
For the remainder of this subsection we give various coupling results concenring $\zeta_{t}^{O}$ the nearest neighbours three state contact process with parameters $(\lambda,\mu)$ started from the standard initial configuration, let $I_{t} = \mathcal{I}(\zeta_{t}^{O})$, $r_{t}= \sup I_{t}$ and $l_{t}= \inf I_{t}$.
We note that both the nearest neighbours assumption as well as the assumption that $\mu \geq \lambda$ in all three of the proofs in this subsection is crucial.
The next lemma will be used repeatedly throughout this paper, its proof given below is a simple extension of a well known result for the nearest neighbours contact process on ${\mathbb{Z}}$, see $\mbox{e.g.}$ $\cite{D80}$.
\[piprendcoup1\] Let $\eta'$ be any configuration such that $\eta'(0)= 1$ and $\eta'(x)= -1$ for all $x\geq1$. Consider $\zeta_{t}^{\eta'}$ with parameters $(\lambda,\mu)$ and let $r_{t}'= \sup\mathcal{I}(\zeta_{t}^{\eta'})$. For $(\lambda,\mu)$ such that $\mu \geq \lambda$, if $\zeta_{t}^{O}$ and $\zeta_{t}^{\eta'}$ are coupled by the graphical construction then the following property holds, for all $t\geq 0$, $$\label{couprend2}
r_{t} = r_{t}' \mbox{ on } \{ I_{t} \not= \emptyset\}.$$
We prove the following stronger statement, for all $t\geq 0$, $$\label{couprend1}
\zeta_{t}^{O}(x) = \zeta_{t}^{\eta'}(x) \mbox{ for all } x \geq l_{t}, \mbox{ on } \{I_{t} \not= \emptyset\}.$$ Clearly (\[couprend1\]) holds for $t=0$, we show that all possible transitions preserve it. An increase of $l_{t}$ (i.e., a recovery mark at $l_{t}\times t$) as well as any transition changing the state of any site $x$ such that $ x\geq l_{t}+1$ preserve $(\ref{couprend1})$. It remains to examine transitions that decrease $l_{t}$, by monotonicity in the initial configuration we have that the possible pairs of $(\zeta_{t}^{O}(l_{t}-1), \zeta_{t}^{\eta'}(l_{t}-1))$ are the following $(-1,0),(-1,1),(0,0),(0,1)$. In the first pair case $(\ref{couprend1})$ is preserved because $\lambda$-arrows are used for transitions $-1 \rightarrow 1$ as well as $ 0 \rightarrow 1$, while in the three remaining cases this is obvious, the proof of $(\ref{couprend1})$ is thus complete.
The next lemma will be used in the proof of the two final parts of Theorem $\ref{THEthm}$, its proof is a simple variant of that of Lemma $\ref{piprendcoup1}$ and thus is omitted.
\[cccoup\] Let $\xi_{t}^{{\mathbb{Z}}}$ be the nearest neighbours contact process with parameter $\mu$ started from ${\mathbb{Z}}$. For $(\lambda,\mu)$ such that $\mu \geq \lambda>0$, if $\zeta_{t}^{O}$ and $\xi_{t}^{{\mathbb{Z}}}$ are coupled by the graphical construction the following property holds, for all $t\geq 0$, $$I_{t} = \xi_{t}^{{\mathbb{Z}}} \cap [l_{t},r_{t}] \mbox{ \textup{on} } \{I_{t} \not= \emptyset\}.$$
\[defetak\]
Our final coupling result will be used in the definition of break points in Subsection $\ref{S31}$. To state the lemma, define the stopping times $\tau_{k} = \inf\{t: r_{t}=k\}$, $k\geq1$, and also $R = \sup_{t\geq0}r_{t}$.
\[Sk\] Let $(\lambda,\mu)$ be such that $\mu \geq \lambda>0$ and consider the graphical construction. Consider also the processes $\zeta_{t}^{[\eta_{k},\tau_{k}]}$, $k\geq1$, started at times $\tau_{k}$ from $\eta_{k}$, as in Definition \[defetak\]. Then, for all $\mbox{ }k=1,\dots,R$ the following property holds, $$\label{eqSk}
\zeta_{t}^{O} \geq \zeta_{t}^{[\eta_{k}, \tau_{k}]}, \mbox{ for all } t \geq \tau_{k}.$$
We have that $\zeta_{\tau_{k}}^{O}(k) =1$, because $\eta_{k}$ is the least infectious configuration such that $\eta_{k}(k)=1$, we also have $\zeta^{O}_{\tau_{k}} \geq \eta_{k}$ $\mbox{ for all }k=1,\dots,R,$ by monotonicity in the initial configuration the proof is complete.
Exponential estimates {#Sexp}
=====================
This section is intended for proving two exponential estimates for three state contact processes that will be needed in Section $\ref{S3}$. The method used is based on a renormalization result of Durrett and Schinazi $\cite{DS}$ that is an extension of the well-known work of Bezuidenhout and Grimmett [@BG].
Subsequent developments require understanding of oriented site percolation. Consider the set of sites, $\mathbb{L} =\{ (y,n) \in {\mathbb{Z}}^{2} : n \geq0 \mbox{ and } y+n \mbox{ is even}\}.$ For each site $(y,n) \in \mathbb{L}$ we associate an independent Bernoulli random variable $w(y,n)\in \{0,1\}$ with parameter $p>0$; if $w(y,n)=1$ we say that $(y,n)$ is *open*. We write $(x,m)\rightarrow(y,n)$ whenever there exists a sequence of open sites $(x,m) \equiv (y_{0}, m),\dots, (y_{n-m}, n) \equiv (y,n)$ such that and $|y_{i} - y_{i-1}|=1$ for all $i=1,\dots, n-m$. Define $(A_{n})_{n\geq0}$ with parameter $p$ as $A_{n} = \{y:(0,0) \rightarrow (y,n)\}$. We write $\{A_{n} \textup{ survives}\}$ as an abbreviation for $\{\forall \hspace{0.5mm} n\geq1: A_{n} \not= \emptyset \}$.
The next proposition is the renormalization result, it is a consequence of Theorem 4.3 in Durrett [@D95], where the comparison assumptions there hold due to Proposition 4.8 of Durrett and Schinazi $\cite{DS}$. For stating it, given constants $L,T$, we define the set of configurations $Z_{y} = \{ \zeta : |\mathcal{I}(\zeta) \cap [-L+2Ly,L+2Ly] | \geq L^{0.6} \}$, for all integers $y$.
\[couplDS\] Let $\eta$ be any configuration such that $\eta \in Z_{0}$, consider $\zeta_{t}^{\eta}$ with parameters $(\lambda, \mu)$ such that $\mu >\mu_{c}$ and $\lambda>0$. For all $p<1$ there exist constants $L,T$ such that $\zeta_{t}^{\eta}$ can be coupled to $A_{n}$ with parameter $p$ so that, $$y \in A_{n} \hspace{1mm} \Rightarrow \hspace{1mm} \zeta_{nT}^{\eta} \in Z_{y}$$ $(y,n) \in \mathbb{L}$. In particular the process survives.
The first of the exponential estimates that we need for Section $\ref{S3}$ is the following.
\[expbounds3scp\] Consider $\zeta_{t}^{O}$ with parameters $(\lambda, \mu)$. Let also $I_{t} = \mathcal{I}(\zeta_{t}^{O})$, $r_{t} = \sup I_{t}$ and $R= \sup_{t\geq0} r_{t}$, further let $\rho = \inf\{t: I_{t} = \emptyset\}$. If $(\lambda, \mu)$ are such that $\mu >\mu_{c}$ and $\mu\geq\lambda>0$ then there exist constants $C$ and $\gamma>0$ such that $$\label{eqexpbounds3scp}
{\mathbf{P}}(R \geq n, \mbox{ }\rho<\infty) \leq Ce^{-\gamma n},$$ for all $n\geq1$.
Consider the graphical construction for $(\lambda, \mu)$ as in the statement. Recall the component-wise partial order on the space of configurations, the property of monotonicity in the initial configuration that were introduced in subsection $\ref{coupling}$ and, the configurations $\eta_{k}$ as in Definition \[defetak\]. By Proposition $\ref{couplDS}$, emulating the proof of Theorem 2.30 (a) of Liggett [@L99], we have that $$\label{rhoconfin}
\mbox{ } {\mathbf{P}}( t < \rho < \infty) \leq Ce^{-\gamma t},$$ for all $t\geq0$; to see that the arguments given there apply in this context note that, by monotonicity in the initial configuration, for any time $s>0$ and any $x\in I_{s}$, considering the process $\zeta_{t}^{ [\eta_{x}, s]}$ we have that $\zeta_{t}^{O} \geq \zeta_{t}^{[\eta_{x}, s]}$ for all $t\geq s$, hence, the proof we referred to applies for $\delta = {\mathbf{P}}(\zeta_{1}^{O} \in Z_{0}) >0$.
For proving $(\ref{eqexpbounds3scp})$, by set theory we have that for all $n \geq 1$, $$\label{RT}
{\mathbf{P}}\left(R > n, \rho < \infty\right) \leq {\mathbf{P}}\left( \frac{n}{\lambda} < \rho< \infty\right) + {\mathbf{P}}\left(\rho < \frac{n}{\lambda}, \mbox{ } R > n\right)$$ the first term on the right hand side decays exponentially in $n$ due to $(\ref{rhoconfin})$, thus, it remains to prove that the probability of the event $\{\sup_{t\leq \frac{n}{\lambda}} r_{t} >n\}$ decays exponentially in $n$, which however is immediate because $\sup_{t \in (0,u]} r_{t}$ is bounded above in distribution by the number of events by time $u$ in a Poisson process at rate $\lambda$ and standard large deviations results for the latter.
The following elementary result for independent site percolation as well as the subsequent geometrical lemma are needed in the proof of Proposition $\ref{shadldev2}$ below.
\[indperc\] Consider $(A_{n})$ with parameter $p$ and define $R_{n} = \sup A_{n}, n\geq0$. For $p$ sufficiently close to 1 there are strictly positive and finite constants $a,\gamma$ and $C$ such that $$\label{Rnspproc}
{\mathbf{P}}( R_{n} < an, \mbox{ } A_{n} \textup{ survives}) \leq Ce^{- \gamma n},$$ for all $n\geq1$.
Define $A'_{n} = \{y: (x,0) \rightarrow (y,n) \mbox{ for some } x \leq0\}$ and let $R'_{n}= \sup A'_{n}$, $n\geq1$. Because $R_{n} = R'_{n}$ on $\{A_{n} \textup{ survives}\}$, it is sufficient to prove that $p$ can be chosen sufficiently close to 1 such that, for some $a>0$, the probability of the event $R'_{n} < an$ decays exponentially in $n \geq 0$. Letting $B_{n}'$ be independent oriented bond percolation on $\mathbb{L}$ with supercritical parameter $\tilde{p}<1$ started from $\{(x,0) \in \mathbb{L} : x\leq0 \}$, the result follows from the corresponding large deviations result for $B_{n}'$ (see Durrett [@D84], (1) in section 11), because for $p= \tilde{p}(2-\tilde{p})$ we have that $B_{n}'$ can be coupled to $A_{n}'$ such that $B_{n}' \subset A_{n}'$ holds, see Liggett [@L99], p.13.
\[geomRR+\] Let $b,c$ be strictly positive constants. For any $a < c$ we can choose sufficiently small $\phi \in (0,1)$, that does not depend on $t \in {\mathbb{R}}\geq0$, such that for all $x \in [-b \phi t,b \phi t]$, $$\label{eq:geom}
[x - c (1-\phi)t, x + c (1-\phi)t] \supseteq [-at,at],$$ $t\geq0$.
Note that it is sufficient to consider $x = brt$; then, simply choose $\phi$ such that $btr -c(1-\phi)t< - at$, $\mbox{i.e.}$ for $\displaystyle{ \phi < \frac{c-a}{c+b}}$, $\phi>0$, equation $(\ref{eq:geom})$ holds.
The other exponential estimate we will need in Section $\ref{S3}$ is the following.
\[shadldev2\] Let $\bar{\eta}$ such that $\bar{\eta}(x) =1$ for all $x \leq 0$ while $\bar{\eta}(x) =-1$ otherwise. Consider $\zeta^{\bar{\eta}}_{t}$ with parameters $(\lambda, \mu)$ and let $\displaystyle{ \bar{r}_{t} =\sup \mathcal{I}(\zeta^{\bar{\eta}}_{t})}$. If $(\lambda, \mu)$ are such that $\mu >\mu_{c}$ and $\mu\geq\lambda>0$ then there exist strictly positive and finite constants $a,\gamma$ and $C$ such that $${\mathbf{P}}\left(\bar{r}_{t} < at\right) \leq Ce^{-\gamma t},$$ for all $t\geq0$.
Consider the graphical construction for $(\lambda, \mu)$ as in the statement. Let $p$ be sufficiently close to $1$ so that Lemma $\ref{indperc}$ is satisfied. Recall the configurations $\eta_{x}$ as in Definition \[defetak\]. By the proof of Theorem 2.30 (a) of Liggett [@L99]—which applies for the reasons explained in the first paragraph of the proof of Proposition $\ref{expbounds3scp}$—, we have that total time $\sigma$ until we get a percolation process $A_{n}$ with parameter $p$ that is coupled to $\zeta_{t}^{[\eta_{\bar{r}_{\sigma}},\sigma]}$ as explained in Proposition $\ref{couplDS}$ (for $\bar{r}_{\sigma} \times (\sigma +1)$ being thought of as the origin) and is conditioned on $\{A_{n} \textup{ survives}\}$, is exponentially bounded. From this, because $\bar{r}_{t}$ is bounded above in distribution by a Poisson process, we have that there exists a constant $\tilde\lambda$ such that the event $\left\{\bar{r}_{\sigma} \times (\sigma +1) \in [ -\tilde{\lambda}t d, \tilde{\lambda}td] \times (0, td]\right\}$, for all $d \in (0,1)$, occurs outside some exponentially small probability in $t$. Finally on this event, by Lemma $\ref{indperc}$ and the coupling in Lemma $\ref{piprendcoup1}$, we have that there exists an $\tilde{a}>0$ such that $\bar{r}_{t} \geq \tilde{a}t - \bar{r}_{\sigma}$, again outside some exponentially small probability in $t$, choosing $\tilde{\lambda}=b$ and $\tilde{a}=c$ in Lemma $\ref{geomRR+}$ completes the proof.
Main Results {#S3}
============
This section is organized as follows. In Subsection \[S31\] we prove Theorem \[THEprop\] stated below; based on this theorem, we prove Theorem \[THEthm\] in Subsection \[S32\].
\[THEprop\] Consider $\zeta_{t}^{O}$ with parameters $(\lambda, \mu)$ and let $r_{t} = \sup\mathcal{I}(\zeta_{t}^{O})$. Suppose $(\lambda, \mu)$ such that $\mu >\mu_{c}$ and $\mu\geq\lambda>0$. On $\{\zeta^{O}_{t} \mbox{\textup{ survives}}\}$ there exist random (but not stopping) times $\tilde{\tau}_{0}:=0 < \tilde{\tau}_{1} < \tilde{\tau}_{2} < \dots$ such that $(r_{\tilde{\tau}_{n}} -r_{\tilde{\tau}_{n-1}}, \tilde{\tau}_{n}- \tilde{\tau}_{n-1})_{n \geq 1}$ are i.i.d. random vectors, where also $r_{\tilde{\tau}_{1}} \geq 1$ and $\displaystyle{r_{\tilde{\tau}_{n}} = \sup_{t\leq \tilde{\tau}_{n}} r_{t}}$. Furthermore, letting $M_{n}=r_{\tilde{\tau}_{n}}-\inf_{t \in[ \tilde{\tau}_{n}, \tilde{\tau}_{n+1})} r_{t}$, $n \geq 0$, we have that $(M_{n})_{n\geq0}$ are i.i.d. random variables, where also $M_{n} \geq0$. Finally, $r_{\tilde{\tau_{1}}}, \tilde{\tau}_{1},M_{0}$ are exponentially bounded.
Break points {#S31}
------------
We first define our break points.
\[definbpts\]
Letting $\tilde{\tau}_{n}:= \tau_{K_{n}}, n\geq0$, in the definition of break points above gives us that for proving Theorem $\ref{THEprop}$ it is sufficient to prove the two propositions following; this subsection is intended for proving these.
\[PROPexpbnd\] $K_{1}$, $\tau_{K_{1}}$ and $M_{0}$ are exponentially.
\[PROPiid\] $(X_{n}, \Psi_{n} ,M_{n-1})_{n \geq 1}$, are independent identically distributed vectors.
We shall denote by $\mathcal{F}_{t}$ the sigma algebra associated to the ensemble of Poisson processes used for producing the graphical construction up to time $t$.
The setting of the following lemma is important to what follows.
\[bptsinf\] Let $\bar{\eta}$ such that $\bar{\eta}(x) =1$ for all $x \leq 0$ while $\bar{\eta}(x) =-1$ otherwise. Consider $\zeta^{\bar{\eta}}_{t}$ with parameters $(\lambda, \mu)$. Define $\bar{r}_{t} =\sup\mathcal{I}(\zeta^{\bar{\eta}}_{t})$, define also, the stopping times $T_{n} = \inf\{t: \bar{r}_{t} = n\}$, $n\geq0$. Let $(\lambda,\mu)$ be such that $\mu \geq \lambda>0$ and $\mu > \mu_{c}$ and consider the graphical construction.
Let $Y_{1}:=1$ and consider $\zeta_{t}^{1}:=\zeta_{t}^{[\eta_{Y_{1}}, T_{1}]}$, we let $\rho_{1} = \inf\{ t\geq T_{1}: \mathcal{I}(\zeta_{t}^{1}) = \emptyset\}$. For all $ n\geq 1$, proceed inductively: On the event $\{\rho_{n} < \infty\}$ let $$Y_{n+1} = 1+ \sup_{t \in [T_{Y_{n}},\rho_{n})} \bar{r}_{t},$$ and consider $\zeta_{t}^{n+1}:= \zeta_{t}^{[\eta_{Y_{n+1}},T_{Y_{n+1}}]}$, we let $\rho_{n+1} = \inf\{t \geq T_{Y_{n+1}}: \mathcal{I}(\zeta_{t}^{n+1}) = \emptyset\}$; on the event that $\{\rho_{n} = \infty\}$ let $\rho_{l} = \infty$ for all $l > n$. Define the random variable $N = \inf\{n\geq1: \rho_{n}= \infty\}$. We have the following expression, $$\label{Yinf}
Y_{N} = \inf\{k\geq1: \zeta_{t}^{[\eta_{k}, T_{k}]} \mbox{ \textup{survives}}\},$$ and also, $$\label{eq:algopiprendcoup}
\bar{r}_{t} = \sup\mathcal{I}(\zeta_{t}^{n}), \mbox{ for all } t \in [T_{Y_{n}},\rho_{n}) \mbox{ and } n\geq1.$$
We further have that $$\label{cbpts1}
(\zeta^{1}_{t+ T_{1}}- 1)_{t \geq 0} \mbox{ is independent of }\mathcal{F}_{T_{1}} \mbox{ and is equal in distribution to } (\zeta_{t}^{O})_{t\geq0},$$ and also, $$\begin{aligned}
\label{cbpts2}
&& \mbox{ conditional on } \{\rho_{n}<\infty, Y_{n+1} = w\}, w \geq 1, (\zeta^{n+1}_{t+ T_{Y_{n+1}}}- w)_{t \geq 0} \nonumber\\
&& \mbox{ is independent of } \mathcal{F}_{T_{Y_{n+1}}} \mbox{ and is equal in distribution to } (\zeta_{t}^{O})_{t\geq0}.\end{aligned}$$
Equation $(\ref{Yinf})$ is a consequence of Lemma $\ref{Sk}$, to see this note that this lemma gives that for all $n\geq1$ on $\{\rho_{n}<\infty\}$, $\rho_{n} \geq \inf\{t\geq T_{k}: \mathcal{I}(\zeta_{t}^{[\eta_{k},T_{k}]})= \emptyset\}$ $\mbox{for all } k=Y_{n}+1,\dots,Y_{n+1}-1$. Equation $(\ref{eq:algopiprendcoup})$ is immediate due to Lemma $\ref{piprendcoup1}$.
Note that from Proposition $\ref{shadldev2}$ we have that $T_{n} < \infty \mbox{ for all } n\geq0$ a.s.. Then, equation $(\ref{cbpts1})$ follows from the strong Markov property at time $T_{1}<\infty$ and translation invariance; while $(\ref{cbpts2})$ is also immediate by applying the strong Markov property at time $T_{Y_{n+1}}<\infty$, where $T_{Y_{n+1}}<\infty$ because from Proposition $\ref{expbounds3scp}$ we have that, conditional on $\rho_{n}<\infty$, $Y_{n+1} < \infty$ $\mbox{a.s.}$.
The connection between the break points and Lemma $\ref{bptsinf}$ comes by the following coupling result.
\[KeqK’\] Let $\eta'$ be any configuration such that $\eta'(0)= 1$ and $\eta'(x)= -1$ for all $x\geq1$. Consider $\zeta_{t}^{\eta'}$ with parameters $(\lambda,\mu)$ and let $r_{t}'= \sup\mathcal{I}(\zeta_{t}^{\eta'})$, let also $\tau'_{k} = \inf\{t\geq 0: r'_{t} = k\}$, $k\geq1$. Define the integers $$K' = \inf\{k\geq 1: \zeta_{t}^{[\eta_{k}, \tau'_{k}]} \mbox{\textup{ survives}}\},$$ and also $M'= \inf_{0 \leq t \leq \tau_{K}'} r_{t}'.$ Consider further $\zeta_{t}^{O}$ with parameters $(\lambda,\mu)$. For $(\lambda,\mu)$ such that $\mu \geq \lambda>0$ and $\mu >\mu_{c}$, if $\zeta_{t}^{O}$ and $\zeta_{t}^{\eta'}$ are coupled by the graphical construction the following property holds, $$\label{coupK1tauK1}
(K', \tau'_{K'},M') = (K_{1}, \tau_{K_{1}}, M_{0}), \mbox{ on } \{\zeta_{t}^{O} \mbox{\textup{ survives}}\},$$ where $K_{1}, \tau_{K_{1}}, M_{0}$ are as in Definition $\ref{definbpts}$.
The proof of Lemma \[KeqK’\] is trivial, it is an immediate consequence of Lemma $\ref{piprendcoup1}$.
Consider the setting of Lemma $\ref{bptsinf}$. By the definition of break points, Definition \[definbpts\], and Lemma $\ref{KeqK'}$ we have that on $\{\zeta_{t}^{O} \mbox{ survives}\}$, $K_{1}=Y_{N}$, $\tau_{K_{1}}= T_{Y_{N}}$ and $M_{0}=\inf_{t \leq T_{Y_{N}}}\bar{r}_{t}$. It is thus sufficient to prove that the random variables $Y_{N}, T_{Y_{N}}, \inf_{t \leq T_{Y_{N}}}\bar{r}_{t}$ are exponentially bounded, merely because an exponentially bounded random variable is again exponentially bounded conditional on any set of positive probability.
We have that $$\label{eq:YN}
Y_{N} = 1+ \sum\limits_{n=2}^{N} (Y_{k} - Y_{k-1}) \mbox{ on } \{N\geq2\},$$ while $Y_{1}:=1$, using this and Proposition \[expbounds3scp\], we will prove that $Y_{N}$ is bounded above in distribution by a geometric sum of $\mbox{i.i.d.}$ exponentially bounded random variables and hence is itself exponentially bounded.
Let $\rho$ and $R$ be as in Proposition \[expbounds3scp\], we define $p_{R}(w)= {\mathbf{P}}(R +1 = w, \rho<\infty)$, and $\bar{p}_{R}(w)= {\mathbf{P}}(R +1= w | \mbox{ }\rho <\infty)$, for all integer $w\geq1$, define also $p= {\mathbf{P}}(\rho = \infty)>0$ and $q=1-p$, where $p>0$ by Proposition \[couplDS\].
By $(\ref{cbpts1})$ of the statement of Lemma $\ref{bptsinf}$, we have that $$\label{eq:Y2}
{\mathbf{P}}( Y_{2} -Y_{1} = w, \rho_{1} < \infty) = p_{R}(w)$$ $w\geq1$; similarly, from $(\ref{cbpts2})$ of the same statement, we have that, for all $n\geq1$, $$\label{eq:rec1}
{\mathbf{P}}(\rho_{n+1} = \infty| \mbox{ } \rho_{n} <\infty, Y_{n+1} = z, \mathcal{F}_{T_{Y_{n+1}}}) = p,$$ and also, $$\label{eq:rec2}
{\mathbf{P}}(Y_{n+1} - Y_{n} = w, \rho_{n}<\infty |\mbox{ } \rho_{n-1} <\infty, Y_{n}=z,\mathcal{F}_{T_{Y_{n}}}) = p_{R}(w) ,$$ for all $w,z\geq1$.
Clearly $\{N=n\} = \{\rho_{k}<\infty \mbox{ for all } k =1,\dots,n-1 \mbox{ and } \rho_{n} = \infty\}$, $n\geq2$, and hence, $$\displaylines{
\left\{\textstyle{\bigcap \limits_{n=1}^{m}} \{Y_{n+1}-Y_{n} = w_{n}\}, N=m+1\right\} = \hfill \cr
= \left\{\textstyle{ \bigcap\limits_{n=1}^{m}} \{Y_{n+1}-Y_{n} = w_{n}, \rho_{n}<\infty\}, \rho_{m+1}=\infty\right\},}$$ for all $m\geq1$, using this, from $(\ref{eq:rec1})$, $(m-1)$ applications of $(\ref{eq:rec2})$, and $(\ref{eq:Y2})$, since $p_{R}(w)= q\bar{p}_{R}(w)$, we have that $$\label{eq:prod-ind}
{\mathbf{P}}\left(\bigcap \limits_{n=1}^{m} \{Y_{n+1}-Y_{n} = w_{n}\}, N=m+1\right)= p q^{m} \prod \limits_{n=1}^{m} \bar{p}_{R}(w_{n}),$$ for all $m\geq1$ and $w_{n}\geq1$. From the last display and $(\ref{eq:YN})$, due to Proposition \[expbounds3scp\], we have that $Y_{N}$ is exponentially bounded by an elementary conditioning argument as follows. Letting $(\tilde{\rho}_{k}, \tilde{R}_{k}), k\geq1$ be independent pairs of random variables each of which is distributed as $(\rho,R)$ and the geometric random variable $\tilde{N} := \inf\{n\geq1: \tilde{\rho_{n}} = \infty\}$, we have that $Y_{N}$ is equal in distribution to $\sum \limits_{k=0}^{\tilde{N}-1} \tilde{R_{k}}$, $\tilde{R}_{0}:=1$.
We proceed to prove that $T_{Y_{N}}$ and $\inf_{t \leq T_{Y_{N}}}\bar{r}_{t}$ are exponentially bounded random variables. By $(\ref{Yinf})$, letting $\bar{x}_{t} = \sup_{s \leq t}\bar{r}_{s}$, we have that $\{T_{Y_{N}} > t\}= \{\bar{x}_{t} \leq Y_{N}\}$; from this and set theory we have that, for any $a>0$ $$\begin{aligned}
\label{bpexm}
{\mathbf{P}}(T_{Y_{N}} > t) &=& {\mathbf{P}}(\bar{x}_{t} \leq Y_{N}) \nonumber\\
&\leq& {\mathbf{P}}(\bar{x}_{t} < at) + {\mathbf{P}}(\bar{x}_{t}\geq at, \bar{x}_{t} \leq Y_{N}) \nonumber\\
&\leq& {\mathbf{P}}(\bar{x}_{t} < at) + {\mathbf{P}}(Y_{N} \geq \lfloor at \rfloor),\end{aligned}$$ for all $t\geq0$, where $\lfloor \cdot\rfloor$ is the floor function; choosing $a>0$ as in Proposition $\ref{shadldev2}$, because $\bar{x}_{t} \geq \bar{r}_{t}$, and since $Y_{N}$ is exponentially bounded, we deduce by (\[bpexm\]) that $T_{Y_{N}}$ is exponentially bounded as well.
Finally, we prove that $M:= \inf_{t \leq T_{Y_{N}}}\bar{r}_{t}$ is exponentially bounded. From set theory, $${\mathbf{P}}(M < -x) \leq {\mathbf{P}}\left(T_{Y_{N}} \geq \frac{x}{\mu} \right) + {\mathbf{P}}\left(T_{Y_{N}} < \frac{x}{\mu}, \{ \bar{r}_{s} \leq -x \mbox{ for some } s \leq T_{Y_{N}}\}\right),$$ because $T_{Y_{N}}$ is exponentially bounded, it is sufficient to prove that the second term of the right hand side decays exponentially. However, recall that $\bar{r}_{_{T_{Y_{N}}}} \geq 1$, hence, $$\displaylines{ {\mathbf{P}}\left(T_{Y_{N}} < \frac{x}{\mu}, \{ \bar{r}_{s} \leq -x \mbox{ for some } s \leq T_{Y_{N}}\}\right) \leq \hfill \cr
\leq {\mathbf{P}}\left((\bar{r}_{t}- \bar{r}_{s}) > x \mbox{ for some } s \leq \frac{x} {\mu} \mbox{ and } t \leq \frac{x} {\mu} \right),}$$ where the term on the right of the last display decays exponentially in $x$, because $(\bar{r}_{t} - \bar{r}_{s})$, $t >s$ is bounded above in distribution by $\Lambda_{\mu}(s,t]$, the number of events of a Poisson process at rate $\mu$ within the time interval $(s,t]$, by use of standard large deviations for Poisson processes, because $\Lambda_{\mu}(s,t] \leq \Lambda_{\mu}( 0,x / \mu]$ for any $s,t \in (0,x / \mu]$.
The next lemma is used in the proof of Proposition $\ref{PROPiid}$ following.
\[Xbdownii\] Consider the setting of the definition of break points, Definition \[definbpts\]. For all $n\geq1$, we have that $$\label{indXn}
\big\{ \textstyle{ \bigcap \limits_{l=1}^{n} } \{(X_{l}, \Psi_{l}, M_{l-1}) = (x_{l}, t_{l}, m_{l-1}) \}, \zeta_{t}^{O} \textup{ survives}\big\} = \{ \zeta_{t}^{[\eta_{z_{n}},w_{n}]}\textup{ survives}, \tau_{z_{n}} = w_{n}, A \},$$ for some event $A \in \mathcal{F}_{w_{n}}$, where $z_{n} = \sum\limits_{l=1}^{n} x_{l}$ and $w_{n} = \sum\limits_{l=1}^{n} t_{l}$.
Considering the setting of Lemma $\ref{bptsinf}$ we trivially have that $$\{(Y_{N},T_{Y_{N}}, \inf_{t \leq T_{Y_{N}}}\bar{r}_{t}) = (x_{1}, t_{1}, m_{0})\} = \{\zeta_{t}^{[\eta_{x_{1}}, t_{1}]} \mbox{ survives}, T_{x_{1}} = t_{1}, B\},$$ for some event $B \in \mathcal{F}_{t_{1}}$; from this and Lemma $\ref{KeqK'}$ we have that $$\displaylines{ \{ (X_{1}, \Psi_{1}, M_{0}) = (x_{1}, t_{1}, m_{0}) ,\zeta_{t}^{O} \mbox{survives}\} \hfill \cr
\hspace{7mm}=\{\zeta_{t}^{[\eta_{x_{1}},t_{1}]} \mbox{ survives},\tau_{x_{1}} = t_{1}, B, \zeta_{t}^{O} \mbox{survives}\} \cr
= \{\zeta_{t}^{[\eta_{x_{1}},t_{1}]} \mbox{ survives},\tau_{x_{1}} = t_{1}, B, I_{t_{1}} \not=\emptyset\}
}$$ for all $x_{1}\geq 1$, $t_{1} \in {\mathbb{R}}_{+}$, $m_{0} \geq0$, because $\{ I_{t_{1}} \not=\emptyset \} \in \mathcal{F}_{t_{1}}$ we have thus proved $(\ref{indXn})$ for $n=1$, for general $n\geq1$ the proof is derived by repeated applications of the last display.
Consider the setting of the definition of break points, Definition \[definbpts\]. Assume that $K_{n}$, $\tau_{K_{n}}$, $M_{n-1}$ are almost surely finite, we will prove that $$\begin{aligned}
\label{Xn}
{\mathbf{P}}\left((X_{n+1},\Psi_{n+1},M_{n}) = (x, t,m)\vline\mbox{ } \textstyle{ \bigcap \limits_{l=1}^{n}} \{(X_{l},\Psi_{l},M_{l-1}) =(x_{l}, t_{l},m_{l-1})\}, \zeta_{t}^{O} \mbox{survives}\right)&& \nonumber\\
=\mbox{ } {\mathbf{P}}\big( (X_{1},\Psi_{1},M_{0})= (x, t,m)|\mbox{ } \zeta_{t}^{O} \mbox{survives} \big)\hspace{10mm}\end{aligned}$$ for all $(x_{l},t_{l},m_{l-1})$, $x_{l} \geq 1, t_{l}\in {\mathbb{R}}_{+}$, $m_{l-1}\geq0$, $l=1,\dots,n$, and hence in particular that $K_{n+1}$, $\tau_{K_{n+1}}$, $M_{n}$ are exponentially bounded. By induction because $K_{1}$ and $\tau_{K_{1}},M_{0}$ are exponentially bounded by Proposition $\ref{PROPexpbnd}$ we have that $(\ref{Xn})$ completes the proof by Bayes’s sequential formula.
It remains to prove $(\ref{Xn})$, rewrite the conditioning event in its left hand side according to $(\ref{indXn})$ in Lemma $\ref{Xbdownii}$ and note that $$\{\tau_{z_{n}} = w_{n}\} \subset \{ \zeta_{w_{n}}^{O}(z_{n})=1 \mbox{ and } \zeta_{w_{n}}^{O}(y)= -1, \mbox{ for all } y \geq z_{n}+1\},$$ thus, applying Lemma $\ref{KeqK'}$, gives the proof by independence of the Poisson processes in disjoint parts of the graphical construction, because $(\zeta_{t+w_{n}}^{[\eta_{z_{n}}, w_{n}]}-z_{n})_{t\geq0}$ is equal in distribution to $(\zeta_{t}^{O})_{t\geq0}$ by translation invariance.
Proof of Theorem $\ref{THEthm}$ {#S32}
-------------------------------
We denote by $\bar{{\mathbf{P}}}$ the probability measure induced by the construction of the process conditional on $\{\zeta_{t}^{O} \mbox{ survives}\}$ and, by $\bar{{\mathbf{E}}}$ the expectation associated to $\bar{{\mathbf{P}}}$. Consider the setting of Theorem $\ref{THEprop}$ and let $\displaystyle{ \alpha = \frac { \bar{{\mathbf{E}}} (r_{\tilde{\tau}_{1}})} { \bar{{\mathbf{E}}} (\tilde{\tau}_{1})}}$, $\alpha \in (0, \infty)$.
Because $r_{\tilde{\tau}_{n}} = \sum\limits_{m = 1}^{n} (r_{\tilde{\tau}_{m}} - r_{\tilde{\tau}_{m-1}})$ and $\tilde{\tau}_{n} = \sum\limits_{m=1}^{n}( \tilde{\tau}_{m} - \tilde{\tau}_{m-1})$, $n\geq1$, using the strong law of large numbers twice gives us that $$\label{spdatbpts}
\bar{{\mathbf{P}}} \left( \lim_{n \rightarrow \infty} \frac{r_{\tilde{\tau}_{n}}}{\tilde{\tau}_{n}} = \alpha \right)=1,$$ we prove that indeed $\displaystyle{\lim_{t\rightarrow \infty}\frac{r_{t}}{t} =\alpha}$, $\bar{{\mathbf{P}}}$ a.s.. From Theorem $\ref{THEprop}$ we have that $$\label{intt}
\frac{r_{\tilde{\tau}_{n}}-M_{n}}{\tilde{\tau}_{n+1}} \leq \frac{r_{t}}{t} \leq \frac{r_{\tilde{\tau}_{n+1}}} {\tilde{\tau}_{n}} , \mbox{ for all } t\in[\tilde{\tau}_{n}, \tilde{\tau}_{n+1}),$$ $n\geq0$. Further, because $(M_{n})_{n\geq0}$, $M_{0}\geq0$, is a sequence of i.i.d. and exponentially bounded random variables we have that $$\label{spdM}
\bar{{\mathbf{P}}} \left(\lim_{n \rightarrow \infty} \frac{M_{n}}{n} = 0 \right) =1,$$ by the 1st Borel-Cantelli lemma. Consider any $ a < \alpha$, by $(\ref{intt})$ we have that $$\label{eq:spdlessa}
\left\{ \frac{r_{t_{k}}}{t_{k}} < a \mbox{ for some } t_{k} \uparrow \infty \right\}\subseteq \left\{ \limsup_{n\rightarrow \infty} \left\{\frac{r_{\tilde{\tau}_{n}}-M_{n}}{\tilde{\tau}_{n+1}} < a \right\}\right\},$$ however $\displaystyle{ \bar{{\mathbf{P}}}\left(\limsup_{n\rightarrow \infty} \left\{\frac{r_{\tilde{\tau}_{n}}-M_{n}}{\tilde{\tau}_{n+1}} < a \right\}\right)=0}$, to see this simply use $(\ref{spdM})$ and $(\ref{spdatbpts})$ to deduce that $\displaystyle{\lim_{n\rightarrow \infty}\frac{r_{\tilde{\tau}_{n}}-M_{n}}{\tilde{\tau}_{n+1}} = \alpha}$, $\bar{{\mathbf{P}}}$ $\mbox{a.s.}$. By use of the upper bound in $(\ref{intt})$ and $(\ref{spdatbpts})$, we also have that for any $a > \alpha$, $\displaystyle{\bar{{\mathbf{P}}}\left( \left\{ \frac{r_{t_{k}}}{t_{k}} > a \mbox{ for some } t_{k} \uparrow \infty \right\}\right) =0}$, this completes the proof of *(i)*.
We will prove that $$\lim_{t\rightarrow \infty}\bar{{\mathbf{P}}} \left( \frac{r_{t} - \alpha t}{\sqrt{t}} \leq x \right) = \Phi\left(\frac{x}{\sigma^{2}}\right),$$ for some $\sigma^{2}>0$, $x \in{\mathbb{R}}$, where $\Phi$ is the standard normal distribution function, $\mbox{i.e.}$, $\displaystyle{ \Phi(y) := \frac{1}{\sqrt{2\pi}} \int_{-\infty}^{y}\exp\left(-\frac{1}{2} z^{2}\right)dz}$, $y\in {\mathbb{R}}$.
Define $N_{t}= \sup\{n: \tilde{\tau}_{n} <t\}$; evoking Lemma 2 in Kuczek [@K], $\mbox{p.}$ 1330–1331, which applies due to Theorem $\ref{THEprop}$, we have that $$\label{kucinitNt}
\lim_{t\rightarrow \infty}\bar{{\mathbf{P}}} \left( \frac{r_{N_{t}} - \alpha t}{\sqrt{t}} \leq x \right) = \Phi\left(\frac{x}{\sigma^{2}}\right),$$ $x \in{\mathbb{R}}$. From this, by standard association of convergence concepts, i.e. Slutsky’s theorem, it is sufficient to show that $$\label{clt0}
\bar{{\mathbf{P}}} \left(\lim_{t \rightarrow \infty } \frac{r_{t} - r_{N_{t}}}{\sqrt{t}}=0 \right)=1,$$ and that $\sigma^{2}$ is strictly positive. Note however that, by Theorem $\ref{THEprop}$ we have that, $$\label{cltbouds}
\frac{ M_{\tilde{N}_{t}}} {\sqrt{t}} \leq \frac{ r_{t} - r_{N_{t}}}{\sqrt{t}} \leq \frac{ r_{\tilde{\tau}_{N_{t}+1}} - r_{\tilde{\tau}_{N_{t}}}}{\sqrt{t}}$$ for all $t\geq0$.
We show that $(\ref{clt0})$ follows from $(\ref{cltbouds})$. Because $(r_{\tilde{\tau}_{n+1}} - r_{\tilde{\tau}_{n}})_{n\geq0}$, $r_{\tilde{\tau}_{1}}\geq1$, are $\mbox{i.i.d.}$ and exponentially bounded, by the 1st Borel-Cantelli lemma, and then the strong law of large numbers, we have that $$\lim_{n\rightarrow \infty} \frac{ \frac{1}{\sqrt{n}}(r_{\tilde{\tau}_{n+1}} - r_{\tilde{\tau}_{n}})}{\sqrt{ \frac{\tilde{\tau}_{n}}{n}}}= 0$$ $\bar{{\mathbf{P}}}$ $\mbox{a.s.}$, from the last display and emulating the argument given in $(\ref{eq:spdlessa})$ we have that $\displaystyle{ \lim_{t\rightarrow \infty} \frac{ r_{\tilde{\tau}_{N_{t}+1}} - r_{\tilde{\tau}_{N_{t}}}}{\sqrt{t}} =0}$, $\bar{{\mathbf{P}}}$ $\mbox{a.s.}$. Similarly, because $(M_{n})_{n\geq0}$, and $M_{0} \geq 0 $, are also $\mbox{i.i.d.}$ and exponentially bounded, we also have that $\displaystyle{\lim_{t\rightarrow \infty}\frac{ M_{\tilde{N}_{t}}}{\sqrt{t}} =0}$, $\bar{{\mathbf{P}}} \mbox{ a.s.}$. Finally, we show that $\sigma^{2}>0$. As in the proof of Corollary 1 in Kuczek [@K], because $\displaystyle{ \alpha = \frac { \bar{{\mathbf{E}}} (r_{\tilde{\tau}_{1}})} { \bar{{\mathbf{E}}} (\tilde{\tau}_{1})}}$, we need to show that $\bar{{\mathbf{E}}}\left( r_{\tilde{\tau}_{1}}\bar{{\mathbf{E}}}(\tilde{\tau}_{1}) - \tilde{\tau}_{1} \bar{{\mathbf{E}}} (r_{\tilde{\tau}_{1}})\right)^{2}>0$. However, because $r_{\tilde{\tau}_{1}} \geq 1$, this follows by Chebyshev’s inequality. This completes the proof of *(ii)*.
For the remainder of the proof consider the graphical construction for $(\lambda, \mu)$ such that $\mu >\mu_{c}$ and $\mu\geq\lambda>0$. Consider $\zeta_{t}^{O}$, let $r_{t} =\sup I_{t}$ and $l_{t} = \inf I_{t}$ be respectively the rightmost and leftmost infected of $I_{t}= \mathcal{I}(\zeta_{t}^{O})$. Consider also $\xi_{t}^{{\mathbb{Z}}}$, the contact process with parameter $\mu$ started from ${\mathbb{Z}}$. By Lemma $\ref{cccoup}$ we have that, for all $t\geq0$, $$\label{coupHtZ}
I_{t} = \xi_{t}^{{\mathbb{Z}}} \cap [l_{t},r_{t}] \mbox{ \textup{on} } \{I_{t} \not= \emptyset\}.$$
Let $\theta= \theta(\mu) >0$ be the density of the upper invariant measure, i.e., $\displaystyle{ \theta= \lim_{t\rightarrow \infty}{\mathbf{P}}( x \in \xi_{t}^{{\mathbb{Z}}})}$. We prove that $\displaystyle{\lim_{t\rightarrow \infty}\frac{|I_{t}|}{t} = 2 \alpha \theta,}$ $\bar{{\mathbf{P}}}$ a.s..
Considering the interval $[\max\{l_{t}, -\alpha t\}, \min\{r_{t},\alpha t\}]$, we have that for all $t\geq0$, $$\label{sllnineq}
\vline \mbox{ } \sum_{x=l_{t}}^{r_{t}} 1_{\{ x \in \xi_{t}^{{\mathbb{Z}}} \}} - \sum_{x =- \alpha t}^{ \alpha t} 1_{\{ x \in \xi_{t}^{{\mathbb{Z}}} \}} \mbox{ }\vline \leq | r_{t} - \alpha t| + | l_{t} + \alpha t|, \mbox{ on }\{I_{t} \not= \emptyset \},$$ where we denote by $1_{E}$ the indicator of event $E$. However, by $(\ref{coupHtZ})$, we have that $\displaystyle{ |I_{t}|= \sum_{x=l_{t}}^{r_{t}} 1_{\{ x \in \xi_{t}^{{\mathbb{Z}}} \}}}$, $\mbox{on } \{I_{t} \not= \emptyset \}$, thus, because $\displaystyle{\lim_{t\rightarrow \infty} \frac{r_{t}}{t} = \alpha}$ and, by symmetry, $\displaystyle{ \lim_{t\rightarrow \infty} \frac{l_{t}}{t} =-\alpha}$, $\bar{{\mathbf{P}}}$ a.s., the proof follows from $(\ref{sllnineq})$ because it is known that, for any $a>0$, $\displaystyle{ \lim_{t\rightarrow \infty} \frac{1}{t}\sum_{|x| \leq a t} 1_{\{x \in \xi_{t}^{{\mathbb{Z}}} \}} = 2 a \theta}$, ${\mathbf{P}}$ a.s. (see equation (19) in the proof of Theorem 9 of Durrett and Griffeath [@DG]).
Let $\rho = \inf\{t\geq0: I_{t} = \emptyset\}$. In the context of set valued processes, by general considerations, see Durrett [@D95], it is known that weak convergence is equivalent to convergence of finite dimensional distributions and that, by inclusion-exclusion, it is equivalent to show that for any finite set of sites $F\subset {\mathbb{Z}}$ $$\label{eq:compconv}
\lim_{t\rightarrow \infty} {\mathbf{P}}(I_{t} \cap F = \emptyset) = {\mathbf{P}}(\rho < \infty) + {\mathbf{P}}(\rho = \infty)\phi_{F}(\emptyset),$$ where $\displaystyle{ \phi_{F}(\emptyset):= \lim_{t\rightarrow \infty} {\mathbf{P}}(\xi_{t}^{{\mathbb{Z}}} \cap F =\emptyset)}$. By set theory we have that it is sufficient to prove $\displaystyle{ \lim_{t\rightarrow \infty} {\mathbf{P}}(I_{t} \cap F = \emptyset, \rho \geq t)= {\mathbf{P}}(\rho = \infty)\phi_{F}(\emptyset)}$, because $\{\rho <t\} \subseteq \{I_{t}\cap F =\emptyset\}$. However, emulating the proof of the respective result for the contact process (see $\mbox{e.g.}$ Theorem 5.1 in Griffeath [@G]), we get $\displaystyle{\lim_{t \rightarrow \infty} {\mathbf{P}}(\xi_{t}^{{\mathbb{Z}}} \cap F = \emptyset, \rho \geq t) = {\mathbf{P}}(\rho =\infty)\phi_{F}(\emptyset)}$, hence, it is sufficient to prove that $$\label{cccoupconseq}
\limsup_{t \rightarrow \infty} {\mathbf{P}}(I_{t} \cap F = \emptyset, \rho \geq t) \leq \lim_{t\rightarrow \infty}{\mathbf{P}}(\xi_{t}^{{\mathbb{Z}}} \cap F = \emptyset, \rho \geq t),$$ because also $\{ I_{t} \cap F = \emptyset, \rho \geq t\} \supseteq \{ \xi_{t}^{{\mathbb{Z}}} \cap F = \emptyset, \rho \geq t\}$, by $(\ref{coupHtZ})$.
It remains to prove $(\ref{cccoupconseq})$. By elementary calculations, $$\label{ccrhoinf}
{\mathbf{P}}(I_{t} \cap F = \emptyset, \rho = \infty) - {\mathbf{P}}(\xi_{t}^{{\mathbb{Z}}} \cap F = \emptyset, \rho \geq t) \leq {\mathbf{P}}( \xi_{t}^{{\mathbb{Z}}} \cap F \supsetneq I_{t} \cap F, \rho = \infty),$$ for all $t\geq0$, where we used that by $(\ref{coupHtZ})$, $I_{t} \subset \xi_{t}^{{\mathbb{Z}}}$ for all $t\geq0$. From the last display above and set theory we have that $$\displaylines{{\mathbf{P}}(I_{t} \cap F = \emptyset, \rho \geq t)- {\mathbf{P}}(\xi_{t}^{{\mathbb{Z}}} \cap F = \emptyset, \rho \geq t) \hfill \cr
\hfill \leq {\mathbf{P}}( \xi_{t}^{{\mathbb{Z}}} \cap F \supsetneq I_{t} \cap F, \rho = \infty) + {\mathbf{P}}( t<\rho <\infty), }$$ for all $t\geq0$, however the limit as $t\rightarrow \infty$ of both terms of the right hand side in the above display is $0$, for the former this comes by $(\ref{coupHtZ})$, because $\displaystyle{\lim_{t\rightarrow \infty} r_{t} = \infty}$ and $\displaystyle{\lim_{t\rightarrow \infty} l_{t} = \infty}$, $\bar{{\mathbf{P}}}$ $\mbox{a.s.}$, while for the latter this is obvious.
[99]{}
and [ Grimmett, G.R.]{} (1990). The critical contact process dies out. *Ann. Probab.* **18** 1462–1482.
(1980). On the growth of one-dimensional contact processes. [*Ann. Probab.*]{} [**8**]{} 890–907.
(1984). Oriented percolation in two dimensions. [*Ann. Probab.*]{} [**12**]{} 999–1040.
(1988). [*Lecture Notes on Particle Systems and Percolation.*]{} Wadsworth.
(1995). [ *Ten lectures on particle systems*]{} Lecture Notes in Math. [**1608**]{}, Springer-Verlag, New York. and [ Griffeath, D.]{} (1983). Supercritical contact processes on ${\mathbb{Z}}$. [*Ann. Probab.*]{} [**11**]{} 1–15. and [ Schinazi, R. ]{}(2000). Boundary modified contact processes. [*J. Theoret. Probab.*]{} [**13**]{} 575-594. (1979). [*Additive and cancelative interacting particle systems.* ]{} Lecture Notes in Math. [**724**]{} Springer-Verlag, Berlin.
(1972). Nearest neighbor Markov interaction processes on multidimensional lattices. [*Adv. in Math.*]{} [**9**]{} 66–89.
(1989). The central limit theorem for the right edge of supercritical oriented percolation. [*Ann. Prob.*]{} [**17**]{} 1322–1332.
(1985). [*Interacting particle systems.*]{} Springer, New York. (1999). [*Stochastic Interacting Systems: Contact, Voter and Exclusion Processes.*]{} Springer, New York. (2003). Partial immunization processes. [*Ann. Appl. Probab.*]{} [**13**]{}, 669–690.
[^1]: [Heriot-Watt University.]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a setting in which a sender wishes to broadcast a block of $K$ data packets to a set of wireless receivers, where each of the receivers has a subset of the data packets already available to it (e.g., from prior transmissions) and wants the rest of the packets. Our goal is to find a linear network coding scheme that yields the minimum average packet decoding delay (APDD), i.e., the average time it takes for a receiver to decode a data packet. Our contributions can be summarized as follows. First, we prove that this problem is NP-hard by presenting a reduction from the hypergraph coloring problem. Next, we show that a random linear network coding (RLNC) provides an approximate solution to this problem with approximation ratio $2$ with high probability. Next, we present a methodology for designing specialized approximation algorithms for this problem that outperform RLNC solutions while maintaining the same throughput. In a special case of practical interest with a small number of wanted packets our solution can achieve an approximation ratio $\frac{4-2/K}{3}$. Finally, we conduct an experimental study that demonstrates the advantages of the presented methodology.'
author:
- |
Mingchao Yu\*, Alex Sprintson$^\dag$, and Parastoo Sadeghi\*\
\
---
Network coding, decoding delay, NP-hardness, approximation algorithm.
Introduction
============
In this paper, we are interested in a wireless broadcast scenario, in which a sender wishes to broadcast a block of $K$ data packets to a set of wireless receivers, such that each of the receivers already has a subset of the data packets available to it (e.g., from prior transmissions) and is interested in obtaining the rest of the packets. Given a packet reception instance, the goal is to design a linear network coding (NC) scheme that minimizes the average packet decoding delay (APDD), which is defined as the average time it takes for a receiver to decode a data packet.
One of the possible solutions to this problem is to employ a random linear network coding (RLNC) technique [@ho:medard:koetter:karger:effros:2006; @nistor:lucani:vinhoza:costa:barros:2011]. In wireless broadcast scenarios, RLNC can achieve an optimal throughput (i.e., minimize the time required to decode all packets by all receivers) with high probability by mixing all data packets in the block together using linear coefficients randomly chosen from a sufficiently large finite field. However, RLNC is suboptimal in terms of APDD, since in general, no data packet can be decoded by a receiver until it receives $K$ linearly independent coded packets.
Many opportunistic NC techniques have been developed with the aim to reduce APDD or some other measures of decoding delay [@katti1;etal:2008; @Rozner_Heuristic_clique; @costa:munaretto:widmer:baros:2008; @keller:drinea:fragouli:2008; @eryilmaz:ozdaglar:medard:ahmed:2008; @barros:costa:munaretto:widmer:2009; @sundararajan:sadeghi:medard:2009; @nguyen:tran:nguyen:bose:2009; @sadeghi:shams:traskov:2010; @sorour:valaee:2010; @nistor:lucani:vinhoza:costa:barros:2011; @li:idnc_video:2011; @athanasiadou2013stable; @yu:parastoo:neda:2014; @parastoo:yu:neda:isita; @fu2014dynamic]. An important technique in this class is *instantly decodable network coding* (IDNC). The IDNC technique has a potential to reduce the APDD by enabling a subset of receivers to instantly decode a data packet after each transmission. IDNC has been shown to outperform RLNC in terms of APDD for a small number of receivers [@yu:parastoo:neda:2014]. However, since in IDNC schemes a single transmission typically benefits only some of the receivers, IDNC is not throughput optimal. As a result, a larger number of transmissions is necessary to finish the broadcast, which increases the decoding delay for some receivers, and, as a result, increases the value of APDD. Indeed, for larger number of receivers, the throughput of IDNC decreases and APDD increases due to lack of coding opportunities. A similar behavior can be observed for other opportunistic coding techniques [@costa:munaretto:widmer:baros:2008].
In summary, there is no clear winner between RLNC, IDNC, and other opportunistic techniques, as each of them prevail in a different parameter region. Moreover, while the APDD of RLNC can be easily calculated (as shown in Section \[sec:approx\]), the achievable APDD of opportunistic NC techniques has not been characterized analytically.
The contributions of this paper is summarized as follows:
- We first prove that it is NP-hard to minimize APDD, by presenting a reduction from the hypergraph coloring problem.
- Next we show that RLNC achieves an approximation ratio of 2 with high probability, i.e., the APDD achieved by RLNC is at most two times the optimal solution.
- We present a methodology for designing specialized approximation algorithms that achieve lower values of APDD than RLNC while maintaining the same optimal throughput. We also present a case study to demonstrate the algorithm design. We conduct extensive simulations to confirm that our methodology outperforms alternative solutions in the broad range of practical settings.
System Model
============
Our model includes a single sender that holds a set of $K$ data packets that belong to $\F_q$, $\P=\{\p_k\}_{k=1}^K$, and a set of $N$ receivers, $\{\r_n\}_{n=1}^N$, each wants a subset $\w_n$ of $\P$ and has the rest. The packet reception instance in our model is represented by a binary $N\times K$ state feedback matrix (SFM) $\mA$, where $\mA(n,k)=1$ means that $\r_n$ wants $\p_k$, and $\mA(n,k)=0$ means that receiver $r_n$ has packet $p_k$ already available to it (e.g., from prior transmissions). We denote by $w_n$ the size of $\w_n$, and by $t_k$ the number of receivers who want $\p_k$. An example of SFM is given in [Fig.\[fig:sfm\_graph\]]{}(b), which has $w_1=3$ and $t_1=2$.
Given $\mA$, the sender performs a linear NC transmission phase. In each NC transmission, the sender encodes data packets in $\M$ together using linear coefficients from a finite field $\F_q$. The corresponding packet $X$ takes the form of: $$X=\sum_{\p_k\in\M}\beta_k\p_k$$ We denote by $\M=\{p_k\in \P| \beta_k\neq 0\}$ the support of $X$ and refer to it as an *coding set* of $X$. When the coefficients $\{\beta_k\}$ are chosen from $\F_q$ uniformly at random, $X$ is called a random-coded packet of $\M$. A receiver will increase its degree of freedom (DoF) by one when it receives a NC packet that is linearly independent of the set of all the packets it already has. The broadcast will be *completed* at a receiver once it decodes all its wanted data packets.
In order to study the global minimum decoding delay of linear NC in wireless broadcast, we assume the following:
1. NC transmissions are erasure-free, so that every transmitted NC packet can be received by all receivers;
2. Receivers have sufficient computational resources to perform NC decoding under any $\F_q$. When random coding is applied, a sufficiently large $\F_q$ will ensure the linear independency among the random-coded packets with high probability.
A set of $U$ coded packets is called a *NC solution* and is denoted by $\S$ if it allows every receiver to decode all its wanted packets. Let $u_{n,k}$ be the index of the NC transmission at which $\r_n$ decodes $\p_k$. The average packet decoding delay (APDD) of $\S$, denoted by $D_\S$, is calculated as: $$D_\S=\frac{1}{\sum_{n=1}^Nw_n}\sum_{n=1}^N\sum_{k=1}^Ku_{n,k}.$$
Our aim in this paper is to study the smallest $D_\S$ over all possible linear NC solutions. We call it the minimum APDD of $\mA$ and denote it by $\Dmin$. The first question we would like to answer is: *Is it hard to find $\Dmin$?*
The Hardness of Finding $\Dmin$
===============================
In this section, we study the hardness of finding $\Dmin$. To this end, we will first introduce the concept of *perfect NC solution*, whose APDD is a lower bound of $\Dmin$. Then, we will prove that deciding whether a perfect solution exists for a given instance $\mA$ of the problem at hand is an NP-hard problem. This implies, in turn, that the problem of finding $\Dmin$ is also NP-hard.
The Perfect Solution and a Bound of $\Dmin$
-------------------------------------------
A NC solution $\S$ is called a perfect solution and is denoted by $\S_p$ if it allows every receiver $n$ to decode a wanted data packet in every transmission of $\S$.
Obviously, $\S_p$ offers the ideal packet decoding scenario. Its APDD is thus a lower bound of $\Dmin$, denoted by $\underline D$, which is calculated as: $$\begin{aligned}
\Du&=\frac{1}{\sum_{n=1}^Nw_n}\sum_{n=1}^N\sum_{u=1}^{w_n} u\\
&=\frac{1}{\sum_{n=1}^Nw_n}\sum_{n=1}^N\frac{(w_n+1)w_n}{2}\\
&=\frac{\sum_{n=1}^Nw_n^2}{2\sum_{n=1}^Nw_n}+\frac{1}{2}\label{eq:du}\end{aligned}$$ It is clear that $\Du$ can only be achieved by $\S_p$ if a perfect solution exists. The natural question =in this context is: *Does a perfect solution $\S_p$ exist for every SFM?* In the next subsection, by using a reduction from the strong hypergraph coloring problem, we will prove that this question is NP-hard to answer.
Hardness of Finding $\S_p$
--------------------------
We first introduce some useful concepts in hypergraphs. A hypergraph $\mH$ is defined by a pair $(\V,\E)$, where $\V$ is the set of vertices, and $\E$ is the set of hyperedges. Every hyperedge $\e\in\E$ is a subset of $\V$ with size $|\e|\geqslant 1$. A hypergraph is $r-$uniform if every hyperedge $\e$ has equal size, i.e., $|\e|=r$. A $k$-*strong* coloring solution of $\mH$ is a partition of $\V$ into $k$ subsets $\{\V_i\}_{i=1}^k$, such that $|\V_i\cap\e|\leqslant 1$ for any $\e\in\E$. In other words, every color appears at most once in every hyperedge. It is well known that the hypergraph coloring problem is intractable.
\[lemma:strong\_color\] It is NP-hard to determine whether an $r-$uniform hypergraph is $r$-strong colorable, for any $r\geqslant3$.
We then build a reduction from the strong hypergraph coloring solution for $r$-uniform hypergraphs to the problem of finding a perfect NC solution for the average delay minimization problem. Given an $r$-uniform hypergraph $\mH(\V,\E)$ we construct an instance to our problem as follows. First, for each vertex $\v_k$ we introduce a data packet $\p_k$, and for each hyperedge $\e_n$ we introduce a receiver $\r_n$ who wants the data packets that correspond to vertices in $\e_n$. Note that in the resulting SFM $\mA$, every receiver wants $r$ data packets. A 3-uniform hypergraph and the corresponding SFM matrix are depicted in [Fig.\[fig:sfm\_graph\]]{}.
First, we prove that an existence of an $r$-strong coloring solution $\{\V_i\}_{i=1}^r$ of $\mH$ implies a perfect solution for our problem. Let $\{\V_i\}_{i=1}^r$ be an $r$-strong coloring of $\mH$. For each $\V_i$, let $\M_i$ be a set of packets that correspond to vertices in $\V_i$. Note that for each each receiver $\r_n$ and each set $\M_i$ it holds that $|\M_i\cap\w_n|=1$. Consider a coding solution $\S$ that includes $r$ transmissions, such that transmission $i$ includes a sum of packets in $\M_i$ (over $\F_q$). Since every receiver can decode a packet at each transmission, $\S$ is a perfect solution to our our problem.
Next, we show that a perfect solution $\S_p$ for the instance of $\mA$ of our problem implies that there exists an $r$-strong coloring solution $\{\V_i\}_{i=1}^r$ of $\mH$. Let $\M_i$ be the coding set that corresponds to the transmission $i$ of $\S_p$ and let $V_i$ be the set of vertices in $\mH$ that correspond to $\M_i$. Note that in order to allow every receiver to decode one data packet in each of the $r$ transmissions, every $\M_i$ must contain one wanted data packet of every receiver, i.e., $|\M_i\cap\w_n|=1$. Thus, $\{\V_i\}_{i=1}^r$ is an $r$-strong coloring solution $\{\V_i\}_{i=1}^r$ of $\mH$.
We conclude that an $r$-uniform hypergraph is $r$-strong colorable if and only if there exists a perfect NC solution of the instance $\mA$ of our problem. We summarize our results in the following lemma:
\[cor:det\_perfect\] It is NP-hard to determine whether there exists a perfect solution for a given instance $\mA$ of minimum APDD problem.
The theorem follows from our construction and Lemma \[lemma:strong\_color\].
The Hardness of Finding $\Dmin$
-------------------------------
Since $\underline{D}$ can only be achieved by a perfect solution $\S_p$, an optimal algorithm that finds $\Dmin$ will be able to determine the existence of a perfect solution by comparing $\Dmin$ with $\underline{D}$. According to Lemma \[cor:det\_perfect\], this decision is NP-hard to made, and thus it is NP-hard to find $\Dmin$:
It is NP-hard to find $\Dmin$ for a given instance $\mA$ of minimum APDD problem .
In addition to NP-hardness, our reduction from the hypergraph coloring also yields an interesting conjecture on the existence of perfect solution for some special instances of $\mA$. It comes from the famous Erdős-Faber-Lovász conjecture in graph theory [@erdos1981combinatorial]:
Consider an with $r$ hyperedges. Each pair of hyperedges have at most one vertex in common. This hypergraph is $r$-strong colorable.
The corresponding conjecture in NC context is as follows:
Consider an instance $\mA$ of our problem with $r$ receivers, each wants $r$ data packets, and each pair of receivers want at most one data packet in common. This $\mA$ has a perfect NC solution $\S_p$.
We showed that the problem of finding a minimum value of $\Dmin$ is intractable. Accordingly, in the next sections we discuss approximation algorithms for this problem.
Approximating $\Dmin$ {#sec:approx}
======================
In this section, we aim at approximating $\Dmin$. An approximation algorithm of $\Dmin$ produces a linear NC solution $\S$ with its APDD obeying $D_\S\leqslant \a\Dmin$. We refer to $\a\geqslant 1$ as *approximation ratio* of the algorithm.
In the next theorem we analyze the approximation ratio of the RLNC technique:
RLNC technique is at most a $2-$approximation algorithm of $\Dmin$.
In every RLNC transmission, the sender sends a random-coded packet of all data packets. With high probability (that asymptotically goes to 1 with the field size), after receiving $w_n$ such packets, receiver $\r_n$ can decode all its wanted data packets by performing block decoding, i.e., solving a set of $w_n$ linear equations. Hence, the APDD offered by RLNC is: $$\begin{aligned}
\Dr=\frac{1}{\sum_{n=1}^Nw_n}\sum_{n=1}^Nw_n^2.\end{aligned}$$ Comparing $\Dr$ with the lower bound $\underline D$ in , we have: $$\frac{\Dr}{\Du}=\frac{\frac{\sum_{n=1}^Nw_n^2}{\sum_{n=1}^Nw_n}}{\frac{\sum_{n=1}^Nw_n^2}{2\sum_{n=1}^Nw_n}+\frac{1}{2}}< 2.$$ Since $\Dmin\geqslant \Du$, $\Dr$ at most doubles $\Dmin$. Thus, RLNC is at most a 2-approximation algorithm of $\Dmin$.
Therefore, RLNC technique offers guaranteed APDD performance. On the other hand, to the best of our knowledge existing opportunistic APDD-reduction techniques are not able to provide provable performance guarantees. For example, let us analyze a well-known APDD-reduction technique called instantly decodable network coding (IDNC).
IDNC has two variations, strict IDNC (S-IDNC) [@Rozner_Heuristic_clique; @sundararajan:sadeghi:medard:2009; @yu:parastoo:neda:2014] and general IDNC (G-IDNC) [@sorour:valaee:2010]. Both of them have been shown to provide lower APDD than RLNC with a small number of receivers, but become worse than RLNC with increasing number of receivers. Due to the absence of the optimal G-IDNC algorithm [@sorour:valaee:2010], we are not able to prove whether G-IDNC approximates $\Dmin$ or not. However, we are able to prove the following statement for S-IDNC:
\[theo:idnc\] S-IDNC does not provide a constant approximation ratio for the minimum APDD problem.
To prove this, it suffices to provide a counter example. Consider a complete graph $\G(\V,\E)$ with $K$ vertices and $K(K-1)/2$ edges. For every vertex $\v_k$ we generate a data packet $\p_k$. For every edge $\e_{i,j}$ that connects $\v_i$ and $\v_j$ we generate a receiver $\r_n$ with $\w_n=\{\p_i,\p_j\}$. In the resultant $\mA$, every receiver wants two data packets.
S-IDNC prohibits to code together any two data packets that are both wanted by any receiver. In other words, every S-IDNC coding set $\M$ must satisfy $|\M\cap\w_n|\leqslant 1$ for any receiver $\r_n$. Given the above $\mA$, this restriction implies that no data packets can be coded together at all. Hence, all $K$ data packets must be broadcast uncoded alone. The resultant APDD is $(K+1)/2$. Note that it is easy to show that the optimal value of APDD is at most 2. Indeed, the value of 2 can be achieved by using the RLNC technique. Thus, S-IDNC fails to provide a constant approximation ratio for the problem at hand.
In conclusion, in this section we proved that RLNC is at least a $2$-approximation algorithm of $\Dmin$. By setting RLNC as a benchmark, we showed that S-IDNC fails to provide a constant approximation ratio for our problem. Indeed, RLNC is the only existing approximation algorithm, to the best of our knowledge. Therefore, the final question we are interested in is: *How to overtake RLNC*?
How to Overtake RLNC
====================
Imagine a linear NC technique that: 1) is throughput optimal as RLNC (i.e., allows every receiver to increase its DoF by one in every transmission [@keller:drinea:fragouli:2008]); and 2) enables early packet decodings rather than block decodings in RLNC. Such an NC technique offers an APDD lower than RLNC, and thus will be an approximation algorithm of $\Dmin$ with a ratio lower than RLNC. To the best of our knowledge, such NC techniques have not been developed in the literature.
In this section, we propose a methodology for the development of such NC techniques. We first construct a hypergraph $\mH$ that corresponds to a given instance $\mA$. The key idea to guarantee optimal throughput and early packet decodings is to find minimal vertex covers in $\mH$. A vertex cover is a subset $\V_C$ of $\V$ such that $|\V_C\cap\e_n|\geqslant 1$ for every hyperedge. It is minimal if it is not the superset of a smaller vertex cover, implying that $|\V_C\cap\e_n|=1$ for at least one hyperedge. Hence, every receiver wants at least one data packet from $\V_C$, and at least one receiver can instantly decode a wanted data packet from $\V_C$.
Initialize: SFM $\mA$; Construct the hypergraph $\mH(\V,\E)$ of $\mA$ by mapping data packets to vertices and receivers to hyperedges; Find a minimal vertex cover $\V_C$ of $\mH$, and send a random-coded packet of data packets in $\V_C$; Update $\mH$ by removing $\V_C$ from it following some strategy; Send random-coded packets of all data packets until all receivers complete broadcast.
\[alg:dmin\]
The core algorithmic structure of our methodology is sketched in Algorithm \[alg:dmin\]. It generates a solution $\S$ with coding sets $\{\V_C^1,\cdots,\V_C^L,\P,\P,\cdots\}$, where $L$ is the total number of minimal vertex covers found by the algorithm. To achieve optimal throughput, the random-coded packet of every $\{\V_C^l\}_{l=1}^L$ must be able to increase every receiver’s DoF by one. To this end, a proper hypergraph update strategy must be applied. The simplest strategy is to completely remove $\V_C^l$ from $\mH$ before finding $\V_C^{l+1}$. By doing so, all the vertex covers will have empty intersections, and thus serve all the receivers with different data packets. The algorithm stops at the $L$-th round when there is at least one empty hyperedge, after which point, optimal throughput is maintained by sending random-coded packets of all data packets in $\P$, as in RLNC. Hence, the solution $\S$ is throughput optimal as RLNC. Moreover, since the minimal vertex covers enable instant packet decodings, the APDD of $\S$ is better than RLNC.
The design of optimal hypergraph vertex cover algorithms and hypergraph update strategies that minimize APDD is still an open problem. However, regardless of whether optimal or heuristic algorithms/strategies are applied, solutions generated by Algorithm 1 are always throughput optimal, while also providing early packet decodings. Thus, they can approximate $\Dmin$ with ratio smaller than RLNC.
[Fig.\[fig:d\_general\]]{} compares the APDD performance of a simple realization of our methodology with RLNC and a heuristic G-IDNC [@sameh:valaee:globecom:2010]. In this realization, we adopt the aforementioned complete $\V_C$ removal strategy and a heuristic hypergraph vertex cover algorithm, which iteratively adds to $\V_C$ the vertex that 1) is not connected to $\V_C$; and 2) has the highest degree[^1]. To generate the SFM instances, we consider a block of $K=15$ data packets and assume that each receiver wants each data packet randomly with a probability of $0.2$. The number of receivers $N\in[5,100]$. The results show that G-IDNC offers the lowest APDD when $N$ is small, but becomes worse than RLNC when $N>65$. Hence, the heuristic G-IDNC is not an approximation algorithm. Our realization (heur. VC in the figure) always outperforms RLNC. Their gap narrows down with increasing $N$.
To gain a deeper insight into the realization and performance analysis of the proposed methodology, we conduct a case study in the next subsection by considering a special type of SFM.
![The APDD of heuristic VC, RLNC, and G-IDNC when $K=20$.[]{data-label="fig:d_general"}](./D_general){width="\linewidth"}
A Case Study {#a-case-study .unnumbered}
------------
In this subsection, we design a NC technique that approximates the $\Dmin$ of a special type of SFM where every receiver wants two data packets from the packet block $\P$. Though seemingly simple, this case is highly nontrivial, because its $\Dmin$ cannot be approximated by existing APDD-reduction techniques such as S-IDNC, as shown in the proof of Theorem \[theo:idnc\].
Given such an SFM $\mA$, we first construct its hypergraph model by mapping data packets into vertices, and mapping receivers into hyperedges. Note that multiple receivers who want the same set of data packets are represented by one hyperedge. In addition, we weight every vertex $\v_k$ with a value of $t_k$, which is the number of receivers who want $\p_k$. We note that since every hyperedge has $|\e|=2$, the resultant hypergraph is indeed a classic graph $\G(\V,\E)$.
We then partition $\V$ into two subsets:
- The first subset is a minimal vertex cover $\V_C$. Since every edge is incident to $\V_C$, every receiver wants at least one data packet from $\V_C$. Denote by $\R_C$ the receivers who want two data packets from $\V_C$, and by $N_C$ their number;
- The second subset is a set $\V_I=\V\setminus\V_C$. It is obvious that $\V_I$ is a maximal independent set, because it contains no edge (otherwise $\V_C$ is not a minimal vertex cover). Hence, every receiver wants at most one data packet from $\V_I$. We denote by $\R_I$ the set of receivers who want one data packet from $\V_I$, and by $N_I$ their number. We have $N_C+N_I=N$.
An example of such partition is demonstrated in [Fig.\[fig:mis\_example\]]{}. It has 4 data packets and 5 receivers, with 4 of them want one data packet from both $\V_I$ and $\V_C$.
![An example of MIS partition.[]{data-label="fig:mis_example"}](./MIS_example){width="0.3\linewidth"}
We then send the following two NC packets:
- In the first transmission, send an RLNC packet of all data packets in $\V_C$. This allows $\R_I$ to decode one data packet, and allows $\R_C$ to increase DoF by one without decoding;
- In the second transmission, send an RLNC packet of all data packets in $\V$. This allows $\R_I$ to decode the other data packet, and allows $\R_C$ to decode two data packets.
We call this technique maximal independent set (MIS) technique. MIS follows Algorithm \[alg:dmin\], and thus is an approximation algorithm of $\Dmin$. Its APDD, denoted by $D_{\mathrm{MIS}}$, is calculated as: $$D_{\mathrm{MIS}}=\frac{N_I \cdot 1+ N_I\cdot2 +2N_C\cdot2}{2N}=2-\frac{N_I}{2N}$$ which is minimized when $N_I$ is maximized. Since $N_I=\sum_{\v_k\in\V_I}t_k$, we need to find the maximum weighted independent set $\V_I$, which is NP-hard [@Graph_theory]. Nevertheless, even a heuristically finding $\V_I$ can offer $N_I>0$. Hence, $D_{\mathrm{MIS}}<2$ regardless of the way $\V_I$ is found.
We now derive the worst approximation rate of MIS by calculating an upper bound on $D_{\mathrm{MIS}}$. The minimum size of $\V_I$ is one, taking place when $\G$ is complete. In this case, the optimal MIS will find the solo vertex with the largest weight. Hence, $N_I/N$ is minimized when all vertices have the same weight. In this case, we have $\frac{N_I}{N}=\frac{2}{K}$. $D_{\mathrm{MIS}}$ is thus upper bounded as $D_{\mathrm{MIS}}\leqslant2-\frac{1}{K}$. Then, by noting that $\Du=1.5$ when every receiver wants 2 data packets, we conclude that:
MIS is at most a $\frac{4-2/K}{3}$-approximation algorithm of $\Dmin$ when every receiver wants 2 out of $K$ packets.
![The APDD of the proposed technique, RLNC, and G-IDNC when $K=20$ and $w=2$.[]{data-label="fig:d_compare"}](./D_compare){width="\linewidth"}
[Fig.\[fig:d\_compare\]]{} compares the simulated APDD performance of MIS with RLNC and the heuristic G-IDNC [@sameh:valaee:globecom:2010]. The packet block size is $K=20$. The number of receivers $N\in[5,100]$. Every receiver randomly chooses two wanted data packets. Since $K=20$, $D_{\mathrm{MIS}}$ is upper bounded by $2-\frac{1}{K}=1.95$. The optimal $D_{\mathrm{MIS}}$ is obtained by exhaustively searching the maximum weighted independent set. Both the performance of heuristic $D_{\mathrm{MIS}}$ and heuristic G-IDNC are obtained by using the heuristic maximum weighted clique (a complete subgraph of a graph) search algorithm proposed in [@sameh:valaee:globecom:2010]. This algorithm can be adapted for MIS because an independent set of $\G$ is a clique of the complementary graph $\overline\G$. According to the results, both the optimal and heuristic $D_{\mathrm{MIS}}$ are well below their upper bound, and are much better than both G-IDNC and RLNC. On the other hand, the APDD of G-IDNC exceeds RLNC when the number of receivers becomes large.
Conclusion
==========
In this paper, we proved that it is NP-hard to minimize the average packet decoding delay (APDD) in packet block based wireless broadcast using linear network coding. But the minimum APDD can be approximated by RLNC with a ratio of at most 2. In order to achieve a lower approximation rario, we proposed a methodology for the design of specialized approximation algorithms that always outperform RLNC.
In the future, we are interested in designing more sophisticated realizations of the proposed NC framework. We are also interested in its extension to more general network settings, for example, when NC transmissions are subject to erasures. Besides, our hypergraph model and delay analysis may be extended to other network models such as cooperative data exchange and distributed data storage, because they also have similar types of demands on data packets.
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
T. Ho, M. M[é]{}dard, R. Koetter, D. Karger, M. Effros, J. Shi, and B. Leong, “[A random linear network coding approach to multicast]{},” *[IEEE]{} Trans. Inf. Theory*, vol. 52, no. 10, pp. 4413–4430, 2006.
M. Nistor, D. E. Lucani, T. T. V. Vinhoza, R. A. Costa, and J. Barros, “On the delay distribution of random linear network coding,” *[IEEE]{} J. Sel. Areas Commun.*, vol. 29, no. 5, pp. 1084–1093, May 2011.
S. Katti, H. Rahul, W. Hu, D. Katabi, M. M[é]{}dard, and J. Crowcroft, “[XORs in the air: practical wireless network coding]{},” *[IEEE/ACM]{} Trans. Netw.*, vol. 16, no. 3, pp. 497–510, 2008.
E. Rozner, A. P. Iyer, Y. Mehta, L. Qiu, and M. Jafry, “[ER]{}: Efficient retransmission scheme for wireless [LAN]{}s,” in *Proc. ACM CoNEXT*, 2007.
R. Costa, D. Munaretto, J. Widmer, and J. Barros, “Informed network coding for minimum decoding delay,” in *Proc. IEEE Int. Conf. on Mobile Ad Hoc and Sensor System, (MASS)*, 2008, pp. 80–91.
L. Keller, E. Drinea, and C. Fragouli, “[Online broadcasting with network coding]{},” in *Proc. Workshop on Network Coding, Theory and Applications (NetCod)*, 2008, pp. 1–6.
A. Eryilmaz, A. Ozdaglar, M. M[é]{}dard, and E. Ahmed, “On the delay and throughput gains of coding in unreliable networks,” *[IEEE]{} Trans. Inf. Theory*, vol. 54, no. 12, pp. 5511–5524, 2008.
J. Barros, R. A. Costa, D. Munaretto, and J. Widmer, “Effective delay control in online network coding,” in *Proc. IEEE Conf. on Comput. Commun. (INFOCOM)*, Apr. 2009, pp. 208–216.
J. K. Sundararajan, P. Sadeghi, and M. M[é]{}dard, “[A feedback-based adaptive broadcast coding scheme for reducing in-order delivery delay]{},” in *Proc. 5th Workshop on Network Coding, Theory, and Applications (NetCod)*, 2009.
D. Nguyen, T. Tran, T. Nguyen, and B. Bose, “Wireless broadcast using network coding,” *[IEEE]{} Trans. Veh. Technol.*, vol. 58, no. 2, pp. 914–925, Feb. 2009.
P. Sadeghi, R. Shams, and D. Traskov, “An optimal adaptive network coding scheme for minimizing decoding delay in broadcast erasure channels,” *EURASIP J. on Wireless Commun. and Netw.*, pp. 1–14, Jan. 2010.
S. Sorour and S. Valaee, “On minimizing broadcast completion delay for instantly decodable network coding,” in *Proc. IEEE Int. Conf. Communications (ICC)*, May 2010, pp. 1–5.
X. Li, C.-C. Wang, and X. Lin, “On the capacity of immediately-decodable coding schemes for wireless stored-video broadcast with hard deadline constraints,” *IEEE J. Sel. Areas Commun.*, vol. 29, no. 5, pp. 1094–1105, 2011.
S. Athanasiadou, M. Gatzianas, L. Georgiadis, and L. Tassiulas, “Stable and capacity achieving xor-based policies for the broadcast erasure channel with feedback,” in *Proc. IEEE Int. Symp. Information Theory (ISIT)*, 2013, pp. 2905–2909.
M. Yu, P. Sadeghi, and N. Aboutorab, “From instantly decodable to random linear network coding,” *[IEEE]{} Trans. Commun.*, vol. 62, no. 11, pp. 3943–3955, Oct. 2014.
P. Sadeghi, M. Yu, and N. Aboutorab, “On throughput-delay tradeoff of network coding for wireless communications,” in *Proc. IEEE Int. Symp. Information Theory and its Applications (ISITA)*, 2014, pp. 689–693.
A. Fu, P. Sadeghi, and M. M[é]{}dard, “Dynamic rate adaptation for improved throughput and delay in wireless network coded broadcast,” *[IEEE/ACM]{} Trans. Netw.*, vol. 22, no. 6, pp. 1715–1728, 2014.
G. Agnarsson and M. M. Halld[ó]{}rsson, “Strong colorings of hypergraphs,” in *Approximation and Online Algorithms*.1em plus 0.5em minus 0.4emSpringer, 2005, pp. 253–266.
P. Erd[ő]{}s, “On the combinatorial problems which [I]{} would most like to see solved,” *Combinatorica*, vol. 1, no. 1, pp. 25–42, 1981.
S. Sorour and S. Valaee, “Minimum broadcast decoding delay for generalized instantly decodable network coding,” in *Proc. IEEE Global Communications Conference (GLOBECOM)*, Dec. 2010, pp. 1–5.
J. M. Harris, J. L. Hirst, and M. J. Mossinghoff, *Combinatorics and Graph Theory, 2nd Edition*.1em plus 0.5em minus 0.4emSpringer Press, 2008.
[^1]: The degree of a vertex is the number of hyperedges incident to it
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the distribution of the resonances near spectral thresholds of Laplace operators on regular tree graphs with $k$-fold branching, $k \geq 1$, perturbed by nonself-adjoint exponentially decaying potentials. We establish results on the absence of resonances which in particular involve absence of discrete spectrum near some sectors of the essential spectrum of the operators.'
address:
- 'Departamento de Matemáticas, Facultad de Matemáticas, Pontificia Universidad Católica de Chile, Vicuña Mackenna 4860, Santiago de Chile'
- 'Departamento de Matemáticas, Facultad de Matemáticas, Pontificia Universidad Católica de Chile, Vicuña Mackenna 4860, Santiago de Chile'
- 'Departamento de Matemáticas, Facultad de Matemáticas, Pontificia Universidad Católica de Chile, Vicuña Mackenna 4860, Santiago de Chile'
author:
- 'Olivier <span style="font-variant:small-caps;">Bourget</span>'
- 'Diomba <span style="font-variant:small-caps;">Sambou</span>'
- 'Amal <span style="font-variant:small-caps;">Taarabt</span>'
title: Resonances on regular tree graphs
---
Introduction {#s1}
============
A great interest has been focused in the last decades on spectral analysis of Laplace operators on regular trees. This includes local perturbations [@al; @af], random settings [@kl; @aw; @ai; @fr; @fr1; @sh] (see also the references therein), and quantum ergodicity regimes [@am]. For complementary results, we refer the reader to the papers [@br; @bk; @ro1; @ro2; @rr], and for the relationships between the Laplace operator on trees and quantum graphs, see [@km]. However, it seems that resonances have not been systematically studied in the context of (regular) trees.
In this paper, we use resonance methods to obtain better understanding of local spectral properties for perturbed Schrödinger operators on regular tree graphs with $k$-fold branching, $k \geq 1$, as we describe below (cf. Section \[secd\]). Our techniques are similar to those used in [@bbr1; @bbr2] (and references therein), where self-adjoint perturbations are considered. Actually, these methods can be extended to nonself-adjoint models, see for instance [@sa]. Here, we are focused on some nonself-adjoint perturbations of the Laplace operator on regular tree graphs. In particular, we shall derive as a by-product, a description of the eigenvalues distribution near the spectral thresholds of the operator.
Since a nonself-adjoint framework is involved in this article, it is convenient to clarify the different notions of spectra we use. Let $T$ be a closed linear operator acting on a separable Hilbert space $\mathscr{H}$, and $z$ be an isolated point of $\sigma(T)$ the spectrum of $T$. If $\gamma$ is a small contour positively oriented containing $z$ as the only point of $\sigma(T)$, we recall that the Riesz projection $P_z$ associated to $z$ is defined by $$\label{eq1,8}
P_z := \frac{1}{2i\pi} \oint_{\gamma} (T - \zeta)^{-1} d\zeta.$$ The algebraic multiplicity of $z$ is then defined by $$\label{eq1,9}
\textup{m}(z) := \text{rank} (P_z),$$ and when it is finite, the point $z$ is called a discrete eigenvalue of the operator $T$. Note that we have the inequality $\mathrm{m} \, (z) \geq \mathrm{dim} \, \big( \text{Ker}
(T - z) \big)$, which is the geometric multiplicity of $z$. The equality holds if $T = T^\ast$. So, we define the discrete spectrum of $T$ as $$\label{eq1,10}
\sigma_{\text{\bf disc}}(T) := \big\lbrace z \in \sigma(T) : z \hspace*{0.1cm}
\textup{is a discrete eigenvalue of $T$} \big\rbrace.$$ We recall that if a closed linear operator has a closed range and both its kernel and cokernel are finite-dimensional, then it is called a Fredholm operator. Hence, we define the essential spectrum of $T$ as $$\label{eq1,11}
\sigma_{\text{\bf ess}}(T) := \big\lbrace z \in {{\mathbb{C}}}: \textup{$T - z$ \textup{is not a
Fredholm operator}} \big\rbrace.$$ Note that $\sigma_{\text{\bf ess}}(T)$ is a closed subset of $\sigma(T)$.
The paper is organized as follows. In Section \[secd\], we present our model. In section \[secp\], we state our main results Theorem \[t1\] and Corollaries \[t2\], \[t3\]. Section \[sec3\] is devoted to preliminary results we need due to Allard and Froese. In Section \[sec4\], we establish a formula giving a kernel representation of the resolvent associated to the operator we consider and which is crucial for our analysis. In Section \[sec5\], we define and characterize the resonances near the spectral thresholds, while in Section \[sec6\] we give the proof of our main results. Section \[sa\] gathers useful tools on the characteristic values concept of finite meromorphic operator-valued functions.
Presentation of the model {#secd}
=========================
We consider an infinite graph $\mathbb{G} = (\mathcal{V},\mathcal{E})$ with vertices $\mathcal{V}$ and edges $\mathcal{E}$, and we let $\ell^2(\mathcal{V})$ be the Hilbert space $$\label{eq1,0}
\ell^2(\mathcal{V}) := \bigg\lbrace \phi : \mathcal{V} \rightarrow {{\mathbb{C}}}:
\Vert \phi \Vert^2 := \sum_{v \in \mathcal{V}} \big\vert \phi(v)
\big\vert^2 < \infty \bigg\rbrace,$$ with the inner product $$\label{eq,ps}
\langle \phi,\psi \rangle := \sum_{v \in \mathcal{V}} \phi(v) \overline{\psi}(v).$$ On $\ell^2(\mathcal{V})$, we consider the symmetric Schrödinger operator ${{-\Delta}}$ defined by $$\label{eq1,1}
({{-\Delta}}\phi)(v) := - \sum_{w \, : \, w \sim v} \Big( \phi(w) - \phi(v) \Big),$$ where $w \sim v$ means that the vertices $w$ and $v$ are connected by an edge. If we define on $\ell^2(\mathcal{V})$ the symmetric operator $L$ by $$\label{eq1,2}
(L \phi)(v) := \sum_{w \, : \, w \sim v} \phi(w),$$ then it is not difficult to see that the operator ${{-\Delta}}$ can be written as $$\label{eq1,3}
{{-\Delta}}= -L + d,$$ where $d$ is the multiplication operator by the function (also) noted $d : \mathcal{V} \rightarrow {{\mathbb{C}}}$, with $d(v)$ denoting the number of edges connected with the vertex $v$. Note that when $d$ is bounded, then so is the symmetric operators ${{-\Delta}}$ and $L$, hence self-adjoint. In a regular rooted tree graph with $k$-fold branching, $k \ge 1$, (see Figure 2.1 for a binary tree graph), we have $d = k + 1 - d_0$ with $$d_0(v) = \begin{cases}
1 & \text{if } v \text{ coincides with the root of the tree}, \\
0 & \text{otherwise}.
\end{cases}$$ This is the same model described in [@af] and we refer to this paper for more details. In , $d$ can be viewed as a perturbation of the operator $L$. It is well know (see Lemma \[l3,2\]) that the spectrum of the operator $L$ is absolutely continuous, coincides with the essential spectrum and is equal to $$\label{eq1,4}
\sigma (L) = \sigma_{\textup{ac}} (L) = \sigma_{\textup{ess}} (L) =
\left[ -2\sqrt{k},2\sqrt{k} \right].$$
On $\ell^2(\mathcal{V})$, we define the perturbed operator $$\label{eq1,5}
{{-\Delta}}_M := {{-\Delta}}+ M,$$ where $M$ is identified with the multiplication operator by the bounded potential function (also) noted $M$. In a regular rooted tree graph with $k$-fold branching, according to above, the operator ${{-\Delta}}_M$ can be written as $$\label{eq1,6}
{{-\Delta}}_M = -L + k + 1 - d_0 + M.$$ In , the degree term $d_0$ can be included in the potential perturbation so that ${{-\Delta}}_M$ can be viewed as a perturbation of the operator $-L + k + 1$. Hence, from now on, the operator ${{-\Delta}}_M$ will be written as $$\label{eq1,60}
{{-\Delta}}_{\widetilde M} = -L + k + 1 + \widetilde M \quad {\rm with} \quad
\widetilde M := - d_0 + M.$$ In the sequel, we set $$\label{eq1,61}
t_\pm(k) := \pm 2\sqrt{k} + k + 1,$$ and we shall simply write $t_\pm$ when no confusion can arise. Then, from , it follows that the spectrum of the operator $-L + k + 1$ satisfies $$\label{eq1,7}
\sigma (-L + k + 1) = \sigma_{\textup{ac}} (-L + k + 1) = \sigma_{\textup{ess}} (-L + k + 1) =
\left[ t_-(k),t_+(k) \right],$$ where the $t_\pm(k)$ play the role of thresholds of this spectrum.
Now, let us choose some vertex $v_0 = 0 \in \mathcal{V}$ as the origin of the graph $\mathbb{G} = (\mathcal{V},\mathcal{E})$. For $v \in \mathcal{V}$, we define $\vert v \vert$ as the length of the shortest path connecting $0$ to $v$. Hence, $\vert v \vert$ defines in the graph the distance from $0$ to $v$. For $r > 0$, let $S_r$ be the sphere of radius $r$ in the graph defined by $$S_r := \big\lbrace v \in \mathcal{V} : \vert v \vert = r \big\rbrace.$$
\[fig 1\]
In this case, we have $$\displaystyle \mathcal{V} = \bigsqcup_{0}^\infty S_r,$$ where $\bigsqcup$ means a disjoint union, so that we have $$\ell^2(\mathcal{V}) = \bigoplus_0^\infty \ell^2(S_r).$$ In this paper, we are interested in the case of regular rooted tree graphs with $k$-fold branching, $k \geq 1$. Moreover, the potential $M$ will be assumed to satisfy the following assumption:
[**Assumption (A)**]{}: For $ v \in \mathcal{V}$, we have $$\label{eq1,7}
\vert M(v)\vert \leq {\rm Const.} \, e^{-\delta \vert v \vert}, \quad {\rm with} \quad
\begin{cases}
\delta > 0 & \text{if } \, k = 1, \\
\delta \geq 6 \ln (k) & {\rm otherwise}.
\end{cases}$$
\[r1,1\] We point out that in Assumption (A) above, there is no restriction on the perturbation potential $M$ concerning its self-adjointness or not. The case $k = 1$ includes in particular the case of the Laplacian on $\ell^2({{\mathbb{N}}},{{\mathbb{C}}})$ without any boundary condition at $0$.
As mentioned above, in this article we investigate the resonances (or eigenvalues) distribution for the operator ${{-\Delta}}_M$ near the spectral thresholds $t_\pm(k)$ given by . As this will be observed, the work of Allard and Froese [@af] will play an important role in our analysis (cf. Section \[sec3\] for more details). More precisely, in order to establish a suitable representation of the resolvent associated to the operator $-L + k + 1$, $k \geq 1$ (cf. Theorem \[p4,1\]). Under Assumption (A), the perturbation potential $\widetilde{M}$ satisfies the decay assumption of $M$. So, if we let $\Lambda_n$ to denote the orthogonal projection onto $\bigoplus_{r=0}^n \ell^2(S_r)$, then with the aid of the Schur lemma, it can be shown that $$\label{eq,l}
\Vert \widetilde{M} - \widetilde{M} \Lambda_n \Vert \underset{n \rightarrow
\infty}{\longrightarrow} 0.$$ Since $\ell^2(S_r)$ is a $k^r$-dimensional space, then implies that the operator $\widetilde{M}$ is the limit in norm of a sequence of finite rank operators. Therefore, $\widetilde{M}$ is a compact operator and in particular it is relatively compact with respect to the operator $-L + k + 1$. Thus, since the operator $-L + k + 1$ is self-adjoint, then by [@gohs Theorem 2.1, p. 373] we have a disjoint union $$\sigma({{-\Delta}}_{\widetilde M}) = \sigma_{\text{\bf ess}}({{-\Delta}}_{\widetilde M}) \bigsqcup
\sigma_{\text{\bf disc}}({{-\Delta}}_{\widetilde M}).$$ Moreover, Weyl’s criterion on the invariance of the essential spectrum implies that $$\sigma_{\textup{\bf ess}} ({{-\Delta}}_{\widetilde M}) = \sigma_{\textup{\bf ess}}
(-L + k + 1) = \left[ t_-(k),t_+(k) \right].$$ However, the (complex) discrete spectrum $\sigma_{\text{\bf disc}}({{-\Delta}}_{\widetilde M})$ generated by the potential $\widetilde{M}$ can only accumulate at the points of $\sigma_{\textup{\bf ess}} ({{-\Delta}}_{\widetilde M})$.
When ${\widetilde M} = {\widetilde M}^\ast$, $\sigma_{\textup{\bf disc}} ({{-\Delta}}_{\widetilde M})$ is just the set of real eigenvalues of ${{-\Delta}}_{\widetilde M}$ respectively from the right and the left of $t_\pm(k)$.
Exploiting the exponential decay of the potential $\widetilde{M}$, we extend (cf. Section \[sec5\]) meromorphically in Banach weighted spaces the resolvent of the operator ${{-\Delta}}_{\widetilde M}$ near $z = t_\pm(k)$, in some two sheets Riemann surfaces $\mathcal{M}_{t_\pm}$ respectively. The first main difficulty to overcome is to establish a good representation of the kernel of the resolvent associated to the operator $-L + k + 1$, $k \ge 1$ (cf. Section \[sec4\]). We thus define the resonances of the operator ${{-\Delta}}_{\widetilde M}$ near $z = t_\pm(k)$ as the poles of the above meromorphic extensions. Notice that this set of resonances contains the eigenvalues of the operator ${{-\Delta}}_{\widetilde M}$ localized near the spectral thresholds $t_\pm(k)$. Otherwise, in the two sheets Riemann surfaces $\mathcal{M}_{t_\pm}$, the resonances will be parametrized by $z_{t_\pm}(\lambda)$ for $\lambda$ sufficiently small for technical reasons. Furthermore, the point $\lambda = 0$ corresponds to the threshold $z = t_\pm(k)$ (cf. Section \[sec5\] for more details). Actually, the resonances verifying $$z_{t_\pm}(\lambda) \in \mathcal{M}_{t_\pm}, \quad {\operatorname{Im}}(\lambda) < 0,$$ live in the non physical plane while those verifying $$z_{t_\pm}(\lambda) \in \mathcal{M}_{t_\pm}, \quad {\operatorname{Im}}(\lambda) \geq 0,$$ coincide with the discrete and the embedded eigenvalues of the operator ${{-\Delta}}_{\widetilde M}$ near $t_\pm$ and are localized in the physical plane. We state Theorem \[t1\] where we establish an absence of resonances of the operator ${{-\Delta}}_{\widetilde M}$ near the spectral thresholds $t_\pm(k)$. In particular, this implies results on the absence of discrete spectrum and embedded eigenvalues near $t_\pm(k)$ (cf. Corollaries \[t2\] and \[t3\]). To prove these results, we first reduce the analysis of resonances near the thresholds $t_\pm(k)$ to that of the noninvertibility of some nonself-adjoint compact operators near $\lambda = 0$ (cf. Propositions \[p5,2\] and \[p5,3\]). This can be seen as a Birman-Schwinger principle in a nonself-adjoint context. Afterwards, the reduction made on the problem is reformulated in terms of characteristic values problems (cf. Propositions \[p6,2\] and \[p6,3\]). This allows us to apply powerful results (cf. Section \[sa\]) on the theory of characteristic values of finite meromorphic operator-valued functions to conclude.
Statement of the main results {#secp}
=============================
Let us first fix some notations. If $\lambda \in {{\mathbb{C}}}$, as usual $\vert \lambda \vert <\!\!<1$ means that $\lambda$ is chosen small enough. The set of resonances of the operator ${{-\Delta}}_{\widetilde M}$ near the spectral thresholds $t_\pm(k)$ given by will be respectively denoted by $${\rm Res}_{t_\pm} \, ({{-\Delta}}_{\widetilde M}).$$ We also recall that near $z = t_\pm(k)$, the resonances are defined in some Riemann surfaces $\mathcal{M}_{t_\pm}$ and coincide with $z_{t_\pm}(\lambda)$, $0 < \vert \lambda \vert <\!\!<1 $. More precisely, they are parametrized respectively by $$\label{p1}
z_{t_\pm}(\lambda) := t_\pm(k) \mp \lambda^2\sqrt{k} \in \mathcal{M}_{t_\pm}.$$ Furthermore, the embedded eigenvalues and the discrete spectrum of the operator ${{-\Delta}}_{\widetilde M}$ near $t_\pm(k)$ are the resonances $z_{t_\pm}(\lambda) \in \mathcal{M}_{t_\pm}$ with ${\operatorname{Im}}(\lambda) \geq 0$.
Now, for $0 < r <\!\!<1$, let us introduce the punctured neighborhood of $\lambda = 0$ $$\Omega_r^\ast := \big \lbrace \lambda \in {{\mathbb{C}}}: 0 < \vert \lambda \vert < r \big \rbrace.$$ We then can state our first main result that gives an absence of resonances of the operator ${{-\Delta}}_{\widetilde M}$ near the thresholds $t_\pm(k)$, in small domains of the form $t_\pm(k) \mp \sqrt{k} \, {\Omega_r^\ast}^2$.
\[t1\] Assume that the potential $M$ satisfies Assumption (A). Then, for any $r > 0$ small enough and any punctured neighborhood $\Omega_r^\ast$, we have $$\# \big \lbrace z_{t_\pm}(\lambda) \in Res_{t_\pm}({{-\Delta}}_{\widetilde M}) : \lambda \in
\Omega_r^\ast \big \rbrace = 0,$$ the resonances being counted with their multiplicity given by and .
Notice that Theorem \[t1\] just says that the operator ${{-\Delta}}_{\widetilde M}$ has no resonances in a punctured neighborhood of $t_\pm(k)$ in the two-sheets Riemann surfaces $\mathcal{M}_{t_\pm}$ where they are defined.
\[fig 1\]
=\[fill=gray!60\] =\[fill=white!100\] =\[fill=gray!20\]
(0,0) – (180:1.58) arc (180:360:1.58) – cycle; (0,0) – (0:1.58) arc (0:180:1.58) – cycle; (0,0) circle (0.5); (-2.5,0) – (2.3,0); (2.3,0) node\[right\] ; (0,-2.5) – (0,2.7); (0.05,2.7) node\[right\] ; (0.26,-0.40) node\[above\] ; (0,0) – (0.24,-0.45); (0,0) – (1.31,0.9); (0.8,0.55) node\[above\] ;
at (-1.4,-1.2) ;
at (0.9,-1.4) ; at (1.8,-0.2) ; at (1.4,-1.2) ;
at (0.9,1.4) ; at (1.4,1.2) ;
at (-1.7,0.4) ; at (-1.4,1.2) ;
at (-0.5,1.7) ;
at (-0.2,1.7) ;
at (2,2.3) ; (1.8,2.15) – (-0.5,1); (1.8,2.15) – (1,-0.5);
at (-4.5,0.5) ; at (-4.5,0.2) ; (-2.1,0.5) – (-1,0.5);
at (-4.67,-0.5) ; at (-4.55,-0.8) ; (-1.8,-0.5) – (-1,-0.5);
Since near $t_\pm(k)$ the discrete spectrum of the operator ${{-\Delta}}_{\widetilde M}$ corresponds to resonance points $z_{t_\pm}(\lambda) \in \mathcal{M}_{t_\pm}$ with ${\operatorname{Im}}(\lambda) \geq 0$, then a first consequence of Theorem \[t1\] is the following result giving a non cluster phenomena of real or non real eigenvalues near $t_\pm(k)$.
\[t2\] Assume that the potential $M$ satisfies Assumption (A). Then, there is no sequence $(\nu_j)_j$ of non real or real eigenvalues of the operator ${{-\Delta}}_{\widetilde M}$ accumulating at $t_\pm(k)$.
Now, thanks to the parametrizations , the embedded eigenvalues of the operator ${{-\Delta}}_{\widetilde M}$ near $t_\pm(k)$ respectively from the left and the right are the resonances $z_{t_\pm}(\lambda) = t_\pm(k) \mp \lambda^2\sqrt{k}$ with $\lambda \in {{\mathbb{R}}}_+$ sufficiently small. Therefore, as a second consequence of Theorem \[t1\] together with [@af Theorem 9], we have the following:
\[t3\] Assume that the potential $M$ satisfies Assumption (A). Then, for any $r > 0$ small enough, the operator ${{-\Delta}}_{\widetilde M}$ has no embedded eigenvalues in $$\left( t_-(k),t_-(k) + r^2 \right) \cup \left( t_+(k) - r^2,t_+(k) \right).$$ In particular, for $M = M^\ast$, the set of embedded eigenvalues of the operator ${{-\Delta}}_{\widetilde M}$ in $\left( t_-(k),t_+(k) \right)$ is finite.
On a diagonalization of the operator $L$ for a regular tree graph {#sec3}
=================================================================
In this section, we summarize some results and tools we need and which are developed in [@af; @al]. We shall essentially follow [@af Section 3] and we refer to the cited papers for more details.
Define the operator $\Pi$ on $\ell^2(\mathcal{V})$ by $$\label{eq3,1}
(\Pi \phi)(v) := \sum_{w \, : \, w \rightarrow v} \phi(w),$$ where for two vertices $v$ and $w$, $v \rightarrow w$ means that they are connected by an edge with $\vert w \vert = \vert v \vert + 1$. Using the inner product defined on the Hilbert space $\ell^2(\mathcal{V})$ by , it can be easily checked that the adjoint operator $\Pi^\ast$ is given by $$\label{eq3,2}
(\Pi^\ast \phi)(v) := \sum_{w \, : \, v \rightarrow w} \phi(w).$$ If we let $L_S$ be the spherical Laplacian defined on $\ell^2(\mathcal{V})$ by $$\label{eq3,3}
(L_S \phi)(v) := \sum_{\substack{w \, : \, w \sim v \\ \vert w \vert = \vert v \vert}} \phi(w),$$ then the operator $L$ given by can be written as $$\label{eq3,4}
L = \Pi + \Pi^\ast + L_S.$$ In a regular rooted tree graph with $k$-fold branching, since there are no edges connecting vertices within each sphere, then $L_S = 0$ so that $$\label{eq3,5}
L = \Pi + \Pi^\ast.$$ To diagonalize the operator $L$ given by , invariant subspaces $M_n$, $n \geq 0$, for $\Pi$ are firstly constructed in [@af]. More precisely, we have the following lemma:
[[@af Lemma 1]]{}\[l3,1\] The Hilbert space $\ell^2(\mathcal{V})$ can be decomposed as an orthogonal direct sum $$\label{eq3,6}
\ell^2(\mathcal{V}) = \bigoplus_{n = 0}^\infty M_n =
\bigoplus_{n = 0}^\infty \bigoplus_{r = n}^\infty Q_{n,r},$$ where the subspaces $M_n$ are $L$-invariant.
By construction in this Lemma \[l3,1\], we have $Q_{0,0} := \ell^2(S_0)$ and $\ell^2(S_r) = \bigoplus_{\ell = 0}^r Q_{\ell,r}$. A schematic interpretation yields a triangular diagram as in Figure 4.1 below.
\[fig 2\]
According to Lemma \[l3,1\], for any $n \geq 0$, the subspace $M_n$ is invariant for the operator $L$. Thus, $L$ can be decomposed as $$\label{eq3,60}
L = \bigoplus_{n = 0}^{\infty} L_n,$$ the operators $L_n$, $n \geq 0$, being the restriction of $L$ to $M_n$. So, in order to diagonalize the operator $L$, it suffices to do it for each operator $L_n$ for $n \geq 0$. Consider a vector $\phi \in M_n$, i.e. $$\label{eq3,7}
\phi = \displaystyle\oplus_{j = 0}^{\infty} \phi_{n,n+j},$$ with $\phi_{n,n+j} \in Q_{n,n+j}$ for any $j \geq 0$. The idea is to construct an isomorphism between the subspace $M_n$ and the space of $Q_{n,n}$-valued sequences, namely the space $\ell^2({{\mathbb{Z}}}^+,Q_{n,n})$. By construction of the $Q_{n,r}$, $n \geq 0$, $r \geq 0$ (see for instance [@af]), for any $j \geq 0$, the operator $\big( \frac{1}{\sqrt{k}} \Pi \big)^j$ defines an isometry between $Q_{n,n+j}$ and $Q_{n,n}$, and can be written as $$\label{eq3,8}
\phi = \oplus_{j = 0}^{\infty} \left( \frac{1}{\sqrt{k}} \Pi \right)^j \chi_{n + j},$$ where $\chi := (\chi_{n + j})_{j \geq 0}$ defines a sequence of vectors lying in $Q_{n,n}$. Therefore, under the above considerations, the operator $$\label{eq3,9}
W : M_n \longrightarrow \ell^2({{\mathbb{Z}}}^+,Q_{n,n}), \quad W \phi = \chi,$$ defines an isomorphism between the spaces $M_n$ and $\ell^2({{\mathbb{Z}}}^+,Q_{n,n})$. Indeed, we have $$\label{eq3,90}
\langle \phi,\phi \rangle = \sum_{j = 0}^{\infty} \langle \phi_{n + j},\phi_{n + j}
\rangle = \sum_{j = 0}^{\infty} \langle \chi_{n + j},\chi_{n + j}
\rangle = \langle \chi,\chi \rangle_{\ell^2({{\mathbb{Z}}}^+,Q_{n,n})}.$$ Now, let ${{\mathbb{T}}}:= {{\mathbb{R}}}/2\pi {{\mathbb{Z}}}$ be the torus and define the unitary operator $$\label{eq3,10}
U : M_n \cong \ell^2({{\mathbb{Z}}}^+,Q_{n,n}) \longrightarrow L^2_{\rm odd} (\mathbb{T},Q_{n,n})$$ acting as $$\label{eq3,11}
U \big( (\chi_n,\chi_{n + 1}, \ldots ) \big) := \frac{1}{\sqrt{\pi}} \sum_{j = 0}^\infty
\chi_{n + j} \sin \big( (j + 1) \theta \big).$$ Notice that the inner product in $L^2_{\rm odd} (\mathbb{T},Q_{n,n})$ is defined by $$\label{eq3,12}
\langle f,g \rangle_{L^2_{\rm odd}} := \int_{\mathbb{T}} \langle f(\theta),g(\theta)
\rangle \, d\theta.$$ Hence, a direct computation shows that $$\label{eq3,13}
\langle U \chi,U \chi \rangle_{L^2_{\rm odd}} =
\langle \chi,\chi \rangle_{\ell^2({{\mathbb{Z}}}^+,Q_{n,n})} = \langle \phi,\phi \rangle,$$ where the last equality corresponds to . Moreover, we have the following lemma:
[[@af Lemma 2]]{}\[l3,2\] For any $n \geq 0$, we have $$\label{eq3,14}
U L_n U^\ast = 2\sqrt{k} \cos (\theta).$$
In particular, Lemma \[l3,2\] shows that for any $n \geq 0$, the spectrum of the operator $L_n$ is equal to $\big[ -2\sqrt{k},2\sqrt{k} \big]$ and is absolutely continuous. Hence, this implies .
Representation of the weighted resolvent $A (-L + k + 1 - z)^{-1} B^\ast$ {#sec4}
=========================================================================
In this section, we give a suitable representation of the weighted resolvent $A (-L + k + 1 - z)^{-1} B^\ast$ which turns to be useful in our analysis, where $A$ and $B$ are bounded operators on $\ell^2(\mathcal{V})$.
For $n \geq 0$, let $P_n$ be the orthogonal projection of $\ell^2(\mathcal{V})$ onto $M_n$, the subspace defined in Lemma \[l3,1\]. Since $M_n$ can be decomposed as $$\label{eq4,1}
M_n = \bigoplus_{j = 0}^{\infty} Q_{n,n + j},$$ then if we let $(E_m^{n,n+j})_{0 \leq m \leq N_j}$ denote an orthonormal basis of the finite-dimensional space $Q_{n,n + j}$ for any $j$ fixed, we have $$\label{eq4,2}
P_n = \sum_{j \, \geq \, 0} \sum_{0 \leq m \leq N_j} \langle \cdot,E_m^{n,n+j} \rangle
E_m^{n,n+j}.$$ Notice that for any $j \geq 0$ fixed, we have $$\label{eq4,3}
1 + N_j = \dim \, Q_{n,n+j} < k^{n + j} = \dim \, \ell^2(S_{n + j}),$$ since $Q_{n,n+j} \subset \ell^2(S_{n + j})$. Furthermore, according to Section \[sec3\], for any $0 \leq m \leq N_j$ fixed, there exists a unique vector $\chi_m^{n,n+j} \in Q_{n,n}$ such that $$\label{eq4,4}
E_m^{n,n+j} = \left( \frac{1}{\sqrt{k}} \Pi \right)^j \chi_m^{n,n + j}.$$ Our goal in this section is to prove the following result:
\[p4,1\] Let $A$ and $B$ be two bounded operators on $\ell^2(\mathcal{V})$. Then, for any $z$ in the resolvent set of the operator $-L + k + 1$ and any $\varphi \in \ell^2(\mathcal{V})$, we have $$\label{eq4,5}
A (-L + k + 1 - z)^{-1} B^\ast
\varphi (v) = \sum_{v' \in \mathcal{V}} K(v,v') \varphi (v'),$$ where the kernel $K(v,v')$ is given by $$\label{eq4,6}
\begin{split}
K(v,v') := & \frac{1}{2 \sqrt{k}} \sqrt{\frac{2}{\pi}} \sum_{n \, \geq \, 0}
\sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\overline{ \Big( B E_m^{n,n+j} \Big) (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \times \left( - \frac{i e^{i(j + \ell + 2)\Phi}}{\sqrt{u}
\sqrt{4 - u}} + \frac{i e^{i\vert j - \ell \vert\Phi}}{\sqrt{u} \sqrt{4 - u}} \right)
\Big( A E_q^{n,n + \ell} \Big) (v) ,
\end{split}$$ throughout the double change of variables $$\label{eq4,60}
\frac{z + 2\sqrt{k} - (k + 1)}{\sqrt{k}} = u = 4 \sin^2 \left( \frac{\Phi}{2} \right),
\quad {\operatorname{Im}}(\Phi) > 0.$$
Let $\varphi \in \ell^2(\mathcal{V})$ and $z \in \rho (-L + k + 1)$ the resolvent set of the operator $-L + k + 1$. Thanks to , we have $$\label{eq4,7}
\begin{split}
(-L + k + 1 - z)^{-1} & = \sum_{n \, \geq \, 0} P_n (-L_n + k + 1 - z)^{-1} P_n \\
& = \sum_{n \, \geq \, 0} P_n U^\ast U (-L_n + k + 1 - z)^{-1} U^\ast U P_n,
\end{split}$$ where $U$ is the unitary operator defined by . Thus, for any vector $\psi \in \ell^2(\mathcal{V})$, we have $$\label{eq4,8}
\begin{split}
& \biggl\langle A (-L + k + 1 - z)^{-1} B^\ast
\varphi,\psi \biggr\rangle \\
& = \sum_{n \, \geq \, 0} \biggl\langle U (-L_n + k + 1 - z)^{-1} U^\ast U P_n
B^\ast \varphi,U P_n A^\ast \psi
\biggr\rangle_{L^2_{\rm odd}},
\end{split}$$ $\langle \cdot,\cdot \rangle_{L^2_{\rm odd}}$ being the inner product defined by . Together with Lemma \[l3,2\], this gives $$\label{eq4,9}
\begin{split}
& \biggl\langle A (-L + k + 1 - z)^{-1} B^\ast
\varphi,\psi \biggr\rangle \\
& = \sum_{n \, \geq \, 0} \int_{\mathbb{T}} \big( -2\sqrt{k} \cos (\theta) + k + 1 - z \big)^{-1}
\biggl\langle U P_n B^\ast \varphi (\theta),U P_n
A^\ast \psi (\theta) \biggr\rangle \, d\theta.
\end{split}$$ From , it follows that for any $\phi \in \ell^2(\mathcal{V})$ and any bounded operator $W$ on $\ell^2(\mathcal{V})$, we have $$\label{eq4,10}
U P_n W \phi (\theta) = \sum_{j \, \geq \, 0}
\sum_{0 \leq m \leq N_j} \biggl\langle \phi,W^\ast E_m^{n,n+j}
\biggr\rangle U E_m^{n,n+j} (\theta).$$ Then, combining and , we obtain $$\label{eq4,11}
\begin{split}
\biggl\langle & A (-L + k + 1 - z)^{-1} B^\ast
\varphi,\psi \biggr\rangle \\
& = \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\biggl\langle \varphi,B E_m^{n,n+j}
\biggr\rangle \overline{\biggl\langle \psi,A E_q^{n,n+\ell}
\biggr\rangle} \\
& \times \int_{\mathbb{T}} \big( -2\sqrt{k} \cos (\theta) + k + 1 - z \big)^{-1}
\biggl\langle U E_m^{n,n+j} (\theta),U E_q^{n,n+\ell} (\theta)
\biggr\rangle \, d\theta \\
& = \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\biggl\langle \biggl\langle \varphi,B E_m^{n,n+j} \biggr\rangle
A E_q^{n,n+\ell},\psi \biggr\rangle \\
& \times \int_{\mathbb{T}} \big( -2\sqrt{k} \cos (\theta) + k + 1 - z \big)^{-1}
\biggl\langle U E_m^{n,n+j} (\theta),U E_q^{n,n+\ell} (\theta)
\biggr\rangle \, d\theta.
\end{split}$$ According to the construction of the unitary operator $U$ and , for $p = m$, $q$ and $p' = j$, $\ell$ respectively, we have $$U E_p^{n,n+p'} (\theta) = \frac{1}{\sqrt{\pi}} \chi_p^{n,n+p'}
\sin \big( (p' + 1) \theta \big).$$ Putting this together with , we obtain $$\label{eq4,12}
\begin{split}
& \biggl\langle A (-L + k + 1 - z)^{-1} B^\ast
\varphi,\psi \biggr\rangle \\
& = \frac{1}{\pi} \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\biggl\langle \biggl\langle \varphi,B E_m^{n,n+j} \biggr\rangle
\langle \chi_m^{n,n+j},\chi_q^{n,n+\ell} \rangle A E_q^{n,n+\ell},\psi \biggr\rangle \\
& \times \int_{\mathbb{T}} \big( -2\sqrt{k} \cos (\theta) + k + 1 - z \big)^{-1}
\sin \big( (j + 1) \theta \big) \sin \big( (\ell + 1) \theta \big) \, d\theta.
\end{split}$$ Now, implies that the action of the operator $A (-L + k + 1 - z)^{-1} B^\ast$ on $\ell^2(\mathcal{V})$ can be described by $$\label{eq4,13}
\begin{split}
& A (-L + k + 1 - z)^{-1} B^\ast \varphi (v) \\
& = \frac{1}{\pi} \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\biggl\langle \varphi,B E_m^{n,n+j} \biggr\rangle
\Big( A E_q^{n,n+\ell} \Big) (v)
\langle \chi_m^{n,n+j},\chi_q^{n,n+\ell} \rangle \\
& \times \int_{\mathbb{T}} \big( -2\sqrt{k} \cos (\theta) + k + 1 - z \big)^{-1}
\sin \big( (j + 1) \theta \big) \sin \big( (\ell + 1) \theta \big) \, d\theta.
\end{split}$$ Since we have $$\biggl\langle \varphi,W E_m^{n,n+j} \biggr\rangle =
\sum_{v' \in \mathcal{V}} \varphi (v') \overline{ \Big( W E_m^{n,n+j} \Big) (v')},$$ $W$ being as above, then it follows from that $$A (-L + k + 1 - z)^{-1} B^\ast
\varphi (v) = \sum_{v' \in \mathcal{V}} K(v,v') \varphi (v'),$$ with $$\label{eq4,14}
\begin{split}
K(v,v') :=
& \frac{1}{\pi} \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\overline{ \Big( B E_m^{n,n+j} \Big) (v')}
\Big( A E_q^{n,n+\ell} \Big) (v)
\langle \chi_m^{n,n+j},\chi_q^{n,n+\ell} \rangle \\
& \times \int_{\mathbb{T}} \big( -2\sqrt{k} \cos (\theta) + k + 1 - z \big)^{-1}
\sin \big( (j + 1) \theta \big) \sin \big( (\ell + 1) \theta \big) \, d\theta.
\end{split}$$ Then, to complete the proof of the theorem, it remains only to show that $$\label{eq4,15}
\begin{split}
\frac{1}{\pi} \int_{\mathbb{T}} & \big( -2\sqrt{k} \cos (\theta) + k + 1 - z \big)^{-1}
\sin \big( (j + 1) \theta \big) \sin \big( (\ell + 1) \theta \big) \, d\theta \\
& = \frac{1}{2\sqrt{k}} \sqrt{\frac{2}{\pi}} \left( - \frac{i e^{i(j + \ell + 2)\Phi}}{\sqrt{u}
\sqrt{4 - u}} + \frac{i e^{i\vert j - \ell \vert\Phi}}{\sqrt{u} \sqrt{4 - u}} \right),
\end{split}$$ where the relation between $z$, $u$ and $\Phi$ is given by . To do this, we have to deal with the discrete Fourier transform $\mathcal{F} : \ell^2({{\mathbb{Z}}},{{\mathbb{C}}})
\rightarrow L^2({{\mathbb{T}}})$, defined for any $x \in \ell^2({{\mathbb{Z}}},{{\mathbb{C}}})$ and $f \in L^2({{\mathbb{T}}})$ by $$(\mathcal{F}x)(\theta) := (2\pi)^{-\frac{1}{2}} \sum_{n \, \in \, {{\mathbb{Z}}}}
e^{-in\theta} x(n), \quad
\big( \mathcal{F}^{-1} f \big)(n) := (2\pi)^{-\frac{1}{2}}
\int_{{{\mathbb{T}}}} e^{in\theta} f(\theta) \, d\theta.
$$ Let $u \in {{\mathbb{C}}}\setminus [0,4]$ and introduce the following change of variables $$u = 4 \sin^2 \left( \frac{\Phi}{2} \right), \quad {\operatorname{Im}}(\Phi) > 0.$$ Then, it can be proved (cf. e.g. [@it Section 2]) that we have $$\Big( \mathcal{F}^{-1} \big( 2 - 2 \cos (\cdot) - u \big)^{-1} \Big)(n)
= \frac{ie^{i \vert n \vert \Phi}}{2 \sin(\Phi)} =
\frac{ie^{i \vert n \vert \Phi}}{\sqrt{u} \sqrt{4 - u}},$$ or equivalently $$\label{eq4,16}
\frac{1}{\pi} \int_{{{\mathbb{T}}}} \big( 2 - 2 \cos(\theta) - u \big)^{-1} e^{in\theta} \, d\theta
= \sqrt{\frac{2}{\pi}} \frac{ie^{i \vert n \vert \Phi}}{\sqrt{u} \sqrt{4 - u}}.$$ Now, with the help of the transformations $\sin(n\theta) = \frac{1}{2i} \left( e^{in\theta} - e^{-in\theta} \right)$ and $$\Big( -2\sqrt{k} \cos(\theta) + k + 1 - z \Big)^{-1} = \frac{1}{\sqrt{k}}
\Big( 2 - 2 \cos(\theta) - u \Big)^{-1},$$ where $u = \frac{z + 2\sqrt{k} - (k + 1)}{\sqrt{k}}$ give immediately . This completes the proof of the theorem.
Resonances near $z = t_\pm(k)$ {#sec5}
==============================
Definition of the resonances
----------------------------
In this subsection, we define the resonances of the operator ${{-\Delta}}_{\widetilde M}$ near the spectral thresholds $t_\pm(k)$ given by . As preparation, preliminary lemmas will be proved firstly.
From now on, the potential perturbation $M$ is assumed to satisfy Assumption (A). Moreover, the following determination of the complex square root $$\label{eq5,1}
{{\mathbb{C}}}\setminus (-\infty,0] \overset{\sqrt{\cdot}}
{\longrightarrow} {{\mathbb{C}}}^+ := \big\lbrace z \in {{\mathbb{C}}}: {\operatorname{Im}}(z) > 0 \big \rbrace$$ will be adopted throughout this paper. For $\varepsilon > 0$ such that $0 < \varepsilon < \frac{\delta}{4}$, we let $D(0,\varepsilon)^\ast$ be the punctured neighborhood of $0$ defined by $$\label{eq5,2}
D(0,\varepsilon)^\ast := \big \lbrace \lambda \in {{\mathbb{C}}}: 0 < \vert \lambda \vert <
\varepsilon \big \rbrace.$$ Thanks to the first change of variables in , to define and to study the resonances of the operator ${{-\Delta}}_{\widetilde M}$ near the spectral thresholds $t_\pm(k)$, it suffices to define and to study them respectively near $u = 0$ and $u = 4$. However, in practice, there is a simple way (see the comments just after Definition \[d5,1\]) allowing to reduce the analysis of the resonances near the second threshold $t_+(k)$ to that of the first one $t_-(k)$. For further use, let $e_{\pm}$ be the multiplication operators by the functions $$\label{eq5,20}
v \longmapsto e_\pm (v) := e^{\pm\frac{\delta}{2}\vert v \vert}.$$ We have the following lemma:
\[l5,1\] Let $z_{t_-}(\lambda)$ be the parametrization defined by . Then, there exists $0 < \varepsilon_0 \le \frac{\delta}{8}$ small enough such that the operator-valued function $$\label{eq5,3}
\lambda \mapsto e_- \big( -L + k + 1 - z_{t_-}(\lambda) \big)^{-1} e_-,$$ admits an extension from $D(0,\varepsilon_0)^\ast \cap {{\mathbb{C}}}^+$ to $D(0,\varepsilon_0)^\ast$, with values in ${{\textbf{\textup{S}}_\infty}}\big( \ell^2(\mathcal{V})
\big)$ the set of compact linear operators on $\ell^2(\mathcal{V})$. Moreover, this extension is holomorphic.
By Theorem \[p4,1\], for $\lambda \in D(0,\varepsilon)^\ast \cap {{\mathbb{C}}}^+$, $0 < \varepsilon < \frac{\delta}{4}$ small enough, the operator $$\label{eq5,30}
e_- \big( -L + k + 1 - z_{t_-}(\lambda) \big)^{-1} e_-$$ admits the kernel $$\label{eq5,4}
K(\lambda,v,v') = \frac{1}{2 \sqrt{k}} \sqrt{\frac{2}{\pi}} \Big( K_1(\lambda,v,v')
+ K_2(\lambda,v,v') \Big),$$ where $$\label{eq5,5}
\begin{split}
K_1(\lambda,v,v') : = \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle
f_1(j,\ell,\lambda) e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v),
\end{split}$$ and $$\label{eq5,6}
\begin{split}
K_2(\lambda,v,v') := \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\
\ell \, \geq \, 0}} \sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle
f_2(j,\ell,\lambda) e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v),
\end{split}$$ with $$\label{eq5,60}
f_1(j,\ell,\lambda) := - \frac{i e^{i(j + \ell + 2) 2\arcsin \frac{\lambda}{2}}}{\lambda
\sqrt{4 - \lambda^2}} , \qquad
f_2(j,\ell,\lambda) := \frac{i e^{i\vert j - \ell \vert 2\arcsin \frac{\lambda}{2}}}{\lambda
\sqrt{4 - \lambda^2}}.$$
a\) We want to prove the convergence of $\displaystyle\sum_{v, v' \in \, \mathcal{V}} \big\vert K(\lambda,v,v') \big\vert^2$ for $\lambda \in D(0,\varepsilon_0)^\ast$ for some $0 < \varepsilon_0 \le \frac{\delta}{8}$ small enough.
We point out that constants are generic, i.e. can change from an estimate to another. By –, we have $$\label{eq5,7}
\begin{split}
\sum_{v, v' \in \, \mathcal{V}} \big\vert K(\lambda,v,v') \big\vert^2 & = \frac{1}{2k\pi}
\sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_1(\lambda,v,v')
+ K_2(\lambda,v,v') \big\vert^2 \\
& \le \frac{1}{k\pi} \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \left( \big\vert
K_1(\lambda,v,v') \big\vert^2 + \big\vert K_2(\lambda,v,v') \big\vert^2 \right).
\end{split}$$ Let us first prove that $\displaystyle\sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_1(\lambda,v,v') \big\vert^2$ converges accordingly to the above claim. Thanks to , the properties (in Section \[sec4\]) of the vectors $E_m^{n,n+j}$, $\chi_m^{n,n+j}$ and , we have $$\label{eq5,9}
\begin{split}
& \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_1(\lambda,v,v') \big\vert^2
= \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \Biggl\vert
\sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell : n + \ell = r}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
e^{-\frac{\delta}{2} r'} \\
& \times \overline{ E_m^{n,n+j} (v')} \langle \chi_m^{n,n + j},\chi_q^{n,n + \ell}
\rangle \left( - \frac{i e^{i(j + \ell + 2) 2\arcsin \frac{\lambda}{2}}}{\lambda
\sqrt{4 - \lambda^2}} \right) e^{-\frac{\delta}{2}r} E_q^{n,n + \ell} (v) \Biggr\vert^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}}
\left( \sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell : n + \ell = r}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
e^{-\frac{\delta}{2} r'} \frac{e^{-(j + \ell + 2) {\operatorname{Im}}\big( 2\arcsin \frac{ \lambda}{2}
\big)}}{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert } e^{-\frac{\delta}{2}r} \right)^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}}
\left( \sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell : n + \ell = r}}
k^{n + j} k^{n + \ell} e^{-\frac{\delta}{2} r'}
\frac{e^{-(j + \ell + 2) {\operatorname{Im}}\big( 2\arcsin \frac{\lambda}{2}
\big)}}{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert} e^{-\frac{\delta}{2}r} \right)^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'} e^{-\delta r}
e^{-\delta r'} \left( \sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell : n + \ell = r}}
\frac{e^{-(j + \ell + 2) {\operatorname{Im}}\big( 2\arcsin \frac{\lambda}{2}
\big)}}{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert} \right)^2,
\end{split}$$ for some constant $C > 0$. Since for $0 < \vert \lambda \vert \ll 1$ we have $2 \arcsin \frac{\lambda}{2} = \lambda + o(\vert \lambda \vert)$, then there exists $0 < \varepsilon_0 \le \frac{\delta}{8}$ small enough such that for each $0 < \vert \lambda \vert \le \varepsilon_0$, we have $$\label{eq5,90}
\frac{e^{-(j + \ell + 2) {\operatorname{Im}}\big( 2\arcsin \frac{\lambda}{2}
\big)}}{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert} \le
\frac{e^{-\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) (j + \ell + 2)}}
{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert}.$$ Then, it follows from that for each $\lambda \in D(0,\varepsilon_0)^\ast$, we have $$\label{eq5,10}
\begin{split}
& \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} \sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_1(\lambda,v,v') \big\vert^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'}
e^{-\delta r} e^{-\delta r'}
\left( \sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell : n + \ell = r}}
\frac{e^{-\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) (j + \ell + 2)}}{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert} \right)^2.
\end{split}$$ Clearly, if $F = F(j,\ell)$ is a function of the variables $j$ and $\ell$, then $$\label{eq5,11}
\sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell : n + \ell = r}} F(j,\ell)
= \sum_{n = 0}^{\min(r,r')} F(r'-n,r-n).$$ This together with imply that $$\label{eq5,12}
\begin{split}
& \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_1(\lambda,v,v') \big\vert^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'}
e^{-\delta r} e^{-\delta r'}
\left( \sum_{n = 0}^{\min(r,r')} \frac{e^{-\big( {\operatorname{Im}}(\lambda)
- \frac{\delta}{8} \big) (r + r' + 2)} e^{2\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big)n }} {\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert}
\right)^2.
\end{split}$$ Since ${\operatorname{Im}}(\lambda) - \frac{\delta}{8} \leq 0$, then it follows from that $$\label{eq5,13}
\begin{split}
& \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_1(\lambda,v,v') \big\vert^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'}
e^{-\delta r} e^{-\delta r'}
\left( \sum_{n = 0}^{\min(r,r')} \frac{e^{-\big( {\operatorname{Im}}(\lambda)
- \frac{\delta}{8} \big) (r + r' + 2)}}{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert} \right)^2 \\
& \leq C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'} e^{-\delta r}
e^{-\delta r'} \frac{e^{-2\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) (r + r' + 2)}}
{\big\vert \lambda^2 (4 - \lambda^2) \big\vert} (r + 1)(r' + 1) \\
& = C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} (r + 1) k^{3r} e^{-\delta r}
\frac{e^{-2\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) (r + 1)} e^{-2\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) (r' + 1)}} {\big\vert \lambda^2 (4 - \lambda^2) \big\vert}
(r' + 1) k^{3r'} e^{-\delta r'} \\
& \leq C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} (r + 1) k^{3(r+1)} e^{-\delta (r+1)}
\frac{e^{-2\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) (r + 1)} e^{-2\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) (r' + 1)}} {\big\vert \lambda^2 (4 - \lambda^2) \big\vert} \\
& \hspace{3cm} \times (r' + 1) k^{3(r'+1)} e^{-\delta (r'+1)} \\
& = C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} (r + 1) e^{-\big( \frac{\delta}{2}
- 3 \ln(k) \big) (r+1)} \frac{e^{-2 \big( {\operatorname{Im}}(\lambda) + \frac{\delta}{8} \big)
(r + 1)} e^{-2\big( {\operatorname{Im}}(\lambda) + \frac{\delta}{8} \big) (r' + 1)}} {\big\vert \lambda^2 (4 - \lambda^2) \big\vert} \\
& \hspace{3cm} \times (r' + 1) e^{-\big( \frac{\delta}{2} - 3 \ln(k) \big) (r'+1)}.
\end{split}$$ Assumption (A) implies that $\frac{\delta}{2} - 3 \ln(k) \geq 0$. Thus, the r.h.s. and then the l.h.s. of is convergent for any $\lambda \in D(0,\varepsilon_0)^\ast$.
Similarly let us prove that $\displaystyle\sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_2(\lambda,v,v') \big\vert^2$ converges. As in , we can show that $$\label{eq5,14}
\begin{split}
\sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} &
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_2(\lambda,v,v') \big\vert^2
= \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \Biggl\vert
\sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell : n + \ell = r}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}} \\
& e^{-\frac{\delta}{2} r'} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle
\left( \frac{i e^{i \vert j - \ell \vert 2\arcsin \frac{\lambda}{2}}}{\lambda
\sqrt{4 - \lambda^2}} \right) e^{-\frac{\delta}{2}r} E_q^{n,n + \ell} (v) \Biggr\vert^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'} e^{-\delta r}
e^{-\delta r'} \left( \sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell
: n + \ell = r}} \frac{e^{- \vert j - \ell \vert {\operatorname{Im}}\big( 2\arcsin \frac{\lambda}{2}
\big)}}{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert} \right)^2.
\end{split}$$ Thus, similarly to , for each $\lambda \in D(0,\varepsilon_0)^\ast$, we have $$\label{eq5,15}
\begin{split}
& \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_2(\lambda,v,v') \big\vert^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'}
e^{-\delta r} e^{-\delta r'}
\left( \sum_{n = 0}^{\min(r,r')} \sum_{\substack{j : n + j = r' \\ \ell : n + \ell = r}} \frac{e^{-\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) \vert j - \ell \vert}}{\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert} \right)^2.
\end{split}$$ In this case, we use to write $$\label{eq5,16}
\begin{split}
& \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}}
\sum_{\substack{v \, \in \, S_r \\ v' \in \, S_{r'}}} \big\vert K_2(\lambda,v,v') \big\vert^2 \\
& \le C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'}
e^{-\delta r} e^{-\delta r'}
\left( \sum_{n = 0}^{\min(r,r')} \frac{e^{-\big( {\operatorname{Im}}(\lambda)
- \frac{\delta}{8} \big) \vert r - r' \vert}} {\big\vert \lambda \sqrt{4 - \lambda^2} \big\vert} \right)^2 \\
& \leq C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} k^{3r} k^{3r'} e^{-\delta r}
e^{-\delta r'} \frac{e^{-2\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) \vert r - r' \vert}}
{\big\vert \lambda^2 (4 - \lambda^2) \big\vert} (r + 1)(r' + 1) \\
& = C \sum_{\substack{r \, \ge \, 0 \\ r' \ge \, 0}} (r + 1) e^{-\big( \frac{\delta}{2}
- 3 \ln(k) \big)r} \frac{e^{- \frac{\delta}{2}r} e^{-2\big( {\operatorname{Im}}(\lambda) - \frac{\delta}{8} \big) \vert r - r' \vert)} e^{- \frac{\delta}{2}r'}} {\big\vert \lambda^2 (4 - \lambda^2) \big\vert} \\
& \hspace*{3cm} \times (r' + 1) e^{-\big( \frac{\delta}{2} - 3 \ln(k) \big)r'}.
\end{split}$$ Assumption (A) implies that $\frac{\delta}{2} - 3 \ln(k) \geq 0$. Thus, the r.h.s. and then the l.h.s. of is convergent for any $\lambda \in D(0,\varepsilon_0)^\ast$.
Now, since $\displaystyle\sum_{v, v' \in \, \mathcal{V}} \big\vert K(\lambda,v,v') \big\vert^2$ is convergent for $\lambda \in D(0,\varepsilon_0)^\ast$, then the operator given by belongs in ${{\textbf{\textup{S}}_2}}\big( \ell^2(\mathcal{V}) \big)$ for $\lambda \in D(0,\varepsilon_0)^\ast$, the class of Hilbert-Schmidt operators on $\ell^2(\mathcal{V})$. Consequently, the operator-valued function defined by can be extended from $D(0,\varepsilon_0)^\ast \cap {{\mathbb{C}}}^+$ to $D(0,\varepsilon_0)^\ast$, with values in ${{\textbf{\textup{S}}_\infty}}\big( \ell^2(\mathcal{V}) \big)$. It remains to prove that this extension is holomorphic.
b\) To simply notation, let us denote this extension by $$D(0,\varepsilon_0)^\ast \ni \lambda \mapsto T(\lambda).$$ Since the kernel of the operator $T(\lambda)$ is given by $K(\lambda,v,v')$ defined by , then to show the claim, it is sufficient to prove it for the maps $$D(0,\varepsilon_0)^\ast \ni \lambda \mapsto T_s(\lambda),$$ where for $s = 1, 2$, $T_s(\lambda)$ is the operator with kernel given by $K_s(\lambda,v,v')$ in and . We give the proof only for the case $s = 1$, the case $s = 2$ being treated in a similar way. So, for $\lambda \in D(0,\varepsilon_0)^\ast$, let $f_1(j,\ell,\lambda)$ be the function defined by and $D_1(\lambda)$ be the operator whose kernel is $$\begin{aligned}
& \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \partial_\lambda f_1(j,\ell,\lambda)
e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v) \\
& = \sum_{n \, \geq \, 0} \sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \times \left( - \frac{i e^{i(j + \ell + 2) 2\arcsin \frac{\lambda}{2}}}{\lambda^2
(4 - \lambda^2)} \right) e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v)
\Bigg( 2i (j + \ell + 2) - \frac{4 - 2\lambda^2}{\sqrt{4 - \lambda^2}} \Bigg).\end{aligned}$$ As in a) above, we can show that $D_1(\lambda) \in {{\textbf{\textup{S}}_2}}\big( \ell^2(\mathcal{V}) \big)$. Therefore, for $\lambda_0 \in D(0,\varepsilon_0)^\ast$, the kernel of the Hilbert-Schmidt operator $\frac{T_1(\lambda) - T_1(\lambda_0)}{\lambda - \lambda_0}
- D_1(\lambda_0)$ is given by $$\label{eq5,160}
\begin{split}
& K_1(\lambda,\lambda_0,v,v') := \sum_{n \, \geq \, 0} \sum_{\substack{j \,
\geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \times \Bigg( \frac{f_1(j,\ell,\lambda) - f_1(j,\ell,\lambda_0)}{\lambda - \lambda_0} -
\partial_\lambda f_1(j,\ell,\lambda_0) \Bigg) e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v).
\end{split}$$ Thus, to conclude the proof of the lemma, we have just to justify that $$\Bigg\Vert \frac{T_1(\lambda) - T_1(\lambda_0)}{\lambda - \lambda_0}
- D_1(\lambda_0) \Bigg\Vert_{{{\textbf{\textup{S}}_2}}(\ell^2(\mathcal{V}))} \longrightarrow 0$$ as $\lambda \rightarrow \lambda_0$. Since we have $$\label{eq5,161}
\bigg\Vert \frac{T_1(\lambda) - T_1(\lambda_0)}{\lambda - \lambda_0}
- D_1(\lambda_0) \bigg\Vert_{{{\textbf{\textup{S}}_2}}(\ell^2(\mathcal{V}))} \le
\sum_{v, v' \in \, \mathcal{V}} \big\vert K_1(\lambda,\lambda_0,v,v') \big\vert^2,$$ then it suffices to prove that the r.h.s. of tends to zero as $\lambda \rightarrow \lambda_0$. The Taylor-Lagrange formula applied to the function $$[0,1] \ni t \mapsto g(t) := f_1 \big(j,\ell,t\lambda + (1-t)\lambda_0 \big)$$ asserts there exists $\theta \in (0,1)$ such that $$\label{eq5,162}
\begin{split}
f_1(j,\ell,\lambda) = f_1(j,\ell,\lambda_0) + (\lambda - \lambda_0)
\partial_\lambda f_1(j,\ell,\lambda_0) + \frac{(\lambda - \lambda_0)^2}{2}
\partial_\lambda^{(2)} f_1 \big(j,\ell, \theta \lambda + (1-\theta)\lambda_0 \big).
\end{split}$$ Then, it follows from and that $K_1(\lambda,\lambda_0,v,v')$ can be represented as $$\label{eq5,163}
\begin{split}
& K_1(\lambda,\lambda_0,v,v') = \frac{\lambda - \lambda_0}{2}
\sum_{n \, \geq \, 0} \sum_{\substack{j \,
\geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \times \partial_\lambda^{(2)} f_1 \big(j,\ell, \theta \lambda + (1-\theta)\lambda_0 \big) e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v).
\end{split}$$ Now, easy but fastidious computations allow to see that there exists a family of holomorphic functions $F_{p,q}$, $0 \le p \le q$, on $D(0,\varepsilon_0)^\ast$ such that $$\partial_\lambda^{(q)} f_1(j,\ell,\lambda) = ie^{i(j+\ell+2) 2\arcsin \frac{\lambda}{2}}
\sum_{p=0}^q F_{p,q}(\lambda) (j + \ell + 2)^{q-p}.$$ In particular, for $q = 2$, we have $$\partial_\lambda^{(2)} f_1(j,\ell,\lambda) = ie^{i(j+\ell+2) 2\arcsin \frac{\lambda}{2}}
\Big( F_{0,2}(\lambda)(j+\ell+2)^2 + F_{1,2}(\lambda) (j+\ell+2) + F_{2,2}(\lambda) \Big).$$ Putting this together with , we obtain $$\begin{aligned}
& \big\vert K_1(\lambda,\lambda_0,v,v') \big\vert \\
& \le \frac{\vert \lambda - \lambda_0 \vert}{2} \big\vert F_{0,2} \big( \theta \lambda + (1-\theta)\lambda_0 \big) \big\vert \sum_{n \, \geq \, 0} \sum_{\substack{j \,
\geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\bigg\vert e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \hspace*{2cm} \times e^{i(j+\ell+2) 2\arcsin \frac{\theta \lambda + (1-\theta)\lambda_0}{2}}
(j+\ell+2)^2 e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v) \bigg\vert \\
& + \frac{\vert \lambda - \lambda_0 \vert}{2} \big\vert F_{1,2} \big( \theta \lambda + (1-\theta)\lambda_0 \big) \big\vert \Big\vert \sum_{n \, \geq \, 0} \sum_{\substack{j \,
\geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\bigg\vert e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \hspace*{2cm} \times e^{i(j+\ell+2) 2\arcsin \frac{\theta \lambda + (1-\theta)\lambda_0}{2}}
(j+\ell+2) e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v) \bigg\vert \\
& + \frac{\vert \lambda - \lambda_0 \vert}{2} \big\vert F_{2,2} \big( \theta \lambda + (1-\theta)\lambda_0 \big) \big\vert \sum_{n \, \geq \, 0} \sum_{\substack{j \,
\geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\bigg\vert e^{-\frac{\delta}{2}\vert v' \vert} \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \hspace*{2cm} \times e^{i(j+\ell+2) 2\arcsin \frac{\theta \lambda + (1-\theta)\lambda_0}{2}}
e^{-\frac{\delta}{2}\vert v \vert} E_q^{n,n + \ell} (v) \bigg\vert \\
& =: \sum_{p=0}^2 Q_p(\lambda,\lambda_0,v,v').\end{aligned}$$ Thus, we have $$\label{eq5,164}
\begin{split}
& \sum_{v, v' \in \, \mathcal{V}} \big\vert K_1(\lambda,\lambda_0,v,v') \big\vert^2 \\
& \le {\rm Const.} \Bigg( \sum_{v, v' \in \, \mathcal{V}} Q_1(\lambda,\lambda_0,v,v')^2
+ \sum_{v, v' \in \, \mathcal{V}} Q_2(\lambda,\lambda_0,v,v')^2 +
\sum_{v, v' \in \, \mathcal{V}} Q_3(\lambda,\lambda_0,v,v')^2 \Bigg).
\end{split}$$ For $p \in \lbrace 0,1,2 \rbrace$, let us show that $\displaystyle\sum_{v, v' \in \, \mathcal{V}} Q_p(\lambda,\lambda_0,v,v')^2 \longrightarrow 0$ as $\lambda \rightarrow \lambda_0$. Since $\theta \lambda + (1-\theta)\lambda_0$ belongs in $ D(0,\varepsilon_0)^\ast$ for $\lambda, \lambda_0 \in D(0,\varepsilon_0)^\ast$, then similarly to we have $$\label{eq5,165}
\begin{split}
e^{-(j + \ell + 2) {\operatorname{Im}}\big( 2\arcsin \frac{\theta \lambda + (1-\theta)\lambda_0}{2}
\big)} & \le e^{-\big( {\operatorname{Im}}(\theta \lambda + (1-\theta)\lambda_0) - \frac{\delta}{8} \big) (j + \ell + 2)} \\
& \le e^{\frac{\delta}{4}(j + \ell + 2)}.
\end{split}$$ Therefore, by arguing as in a) above, we obtain that there exists a uniform constant Const. in $\lambda$ and $\lambda_0$ such that for $q \in \lbrace 0,1,2 \rbrace$, $$\sum_{v, v' \in \, \mathcal{V}} Q_p(\lambda,\lambda_0,v,v')^2 \le {\rm Const.}
\frac{\vert \lambda - \lambda_0 \vert^2}{4} \big\vert F_{q,2}(\theta \lambda + (1-\theta)\lambda_0) \big\vert^2 \underset{\lambda \rightarrow \lambda_0}{\longrightarrow} 0,$$ implying by and that $\Big\Vert \frac{T_1(\lambda) - T_1(\lambda_0)}{\lambda - \lambda_0} - D_1(\lambda_0) \Big\Vert_{{{\textbf{\textup{S}}_2}}(\ell^2(\mathcal{V}))} \longrightarrow 0$ as $\lambda \rightarrow \lambda_0$. Thus, the operator-valued function $D(0,\varepsilon_0)^\ast \ni \lambda \mapsto T_1(\lambda)$ is holomorphic with derivative $\partial_\lambda T_1(\lambda) = D_1(\lambda)$. Similarly, $D(0,\varepsilon_0)^\ast \ni \lambda \mapsto T_2(\lambda)$ is holomorphic, and then $D(0,\varepsilon_0)^\ast \ni \lambda \mapsto T(\lambda)$. This concludes the proof of the lemma.
It follows from the identities $$\label{eq5,17}
\Big( {{-\Delta}}_{\pm \widetilde M} - z \Big)^{-1} \left( I \pm {\widetilde M} (-L + k + 1 - z)^{-1}
\right) = (-L + k + 1 - z)^{-1},$$ that $$\label{eq5,18}
\begin{split}
e_- \Big( & {{-\Delta}}_{\pm \widetilde M} - z \Big)^{-1} e_- \\
& = e_- (-L + k + 1 - z)^{-1} e_- \left( I \pm e_+ {\widetilde M}(-L + k + 1 - z)^{-1}
e_- \right)^{-1}.
\end{split}$$ Assumption (A) on the potential perturbation $M$ implies that $$\label{eq5,19}
e_+ {\widetilde M} = \mathscr{\widetilde M} e_-$$ for some bounded operator $\mathscr{\widetilde M}$ on $\ell^2(\mathcal{V})$. Thus, combining and Lemma \[l5,1\], we obtain that the operator-valued functions $$\label{eq5,20}
\lambda \longmapsto \pm e_+ {\widetilde M} \big( -L + k + 1 - z_{t_-}(\lambda) \big)^{-1} e_-$$ are holomorphic in $D(0,\varepsilon)^\ast$, with values in ${{\textbf{\textup{S}}_\infty}}\big( \ell^2(\mathcal{V})
\big)$. Therefore, by the analytic Fredholm extension theorem, the operator-valued functions $$\label{eq5,21}
\lambda \longmapsto \left( I \pm e_+ {\widetilde M} \big( -L + k + 1 - z_{t_-}(\lambda)
\big)^{-1} e_- \right)^{-1}$$ admit meromorphic extensions from $D(0,\varepsilon_0)^\ast \cap {{\mathbb{C}}}^+$ to $D(0,\varepsilon_0)^\ast$. Defining the Banach spaces $$\label{eq5,22}
\ell_{\pm \delta}^{2}(\mathcal{V}) := e^{\pm \frac{\delta}{2} \vert v \vert} \ell^2(\mathcal{V}),$$ we then get the following proposition:
\[p5,1\] The operator-valued functions $$\label{eq5,23}
\lambda \longmapsto \Big( {{-\Delta}}_{\pm \widetilde M} - z_{t_-}(\lambda) \Big)^{-1} \in
\mathscr{L} \left( \ell_{-\delta}^{2}(\mathcal{V}),\ell_\delta^{2}(\mathcal{V}) \right),$$ admit meromorphic extensions from $D(0,\varepsilon_0)^\ast \cap {{\mathbb{C}}}^+$ to $D(0,\varepsilon_0)^\ast$. These extensions will be denoted by $R_{\pm \widetilde M}
\big( z_{t_-}(\lambda) \big)$ respectively.
As in , Assumption (A) on $M$ implies that there exists a bounded operator $\mathscr{B}$ on $\ell^2(\mathcal{V})$ such that $\sqrt{\vert \widetilde M \vert} = \mathscr{B} e_-$. Together with Lemma \[l5,1\], this gives the following lemma:
\[l5,2\] Let $J$ be defined by the polar decomposition ${\widetilde M} = J \vert \widetilde M \vert$ of the potential perturbation ${\widetilde M}$. Then, the operator-valued functions $$\label{eq5,24}
\lambda \longmapsto \mathcal{T}_{\pm \widetilde M} \big( z_{t_-}(\lambda) \big) := \pm J
\sqrt{\vert \widetilde M \vert} \big( -L + k + 1 - z_{t_-}(\lambda) \big)^{-1}
\sqrt{\vert \widetilde M \vert},$$ admit holomorphic extensions from $D(0,\varepsilon_0)^\ast \cap {{\mathbb{C}}}^+$ to $D(0,\varepsilon_0)^\ast$, with values in ${{\textbf{\textup{S}}_\infty}}\big( \ell^2(\mathcal{V}) \big)$.
We are now in position to define the resonances of the operator ${{-\Delta}}_{\widetilde M}$ near the spectral thresholds $z = t_{\pm}(k)$. Note that in the next definitions, the quantity $Ind_{\gamma}(\cdot)$ is defined in the appendix by .
\[d5,1\] We define the resonances of the operator ${{-\Delta}}_{\widetilde M}$ near $t_-(k)$ as the poles of the meromorphic extension $R_{\widetilde M}(z)$, of the resolvent $\Big( {{-\Delta}}_{\widetilde M} - z \Big)^{-1}$ in $\mathscr{L} \left( \ell_{-\delta}^{2}(\mathcal{V}),\ell_{\delta}^{2}(\mathcal{V}) \right)$. The multiplicity of a resonance $$z_{t_-} := z_{t_-}(\lambda) = t_-(k) + \lambda^2\sqrt{k},$$ is defined by $$\label{eq5,25}
\textup{mult} \big( z_{t_-} \big) := Ind_{\gamma} \, \biggl( I +
\mathcal{T}_{\widetilde M} \Big( z_{t_-}(\cdot) \Big) \biggr),$$ where $\gamma$ is a small contour positively oriented containing $\lambda$ as the only point satisfying that $z_{t_-}(\lambda)$ is a resonance of ${{-\Delta}}_{\widetilde M}$.
As mentioned previously, to define the resonances of the operator ${{-\Delta}}_{\widetilde M}$ near $t_+(k)$, there exists a specific reduction which exploits a simple relation between the two thresholds $t_{\pm}(k)$. Indeed, define the self-adjoint unitary operator $\Theta$ on $\ell^2(\mathcal{V})$ by $$\label{eq5,26}
(\Theta \varphi)(v) := (-1)^{\vert v \vert} \varphi(v).$$ We thus have
- $\Theta^2 = I$,
- $\Theta L \Theta^{-1} = -L$,
- $\Theta \widetilde{M} \Theta^{-1} = \widetilde{M}$.
In the last point, we have used the fact that $\widetilde{M}$ is the multiplication operator by the function $\widetilde{M}$. Thus, it can be easily verified that we have $$\label{eq5,27}
\Theta \Big( -L + k + 1 + \widetilde{M} - z \Big) \Theta^{-1} =
-(-L + k + 1) + \widetilde{M} + 2(k + 1) - z,$$ so that $$\label{eq5,28}
\Theta e_- \Big( {{-\Delta}}_{\widetilde M} - z \Big)^{-1} e_- \Theta^{-1}
= - e_- \Big( {{-\Delta}}_{-\widetilde M} - \big( 2(k + 1) - z \big) \Big)^{-1} e_-.$$ Set $$\label{eq5,29}
\omega := 2(k + 1) - z.$$ Since $\omega$ is near $t_-(k)$ for $z$ near $t_+(k)$, then using relation , we can define the resonances of the operator ${{-\Delta}}_{\widetilde M}$ near $t_+(k)$ as the poles of the meromorphic extension of the resolvent $$\label{eq5,300}
- \Big( {{-\Delta}}_{-\widetilde M} - \omega \Big)^{-1} : \ell_{-\delta}^2(\mathcal{V})
\rightarrow \ell_{\delta}^2(\mathcal{V}),$$ near $\omega = t_-(k)$, similarly to Definition \[d5,1\]. More precisely, we have the following definition:
\[d5,2\] We define the resonances of the operator ${{-\Delta}}_{\widetilde M}$ near $t_+(k)$ as the poles of the meromorphic extension $R_{-\widetilde M}(\omega)$, of the resolvent $\Big( {{-\Delta}}_{-\widetilde M} - \omega \Big)^{-1}$ in $\mathscr{L} \left( \ell_{-\delta}^{2}(\mathcal{V}),\ell_{\delta}^{2}(\mathcal{V}) \right)$, for $\omega$ given by near $t_-(k)$. The multiplicity of a resonance $$z_{t_+} := z_{t_+}(\lambda) = 2(k+ 1) - \Big( t_-(k) + \lambda^2\sqrt{k} \Big)
= t_+(k) - \lambda^2\sqrt{k},$$ is defined by $$\label{eq5,31}
\textup{mult}(z_{t_+}) := Ind_{\gamma} \, \biggl( I +
\mathcal{T}_{-\widetilde M} \Big( 2(k + 1) - z_{t_+}(\cdot) \Big) \biggr),$$ where $\gamma$ is a small contour positively oriented containing $\lambda$ as the only point satisfying that $2(k + 1) - z_{t_+}(\lambda)$ is a pole of $R_{-\widetilde M}(\omega)$.
Notice that the resonances $z_{t_\pm}(\lambda)$ near the spectral thresholds $t_\pm(k)$ are defined in some two-sheets Riemann surfaces $\mathcal{M}_{t_\pm}$ respectively. Otherwise, the discrete eigenvalues of the operator ${{-\Delta}}_{\widetilde M}$ near $t_\pm(k)$ are resonances. Moreover, the algebraic multiplicity of a discrete eigenvalue coincides with its multiplicity as a resonance near $t_\pm(k)$ respectively given by and . Let us give the proof only for the equality $\eqref{eq1,9} = \eqref{eq5,25}$, the equality $\eqref{eq1,9} = \eqref{eq5,31}$ could be treated in a similar fashion. Let $z_{t_-} := z_{t_-}(\lambda) \in {{\mathbb{C}}}\setminus
[t_-(k),t_+(k)]$ be a discrete eigenvalue of ${{-\Delta}}_{\widetilde M}$ near $t_-(k)$. Firstly, observe that Assumption (A) on $M$ implies that ${\widetilde M}$ is of trace-class. In this case, it is is well know (see e.g. [@si Chap. 9]) that $z_{t_-} \in \sigma_{\textup{\textbf{disc}}} \big( {{-\Delta}}_{\widetilde M} \big)$ if and only if $h(z_{t_-}) = 0$, where for $z \in {{\mathbb{C}}}\setminus [t_-(k),t_+(k)]$, $h$ is the holomorphic function defined by $$\begin{aligned}
h(z) := \textup{det} \Bigl( I + {\widetilde M} \big( -L + k + 1 - z \big)^{-1} \Bigr)
= \det \Bigg( I + J \sqrt{\vert \widetilde M \vert} \big( -L + k + 1 - z \big)^{-1}
\sqrt{\vert \widetilde M \vert} \Bigg).\end{aligned}$$ Moreover, the algebraic multiplicity of $z_{t_-}$ is equal to its order as zero of the function $h$. Namely, by the residues theorem, $$\textup{m}(z_{t_-}) = ind_{\gamma'} h :=
\frac{1}{2i\pi} \int_{\gamma'} \frac{h'(z)}{h(z)} dz,$$ where $\gamma'$ is a small circle positively oriented containing $z_{t_-}$ as the only zero of $h$. Then, the claim follows directly from the equality $$ind_{\gamma'} h = Ind_{\gamma} \, \biggl( I + \mathcal{T}_{\widetilde M} \Big( z_{t_-}(\cdot) \Big) \biggr),$$ see for instance [@bbr2 Identity (6)] for more details.
Characterization of the resonances
----------------------------------
In this subsection, we give a simple characterization of resonances of ${{-\Delta}}_{\widetilde M}$ near the spectral thresholds $t_\pm(k)$. The first one concerns the resonances near $z = t_-(k)$.
\[p5,2\] The following assertions are equivalent:
- $z_{t_-} = z_{t_-}(\lambda) \in \mathcal{M}_{t_-}$ is a resonance,
- $z_{t_-}$ is a pole of $R_{\widetilde M}(z)$,
- $-1$ is an eigenvalue of $\mathcal{T}_{\widetilde M} \big( z_{t_-}(\lambda) \big)$.
$(a) \Longleftrightarrow (b)$ is just the Definition \[d5,1\], while $(b) \Longleftrightarrow (c)$ is a consequence of the identity $$\left( I + J \sqrt{\vert \widetilde M \vert} (-L + k + 1 - z)^{-1}
\sqrt{\vert \widetilde M \vert} \right)
\left( I - J \sqrt{\vert \widetilde M \vert} \Big( {{-\Delta}}_{\widetilde M} - z \Big)^{-1} \sqrt{\vert \widetilde M \vert} \right) = I,$$ coming from the resolvent equation.
Similarly, we have the following proposition:
\[p5,3\] The following assertions are equivalent:
- $z_{t_+} = z_{t_+}(\lambda) \in \mathcal{M}_{t_+}$ is a resonance,
- $2(k + 1) - z_{t_+}(\lambda)$ is a pole of $R_{-\widetilde M}(\omega)$ for $\omega$ given by near $t_-(\lambda)$,
- $-1$ is an eigenvalue of $\mathcal{T}_{-\widetilde M} \Big( 2(k + 1) - z_{t_+}(\lambda) \Big)$.
Proof of Theorem \[t1\] {#sec6}
=======================
This section is devoted to the proof of Theorem \[t1\]. It will be divided into tree steps.
A preliminary result {#ss6,1}
--------------------
The first step consists on refining the representations of the sandwiched resolvents $\mathcal{T}_{\widetilde M} \big( z_{t_-}(\lambda) \big)$ and $\mathcal{T}_{-\widetilde M} \big( 2(k + 1) - z_{t_+}(\lambda) \big)$ near the spectral thresholds $z = t_{\pm}(k)$. Notice that $$\label{eq6,0}
2(k + 1) - z_{t_+}(\lambda) = z_{t_-}(\lambda),$$ so that our analysis will be just reduced to the operators $\mathcal{T}_{\pm \widetilde M} \big( z_{t_-}(\lambda) \big)$.
Recall that $\mathcal{T}_{\pm \widetilde M} \big( z_{t_-}(\lambda) \big) = \pm J
\sqrt{\vert \widetilde M \vert} \big( -L + k + 1 - z_{t_-}(\lambda) \big)^{-1}
\sqrt{\vert \widetilde M \vert}$, and let us set $$\label{eq6,1}
\gamma(\lambda) := \frac{\left( e^{i (j + \ell + 2) 2\arcsin \frac{\lambda}{2}}
- 1 \right)} {\lambda \sqrt{4 - \lambda^2}} \quad \text{and} \quad
\beta(\lambda) := \frac{\left( e^{i \vert j - \ell \vert
2\arcsin \frac{\lambda}{2}} - 1 \right)} {\lambda \sqrt{4 - \lambda^2}}, \,
j, \ell \ge 0.$$ By construction, as shows the proof of Lemma \[l5,1\], for $\lambda \in D(0,\varepsilon_0)^\ast$ the operator $$\label{eq6,2}
\sqrt{\vert \widetilde M \vert} \big( -L + k + 1 - z_{t_-}(\lambda) \big)^{-1}
\sqrt{\vert \widetilde M \vert}$$ admits the integral kernel $$\label{eq6,3}
\pm \frac{1}{2 \sqrt{k}} \sqrt{\frac{2}{\pi}}
\Big( \mathcal{K}_1^{(\lambda)}(v,v') + \mathcal{K}_2^{(\lambda)}(v,v') \Big),$$ where $$\label{eq6,4}
\begin{split}
\mathcal{K}_1^{(\lambda)}(v,v') : = & \sum_{n \, \geq \, 0}
\sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\sqrt{\vert \widetilde M \vert} (v') \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \times i \left( - \gamma(\lambda) - \frac{1}{\lambda \sqrt{4 - \lambda^2}} \right) \sqrt{\vert \widetilde M \vert} (v) E_q^{n,n + \ell} (v),
\end{split}$$ and $$\label{eq6,5}
\begin{split}
\mathcal{K}_2^{(\lambda)}(v,v') : = & \sum_{n \, \geq \, 0}
\sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\sqrt{\vert \widetilde M \vert} (v') \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \times i \left( \beta(\lambda) + \frac{1}{\lambda \sqrt{4 - \lambda^2}} \right) \sqrt{\vert \widetilde M \vert} (v) E_q^{n,n + \ell} (v).
\end{split}$$ Since $\gamma$ and $\beta$ can be extended to holomorphic functions on the open disk $D(0,\varepsilon_0)^\ast \cup \lbrace 0 \rbrace$, then by combining identities -, we get the following result:
\[p6,1\] For $\lambda \in D(0,\varepsilon_0)^\ast \cup \lbrace 0 \rbrace$, we have $$\label{eq6,6}
\mathcal{T}_{\pm \widetilde M} \big( z_{t_-}(\lambda) \big) =
\pm J \sqrt{\frac{2}{\pi}} \, {\rm Hol}(\lambda),$$ where ${\rm Hol}(\lambda)$ defines a holomorphic operator on $D(0,\varepsilon_0)^\ast \cup \lbrace 0 \rbrace$ with values in ${{\textbf{\textup{S}}_\infty}}\big( \ell^2(\mathcal{V}) \big)$, and with kernel given by $$\label{eq6,7}
\begin{split}
& \frac{i}{2 \sqrt{k}} \sum_{n \, \geq \, 0}
\sum_{\substack{j \, \geq \, 0 \\ \ell \, \geq \, 0}}
\sum_{\substack{0 \leq m \leq N_j \\ 0 \leq q \leq N_\ell}}
\sqrt{\vert \widetilde M \vert} (v') \overline{ E_m^{n,n+j} (v')}
\langle \chi_m^{n,n + j},\chi_q^{n,n + \ell} \rangle \\
& \hspace*{1cm} \times \Big( \beta(\lambda) - \gamma(\lambda) \Big) \sqrt{\vert \widetilde M \vert} (v) E_q^{n,n + \ell} (v).
\end{split}$$
Reformulation of the problem
----------------------------
Let $\mathscr{H}$ be a separable Hilbert space, $\mathcal{D} \subseteq {{\mathbb{C}}}$ be a domain containing $0$, and ${{\textbf{\textup{S}}_\infty}}(\mathscr{H})$ denote the set of compact linear operators in $\mathscr{H}$. For a holomorphic operator-valued function $$K : \mathcal{D} \setminus \lbrace 0 \rbrace \longrightarrow {{\textbf{\textup{S}}_\infty}}(\mathscr{H}),$$ and a subset $\Omega \subseteq \mathcal{D} \setminus \lbrace 0 \rbrace$, a complex number $\lambda \in \Omega$ is said to be a [*characteristic value*]{} of the operator-valued function $$\lambda \longmapsto I + K(\lambda),$$ if the operator $I + K(\lambda)$ is not invertible (cf. Section \[sa\] for more details about the concept of characteristic value). By abuse of language, we shall sometimes say that $\lambda$ is a characteristic value of the operator $I + K(\lambda)$. Once there exists $\lambda_{0} \in \Omega$ such that $I + K(\lambda_0)$ is invertible, then by the analytic Fredholm theorem, the set of characteristic values $\lambda \in \Omega$ of $I + K(\cdot)$ is discrete. Moreover, according to Definition \[d,a1\] and , the multiplicity of a characteristic value $\lambda$ is defined by $$\label{eqa,11}
\textup{mult}(\lambda) := Ind_{\gamma} \big( I + K(\cdot) \big),$$ $\gamma$ being a small contour positively oriented which contains $\lambda$ as the only point satisfying $I + K(z)$ is not invertible, and with $I + K(\cdot)$ not vanishing on $\gamma$. We then can reformulate Propositions \[p5,2\] and \[p5,3\] in the following way:
\[p6,2\] For $\lambda \in D(0,\varepsilon_0)^\ast$, the following assertions are equivalent:
- $z_{t_-} = z_{t_-}(\lambda) \in \mathcal{M}_{t_-}$ is a resonance,
- $\lambda$ is a characteristic value of $I + \mathcal{T}_{\widetilde M} \big( z_{t_-}(\cdot) \big)$.
Moreover, thanks to , the multiplicity of the resonance $z_{t_-}(\lambda)$ coincides with that of the characteristic value $\lambda$.
\[p6,3\] For $\lambda \in D(0,\varepsilon_0)^\ast$, the following assertions are equivalent:
- $z_{t_+} = z_{t_+}(\lambda) \in \mathcal{M}_{t_+}$ is a resonance,
- $\lambda$ is a characteristic value of $I + \mathcal{T}_{-\widetilde M} \Big( 2(k + 1) - z_{t_+}(\cdot) \Big)$.
Moreover, thanks to , the multiplicity of the resonance $z_{t_+}(\lambda)$ coincides with that of the characteristic value $\lambda$.
End of the proof of Theorem \[t1\]
----------------------------------
From Propositions \[p6,2\], \[p6,3\] and \[p6,1\] together with the identity , it follows that $z_{t_\pm}(\lambda)$ is a resonance of ${{-\Delta}}_{\widetilde M}$ near $t_\pm(k)$ if and only if $\lambda$ is a characteristic value of $$I + \mathcal{T}_{\pm \widetilde M} \big( z_{t_-}(\lambda) \big) = I \pm
J \sqrt{\frac{2}{\pi}} \, {\rm Hol}(\lambda).$$ Since the operator ${\rm Hol}(\lambda)$ is holomorphic in the open disk $D(0,\varepsilon_0)^\ast \cup \lbrace 0 \rbrace$ with values in ${{\textbf{\textup{S}}_\infty}}\big( \ell^2(\mathcal{V})
\big)$, then Theorem \[t1\] holds by applying Proposition \[p,a1\] with
- $\mathcal{D} = \Omega_r^\ast \cup \lbrace 0 \rbrace$, $\Omega_r^\ast \subseteq D(0,\varepsilon_0)^\ast$,
- $Z = \lbrace 0 \rbrace$,
- $F = I + \mathcal{T}_{\pm \widetilde M} \big( z_{t_-}(\cdot) \big)$.
This concludes the proof of Theorem \[t1\].
Appendix {#sa}
========
We recall some tools we need on characteristic values of finite meromorphic operator-valued functions. For more details on the subject, we refer for instance to [@go] and the book [@goh Section 4]. The content of this section follows [@goh Section 4].
Let $\mathscr{H}$ be separable Hilbert space, and let $\mathscr{L}(\mathscr{H})$ (resp. ${\rm GL}(\mathscr{H})$) denote the set of bounded (resp. invertible) linear operators in $\mathscr{H}$.
Let $\mathcal{U}$ be a neighborhood of a fixed point $w \in {{\mathbb{C}}}$, and $F : \mathcal{U} \setminus \lbrace w \rbrace \longrightarrow \mathscr{L}(\mathscr{H})$ be a holomorphic operator-valued function. The function $F$ is said to be finite meromorphic at $w$ if its Laurent expansion at $w$ has the form $$F(z) = \sum_{n = m}^{+\infty} (z - w)^n A_n, \quad m > - \infty,$$ where (if $m < 0$) the operators $A_m, \ldots, A_{-1}$ are of finite rank. Moreover, if $A_0$ is a Fredholm operator, then the function $F$ is said to be Fredholm at $w$. In that case, the Fredholm index of $A_0$ is called the Fredholm index of $F$ at $w$.
We have the following proposition:
[[@goh Proposition 4.1.4]]{}\[p,a1\] Let $\mathcal{D} \subseteq \mathbb{C}$ be a connected open set, $Z \subseteq \mathcal{D}$ be a closed and discrete subset of $\mathcal{D}$, and $F : \mathcal{D} \longrightarrow
\mathscr{L}(\mathscr{H})$ be a holomorphic operator-valued function in $\mathcal{D} \backslash
Z$. Assume that:
- $F$ is finite meromorphic on $\mathcal{D}$ (i.e. it is finite meromorphic near each point of $Z$),
- $F$ is Fredholm at each point of $\mathcal{D}$,
- there exists $w_0 \in \mathcal{D} \backslash Z$ such that $F(w_0)$ is invertible.
Then, there exists a closed and discrete subset $Z'$ of $\mathcal{D}$ such that:
- $Z \subseteq Z'$,
- $F(z)$ is invertible for each $z \in \mathcal{D} \backslash Z'$,
- $F^{-1} : \mathcal{D} \backslash Z' \longrightarrow {\rm GL}(\mathscr{H})$ is finite meromorphic and Fredholm at each point of $\mathcal{D}$.
In the setting of Proposition \[p,a1\], we define the characteristic values of $F$ and their multiplicities as follows:
\[d,a1\] The points of $Z'$ where the function $F$ or $F^{-1}$ is not holomorphic are called the characteristic values of $F$. The multiplicity of a characteristic value $w_0$ is defined by $${\rm mult}(w_0) := \frac{1}{2i\pi} \textup{Tr} \oint_{\vert w - w_0 \vert = \rho}
F'(z)F(z)^{-1} dz,$$ where $\rho > 0$ is chosen small enough so that $\big\lbrace w \in {{\mathbb{C}}}: \vert w -
w_0 \vert \leq \rho \big\rbrace \cap Z' = \lbrace w_0 \rbrace$.
According to Definition \[d,a1\], if the function $F$ is holomorphic in $\mathcal{D}$, then the characteristic values of $F$ are just the complex numbers $w$ where the operator $F(w)$ is not invertible. Then, results of [@go] and [@goh Section 4] imply that ${\rm mult}(w)$ is an integer.
Let $\Omega \subseteq \mathcal{D}$ be a connected domain with boundary $\partial \Omega$ not intersecting $Z'$. The sum of the multiplicities of the characteristic values of the function $F$ lying in $\Omega$ is called [*the index of $F$ with respect to the contour $\partial \Omega$*]{} and is defined by $$\label{eqa,2}
Ind_{\partial \Omega} \hspace{0.5mm} F := \frac{1}{2i\pi} \textup{Tr}
\oint_{\partial \Omega} F'(z)F(z)^{-1} dz = \frac{1}{2i\pi} \textup{Tr}
\oint_{\partial \Omega} F(z)^{-1} F'(z) dz.$$
[**Acknowledgements:**]{} O. Bourget is supported by the Chilean Fondecyt Grant $1161732$. D. Sambou is supported by the Chilean Fondecyt Grant $3170411$.
The authors express their gratitude to S. Golénia for bringing to their attention the paper [@af], and to V. Bruneau and S. Kupin for their helpful discussions and valuable suggestions.
[99]{}
<span style="font-variant:small-caps;">M. Aizenman, S. Warzel</span>, *Absence of mobility edge for the Anderson random potential on tree graphs at weak disorder*, EPL [**96**]{} 37004 (2011).
<span style="font-variant:small-caps;">M. Aizenman, S. Warzel</span>, *Resonant delocalization for random Schrödinger operators on tree graphs*, J. Eur. Math. Soc. [**15**]{} (2013) 1167-1222.
<span style="font-variant:small-caps;">C. Allard</span>, *Asymptotic Completeness via Mourre Theory for a Schrödinger Operator on a Binary Tree Graph*, Master’s thesis, University of British Columbia, April 1997.
<span style="font-variant:small-caps;">C. Allard, R. Froese</span>, *A Mourre estimate for a Schrödinger operator on a binary tree*, Rev. in Math. Phys. **12** (12) (2000), 1655-1667.
<span style="font-variant:small-caps;">N. Anantharaman, E. Le Masson</span>, *Quantum ergodicity on large regular graphs*, Duke Math. J. **164** (4) (2015), 723-765.
<span style="font-variant:small-caps;">J.-F. Bony, V. Bruneau, G. Raikov</span>, *Resonances and Spectral Shift Function near the Landau levels*, Ann. Inst. Fourier, **57** (2) (2007), 629-671.
<span style="font-variant:small-caps;">J.-F. Bony, V. Bruneau, G. Raikov</span>, *Counting function of characteristic values and magnetic resonances*, Commun. PDE. **39** (2014), 274-305.
<span style="font-variant:small-caps;">J. Breuer</span>, *Singular continuous spectrum for the Laplacian on certain sparse trees*, Commun. Math. Phys. **269** (2007), 851-857.
<span style="font-variant:small-caps;">J. Breuer, M. Keller</span>, *Spectral analysis of certain spherically homogeneous graphs*, Operators and Matrices **7** (4) (2013), 825-847.
<span style="font-variant:small-caps;">R. Froese, D. Lee, C. Sadel, W. Spitzer, G. Stolz</span>, *Localization for transversally periodic random potentials on binary trees*, To appear in Journal of Spectral Theory, arXiv:1408.3961
<span style="font-variant:small-caps;">R. Froese, F. Halasan, D. Hasler</span>, *Absolutely continuous spectrum for the Anderson model on a product of a tree with a finite graph*, J. Func. Anal., [**262**]{} (3) (2012), 1011-1042.
<span style="font-variant:small-caps;">I. Gohberg, E. I. Sigal</span>, *An operator generalization of the logarithmic residue theorem and Rouché’s theorem*, Mat. Sb. (N.S.) **84** (126) (1971), 607-629.
<span style="font-variant:small-caps;">I. Gohberg, S. Goldberg, M. A. Kaashoek</span>, *Classes of Linear Operators*, Operator Theory, Advances and Applications, vol. I Birkhäuser Verlag, Bassel, 1990.
<span style="font-variant:small-caps;">I. Gohberg, J. Leiterer</span>, *Holomorphic operator functions of one variable and applications*, Operator Theory, Advances and Applications, vol. **192** Birkhäuser Verlag, 2009, Methods from complex analysis in several variables.
<span style="font-variant:small-caps;">I. Gohberg, S. Goldberg, N. Krupnik</span>, *Traces and Determinants of Linear Operators*, Operator Theory, Advances and Applications, vol. **116** Birkhäuser Verlag, 2000.
<span style="font-variant:small-caps;">K. Ito, A. Jensen</span>, *A complete classification of threshold properties for one-dimensional discrete Schrödinger operators*, Rev. in Math. Phys. [**27**]{} (1) (2015), 1550002 (45 pages).
<span style="font-variant:small-caps;">A. Klein</span>, *Extended states in the Anderson model on the Bethe lattice*, Adv. Math. [**133**]{} (1) (1998), 163–184.
<span style="font-variant:small-caps;">A. S. Kostenko, M. M. Malamud, H. Neidhardt, P. Exner</span>, *Infinite Quantum Graphs*, Doklady Mathematics [**95**]{} (2017) (1), 31-36.
<span style="font-variant:small-caps;">D. Sambou</span>, *On eigenvalue accumulation for non-self-adjoint magnetic operators*, J. Maths Pures et Appl. **108** (2017), 306-332.
<span style="font-variant:small-caps;">O. Rojo</span>, *On the spectra of certain rooted trees*, Linear Agebra Appli. **414** (2006), 218-243.
<span style="font-variant:small-caps;">O. Rojo</span>, *The spectra of some trees and bounds for the largest eigenvalue of any tree*, Linear Agebra Appli. **414** (2006), 199-217.
<span style="font-variant:small-caps;">O. Rojo, M. Robbiano</span>, *An explicit formula for eigenvalues of Bethe trees and upper bounds on the largest eigenvalue of any tree*, Linear Agebra Appli. **427** (2007), 138-150.
<span style="font-variant:small-caps;">M. Shamis</span>, *Resonant delocalization on the Bethe strip*, Ann. Henri Poincaré **15** (8) (2014), 1549-1567.
<span style="font-variant:small-caps;">B. Simon</span>, *Trace ideals and their applications*, Lond. Math. Soc. Lect. Not. Series, **35** (1979), Cambridge University Press.
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'auto\_generated.bib'
title: 'Search for supersymmetry with razor variables in pp collisions at sqrt(s)=7 TeV'
---
=1
$Revision: 262505 $ $HeadURL: svn+ssh://svn.cern.ch/reps/tdr2/papers/SUS-12-005/trunk/SUS-12-005.tex $ $Id: SUS-12-005.tex 262505 2014-09-30 15:51:22Z alverson $
Introduction\[sec:prd-intro\]
=============================
Extensions of the standard model (SM) with softly broken supersymmetry (SUSY) [@Ramond; @Golfand; @Volkov; @Wess; @Fayet] predict new fundamental particles that are superpartners of the SM particles. Under the assumption of $R$-parity [@bib-rparity] conservation, searches for SUSY particles at the fermilab Tevatron [@:2007ww; @Aaltonen:2008rv] and the CERN LHC [@Collaboration:2011xk; @daCosta:2011qk; @Aad:2011hh; @Aad:2011xm; @RA2; @alphaT; @:2011wb; @Chatrchyan:2011bz; @Chatrchyan:2012qka; @Chatrchyan:2012mea; @:2012th; @:2012jx; @ATLAS:2012ai; @ATLAS:2012ag; @Aad:2012rt; @:2012cwa; @Aad:2012cz] have focused on event signatures with energetic hadronic jets and leptons from the decays of pair-produced squarks $\PSq$ and gluinos $\PSg$. Such events frequently have large missing transverse energy () resulting from the stable weakly interacting superpartners, one of which is produced in each of the two decay chains.
In this paper, we present the detailed methodology of an inclusive search for SUSY based on the razor kinematic variables [@razor2010; @rogan]. A summary of the results of this search, based on 4.7of $\Pp\Pp$ collision data at $\sqrt{s} =
7$collected with the CMS detector at the LHC, can be found in Ref. [@PhysRevLett.111.081802]. The search is sensitive to the production of pairs of heavy particles, provided that the decays of these particles produce significant . The jets in each event are cast into two disjoint sets, referred to as “megajets”.
The razor variables ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$, defined in Section \[sec:prd-razor\], are calculated from the four-momenta of these megajets event-by-event, and the search is performed by determining the expected distributions of SM processes in the two-dimensional (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) razor plane. A critical feature of the razor variables is that they are computed in the approximate center-of-mass frame of the produced superpartner candidates.
The megajets represent the visible part of the decay chain of pair-produced superpartners, each of which decays to one or more visible SM particles and one stable, weakly interacting lightest SUSY particle (LSP), here taken to be the lightest neutralino $\chiz_1$. In this framework the reconstructed products of the decay chain of each originally produced superpartner are collected into one megajet. Every topology can then be described kinematically by the simplest example of squark-antisquark production with the direct two-body squark decay $\PSq \to \Pq \chiz_1$, denoted a “dijet plus " final state, to which the razor variables strictly apply.
The strategy and execution of the search is summarized as follows:
1. Events with two reconstructed jets at the hardware-based first level trigger (L1) are processed by a dedicated set of algorithms in the high-level trigger (HLT). From the jets and leptons reconstructed at the HLT level, the razor variables ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ are calculated and their values are used to determine whether to retain the event for further offline processing. A looser kinematic requirement is applied for events with electrons or muons, due to the smaller rate of SM background for these processes. The correspondence between the HLT and offline reconstruction procedures allows events of interest to be selected more efficiently than is possible with an inclusive multipurpose trigger.
2. In the offline environment, leptons and jets are reconstructed, and a tagging algorithm is applied to identify those jets likely to have originated from a bottom-quark jet ($\cPqb$ jet).
3. The reconstructed objects in each event are combined into two megajets, which are used to calculate the variables ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$. Several baseline kinematic requirements are applied to reduce the number of misreconstructed events and to ensure that only regions of the razor plane where the trigger is efficient are selected.
4. Events are assigned to final state “boxes” based on the presence or absence of a reconstructed lepton. This box partitioning scheme allows us to isolate individual SM background processes based on the final-state particle content and kinematic phase space; we are able to measure the yield and the distribution of events in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) razor plane for different SM backgrounds. Events with at least one tagged $\cPqb$ jet are considered in a parallel analysis focusing on a search for the superpartners of third-generation quarks. In total, we consider 12 mutually exclusive final-state boxes: dielectron events (ELE-ELE), electron-muon events (ELE-MU), dimuon events (MU-MU), single-electron events (ELE), single-muon events (MU), and events with no identified electron or muon (HAD), each inclusive or with a $\cPqb$-tagged jet.
5. For each box we use the low (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) region of the razor plane, where negligible signal contributions are expected, to determine the shape and normalization of the various background components. An analytic model constructed from these results is used to predict the SM background over the entire razor plane.
6. The data are compared with the prediction for the background in the sensitive regions of the razor plane and the results are used to constrain the parameter space of SUSY models.
This paper is structured as follows. The definition of the razor variables is given in Section \[sec:prd-razor\]. The trigger and offline event selection are discussed in Section \[sec:datana\]. The features of the signal and background kinematic distributions are described in Section \[sec:mc-ana\]. In Section \[sec:BKG2011\] we describe the sources of SM background, and in Section \[sec:fits\] the analytic model used to characterize this background in the signal regions. Systematic uncertainties are discussed in Section \[sec:systematics\]. The interpretation of the results is presented in Section \[sec:INTERP2011\] in terms of exclusion limits on squark and gluino production in the context both of the constrained minimal SUSY model (CMSSM) [@Chamseddine:1982jx; @Barbieri:1982eh; @Hall:1983iz] and for some simplified model spectra (SMS) [@ArkaniHamed:2007fw; @Alwall-2; @Alwall-1; @Alwall:2008va; @Alves:2011wf]. Section \[sec:summary\] contains a summary. For the CMSSM, exclusion limits are provided as a function of the universal scalar and fermion mass values at the unification scale, respectively denoted ${\ensuremath{m_0}\xspace}$ and ${\ensuremath{m_{1/2}}\xspace}$. For the SMS, limits are provided in terms of the masses of the produced SUSY partner and the LSP.
The razor approach\[sec:prd-razor\]
===================================
The razor kinematic variables are designed to be sensitive to processes involving the pair-production of two heavy particles, each decaying to an unseen particle plus jets. Such processes include SUSY particle production with various decay chains, the simplest example of which is the pair production of squarks, where each squark decays to a quark and the LSP, with the LSP assumed to be stable and weakly interacting. In processes with two or more undetected energetic final-state particles, it is not possible to fully reconstruct the event kinematics. Event-by-event, one cannot make precise assignments of the reconstructed final-state particles (leptons, jets, and undetected neutrinos and LSPs) to each of the original superpartners produced. For a given event, there is not enough information to determine the mass of the parent particles, the subprocess center-of-mass energy $\sqrt{\hat{s}}$, the center-of-mass frame of the colliding protons, or the rest frame of the decay of either parent particle. As a result, it is challenging to distinguish between SUSY signal events and SM background events with energetic neutrinos, even though the latter involve different topologies and mass scales. It is also challenging to identify events with instrumental sources of that can mimic the signal topology.
The razor approach [@razor2010; @rogan] addresses these challenges through a novel treatment of the event kinematics. The key points of this approach are listed below.
- The visible particles (leptons and jets) are used to define two *megajets*, each representing the visible part of a parent particle decay. The megajet reconstruction ignores details of the decay chains in favor of obtaining the best correspondence between a signal event candidate and the presumption of a pair-produced heavy particle that undergoes two-body decay.
- Lorentz-boosted reference frames are defined in terms of the megajets. These frames approximate, event-by-event, the center-of-mass frame of the signal subprocess and the rest frames of the decays of the parent particles.The kinematic quantities in these frames can be used to extract the relevant SUSY mass scales.
- The razor variables, ${\ensuremath{M_R}\xspace}$, ${\ensuremath{M_\mathrm{T}^R}\xspace}$, and ${\ensuremath{R}\xspace}\equiv {\ensuremath{M_\mathrm{T}^R}\xspace}/{\ensuremath{M_R}\xspace}$, are computed from the megajet four-momenta and the in the event. The ${\ensuremath{M_R}\xspace}$ variable is an estimate of an overall mass scale, which in the limit of massless decay products equals the mass of the heavy parent particle. It contains both longitudinal and transverse information, and its distribution peaks at the true value of the new-physics mass scale. The razor variable ${\ensuremath{M_\mathrm{T}^R}\xspace}$ is defined entirely from transverse information: the transverse momenta ($\pt$) of the megajets and the . This variable has a kinematic endpoint at the same underlying mass scale as the ${\ensuremath{M_R}\xspace}$ mean value. The ratio ${\ensuremath{R}\xspace}$ quantifies the flow of energy in the plane perpendicular to the beam and the partitioning of momentum between visible and invisible particles.
- The shapes of the distributions in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane are described for the SM processes. Razor variable distributions exhibit peaks for most SM backgrounds, as a result of turn-on effects from trigger and selection thresholds as well as of the relevant heavy mass scales for SM processes, namely the top quark mass and the $\PW$ and $\cPZ$ boson masses. However, compared with signals involving heavier particles and new-physics sources of , the SM distributions peak at smaller values of the razor variables. For values of the razor variables above the peaks, the SM background distributions (and also the signal distributions) exhibit exponentially falling behavior in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane. Hence, the asymptotic behavior of the razor variables is determined by a combination of the parton luminosities and the intrinsic sources of . The multijet background from processes described by quantum chromodynamics (QCD), which contains the smallest level of intrinsic amongst the major sources of SM background, has the steepest exponential fall-off. Backgrounds with energetic neutrinos from $\PW/\cPZ$ boson and top-quark production exhibit a slower fall-off and resemble each other closely in the asymptotic regime. Thus, razor signals are characterized by peaks in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane on top of exponentially falling SM background distributions. Any SUSY search based on razor variables is then more similar to a “bump-hunt”, e.g., a search for heavy resonances decaying to two jets [@Harris:2011bh], than to a traditional SUSY search. This justifies the use of a shape analysis, based on an analytic fit of the background, as described in Section \[sec:fits\].
Razor megajet reconstruction
----------------------------
The razor megajets are defined by dividing the reconstructed jets of each event into two partitions. Each partition contains at least one jet. The megajet four-momenta are defined as the sum of the four-momenta of the assigned jets. Of all the possible combinations, the one that minimizes the sum of the squared-invariant-mass values of the two megajets is selected. In simulated event samples, this megajet algorithm is found to be stable against variations in the jet definition and it provides an unbiased description of the visible part of the two decay chains in SUSY signal events. The inclusive nature of the megajets allows an estimate of the SM background in the razor plane.
Reconstructed leptons in the final state can be included as visible objects in the reconstruction of the megajets, or they can be treated as invisible, i.e., as though they are escaping weakly interacting particles [@razor2010]. For SM background processes such as $\PW(\ell\cPgn)$+jets, the former choice yields more transversely balanced megajets and lower values of ${\ensuremath{R}\xspace}$. If the leptons are treated as invisible in these processes, the corresponds to the entire $\PW$ boson $\pt$ value, similar to the case of $\cPZ (\cPgn
\cPagn)$+jets events.
Razor variables
---------------
To the extent that the reconstructed pair of megajets accurately reflects the visible portion of the underlying parent particle decays, the kinematics of the event are equivalent to that of the pair production of heavy squarks $\PSq_1$, $\PSq_2$, with $\PSq_i\to \cPq_i \PSGczDo$, where $\PSGczDo$ denotes the LSP and $\cPq_i$ denotes the visible products of the decays as represented by the megajets.
The razor analysis approximates the unknown center-of-mass and parent particle rest frames with a razor frame defined unambiguously from measured quantities in the laboratory frame. Two observables ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{M_\mathrm{T}^R}\xspace}$ estimate the heavy mass scale ${\ensuremath{M_\Delta}\xspace}$. Consider the two visible four-momenta written in the rest frame of the respective parent particles: $$\begin{aligned}
p_{\cPq_1} &= \left(\frac{m_{\PSq}^2+m_{\cPq_1}^2-m_{\chiz_1}^2}{2m_{\PSq}},\; \frac{{\ensuremath{M_\Delta}\xspace}}{2}\,\hat{u}_{\cPq_1}\right),\\
p_{\cPq_2} &= \left(\frac{m_{\PSq}^2+m_{\cPq_2}^2-m_{\chiz_1}^2}{2m_{\PSq}},\; \frac{{\ensuremath{M_\Delta}\xspace}}{2}\,\hat{u}_{\cPq_2}\right).\end{aligned}$$ where $\hat{u}_{\cPq_i}$ (${i} = 1,2$) is a unit three-vector and $m_\cPq{}_i$ represents the mass corresponding to the megajet, e.g., the top-quark mass for $\PSQt \to \cPqt \PSGczDo$. Here we have parameterized the magnitude of the three-momenta by the mass scale ${\ensuremath{M_\Delta}\xspace}$, where $${\ensuremath{M_\Delta}\xspace}^2 \equiv \frac{\left[m_{\PSq}^2-(m_\cPq+m_{\chiz_1})^2\right]\left[m_{\PSq}^2-(m_\cPq-m_{\chiz_1} )^2\right]}{m_{\PSq}^2}.$$ In the limit of massless megajets we then have ${\ensuremath{M_\Delta}\xspace}= (m_{\PSq}^2
- m_{\chiz_1}^2)/m_{\PSq}$ and the four-momenta reduce to $$\begin{aligned}
p_{\cPq_1} &= \frac{{\ensuremath{M_\Delta}\xspace}}{2}(1,\; \hat{u}_{\cPq_1}),\\
p_{\cPq_2} &= \frac{{\ensuremath{M_\Delta}\xspace}}{2}(1,\; \hat{u}_{\cPq_2}).\end{aligned}$$
The razor variable ${\ensuremath{M_R}\xspace}$ is defined in terms of the momenta of the two megajets by $$\label{eq:MRstar}
{\ensuremath{M_R}\xspace}\equiv
\sqrt{(\abs{\vec{p}_{\cPq_{1}}}+\abs{\vec{p}_{\cPq_{2}}})^2 -(p^{\cPq_1}_z+p^{\cPq_2}_z)^2}.$$ where $\vec{p}_{\cPq_i}$ is the momentum of megajet $\cPq_i$ (${i} = 1,2$) and $p^{\cPq_i}_z$ is its component along the beam direction.
For massless megajets, ${\ensuremath{M_R}\xspace}$ is invariant under a longitudinal boost. It is always possible to perform a longitudinal boost to a razor frame where $p^{\cPq_1}_z+p^{\cPq_2}_z$ vanishes, and ${\ensuremath{M_R}\xspace}$ becomes just the scalar sum of the megajet three-momenta added in quadrature. For heavy particle production near threshold, the three-momenta in this razor frame do not differ significantly from the three-momenta in the actual parent particle rest frames. Thus, for SUSY signal events, ${\ensuremath{M_R}\xspace}$ is an estimator of ${\ensuremath{M_\Delta}\xspace}$, and for simulated samples we find that the distribution of ${\ensuremath{M_R}\xspace}$ indeed peaks around the true value of ${\ensuremath{M_\Delta}\xspace}$. This definition of ${\ensuremath{M_R}\xspace}$ is improved with respect to the one used in Ref. [@razor2010], to avoid configurations where the razor frame is unphysical.
The razor observable ${\ensuremath{M_\mathrm{T}^R}\xspace}$ is defined as $${\ensuremath{M_\mathrm{T}^R}\xspace}\equiv \sqrt{ \frac{\ETmiss(\pt^{\cPq_1}+\pt^{\cPq_2}) -
\VEtmiss {\cdot}
(\ptvec^{\,\cPq_1}+\ptvec^{\,\cPq_2})}{2}},$$ where $\ptvec^{\,\cPq_i}$ is the transverse momentum of megajet ${\cPq_i}$ ($i = 1,2$) and $\pt^{\cPq_i}$ is the corresponding magnitude; similarly, $\VEtmiss$ is the missing transverse momentum in the event and its magnitude.
Given a global estimator ${\ensuremath{M_R}\xspace}$ and a transverse estimator ${\ensuremath{M_\mathrm{T}^R}\xspace}$, the razor dimensionless ratio is defined as $${\ensuremath{R}\xspace}\equiv \frac{{\ensuremath{M_\mathrm{T}^R}\xspace}}{{\ensuremath{M_R}\xspace}}.$$
For signal events, ${\ensuremath{M_\mathrm{T}^R}\xspace}$ has a maximum value (a kinematic endpoint) at ${\ensuremath{M_\Delta}\xspace}$, so ${\ensuremath{R}\xspace}$ has a maximum value of approximately one. Thus, together with the shape of ${\ensuremath{M_R}\xspace}$ peaking at ${\ensuremath{M_\Delta}\xspace}$, this behavior is in stark contrast with, for example, QCD multijet background events, whose distributions in both ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ fall exponentially. These properties allow us to identify a region of the two-dimensional (2D) razor space where the contributions of the SM background are reduced while those of signal events are enhanced.
{width="48.00000%"} {width="48.00000%"} {width="48.00000%"} {width="48.00000%"}
SUSY and SM in the razor plane
------------------------------
The expected distributions of the main SM backgrounds in the razor plane, based on simulation, are shown in Fig. \[fig:raz-plane\], along with the results from the CMSSM low-mass benchmark model LM6 [@PTDR2], for which ${\ensuremath{M_\Delta}\xspace}= 831$. The peaking behavior of the signal events at ${\ensuremath{M_R}\xspace}\approx {\ensuremath{M_\Delta}\xspace}$, and the exponential fall-off of the SM distributions with increasing ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$, are to be noted. For both signal and background processes, events with small values of ${\ensuremath{M_R}\xspace}$ are suppressed because of a requirement that there be at least two jets above a certain threshold in (Section \[sec:BOX2011\]).
In the context of SMS, we refer to the pair production of squark pairs $\PSq$, $\PSq^*$, followed by $\PSq \to \Pq~\chiz_1$, as “T2" scenarios [@SUS-11-016], where the $\tilde\cPq^*$ state is the charge conjugate of the $\tilde\cPq$ state. Figure \[fig:T2-tutorial\] (a) shows a diagram for heavy-squark pair production. The distributions of ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ for different LSP masses are shown in Figs. \[fig:T2-tutorial\] (b) and (c). Figure \[fig:T2-tutorial\] (d) shows the distribution of signal events in the razor plane. The colored bands running from top left to bottom right show the approximate SM background constant-yield contours. The associated numbers indicate the SM yield suppression relative to the reference line marked “1”. Based on these kinematic properties, a 2D analytical description of the SM processes in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane is developed.
{width="48.00000%"} {width="48.00000%"} {width="48.00000%"} {width="48.00000%"}
Data taking and event selection\[sec:datana\]
=============================================
The CMS apparatus\[sec:cmsdet\]
-------------------------------
A hallmark of the CMS detector [@CMS:2008zzk] is its superconducting solenoid magnet, of 6 m internal diameter, providing a field of 3.8 T. The silicon pixel and strip tracker, the crystal electromagnetic calorimeter (ECAL), and the brass/scintillator hadron calorimeter (HCAL) are contained within the solenoid. Muons are detected in gas-ionization detectors embedded in the steel flux-return yoke, based on three different technologies: drift tubes, resistive plate chambers, and cathode strip chambers (CSCs). The ECAL has an energy resolution better than 0.5% above 100. The combination of the HCAL and ECAL provides jet energy measurements with a resolution $\Delta E/E \approx 100\% / \sqrt{E/\GeV} \oplus 5\%$.
The CMS experiment uses a coordinate system with the origin located at the nominal collision point, the $x$ axis pointing towards the center of the LHC ring, the $y$ axis pointing up (perpendicular to the plane containing the LHC ring), and the $z$ axis along the counterclockwise beam direction. The azimuthal angle, $\phi$, is measured with respect to the $x$ axis in the $(x,y)$ plane, and the polar angle, $\theta$, is defined with respect to the $z$ axis. The pseudorapidity is $\eta =
-\ln [\tan(\theta / 2)]$.
For the data used in this analysis, the peak luminosity of the LHC increased from to over . For the data collected between , the increase was achieved by increasing the number of bunches colliding in the machine, keeping the average number of interactions per crossing at about 7. For the rest of the data, the increase in the instantaneous luminosity was achieved by increasing the number and density of the protons in each bunch, leading to an increase in the average number of interactions per crossing from around 7 to around 17. The presence of multiple interactions per crossing was taken into account in the CMS Monte Carlo (MC) simulation by adding a random number of minimum bias events to the hard interactions, with the multiplicity distribution matching that in data.
Trigger selection\[sec:TRIGGER2011\]
------------------------------------
The CMS experiment uses a two-stage trigger system, with events flowing from the L1 trigger at a rate up to 100. These events are then processed by the HLT computer farm. The HLT software selects events for storage and offline analysis at a rate of a few hundred Hz. The HLT algorithms consist of sequences of offline-style reconstruction and filtering modules.
The 2010 CMS razor-based inclusive search for SUSY [@razor2010] used triggers based on the scalar sum of jet $\pt$, $\HT$, for hadronic final states and single-lepton triggers for leptonic final states. Because of the higher peak luminosity of the LHC in 2011, the corresponding triggers for 2011 had higher thresholds. To preserve the high sensitivity of the razor analysis, CMS designed a suite of dedicated razor triggers, implemented in the spring of 2011. The total integrated luminosity collected with these triggers was 4.7at $\sqrt{s} = 7$.
The razor triggers apply thresholds to the values of ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R}\xspace}$ driven by the allocated bandwidth. The algorithms used for the calculation of ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R}\xspace}$ are based on calorimetric objects. The reconstruction of these objects is fast enough to satisfy the stringent timing constraints imposed by the HLT.
Three trigger categories are used: hadronic triggers, defined by applying moderate requirements on ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R}\xspace}$ for events with at least two jets with $\pt>$ 56; electron triggers, similar to the hadronic triggers, but with looser requirements for ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R}\xspace}$ and requiring at least one electron with $\pt>$10satisfying loose isolation criteria; and muon triggers, with similar ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R}\xspace}$ requirements and at least one muon with $\abs{\eta}<2.1$ and $\pt>$ 10. All these triggers have an efficiency of $(98 \pm 2)\%$ in the kinematic regions used for the offline selection.
In addition, control samples are defined using several non-razor triggers. These include prescaled inclusive hadronic triggers, hadronic multijet triggers, hadronic triggers based on $\HT$, and inclusive electron and muon triggers.
Physics object reconstruction\[sec:RECO2011\]
---------------------------------------------
Events are required to have at least one reconstructed interaction vertex [@TRK-10-005]. When multiple vertices are found, the one with the highest scalar sum of charged track $\pt^2$ is taken to be the event interaction vertex. Jets are reconstructed offline from calorimeter energy deposits using the infrared-safe anti- [@antikt] algorithm with a distance parameter ${\ensuremath{R}\xspace}=0.5$. Jets are corrected for the non-uniformity of the calorimeter response in energy and $\eta$ using corrections derived from data and simulations and are required to have $\pt> 40$and $\abs{\eta} < 3.0$ [@JES]. To match the trigger requirements, the $\pt$ of the two leading jets is required to be greater than 60. The jet energy scale uncertainty for these corrected jets is 5% [@JES]. The $\ETmiss$ is defined as the negative of the vector sum of the transverse energies () of all the particles found by the particle-flow algorithm [@PFMET].
Electrons are reconstructed using a combination of shower shape information and matching between tracks and electromagnetic clusters [@CMS_e]. Muons are reconstructed using information from the muon detectors and the silicon tracker and are required to be consistent with the reconstructed primary vertex [@CMS_mu].
The selection criteria for electrons and muons are considered to be [*tight*]{} if the electron or muon candidate is isolated, satisfies the selection requirements of Ref. [@CMS_WZ], and lies within $\abs{\eta}<2.5$ and $\abs{\eta}<2.1$, respectively. [*Loose*]{} electron and muon candidates satisfy relaxed isolation requirements.
Selection of good quality data
------------------------------
The 4.7integrated luminosity used in this analysis is certified as having a fully functional detector. Events with various sources of noise in the ECAL or HCAL detectors are rejected using either topological information, such as unphysical charge sharing between neighboring channels, or timing and pulse shape information. The last requirement exploits the difference between the shapes of the pulses that develop from particle energy deposits in the calorimeters and from noise events [@METJINST]. Muons produced from proton collisions upstream of the detector (beam halo) can mimic proton-proton collisions with large $\ETm$ and are identified using information obtained from the CSCs. The geometry of the CSCs allows efficient identification of beam halo muons, since halo muons that traverse the calorimetry will mostly also traverse one or both CSC endcaps. Events are rejected if a significant amount of energy is lost in the masked crystals that constitute approximately 1% of the ECAL, using information either from the separate readout of the L1 hardware trigger or by measuring the energy deposited around the masked crystals. We select events with a well-reconstructed primary vertex and with the scalar $\sum \pt$ of tracks associated to it greater than 10% of the scalar $\sum \pt$ of all jet transverse momenta. These requirements reject 0.003% of an otherwise good inclusive sample of proton-proton interactions (minimum bias events).
Event selection and classification\[sec:BOX2011\]
-------------------------------------------------
Electrons enter the megajet definition as ordinary jets. Reconstructed muons are not included in the megajet grouping because, unlike electrons, they are distinguished from jets in the HLT. This choice also allows the use of $\PW(\Pgm\cPgn)$+jets events to constrain and study the shape of $\cPZ(\cPgn\cPagn)$+jets events in fully hadronic final states.
The megajets are constructed as the sum of the four-momenta of their constituent objects. After considering all possible partitions into two megajets, the combination is selected that has the smallest sum of megajet squared-invariant-mass values.
The variables ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ are calculated from the megajet four-momenta. The events are assigned to one of the six final state [*boxes*]{} according to whether the event has zero, one, or two isolated leptons, and according to the lepton flavor (electrons and muons), as shown in Table \[tab:boxes\]. The lepton $\pt$, ${\ensuremath{M_R}\xspace}$, and ${\ensuremath{R^2}\xspace}$ thresholds for each of the boxes are chosen so that the trigger efficiencies are independent of ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$.
[l c]{}\
ELE-MU (loose-tight) & $\pt > 20$, $\pt > 15$\
MU-MU (loose-loose) & $\pt > 15$, $\pt > 10$\
ELE-ELE (loose-tight) & $\pt > 20$, $\pt > 10$\
MU (tight) & $\pt > 12$\
ELE (loose) & $\pt > 20$\
\
The requirements given in Table \[tab:boxes\] determine the [ *full analysis regions*]{} of the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane for each box. These regions are large enough to allow an accurate characterization of the background, while maintaining efficient triggers. To prevent ambiguities when an event satisfies the selection requirements for more than one box, the boxes are arranged in a predefined hierarchy. Each event is uniquely assigned to the first box whose criteria the event satisfies. Table \[tab:boxes\] shows the box-filling order followed in the analysis.
Six additional boxes are formed with the requirement that at least one of the jets with $\pt> 40$and $\abs{\eta} < 3.0$ be tagged as a b jet, using an algorithm that orders the tracks in a jet by their impact parameter significance and discriminates using the track with the second-highest significance [@BTAG]. This algorithm has a tagging efficiency of about 60%, evaluated using $\cPqb$ jets containing muons from semileptonic decays of $\cPqb$ hadrons in data, and a misidentification rate of about 1% for jets originating from u, d, and s quarks or from gluons, and of about 10% for jets coming from c quarks [@BTAG]. The combination of these six boxes defines an inclusive event sample with an enhanced heavy-flavor content.
Signal and standard model background modeling\[sec:mc-ana\]
===========================================================
The razor analysis is guided by studies of MC event samples generated with the v6.426 [@pythia] (with Z2 tune) and [v4.22]{} [@Maltoni:2002qb] programs, using the CTEQ6 parton distribution functions (PDF) [@Pumplin:2002vw]. Events generated with are processed with [@pythia] to provide parton showering, hadronization, and the underlying event description. The matrix element/parton shower matching is performed using the approach described in Ref. [@mlm]. Generated events are processed with the [@G4] based simulation of the CMS detector, and then reconstructed with the same software used for data.
The simulation of the $\ttbar$, $\PW$+jets, $\cPZ$+jets, single-top ($s$, $t$, and $\cPqt$–$\PW$ channels), and diboson samples is performed using . The events containing top-quark pairs are generated accompanied by up to three extra partons in the matrix-element calculation [@Alwall:2007st]. Multijet samples from QCD processes are produced using .
To generate SUSY signal MC events in the context of the CMSSM, the mass spectrum is first calculated with the <span style="font-variant:small-caps;">softsusy</span> program [@softsusy] and the decays with the <span style="font-variant:small-caps;">sus-hit</span> [@Susyhit] package. The generator is used with the SUSY Les Houches Accord (SLHA) interface [@SLHA] to generate the events. The generator-level cross sections and the K-factors for the next-to-leading-order (NLO) cross sections are computed using <span style="font-variant:small-caps;">prospino</span> [@prospino].
We also use SMS MC simulations in the interpretation of the results. In an SMS simulation, a limited set of hypothetical particles is introduced to produce a given topological signature. The amplitude describing the production and decay of these particles is parameterized in terms of the particle masses. Compared with the constrained SUSY models, SMS provide benchmarks that focus on one final-state topology at a time, with a broader variation in the masses determining the final-state kinematics. The SMS are thus useful for comparing search strategies as well as for identifying challenging areas of parameter space where search methods may lack sensitivity. Furthermore, by providing a tabulation of both the signal acceptance and the 95% confidence level () exclusion limit on the signal cross section as a function of the SMS mass parameters, SMS results can be used to place limits on a wide variety of theoretical models beyond SUSY.
The considered SMS scenarios produce multijet final states with or without leptons and $\cPqb$-tagged jets [@SUS-11-016]. While the SUSY terminology is employed, interpretations of SMS scenarios are not restricted to SUSY scenarios.
In the SMS scenarios considered here, each produced particle decays directly to the LSP and SM particles through a two-body or three-body decay. Simplified models that are relevant to inclusive hadronic jets+analyses are gluino pair production with the direct three-body decay $\sGlu \to \cPq \cPaq \chiz_1$ (T1), and squark-antisquark production with the direct two-body decay $\sQua \to \cPq
\chiz_1$ (T2). For $\cPqb$-quark enriched final states, we have considered two additional gluino SMS scenarios, where each gluino is forced into the three-body decay $\sGlu \to \cPqb \cPaqb
\chiz_1$ with 100% branching fraction (T1bbbb), or where each gluino decays through $\sGlu \to \cPqt \cPaqt \chiz_1$ (T1tttt). For $\cPqb$-quark enriched final states we also consider SMS that describe the direct pair production of bottom or top squarks, with the two-body decays $\sBot \to \cPqb \chiz_1$ (T2bb) and $\sTop \to
\cPqt \chiz_1$ (T2tt).
Note that first-generation $\sQua\sQua$ production (unlike $\sQua\sQua^*$ production) is not part of the simplified models used for the interpretation of the razor results, even though it is often the dominant process in the CMSSM for low values of the scalar-mass parameter $m_0$. This is because of the additional complication that the production rate depends on the gluino mass. However, the acceptance for $\sQua\sQua$ production is expected to be somewhat higher than for $\sQua\sQua^*$, so the limits from T2 can be conservatively applied to $\sQua\sQua$ production with analogous decays.
For each SMS, simulated samples are generated for a range of masses of the particles involved, providing a wider spectrum of mass spectra than allowed by the CMSSM. A minimum requirement of ${\cal
O}(100~\GeV)$ on the mass difference between the mother particle and the LSP is applied, to remove phase space where the jets from superpartner decays become soft and the signal is detected only when it is given a boost by associated jet production. By restricting attention to SMS scenarios with large mass differences, we avoid the region of phase space where accurate modeling of initial- and final-state radiation from quarks and gluons is required, and where the description of the signal shape has large uncertainties.
The production of the primary particles in each SMS is modeled with SUSY processes in the appropriate decoupling limit of the other superpartners. In particular, for $\sQua\,\sQua^*$ production, the gluino mass is set to a very large value so that it has a minimal effect on the kinematics of the squarks. The mass spectrum and decay modes of the particles in a specific SMS point are fixed using the SLHA input files, which are processed with v6.426 with Tune Z2 [@Field:2010bc; @UEPAS] to produce signal events as an input to a parameterized fast simulation of the CMS detector [@CMS-DP-2010-039], resulting in simulated samples of reconstructed events for each choice of masses for each SMS. These samples are used for the direct calculation of the signal efficiency, and together with the background model are used to determine the 95% upper bound on the allowed production cross section.
Standard model backgrounds in the (MR MR2) razor plane\[sec:BKG2011\]
=====================================================================
The distributions of SM background events in both the MC simulations and the data are found to be described by the sum of exponential functions of ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ over a large part of the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane. Spurious instrumental effects and QCD multijet production are challenging backgrounds due to difficulties in modeling the high $\pt$ and tails. Nevertheless, these event classes populate predictable regions of the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane, which allows us to study them and reduce their contribution to negligible levels. The remaining backgrounds in the lepton, dilepton, and hadronic boxes are processes with genuine due to energetic neutrinos and charged leptons from vector boson decay, including $\PW$ bosons from top-quark and diboson production. The analysis uses simulated events to characterize the shapes of the SM background distributions, determine the number of independent parameters needed to describe them, and to extract initial estimates of the values of these parameters. Furthermore, for each of the main SM backgrounds a control data sample is defined using ${\approx}250$of data collected at the beginning of the run. These events cannot be used in the search, as the dedicated razor triggers were not available. Instead, events in this run period were collected using inclusive non-razor hadronic and leptonic triggers, thus defining kinematically unbiased data control samples. We use these control samples to derive a data-driven description of the shapes of the background components and to build a background representation using statistically independent data samples; this is used as an input to a global fit of data selected using the razor triggers in a signal-free region of the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) razor plane.
The two-dimensional probability density function $P_{j}({\ensuremath{M_R}\xspace},{\ensuremath{R^2}\xspace})$ describing the ${\ensuremath{R^2}\xspace}$ versus ${\ensuremath{M_R}\xspace}$ distribution of each considered SM process $j$ is found to be well approximated by the same family of functions $F_{j}({\ensuremath{M_R}\xspace},{\ensuremath{R^2}\xspace})$: where $k_{j}$, ${\ensuremath{M}\xspace}^0_{{\ensuremath{R}\xspace},j}$, and ${\ensuremath{R}\xspace}_{0,j}^2$ are free parameters of the background model. After applying a baseline selection in the razor kinematic plane, ${\ensuremath{M_R}\xspace}>{{\ensuremath{M_R}\xspace}^\text{min}}$ and ${\ensuremath{R^2}\xspace}>
{{\ensuremath{R^2}\xspace}_\text{min}}$, this function exhibits an exponential behavior in ${\ensuremath{R^2}\xspace}$ (${\ensuremath{M_R}\xspace}$), when integrated over ${\ensuremath{M_R}\xspace}$ (${\ensuremath{R^2}\xspace}$): $$\begin{aligned}
\int_{{{\ensuremath{R^2}\xspace}_\text{min}}}^{+\infty} F_{j}({\ensuremath{M_R}\xspace}, {\ensuremath{R^2}\xspace}) \rd{\ensuremath{R^2}\xspace}&\sim
\re^{-(a+b\times {{\ensuremath{R^2}\xspace}_\text{min}}){\ensuremath{M_R}\xspace}}, \\
\int_{{{\ensuremath{M_R}\xspace}^\text{min}}}^{+\infty} F_{j}({\ensuremath{M_R}\xspace}, {\ensuremath{R^2}\xspace})
\rd{}M_{{\ensuremath{R}\xspace}} &\sim \re^{-(c+d\times {{\ensuremath{M_R}\xspace}^\text{min}}){\ensuremath{R^2}\xspace}},\end{aligned}$$ where $a = -k_j \times {\ensuremath{R}\xspace}_{0,j}^2$, $c = -k_j \times
M^0_{{\ensuremath{R}\xspace},j}$, and $b = d = k_j$. The fact that the function in Eq. (\[eqn:2dpdf\]) depends on ${\ensuremath{R^2}\xspace}$ and not simply on ${\ensuremath{R}\xspace}$ motivates the choice of ${\ensuremath{R^2}\xspace}$ as the kinematic variable quantifying the transverse imbalance. The values of ${\ensuremath{M}\xspace}^0_{{\ensuremath{R}\xspace},j}$, ${\ensuremath{R}\xspace}_{0,j}$, $k_{j}$, and the normalization constant are floated when fitting the function to the data or simulation samples.
The function of Eq. (\[eqn:2dpdf\]) describes the QCD multijet, the lepton+jets (dominated by $\PW$+jets and events), and the dilepton+jets (dominated by and Z+jets events) backgrounds in the simulation and data control samples. The initial filtering of the SM backgrounds is performed at the trigger level and the analysis proceeds with the analytical description of the SM backgrounds.
QCD multijet background\[sec:qcd\]
----------------------------------
The QCD multijet control sample for the hadronic box is obtained using events recorded with prescaled jet triggers. The trigger used in this study requires at least two jets with average uncorrected $\pt$ thresholds of 60. The QCD multijet background samples provide $\ga 95\%$ of the events with low ${\ensuremath{M_R}\xspace}$, allowing the study of the ${\ensuremath{M_R}\xspace}$ shapes with different thresholds on ${\ensuremath{R^2}\xspace}$, which we denote ${{\ensuremath{R^2}\xspace}_\text{min}}$. The study was repeated using datasets collected with many jet trigger thresholds and prescale factors during the course of the 2011 LHC data taking, with consistent results.
The ${\ensuremath{M_R}\xspace}$ distributions for events satisfying the HAD box selection in this multijet control data sample are shown for different values of the ${\ensuremath{R}\xspace}_\text{min}$ threshold in Fig. \[fig:QCD\_calo1\] (a). The ${\ensuremath{M_R}\xspace}$ distribution is exponentially falling, except for a turn-on at low ${\ensuremath{M_R}\xspace}$ resulting from the $\pt$ threshold requirement on the jets entering the megajet calculation. The exponential region of these distributions is fitted for each value of ${{\ensuremath{R^2}\xspace}_\text{min}}$ to extract the absolute value of the coefficient in the exponent, denoted ${\ensuremath{S}\xspace}$. The value of ${\ensuremath{S}\xspace}$ that maximizes the likelihood in the exponential fit is found to be a linear function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, as shown in Fig \[fig:QCD\_calo1\] (b). Fitting ${\ensuremath{S}\xspace}$ to the form ${\ensuremath{S}\xspace}= -a -
b{{\ensuremath{R^2}\xspace}_\text{min}}$ determines the values of $a$ and $b$.
![(a) The ${\ensuremath{R^2}\xspace}$ distributions for different values of ${{\ensuremath{M_R}\xspace}^\text{min}}$ for events in data selected in the HAD box of a multijet control sample, fit to an exponential function. (b) The coefficient in the exponent ${\ensuremath{S}\xspace}'$ from fits to the ${\ensuremath{R^2}\xspace}$ distributions, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$. \[fig:QCD\_calo2\]](qcd-mr-prd.pdf "fig:"){width="49.50000%"} ![(a) The ${\ensuremath{R^2}\xspace}$ distributions for different values of ${{\ensuremath{M_R}\xspace}^\text{min}}$ for events in data selected in the HAD box of a multijet control sample, fit to an exponential function. (b) The coefficient in the exponent ${\ensuremath{S}\xspace}'$ from fits to the ${\ensuremath{R^2}\xspace}$ distributions, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$. \[fig:QCD\_calo2\]](qcd-slopeMR-prd.pdf "fig:"){width="49.50000%"}
![(a) The ${\ensuremath{R^2}\xspace}$ distributions for different values of ${{\ensuremath{M_R}\xspace}^\text{min}}$ for events in data selected in the HAD box of a multijet control sample, fit to an exponential function. (b) The coefficient in the exponent ${\ensuremath{S}\xspace}'$ from fits to the ${\ensuremath{R^2}\xspace}$ distributions, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$. \[fig:QCD\_calo2\]](qcd-rsq-prd.pdf "fig:"){width="49.50000%"} ![(a) The ${\ensuremath{R^2}\xspace}$ distributions for different values of ${{\ensuremath{M_R}\xspace}^\text{min}}$ for events in data selected in the HAD box of a multijet control sample, fit to an exponential function. (b) The coefficient in the exponent ${\ensuremath{S}\xspace}'$ from fits to the ${\ensuremath{R^2}\xspace}$ distributions, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$. \[fig:QCD\_calo2\]](qcd-slopeR-prd.pdf "fig:"){width="49.50000%"}
The ${\ensuremath{R}\xspace}_\text{min}^2$ distributions are shown for different values of the ${\ensuremath{M_R}\xspace}$ threshold in Fig. \[fig:QCD\_calo2\] (a). The ${\ensuremath{R^2}\xspace}$ distribution is exponentially falling, except for a turn-on at low ${\ensuremath{R^2}\xspace}$. The exponential region of these distributions is fitted for each value of ${{\ensuremath{M_R}\xspace}^\text{min}}$ to extract the absolute value of the coefficient in the exponent, denoted by ${\ensuremath{S}\xspace}'$. The value of ${\ensuremath{S}\xspace}'$ that maximizes the likelihood in the exponential fit is found to be a linear function of ${{\ensuremath{M_R}\xspace}^\text{min}}$ as shown in Fig. \[fig:QCD\_calo2\] (b). Fitting ${\ensuremath{S}\xspace}'$ to the form ${\ensuremath{S}\xspace}' = - c -
dM^\text{min}_{\ensuremath{R}\xspace}$ determines the values of $c$ and $d$. The slope $d$ is found to be equal to the slope $b$ to within a few per cent, as seen from the values of these parameters listed in Figs. \[fig:QCD\_calo1\] (b) and \[fig:QCD\_calo2\] (b), respectively. The equality $d=b$ is essential for building the 2D probability density function that analytically describes the ${\ensuremath{R^2}\xspace}$ versus ${\ensuremath{M_R}\xspace}$ distribution, as it reduces the number of possible 2D functions to the function given in Eq. (\[eqn:2dpdf\]). Note that in Eq. (\[eqn:2dpdf\]) the $k_j$ parameters are the $b_j$, $d_j$ parameters used in the description of the SM backgrounds.
Lepton+jets backgrounds {#V+jetsBKG}
-----------------------
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for events in the MU box, with the requirement of zero $\cPqb$-tagged jets.\[fig:data\_W\_MRs\]](data-Wj-mr-prd.pdf){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for events in the MU box, with the requirement of zero $\cPqb$-tagged jets.\[fig:data\_W\_MRs\]](data-Wj-slope1MR.pdf "fig:"){width="49.50000%"} ![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for events in the MU box, with the requirement of zero $\cPqb$-tagged jets.\[fig:data\_W\_MRs\]](data-Wj-slope2MR.pdf "fig:"){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for simulated $\PW$+jets events in the MU box with the requirement of zero $\cPqb$-tagged jets.\[fig:MC\_W\_MRs\]](MC-Wj-mr-prd.pdf){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for simulated $\PW$+jets events in the MU box with the requirement of zero $\cPqb$-tagged jets.\[fig:MC\_W\_MRs\]](MC-Wj-slope1MR.pdf "fig:"){width="49.50000%"} ![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for simulated $\PW$+jets events in the MU box with the requirement of zero $\cPqb$-tagged jets.\[fig:MC\_W\_MRs\]](MC-Wj-slope2MR.pdf "fig:"){width="49.50000%"}
The major SM backgrounds with leptons and jets in the final state are ($\PW/\cPZ$)+jets, $\ttbar$, and single-top-quark production. These events can also contain genuine . In both the simulated and the data events in the MU and ELE razor boxes, the ${\ensuremath{M_R}\xspace}$ distribution is well described by the sum of two exponential components. One component, which we denote the “first" component, has a steeper slope than the other, “second" component, i.e., $\abs{{\ensuremath{S}\xspace}_1} > \abs{{\ensuremath{S}\xspace}_2}$, and thus the second component is dominant in the high-${\ensuremath{M_R}\xspace}$ region. The relative normalization of the two components is considered as an additional degree of freedom. Both the ${\ensuremath{S}\xspace}_1$ and ${\ensuremath{S}\xspace}_2$ values, along with their relative and absolute normalizations, are determined in the fit. The ${\ensuremath{M_R}\xspace}$ distributions are shown as a function of ${\ensuremath{R}\xspace}_\text{min}^2$ in Fig. \[fig:data\_W\_MR\] for the zero $\cPqb$-jet MU data, which is dominated by W+jets events. The dependence of ${\ensuremath{S}\xspace}_1$ and ${\ensuremath{S}\xspace}_2$ on ${\ensuremath{R}\xspace}_\text{min}^2$ is shown in Fig. \[fig:data\_W\_MRs\].
The corresponding results from simulation are shown in Figs. \[fig:MC\_W\_MR\] and \[fig:MC\_W\_MRs\]. It is seen that the values of the slope parameters $b_1$ and $b_2$ from simulation, given in Fig. \[fig:MC\_W\_MRs\], agree within the uncertainties with the results from data, given in Fig. \[fig:data\_W\_MRs\].
The ${\ensuremath{R^2}\xspace}$ distributions as a function of ${\ensuremath{M_R}\xspace}^\text{min}$ for the data are shown in Fig. \[fig:data\_W\_R\] for the MU box with the requirement of zero b-tagged jets. The ${\ensuremath{S}\xspace}'_1$ and ${\ensuremath{S}\xspace}'_2$ parameters characterizing the exponential behavior of the first and second $\PW(\mu\nu)$+jets components are shown in Fig. \[fig:data\_W\_Rs\]. The corresponding results from simulation are shown in Figs. \[fig:MC\_W\_R\] and \[fig:MC\_W\_Rs\]. The results for the slopes $d_1$ and $d_2$ from simulation, listed in Fig. \[fig:MC\_W\_Rs\], are seen to be in agreement with the measured results, listed in Fig. \[fig:data\_W\_Rs\]. Furthermore, the extracted values of $d_1$ and $d_2$ are in agreement with the extracted values of $b_1$ and $b_2$, respectively. This is the essential ingredient to build a 2D template for the (${\ensuremath{M_R}\xspace}$,${\ensuremath{R^2}\xspace}$) distributions, starting with the function of Eq. (\[eqn:2dpdf\]).
The corresponding distributions for the $\ttbar$ MC simulation with $\ge$1 $\cPqb$-tagged jet are presented in Appendix \[sec:ttbarAppendix\], for events selected in the HAD box.
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for events in the MU box, with the requirement of zero $\cPqb$-tagged jets.\[fig:data\_W\_Rs\]](data-Wj-Rsq-prd.pdf){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for events in the MU box, with the requirement of zero $\cPqb$-tagged jets.\[fig:data\_W\_Rs\]](data-Wj-slope1R.pdf "fig:"){width="49.50000%"} ![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for events in the MU box, with the requirement of zero $\cPqb$-tagged jets.\[fig:data\_W\_Rs\]](data-Wj-slope2R.pdf "fig:"){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for $\PW$+jets simulated events in the MU box with the requirement of zero $\cPqb$-tagged jets.\[fig:MC\_W\_Rs\]](MC-Wj-Rsq-prd.pdf){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for $\PW$+jets simulated events in the MU box with the requirement of zero $\cPqb$-tagged jets.\[fig:MC\_W\_Rs\]](MC-Wj-slope1R.pdf "fig:"){width="49.50000%"} ![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for $\PW$+jets simulated events in the MU box with the requirement of zero $\cPqb$-tagged jets.\[fig:MC\_W\_Rs\]](MC-Wj-slope2R.pdf "fig:"){width="49.50000%"}
Dilepton backgrounds
--------------------
The SM contributions to the ELE-ELE and MU-MU boxes are expected to be dominated by $\cPZ$+jets events, and the SM contribution to the ELE-MU box by $\ttbar$ events, all at the level of $\ga$95%. We find that the ${\ensuremath{M_R}\xspace}$ distributions as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, and the ${\ensuremath{R^2}\xspace}$ distribution as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, are independent of the lepton-flavor combination for both the ELE-ELE and MU-MU boxes, as determined using simulated $\ttbar(2\ell 2\nu$+jets$)$ events. In addition, the asymptotic second component is found to be process-independent.
Background model and fits {#sec:fits}
=========================
As described earlier, the full 2D SM background representation is built using statistically independent data control samples. The parameters of this model provide the input to the final fit performed in the [*fit region*]{} (FR) of the data samples, defining an extended, unbinned maximum likelihood (ML) fit with the <span style="font-variant:small-caps;">RooFit</span> fitting package [@Verkerke:2003ir]. The fit region is defined for each of the razor boxes as the region of low ${\ensuremath{M_R}\xspace}$ and small ${\ensuremath{R^2}\xspace}$, where signal contamination is expected to have negligible impact on the shape fit. The 2D model is extrapolated to the rest of the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane, which is sensitive to new-physics signals and where the search is performed.
For each box, the fit is conducted in the signal-free FR of the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane; their definition can be found in Figs. \[fig:had-blue-plot\], \[fig:ele-blue-plot\], \[fig:mu-blue-plot\], \[fig:eleele-blue-plot\], \[fig:mumu-blue-plot\], and \[fig:muele-blue-plot\]. These regions are used to provide a full description of the SM background in the entire (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane in each box. The likelihood function for a given box is written as [@Barlow:1990vc]: $$\label{eq:Lb}
\mathcal{L}_{b} = \frac{\re^{-(\sum_{j \in SM} N_{j})}}{N !}
\prod_{i=1}^{N} \left[\sum_{j \in SM} N_{j} P_{j}(M_{R,i},{\ensuremath{R^2}\xspace}_i)\right],$$ where $N$ is the total number of events in the FR region of the box, the sum runs over all the SM processes relevant for that box, and the $N_{j}$ are normalization parameters for each SM process involved in the considered box.
We find that each SM process in a given final state box is well described in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane by the function $P_{j}$ defined as $$P_{j}(M_R, {\ensuremath{R^2}\xspace}) = (1-f^{j}_2)\times F^{1st}_{j}(M_R, {\ensuremath{R^2}\xspace}) +
f^{j}_2\times F^{2nd}_{j}(M_R, {\ensuremath{R^2}\xspace}),
\label{eq:twocomponents}$$ where the [*first*]{} ($F^{1st}_{j}$) and [*second*]{} ($F^{2nd}_{j}$) components are defined as in Eq. (\[eqn:2dpdf\]), and $f^{j}_2$ is the normalization fraction of the second component with respect to the total. When fitting this function to the data, the shape parameters of each $F_{j}(M_R, {\ensuremath{R^2}\xspace})$ function, the absolute normalization, and the relative fraction $f^{j}_2$ are floated in the fit. Studies of simulated events and fits to data control samples with either a $\cPqb$-jet requirement or a $\cPqb$-jet veto indicate that the parameters corresponding to the first components of these backgrounds (with steeper slopes at low ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$) are box-dependent. The parameters describing the second components are box-independent, and at the current precision of the background model, they are identical between the dominant backgrounds considered in these final states.
We validate the choice of the background shape by use of a sample of $\ttbar$ MC simulated events corresponding to an integrated luminosity of 10. Besides being the dominant background in the $\ge$1 $\cPqb$-tag search, $\ttbar$ events are the dominant background for the inclusive search for large values of ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$. The result for the HAD box in the inclusive razor path is shown in Fig. \[fig:had-tthstat\] expressed as the projection of the 2D fit on ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$. As the same level of agreement is found in all boxes both in the inclusive and in the $\ge$1 $\cPqb$-tagged razor path, we proceed to fit all the SM processes with this shape.
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the HAD box in $\ttbar$ MC simulation. The continuous histogram is the 2D model prediction obtained from a single pseudo-experiment based on the 2D fit. The fit is performed in the (${\ensuremath{M_R}\xspace}$,${\ensuremath{R^2}\xspace}$) fit region and projected into the full analysis region. Only the statistical uncertainty band in the background prediction is drawn in these projections. The points show the distribution for the MC simulated events.\[fig:had-tthstat\]](tthstat-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the HAD box in $\ttbar$ MC simulation. The continuous histogram is the 2D model prediction obtained from a single pseudo-experiment based on the 2D fit. The fit is performed in the (${\ensuremath{M_R}\xspace}$,${\ensuremath{R^2}\xspace}$) fit region and projected into the full analysis region. Only the statistical uncertainty band in the background prediction is drawn in these projections. The points show the distribution for the MC simulated events.\[fig:had-tthstat\]](tthstat-Rsq-prd.pdf "fig:"){width="49.50000%"}
Fit results and validation {#sec:datafits}
--------------------------
The shape parameters in Eq. (\[eqn:2dpdf\]) are determined for each box via the 2D fit. The likelihood of Eq. (\[eq:Lb\]) is multiplied by Gaussian [*penalty terms*]{} [@penalty] to account for the uncertainties of the shape parameters $k_{j}$, ${\ensuremath{M}\xspace}^0_{{\ensuremath{R}\xspace},j}$, and ${\ensuremath{R}\xspace}_{0,j}^2$. The central values of the Gaussians are derived from analogous 2D fits in the low-statistics data control sample. The penalty terms pull the fit to the local minimum closer to the shape derived from the data control samples. Using pseudo-experiments, we verified that this procedure does not bias the determination of the background shape. As an example, the $k_{j}$ parameter uncertainties are typically $\sim$30%. Additional background shape uncertainties due to the choice of the functional form were considered and found to be negligible, as discussed in Appendix \[altBKG\].
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive HAD box. The continuous histogram is the total SM prediction. The dash-dotted and dashed histograms are described in the text. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:had-blue-plot\]) and projected into the full analysis region. The full uncertainty in the total background prediction is drawn in these projections, including the one due to the variation of the background shape parameters and normalization.\[fig:had-box\]](had-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive HAD box. The continuous histogram is the total SM prediction. The dash-dotted and dashed histograms are described in the text. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:had-blue-plot\]) and projected into the full analysis region. The full uncertainty in the total background prediction is drawn in these projections, including the one due to the variation of the background shape parameters and normalization.\[fig:had-box\]](had-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The fit region, FR, and signal regions, SR$i$, are defined in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane for the HAD box. The color scale gives the $p$-values corresponding to the observed number of events in each SR$i$, computed from background parameterization derived in the FR. The $p$-values are also given in the table, together with the observed number of events, the median and the mode of the yield distribution, and a 68% interval.\[fig:had-blue-plot\]](p-had-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive ELE box. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:ele-blue-plot\]) and projected into the full analysis region. The histograms are described in the text.\[fig:ele-box\]](ele-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive ELE box. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:ele-blue-plot\]) and projected into the full analysis region. The histograms are described in the text.\[fig:ele-box\]](ele-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The fit region, FR, and signal regions, SR$i$, are defined in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane for the ELE box. The color scale gives the $p$-values corresponding to the observed number of events in each SR$i$. Further explanation is given in the Fig. \[fig:had-blue-plot\] caption.\[fig:ele-blue-plot\]](p-ele-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive MU box. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:mu-blue-plot\]) and projected into the full analysis region. The histograms are described in the text.\[fig:mu-box\]](mu-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive MU box. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:mu-blue-plot\]) and projected into the full analysis region. The histograms are described in the text.\[fig:mu-box\]](mu-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The fit region, FR, and signal regions, SR$i$, are defined in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane for the MU box. The color scale gives the $p$-values corresponding to the observed number of events in each SR$i$. Further explanation is given in the Fig. \[fig:had-blue-plot\] caption.\[fig:mu-blue-plot\]](p-mu-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the ELE-ELE box. The continuous histogram is the total standard model prediction. The histogram is described in the text.\[fig:eleele-box\]](eleele-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the ELE-ELE box. The continuous histogram is the total standard model prediction. The histogram is described in the text.\[fig:eleele-box\]](eleele-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The fit region, FR, and signal regions, SR$i$, are defined in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane for the ELE-ELE box. The color scale gives the $p$-values corresponding to the observed number of events in each SR$i$. Further explanation is given in the Fig. \[fig:had-blue-plot\] caption.\[fig:eleele-blue-plot\]](p-eleele-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive MU-MU box. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:mumu-blue-plot\]) and projected into the full analysis region. The histogram is described in the text.\[fig:mumu-box\]](mumu-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive MU-MU box. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:mumu-blue-plot\]) and projected into the full analysis region. The histogram is described in the text.\[fig:mumu-box\]](mumu-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The fit region, FR, and signal regions, SR$i$, are defined in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane for the MU-MU box. The color scale gives the $p$-values corresponding to the observed number of events in each SR$i$. Further explanation is given in the Fig. \[fig:had-blue-plot\] caption.\[fig:mumu-blue-plot\]](p-mumu-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive MU-ELE box. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:muele-blue-plot\]) and projected into the full analysis region. The histogram is described in the text.\[fig:muele-box\]](muele-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the inclusive MU-ELE box. The fit is performed in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) fit region (shown in Fig. \[fig:muele-blue-plot\]) and projected into the full analysis region. The histogram is described in the text.\[fig:muele-box\]](muele-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The fit region, FR, and signal regions, SR$i$, are defined in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane for the MU-ELE box. The color scale gives the $p$-values corresponding to the observed number of events in each SR$i$. Further explanation is given in the Fig. \[fig:had-blue-plot\] caption. \[fig:muele-blue-plot\]](p-muele-prd.pdf){width="\cmsFigWidthBox"}
The result of the ML fit projected on ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ is shown in Fig. \[fig:had-box\] for the inclusive HAD box. No significant discrepancy is observed between the data and the fit model for any of the six boxes. In order to establish the compatibility of the background model with the observed dataset, we define a set of signal regions (SR${i}$) in the tail of the SM background distribution. Using the 2D background model determined using the ML fit, we derive the distribution of the expected yield in each SR${i}$ using pseudo-experiments, accounting for correlations and uncertainties in the parameters describing the background model. In order to correctly account for the uncertainties in the parameters describing the background model and their correlations, the shape parameters used to generate each pseudo-experiment dataset are sampled from the covariance matrix returned by the ML fit performed on the actual dataset. The actual number of events in each dataset is drawn from a Poisson distribution centered on the yield returned by the covariance matrix sampling. For each pseudo-experiment dataset, the number of events in the SR$i$ is found. For each of the SR${i}$, the distribution of the number of events derived by the pseudo-experiments is used to calculate a two-sided $p$-value (as shown for the HAD box in Fig. \[fig:had-blue-plot\]), corresponding to the probability of observing an equal or less probable outcome for a counting experiment in each signal region. The result of the ML fit and the corresponding $p$-values are shown in Figs. \[fig:ele-box\] and \[fig:ele-blue-plot\] for the ELE box, Figs. \[fig:mu-box\] and \[fig:mu-blue-plot\] for the MU box, Figs. \[fig:eleele-box\] and \[fig:eleele-blue-plot\] for the ELE-ELE box, Figs.\[fig:mumu-box\] and \[fig:mumu-blue-plot\] for the MU-MU box, and Figs. \[fig:muele-box\] and \[fig:muele-blue-plot\] for the MU-ELE box. We note that the background shapes in the single-lepton and hadronic boxes are well described by the sum of two functions: a single-component function with a steeper-slope component, denoted as the $\mathrm{V}$+jets first component, obtained by fixing $f_2=0$ in Eq. (\[eq:twocomponents\]); and a two-component function as in Eq. (\[eq:twocomponents\]), with the first component describing the steeper-slope core of the and single-top background distributions (generically referred to as ), and the [*effective*]{} second component modeling the sum of the indistinguishable tails of different SM background processes. In the dilepton boxes we show the total SM background, which is composed of $\mathrm{V}$+jets and events in the ELE-ELE and MU-MU boxes and of events in the MU-ELE boxes. The corresponding results for the $\geq$1 $\cPqb$-tagged samples are presented in Appendix \[sec:datafitsBTAGAppendix\].
Signal systematic uncertainties {#sec:systematics}
===============================
We evaluate the impact of systematic uncertainties on the shape of the signal distributions, for each point of each SUSY model, using the simulated signal event samples. The following systematic uncertainties are considered, with the approximate size of the uncertainty given in parentheses: (i) PDFs (up to 30%, evaluated point-by-point)) [@Bourilkov:2006cj]; (ii) jet-energy scale (up to 1%, evaluated point-by-point) [@JES]; (iii) lepton identification, using the “tag-and-probe” technique based on $\cPZ
\to \ell \ell$ events [@CMS:2011aa] ($\ell=\Pe,\mu$, 1% per lepton). In addition, the following uncertainties, which affect the signal yield, are considered: (i) luminosity uncertainty [@lumi-moriond] (2.2%); (ii) theoretical cross section [@Kramer:2012bx] (up to 15%, evaluated point-by-point); (iii) razor trigger efficiency (2%); (iv) lepton trigger efficiency (3%). An additional systematic uncertainty is considered for the $\cPqb$-tagging efficiency [@BTAG] (between 6% and 20% in $\pt$ bins). We consider variations of the function modeling, the signal uncertainty (log-normal versus Gaussian), and the binning, and find negligible deviations in the results. The systematic uncertainties are included using the best-fit shape to compute the likelihood values for each pseudo-experiment, while sampling the same pseudo-experiment from a different function, derived from the covariance matrix of the fit to the data. This procedure is repeated for both the background and signal probability density functions.
Interpretation of the results\[sec:INTERP2011\]
===============================================
In order to evaluate exclusion limits for a given SUSY model, its parameters are varied and an excluded cross section at the 95% is associated with each configuration of the model parameters, using the hybrid version of the [$\mathrm{CL}_\mathrm{s}$]{}method [@CLs1; @CLs2; @CLs3], described below.
For each box, we consider the test statistic given by the logarithm of the likelihood ratio $\ln Q = \ln[\mathcal{L}(s+b|H_i)/\mathcal{L}(b|H_i)]$, where $H_i~(i=0,2)$ is the hypothesis under test: $H_1$ (signal-plus-background) or $H_0$ (background-only). The likelihood function for the background-only hypothesis is given by Eq. (\[eq:Lb\]). The likelihood corresponding to the signal-plus-background hypothesis is written as where $\sigma$ is the signal cross section, , the parameter of interest; $L$ is the integrated luminosity; $\epsilon$ is the signal acceptance times efficiency; and $P_S (M_{R,i},{\ensuremath{R^2}\xspace}_i)$ is the two-dimensional probability density function for the signal, computed numerically from the distribution of simulated signal events. The signal and background shape parameters, and the normalization factors $L$ and $\epsilon$, are the nuisance parameters.
For each analysis (inclusive razor or inclusive $\cPqb$-jet razor) we sum the test statistics of the six corresponding boxes to compute the combined test statistic.
The distribution of $\ln Q$ is derived numerically with a MC technique. The values of the nuisance parameters in the likelihood are randomized for each iteration of the MC generation, to reflect the corresponding uncertainty. Once the likelihood is defined, a sample of events is generated according to the signal and background probability density functions. The value of $\ln Q$ for each generated sample is then evaluated, fixing each signal and background parameter to its expected value. This procedure corresponds to a numerical marginalization of the nuisance parameters.
Given the distribution of $\ln Q$ for the background-only and the signal-plus-background pseudo-experiments, and the value of $\ln Q$ observed in the data, we calculate [$\mathrm{CL}_\mathrm{s+b}$]{}and $1-{\ensuremath{\mathrm{CL}_\mathrm{b}}\xspace}$ [@CLs1]. From these values, ${\ensuremath{\mathrm{CL}_\mathrm{s}}\xspace}={\ensuremath{\mathrm{CL}_\mathrm{s+b}}\xspace}/{\ensuremath{\mathrm{CL}_\mathrm{b}}\xspace}$ is computed for that model point. The procedure is independently applied to each of the two analyses (inclusive razor and inclusive $\cPqb$-jet razor).
The CMSSM model is studied in the (${m_{0}}$, ${\ensuremath{m_{1/2}}\xspace}$) plane, fixing $\tan\beta=10$, $A_{0} = 0$, and $\sgn(\mu) = +1$. A point in the plane is excluded at the 95% if ${\ensuremath{\mathrm{CL}_\mathrm{s}}\xspace}< 0.05$. The result obtained for the inclusive razor analysis is shown in Fig. \[fig:all-result\] (a). The shape of the observed exclusion curves reflects the changing relevant SUSY strong-production processes across the parameter space, with squark-antisquark and gluino-gluino production dominating at low and high ${\ensuremath{m_0}\xspace}$, respectively. The observed limit is less constraining than the median expected limit at lower ${\ensuremath{m_0}\xspace}$ due to a local excess of events at large ${\ensuremath{R^2}\xspace}$ in the hadronic box.
For large values of ${\ensuremath{m_0}\xspace}$, boxes with leptons in the final state have a sensitivity comparable to that of the hadronic boxes, as cascade decays of gluinos yield leptons production. Figure \[fig:all-result\] parts (b)-(d) show the CMSSM exclusion limits based on the HAD box only and on the leptonic boxes only.
{width="49.50000%"} {width="49.50000%"} {width="49.50000%"} {width="49.50000%"}
The results are also interpreted as cross section limits on a number of simplified models [@ArkaniHamed:2007fw; @Alwall-2; @Alwall-1; @Alwall:2008va; @Alves:2011wf] where a limited set of hypothetical particles and decay chains are introduced to produce a given topological signature. For each model studied, we derive the maximum allowed cross section at the 95% as a function of the mass of the produced particles (gluinos or squarks, depending on the model) and the LSP mass, as well as the exclusion limit corresponding to the SUSY cross section. We study several SMS benchmark scenarios [@SUS-11-016]:
- gluino-gluino production with four light-flavor jets+in the superpartner decays, T1 in Fig. \[fig:SMSdecay\].
- squark-antisquark production with two light-flavor jets+ in the superpartner decays, T2 in Fig. \[fig:SMSdecay\].
- gluino-gluino production with four $\cPqb$ jets+in the superpartner decays, T1bbbb in Fig. \[fig:SMSdecay\].
- squark-antisquark production with two $\cPqb$ jets+in the superpartner decays, T2bb in Fig. \[fig:SMSdecay\].
- gluino-gluino production with four top quarks+in the superpartner decays, T1tttt in Fig. \[fig:SMSdecay\].
- squark-antisquark production with two top quarks+in the superpartner decays, T2tt in Fig. \[fig:SMSdecay\].
In all cases, additional jets in the final state can arise from initial- and final-state radiation (ISR and FSR), simulated by . We show in Figs. \[fig:sms-incl0\] and \[fig:sms-bincl0\] the excluded cross section at 95% as a function of the mass of the produced particle (gluinos or squarks, depending on the model) and the LSP mass, as well as the exclusion curve corresponding to the NLO+NLL SUSY cross section [@NLONLL1; @NLONLL2; @NLONLL3; @NLONLL4; @NLONLL5], where NLL indicates the next-to-leading-logarithmic. A result is not quoted for the region of the SMS plane in which the signal efficiency strongly depends on the ISR and FSR modeling (gray area), as a consequence of the small mass difference between the produced superpartner and the LSP and the consequent small for the jets produced in the cascade.
In Fig. \[fig:sum-sms\], we present a summary of the 95% excluded largest parent mass for various LSP masses in each of the simplified models studied, showing separately the results from the inclusive razor analysis and the inclusive $\cPqb$-jet razor analysis. A comparison of the razor results with those obtained from other approaches is given in Ref. [@Mahbubani:2012qq].
{width="45.00000%"} {width="45.00000%"} {width="45.00000%"} {width="45.00000%"} {width="45.00000%"} {width="45.00000%"}
{width="49.50000%"} {width="49.50000%"} {width="49.50000%"} {width="49.50000%"}
{width="49.50000%"} {width="49.50000%"} {width="49.50000%"} {width="49.50000%"}
![Summary of the 95% excluded largest parent mass for each of the simplified models studied, for various LSP masses. The results from the $\cPqb$-jet razor analysis are shown immediately below those from the inclusive razor analysis for each of the four categories of events indicated.\[fig:sum-sms\]](sum-sms.pdf){width="0.90\columnwidth"}
Summary {#sec:summary}
=======
Using a data sample of $\sqrt{s}$ = 7proton-proton collisions collected by the CMS experiment at the LHC in 2011, corresponding to an integrated luminosity of 4.7, we have performed a search for pair-produced supersymmetric particles such as squarks and gluinos in the razor-variable plane. A 2D shape description of the relevant standard model processes determined from data control samples and validated with simulated events has been used, and no significant excess over the background expectations has been observed. The results are presented as a 95% limit in the (${\ensuremath{m_0}\xspace}$, ${\ensuremath{m_{1/2}}\xspace}$) CMSSM parameter space. We exclude squark and gluino masses up to 1350for $m(\PSq)\sim m(\PSg)$, while for $m(\PSq)>m(\PSg)$ we exclude gluino masses up to 800. For simplified models, we exclude gluino masses up to 1000, and first- and second- generation squark masses up to 800. The direct production of top or bottom squarks is excluded for squark masses up to 400.
We congratulate our colleagues in the CERN accelerator departments for the excellent performance of the LHC and thank the technical and administrative staffs at CERN and at other CMS institutes for their contributions to the success of the CMS effort. In addition, we gratefully acknowledge the computing centers and personnel of the Worldwide LHC Computing Grid for delivering so effectively the computing infrastructure essential to our analyses. Finally, we acknowledge the enduring support for the construction and operation of the LHC and the CMS detector provided by the following funding agencies: BMWF and FWF (Austria); FNRS and FWO (Belgium); CNPq, CAPES, FAPERJ, and FAPESP (Brazil); MES (Bulgaria); CERN; CAS, MoST, and NSFC (China); COLCIENCIAS (Colombia); MSES and CSF (Croatia); RPF (Cyprus); MoER, SF0690030s09 and ERDF (Estonia); Academy of Finland, MEC, and HIP (Finland); CEA and CNRS/IN2P3 (France); BMBF, DFG, and HGF (Germany); GSRT (Greece); OTKA and NIH (Hungary); DAE and DST (India); IPM (Iran); SFI (Ireland); INFN (Italy); NRF and WCU (Republic of Korea); LAS (Lithuania); MOE and UM (Malaysia); CINVESTAV, CONACYT, SEP, and UASLP-FAI (Mexico); MBIE (New Zealand); PAEC (Pakistan); MSHE and NSC (Poland); FCT (Portugal); JINR (Dubna); MON, RosAtom, RAS and RFBR (Russia); MESTD (Serbia); SEIDI and CPAN (Spain); Swiss Funding Agencies (Switzerland); NSC (Taipei); ThEPCenter, IPST, STAR and NSTDA (Thailand); TUBITAK and TAEK (Turkey); NASU (Ukraine); STFC (United Kingdom); DOE and NSF (USA).
Individuals have received support from the Marie-Curie program and the European Research Council and EPLANET (European Union); the Leventis Foundation; the A. P. Sloan Foundation; the Alexander von Humboldt Foundation; the Belgian Federal Science Policy Office; the Fonds pour la Formation à la Recherche dans l’Industrie et dans l’Agriculture (FRIA-Belgium); the Agentschap voor Innovatie door Wetenschap en Technologie (IWT-Belgium); the Ministry of Education, Youth and Sports (MEYS) of Czech Republic; the Council of Science and Industrial Research, India; the Compagnia di San Paolo (Torino); the HOMING PLUS programme of Foundation for Polish Science, cofinanced by EU, Regional Development Fund; and the Thalis and Aristeia programmes cofinanced by EU-ESF and the Greek NSRF.
Appendices {#appendices .unnumbered}
==========
Additional standard model backgrounds in the (MR MR2) razor plane\[sec:ttbarAppendix\]
======================================================================================
Figure \[fig:MC\_ttj\_MR\] shows the ${\ensuremath{M_R}\xspace}$ distribution as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$ for $\ttbar$ MC events with $\ge$1 $\cPqb$-tagged jets in the HAD box. The ${\ensuremath{S}\xspace}_1$ and ${\ensuremath{S}\xspace}_2$ parameters characterizing the exponential behavior of the first and second $\PW(\mu\nu)$+jets components are shown in Fig. \[fig:MC\_ttj\_MRs\]. The corresponding distributions for ${\ensuremath{R^2}\xspace}$, and for the ${\ensuremath{S}\xspace}'_1$ and ${\ensuremath{S}\xspace}'_2$ parameters, are shown in Figs. \[fig:MC\_ttj\_R\] and \[fig:MC\_ttj\_Rs\], respectively. The conclusions derived from the data and MC studies of Section \[sec:BKG2011\] hold also for $\ttbar$ MC events .
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for $\ttbar$ simulated events in the HAD box with the requirement of $\ge$1 $\cPqb$-tagged jets.\[fig:MC\_ttj\_MRs\]](MC-ttj-mr-prd.pdf){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for $\ttbar$ simulated events in the HAD box with the requirement of $\ge$1 $\cPqb$-tagged jets.\[fig:MC\_ttj\_MRs\]](MC-ttj-slope1MR.pdf "fig:"){width="49.50000%"} ![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}_2$, from fits to the ${\ensuremath{M_R}\xspace}$ distribution, as a function of ${{\ensuremath{R^2}\xspace}_\text{min}}$, for $\ttbar$ simulated events in the HAD box with the requirement of $\ge$1 $\cPqb$-tagged jets.\[fig:MC\_ttj\_MRs\]](MC-ttj-slope2MR.pdf "fig:"){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for $\ttbar$ simulated events in the HAD box with the requirement of $\ge$1 $\cPqb$-tagged jets.\[fig:MC\_ttj\_Rs\]](MC-ttj-Rsq-prd.pdf){width="49.50000%"}
![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for $\ttbar$ simulated events in the HAD box with the requirement of $\ge$1 $\cPqb$-tagged jets.\[fig:MC\_ttj\_Rs\]](MC-ttj-slope1R.pdf "fig:"){width="49.50000%"} ![Value of (a) the coefficient in the first exponent, ${\ensuremath{S}\xspace}'_1$, and (b) the coefficient in the second exponent, ${\ensuremath{S}\xspace}'_2$, from fits to the ${\ensuremath{R^2}\xspace}$ distribution, as a function of ${{\ensuremath{M_R}\xspace}^\text{min}}$, for $\ttbar$ simulated events in the HAD box with the requirement of $\ge$1 $\cPqb$-tagged jets.\[fig:MC\_ttj\_Rs\]](MC-ttj-slope2R.pdf "fig:"){width="49.50000%"}
Alternative background shape analysis {#altBKG}
=====================================
In order to quantify a systematic uncertainty associated with the choice of the fit function, we first generalize our 2D function to allow for deviations from the exponential behavior, once projected onto ${\ensuremath{M_R}\xspace}$ or ${\ensuremath{R^2}\xspace}$. To do this, we i) identify a set of functions that describe the data, ii) use one as a default description, iii) use the rest to quantify the systematic variation, iv) randomly choose one of the three functions when generating the pseudo-experiments used to set limits, and v) use the nominal function when evaluating the likelihood.
For a 1D fit of the ${\ensuremath{M_R}\xspace}$ distribution, an obvious choice is $$({\ensuremath{M_R}\xspace}) = A \re^{-b {\ensuremath{M_R}\xspace}^\beta},$$ where $\beta \neq 1$ accounts for deviations from the exponential function. In this analysis, we need a 2D function of ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ that allows us to measure the deviation from the nominal shape on the projections. For this purpose, we introduce a generalization of the razor 2D function: which has the two following properties: $$\begin{aligned}
\int_{{{\ensuremath{R^2}\xspace}_\text{min}}}^{+\infty} F_\mathrm{SYS}({\ensuremath{M_R}\xspace}, {\ensuremath{R^2}\xspace})\, \rd{\ensuremath{R^2}\xspace}&\sim \re^{-k_{{\ensuremath{M_R}\xspace}}({\ensuremath{M_R}\xspace}- {\ensuremath{M_R}\xspace}^0) ^{1/n} }, \\
\int_{{{\ensuremath{M_R}\xspace}^\text{min}}}^{+\infty} F_\mathrm{SYS}({\ensuremath{M_R}\xspace}, {\ensuremath{R^2}\xspace})\, \rd{\ensuremath{M_R}\xspace}&\sim \re^{-k_{{\ensuremath{R^2}\xspace}}({\ensuremath{R^2}\xspace}- {\ensuremath{R^2}\xspace}_0) ^{1/n}},\end{aligned}$$ where $$\begin{aligned}
k_{{\ensuremath{M_R}\xspace}} &= (k^0_{{\ensuremath{M_R}\xspace}} + b {{\ensuremath{R^2}\xspace}_\text{min}})^{1/n}, \\
k_{{\ensuremath{R^2}\xspace}} &= (k^0_{{\ensuremath{R^2}\xspace}} + b {{\ensuremath{M_R}\xspace}^\text{min}})^{1/n} ,\end{aligned}$$ with ${{\ensuremath{M_R}\xspace}^\text{min}}$ and ${{\ensuremath{R^2}\xspace}_\text{min}}$ respectively the thresholds applied on ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ before projecting onto ${\ensuremath{R^2}\xspace}$ and ${\ensuremath{M_R}\xspace}$. Using this function to evaluate systematic uncertainties corresponds to the 2D generalization needed here. We proceed as follows:
- we repeat the fit in the fit region of each box, using $F_\mathrm{SYS}({\ensuremath{M_R}\xspace}, {\ensuremath{R^2}\xspace})$ rather than $F({\ensuremath{M_R}\xspace},
{\ensuremath{R^2}\xspace})$ for the second component of the background model (the one that extrapolates to the signal region), with $n$ floated in the fit. We determine $n_\text{fit} \pm \sigma_n$ in this fit.
- we assign an allowed range to the difference $n-1$ taking the larger of $n_\text{fit}-1$ and $\sigma_n$, which we refer to as $[n_\text{min},n_\text{max}]$.
- we repeat the fit in the fit region fixing $n$ to first to $n_\text{min}$ and then to $n_\text{max}$ and we take these fits as the alternative background descriptions.
In particular, we find that the fit returns values of $n_\text{fit}$ that are very close to $n$. Following the prescription outlined above, we take the fit uncertainty as the shift in $n$.
The main conclusion of the study is that the systematic uncertainty in the choice of the function is already covered by the large uncertainty in the fit parameters and that the effect corresponds to an increase of about 15% in the 68% range, once this contribution is summed in quadrature with the already quoted uncertainty.
As an example, we present the results of the above procedure for the bins in the HAD box. Fig. \[fig:fitHADsys\] shows the fit result with $n$ floated in the full region of the HAD box, projected onto ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$. The quality of the fit is similar to that of the nominal procedure. We find $n= 0.96 \pm 0.04$. We then take $n_\text{min} = 0.96$ and $n_\text{max} = 1.04$. We show in Table \[tab:predSYSHad\] the bin-by-bin background prediction for the nominal fit and the two alternative fits. We use a finer binning than the one used to compute the $p$-values in the nominal analysis. For comparison, we also show the values obtained with $n$ floated in the fit. For all cases, we quote the predicted background as the center of the 68% probability range and the associated uncertainty corresponds to half the range. The range is defined by integrating the background distribution (derived from the pseudo-experiments) using the probability value as the ordering algorithm. Similar results are obtained for all boxes.
![\[fig:fitHADsys\] Projection of the fit result on the (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ axis for the HAD box, obtained as explained in the text.](nhad-MR-prd.pdf "fig:"){width="48.00000%"} ![\[fig:fitHADsys\] Projection of the fit result on the (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ axis for the HAD box, obtained as explained in the text.](nhad-Rsq-prd.pdf "fig:"){width="48.00000%"}
[lxxxx]{} Bin & & & &\
HAD\_1\_1 & 1558,69 & 1527,109 & 1509,111 & 1511,126\
HAD\_1\_2 & 2898,80 & 2888,89 & 2868,98 & 2866,99\
HAD\_1\_3 & 711,35 & 729,45 & 714,43 & 726,49\
HAD\_1\_4 & 329,37 & 338,31 & 328,32 & 337,34\
HAD\_2\_1 & 1785,64 & 1787,75 & 1759,69 & 1774,67\
HAD\_2\_2 & 3301,82 & 3336,104 & 3313,112 & 3349,118\
HAD\_2\_3 & 945,46 & 957,47 & 957,47 & 964,48\
HAD\_2\_4 & 432,36 & 423,35 & 454,37 & 424,38\
HAD\_3\_1 & 251,26 & 263,28 & 259,31 & 260,29\
HAD\_3\_2 & 537,47 & 544,45 & 561,50 & 550,49\
HAD\_3\_3 & 173,36 & 157,29 & 182,33 & 162,34\
HAD\_3\_4 & 58,18 & 52,17 & 66,19 & 51,18\
HAD\_4\_1 & 39,9 & 37,11 & 43,9 & 38,9\
HAD\_4\_2 & 86,23 & 74,17 & 90,24 & 76,21\
HAD\_4\_3 & 20,7 & 14,6 & 22,9 & 14,7\
HAD\_4\_4 & 4.2,2.9 & 2.7,2.3 & 4.9,3.1 & 2.4,2.4\
HAD\_5\_1 & 4.7,2.8 & 3.9,2.5 & 5.3,3.1 & 4.1,2.9\
HAD\_5\_2 & 8.3,4.7 & 6.0,3.7 & 9.5,4.7 & 5.9,4.0\
HAD\_5\_3 & 1.2,1.2 & 0.8,0.8 & 1.5,1.5 & 0.8,0.8\
HAD\_5\_4 & 0.4,0.4 & 0.4,0.4 & 0.5,0.5 & 0.4,0.4\
HAD\_6\_1 & 0.8,0.8 & 0.6,0.6 & 0.9,0.9 & 0.6,0.6\
HAD\_6\_2 & 1.0,1.0 & 0.7,0.7 & 1.2,1.2 & 0.8,0.8\
HAD\_6\_3 & 0.4,0.4 & 0.3,0.3 & 0.4,0.4 & 0.4,0.4\
Fit results and validations for >1 b-tagged events {#sec:datafitsBTAGAppendix}
=====================================================
Figures \[fig:bhad-box\]-\[fig:bmuele-blue-plot\] show the results for the $\ge$1 $\cPqb$-tagged jet analysis corresponding to the results presented in Section \[sec:datafits\] for the inclusive analysis.
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the HAD box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:bhad-box\]](bhad-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the HAD box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:bhad-box\]](bhad-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The $p$-values corresponding to the observed number of events in the $\ge$1 $\cPqb$-tag HAD box signal regions (SR$i$). \[fig:bhad-blue-plot\]](bp-had-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the ELE box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:bele-box\]](bele-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the ELE box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:bele-box\]](bele-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The $p$-values corresponding to the observed number of events in the $\ge$1 $\cPqb$-tag ELE box signal regions (SR$i$). \[fig:bele-blue-plot\]](bp-ele-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the MU box in the $\ge$1 $\cPqb$-tag analysis path.\[fig:bmu-box\]](bmu-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the MU box in the $\ge$1 $\cPqb$-tag analysis path.\[fig:bmu-box\]](bmu-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The $p$-values corresponding to the observed number of events in the $\ge$1 $\cPqb$-tag MU box signal regions (SR$i$). \[fig:bmu-blue-plot\]](bp-mu-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the ELE-ELE box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:beleele-box\]](beleele-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the ELE-ELE box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:beleele-box\]](beleele-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The $p$-values corresponding to the observed number of events in the $\ge$1 $\cPqb$-tag ELE-ELE box signal regions (SR$i$). \[fig:beleele-blue-plot\]](bp-eleele-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the MU-MU box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:bmumu-box\]](bmumu-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the MU-MU box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:bmumu-box\]](bmumu-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The $p$-values corresponding to the observed number of events in the $\ge$1 $\cPqb$-tag MU-MU box signal regions (SR$i$). \[fig:bmumu-blue-plot\]](bp-mumu-prd.pdf){width="\cmsFigWidthBox"}
![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the MU-ELE box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:bmuele-box\]](bmuele-MR-prd.pdf "fig:"){width="49.50000%"} ![Projection of the 2D fit result on (a) ${\ensuremath{M_R}\xspace}$ and (b) ${\ensuremath{R^2}\xspace}$ for the MU-ELE box in the $\ge$1 $\cPqb$-tag analysis path. \[fig:bmuele-box\]](bmuele-Rsq-prd.pdf "fig:"){width="49.50000%"}
![The $p$-values corresponding to the observed number of events in the $\ge$1 $\cPqb$-tag MU-ELE box signal regions (SR$i$). \[fig:bmuele-blue-plot\]](bp-muele-prd.pdf){width="\cmsFigWidthBox"}
Guide on emulating the razor analysis for additional studies
============================================================
In this appendix, we provide a guide to facilitate use of the razor analysis results for the interpretation of signal scenarios not considered here. We assume the existence of an event generator that can simulate LHC collisions for a given theoretical model. We also assume that this event generator is interfaced to a parton shower simulation, such that a list of produced particles at the generator level is available. The procedure described in this appendix represents a simplification of the analysis, giving conservative limits within the $\pm 1$ standard deviation band of the nominal result.
The following classes of stable particles are relevant to this analysis: i) invisible particles (neutrinos and any weakly interacting stable new particles, for example the LSP in SUSY models); ii) electrons; iii) muons; iv) all other stable electrically charged SM particles; and v) all other stable electrically neutral SM particles. It is possible to emulate the razor analysis as follows:
- all the visible stable particles are clustered into generator-level jets using the anti-algorithm with a distance parameter of 0.5.
- the generator-level is computed as $E^\text{miss}_\text{T, Gen} =
-\sum_p \pt^{p}$, where the sum runs over all the visible stable particles $p$.
- the detector resolution is applied to electrons and muons according to a simplified Gaussian resolution function. The RMS of the Gaussian smearing depends on the $\eta$ and $\pt$ values of the lepton, as well as its flavor. Similarly, the and jet momenta are smeared according to a Gaussian response model.
- the detector efficiency is applied to electrons and muons generating unweighted events from the reconstruction efficiency, interpreted as a probability (see Section \[sec:appemu\]). The efficiency depends on the $\eta$ and $\pt$ values of the lepton, its flavor, and its generator-level isolation, as computed from the stable particles in the event.
- the analysis selection and box classification is applied.
This procedure allows us to estimate the ${\ensuremath{R^2}\xspace}$ versus ${\ensuremath{M_R}\xspace}$ distribution for a signal model and the efficiency in each box. This is the information that is needed to associate a 95% upper limit to a given input model. The procedure matches the full simulation of CMS to within 20% and in general provides a result that is yet closer to the CMS full simulation. The result is in general conservative, since the computation of the upper limit starts from a simplified binned likelihood, which reduces the sensitivity to a signal. This procedure is not expected to correctly simulate the special case of slowly moving electrically charged particles (e.g., staus). The remainder of this appendix describes each step of the razor emulation in more detail, including the calculation of the exclusion limit.
Emulation of reconstructed electrons and muons {#sec:appemu}
----------------------------------------------
The emulation of reconstructed electrons and muons consists of two independent steps: the accounting for the detector resolution and for the reconstruction efficiency.
The effects of detector resolution can be incorporated through a Gaussian smearing of the genuine $\pt$ of a given lepton, while the lepton $\eta$ and $\phi$ can be considered to be unaffected by the detector resolution. The generated lepton is then replaced by the reconstructed one, having the same flight direction with a $\pt$ value randomly extracted according to a Gaussian distribution centered at $\pt^\text{Gen}$ and with $\sigma(\pt^\text{Gen})$ taken from Fig. \[fig:leptonResolution\]. Any lepton outside the two $\eta$ ranges considered in Fig. \[fig:leptonResolution\] should be discarded from the analysis.
![\[fig:leptonResolution\] Momentum resolution for () electrons and () muons within the barrel region of the CMS detector (squares) and in the endcaps (triangles).](Electron.pdf "fig:"){width="49.00000%"} ![\[fig:leptonResolution\] Momentum resolution for () electrons and () muons within the barrel region of the CMS detector (squares) and in the endcaps (triangles).](Muon.pdf "fig:"){width="49.00000%"}
To account for the reconstruction efficiency of a given lepton, the generator-level isolation is computed as follows: $$\text{GenIso}(\ell) = \frac{\sum_{p \neq \ell} \pt^{p}}{\pt^\ell},$$ where the sum runs over all the stable charged and neutral visible particles $p$ within a distance $\Delta R
=\sqrt{(\Delta\eta)^2+(\Delta\phi)^2}<0.5$ from the lepton.
Figure \[fig:electronReco\] shows the reconstruction probability versus the generated electron $\pt$ (before accounting for the detector resolution) for three ranges of $\text{GenIso}$ in the ECAL barrel ($\abs{\eta}< 1.4442$) and endcap ($1.5660<\abs{\eta}<2.5000$) regions. Different values are obtained for the tight and the loose electrons used to define the boxes.
{width="49.00000%"} {width="49.00000%"} {width="49.00000%"} {width="49.00000%"}
Similarly, the reconstruction efficiency for the tight muons is shown in Fig. \[fig:muonReco\]. The reconstruction of loose muons can be considered to be fully efficient for muons with $\pt>10$, since no isolation requirement is applied.
![\[fig:muonReco\] Muon reconstruction efficiency for tight muons pointing to the () barrel and () endcaps, estimated from the CMS MC simulation of $\ttbar$ events. The muon reconstruction is described in Ref. [@CMS_mu].](MuTightEff_0p0_eta_1p4.pdf "fig:"){width="49.00000%"} ![\[fig:muonReco\] Muon reconstruction efficiency for tight muons pointing to the () barrel and () endcaps, estimated from the CMS MC simulation of $\ttbar$ events. The muon reconstruction is described in Ref. [@CMS_mu].](MuTightEff_1p4_eta_2p1.pdf "fig:"){width="49.00000%"}
Once the lepton reconstruction probability is found, the detector efficiency effects can be imposed numerically: the lepton is rejected if a uniformly distributed random number in the range \[0,1\] is found to be larger than the reconstruction efficiency.
Emulation of reconstructed jets and missing ET
----------------------------------------------
The reconstruction of jets and can be emulated by applying a Gaussian resolution to the generator-level quantities. We show in Fig. \[fig:resoJetMET\] the dependence of the Gaussian $\sigma_\text{jet}$ on the jet $\pt$ (for the two relevant bins of $\eta$) and the . The dependence on $\eta$ or other quantities can be safely neglected. One should apply the resolution function to all the reconstructed jets and to the and then impose the acceptance selection on the reconstructed jets.
![\[fig:resoJetMET\]Transverse energy resolution for jets and , in the CMS MC simulation of events.](JetMET.pdf){width="49.00000%"}
Building the 2D templates
-------------------------
Once detector effects have been accounted for, jets are clustered in two megajets. The razor variables can be computed from the four-momenta of the two megajets.
Figure \[fig:razorT1tttt\] (left) and (middle) shows the ${\ensuremath{M_R}\xspace}$ and ${\ensuremath{R^2}\xspace}$ distributions for a sample of pair-produced gluinos of mass 800, where each gluino decays to a $\ttbar$ pair and a LSP of mass 300, obtained with the CMS fast simulations program and with the emulation described in this appendix. The efficiencies obtained for the six boxes are compared in Fig. \[fig:razorT1tttt\] (right).
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}
Evaluating the exclusion limit
------------------------------
The exclusion limit can be computed from the 2D signal templates and the box efficiencies, starting with the observed yield and the expected background. We consider a simplified likelihood obtained by defining bins in the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) plane. Each bin $i$ requires the observed yield $n_i$ and the expected background $\bar{b}_i \pm \delta_i$ computed by integrating the background model and taking into account the uncertainty in shape. The likelihood in a given box is then written as: where $\epsilon_i$ is the signal efficiency in that bin, $L$ is the luminosity, and $\sigma$ is the signal cross section; ${\mathop{\mathrm{log}\mathcal{N}}\nolimits}(b_i|\bar{b}_i,\delta_i)$ is the log-normal distribution describing the uncertainty in the background. ${\mathop{\mathrm{log}\mathcal{N}}\nolimits}(\rho| 1,
\delta_\rho)$ is the distribution describing the uncertainty in the signal efficiency. A value $\delta_\rho \sim 0.20$ (including the uncertainty in the integrated luminosity) is large enough to account for the use of a simplified detector emulation and the typical systematic uncertainty quoted in the analysis. Once this uncertainty is included, the uncertainty in the luminosity can be neglected to a good level of precision. Similarly, the total likelihood can be written as: In this case, the signal systematic parameter $\rho$ is common to the six boxes. A Bayesian upper limit (UL) on the cross section can then be computed assuming a flat prior distribution in $\sigma$: $$\frac{\int_0^{\sigma \mathrm{UL}} \rd\sigma \int \rd\rho\, \rd\vec{b}\, \mathcal{L}_{\text{TOT}}(\vec n|\sigma,\vec{b},\rho)}{\int_0^{+\infty}
\rd\sigma \int \rd\rho\, \rd{}\vec{b} \, \mathcal{L}_{\text{TOT}}(\vec n|\sigma,\vec{b},\rho)} = 0.95.$$ An implementation of this simplified limit calculator is together with the values of $n$, $\cPqb$, and $\delta$ for each bin in each box.
Limit on simplified models
--------------------------
Figure \[fig:smsemu\] shows the limit on the T2tt and T1tttt models, obtained by applying the simplified procedure described in this appendix. We generate a sample of SUSY events using the [8]{} [@sjostrand2007brief] program, scanning the two SMS planes. We then emulate the detector effects as described in this appendix to derive the efficiency and the (${\ensuremath{M_R}\xspace}$, ${\ensuremath{R^2}\xspace}$) signal probability density functions. We use this information to compute the excluded cross section for each point in the SMS plane.
{width="49.00000%"} {width="49.00000%"}
The CMS Collaboration \[app:collab\]
====================================
=5000=500=5000
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the results of a statistical study of the star formation rates (SFR) derived from the [*Galaxy Evolution Explorer*]{} ([*GALEX*]{}) observations in the ultraviolet continuum and in the H$\alpha$ emission line for a sample of about 800 luminous compact galaxies (LCGs). Galaxies in this sample have a compact structure and include one or several regions of active star formation. Global galaxy characteristics (metallicity, luminosity, stellar mass) are intermediate between ones of the nearby blue compact dwarf (BCD) galaxies and Lyman-break galaxies (LBGs) at high redshifts $z>$ 2 - 3. SFRs were corrected for interstellar extinction which was derived from the optical Sloan Digital Sky Survey (SDSS) spectra. We find that SFRs derived from the galaxy luminosities in the far ultraviolet (FUV) and near ultraviolet (NUV) ranges vary in a wide range from 0.18 $M_\odot$ yr$^{-1}$ to 113 $M_\odot$ yr$^{-1}$ with median values of 3.8 $M_\odot$ yr$^{-1}$ and 5.2 $M_\odot$ yr$^{-1}$, respectively. Simple regression relations are found for luminosities $L($H$\alpha )$ and $L$(UV) as functions of the mass of the young stellar population, the starburst age, and the galaxy metallicity. We consider the evolution of $L$(H$\alpha$), $L$(FUV) and $L$(NUV) with a starburst age and introduce new characteristics of star formation, namely the initial H$\alpha$, FUV and NUV luminosities at zero starburst age.'
author:
- 'S. L. Parnovsky'
- 'I. Y. Izotova'
- 'Y. I. Izotov'
title: 'H$\alpha$ and UV luminosities and star formation rates in a large sample of luminous compact galaxies'
---
Introduction {#s:Introduction}
============
@C09 first draw attention to galaxies at redshifts $z$ = 0.112 - 0.360 which were named “green peas” because of their compact structure and green colour on the $gri$ composite Sloan Digital Sky Survey (SDSS) images. Specific colours of these galaxies are mainly caused by the very strong \[O [iii]{}\] $\lambda$5007Å optical emission line. The equivalent widths EW($\lambda$5007) of this line in “green peas” redshifted into the SDSS $r$ band can be as high as $\sim$ 1000Å, resulting in a green colour on SDSS images. @C09 studied a sample of 251 colour-selected galaxies. Some of the galaxies from this sample are active galactic nuclei (AGN). However, most of “green pea” galaxies are found to be strongly star-forming ones with high star formation rates (SFR) of $\sim$ 10 $M_\odot$ yr$^{-1}$. These galaxies are characterised by low metallicity, stellar mass of $M_*$ $\sim$ 10$^{8.5}$ - 10$^{10}$ $M_\odot$, high specific SFR (SSFR) (up to $\sim$ 10$^{-8}$ yr$^{-1}$) which place them between nearby blue compact dwarf (BCD) galaxies and high-redshift ($z > 2 - 3$) UV-luminous Lyman-break galaxies [LBGs, see @G02 for a review]. The available [*Hubble Space Telescope*]{} ([*HST*]{}) high-angular resolution images of a few “green peas” reveal complex morphology on small spatial scales with several regions of star formation and an extended stellar component likely consisting of older stars [@C09; @A12]. @C09 suggested that “green pea” galaxies may be occurrences of the star formation mode prevailing in the early Universe. This galaxy class therefore may provide an excellent opportunity to understand in great detail many processes under physical conditions approaching to those in high-redshift galaxies.
The oxygen and nitrogen chemical abundances in star-forming “green peas” were studied by @A10. These galaxies are revealed to be genuine metal-poor galaxies with mean oxygen abundances of $\sim$ 20% solar. The N/O ratios are found to be unusually high for galaxies of the same metallicity. Detailed study lead @A10 to the conclusion that known general properties of “green peas”, namely high SSFR, extreme compactness and stellar mass, seem to be uncommon in the nearby universe, suggesting a short and extreme phase of their evolution. The possible action of both recent and massive interaction-induced inflow of gas, as well as selective metal-rich gas loss driven by supernova winds are discussed here as main drivers of the starburst activity in “green peas” and their oxygen and nitrogen abundances.
The first direct radio detection with low frequency Giant Metrewave Radio Telescope (GRMT) observations and discussion of the “green peas” properties comprising properties of a new class of sub-mJy sources were reported by @C12. It was shown that this detection may imply large magnetic fields ( $ \gtrsim 30 \,
\mu $G) in “green peas” under reasonable assumption about cosmic ray diffusion and total energy consideration. @C12 concluded that seed fields were amplified significantly (up to $\mu$G) because of turbulence as protogalactic and similar structures formed.
Detailed examination of a large sample of 803 star-forming luminous compact galaxies (LCGs) in the redshift range $z$ = 0.02 - 0.63 was carried out by @I11. These galaxies were selected from the SDSS Data Release 7 (DR7) [@A09] and comprise a complete spectroscopic SDSS sample of strongly star-forming LCGs with reliably derived chemical abundances. Their global properties are similar to those of the star-forming “green pea” galaxies. However, in contrast to “green pea” galaxies, the LCGs are selected on the base of the both their spectroscopic and photometric properties. Applied selection results in a $\sim$ 10 times larger sample, with galaxies spanning a redshift range about $\sim$ 2 times larger as compared to “green pea” sample [@C09]. For LCGs, the oxygen abundances 12 + log O/H are found to be in the range 7.6 - 8.4 with the median value of $\sim$ 8.11 confirming the results by @A10 for a subset of the “green pea” sample of @C09. The ranges of oxygen abundances and heavy element abundance ratios in LCGs are similar to those of nearby low-metallicity BCDs. In the \[O [iii]{}\]$\lambda$5007/H$\beta$ vs. \[N [ii]{}\]$\lambda$6583/H$\alpha$ diagnostic diagram [@K03] the LCGs are shown to occupy the region of high-excitation star-forming galaxies. The SFRs, derived from the H$\alpha$ line emission in the LCGs are revealed to vary in the large range of 0.7 - 60 $M_\odot$ yr$^{-1}$, with a median value of $\sim$ 4 $M_\odot$ yr$^{-1}$ which is about 3 times lower as compared to star-forming LBGs at $z$ $\sim$ 3 [@P01]. The SSFR in LCGs is extremely high and it varies in the range $\sim$ 10$^{-7}$ - 10$^{-9}$ yr$^{-1}$. All these properties imply that LCGs are likely the closest local counterparts of the high-redshift LBGs and Ly$\alpha$-emitting galaxies.
@G11 carried out the spectroscopic analysis of HG 031203, the host galaxy of a long-duration gamma-ray burst (GRB). The galaxy properties such as the oxygen abundance 12 + log O/H $=8.20\pm 0.03$, extinction-corrected H$\alpha$ luminosity $L$(H$\alpha$) = 7.27$\times$10$^{41}$ erg s$^{-1}$, stellar mass $M_*$ = 2.5$\times$10$^8$ $M_\odot$, SFR(H$\alpha$) = 5.74 $M_\odot$ yr$^{-1}$ and SSFR(H$\alpha$) = 2.3$\times$10$^{-8}$ yr$^{-1}$ in HG 031203 are found to be in the range covered by the LCGs properties. This fact implies that the LCGs with extreme star-formation, that also comprise “green peas” as a subclass, may harbor GRB.
@P12 analysed the oxygen and nitrogen abundances derived from global emission-line SDSS spectra of galaxies using the direct method based on the electron temperature determination from the \[O [iii]{}\] $\lambda$4363/($\lambda$4959 + $\lambda$5007) emission-line flux ratio and the two strong line O/N and N/S calibrations. Three samples of objects were compared, including the sample of “green pea" galaxies by @C09 with the detected \[O [iii]{}\] $\lambda$4363Å auroral line. @P12 concluded that the high nitrogen-to-oxygen abundance ratios derived in some “green pea" galaxies may be due to the fact that their SDSS spectra are the ones of composite nebulae made up of several components with different physical properties.
The local analogues of the strong Halpha Emitters (HAEs) dominated the $z\sim 4$ LBG population are identified by @SC12. Using the SDSS spectra authors show that at $z<0.4$ only 0.04% of galaxies are classified as HAEs with equivalent widths EW(H$\alpha$) of $>500$ Å, comparable to that of $z\sim 4$ HAEs. Local HAEs have lower stellar masses and lower UV luminosities than the $z\sim 4$ HAEs. On the other hand, their H$\alpha$-to-UV luminosity ratios and SSFRs are consistent with those of $z\sim 4$ HAEs indicating that the local analogues are the scaled-down versions of high-$z$ star-forming galaxies. Compared to the previously studied local Lyman-break analogs (LBAs) of the $z\sim 2$ LBGs which were selected using rest-frame UV fluxes [@H05], the local HAEs show similar UV luminosity surface densities, but lower metallicities and lower stellar masses. This supports the idea that local HAEs are less evolved galaxies than the traditional LBAs. Local HAEs show a strong He [ii]{} $\lambda$4686 Åemission line in the stacked spectrum, implying a population of hot young ($<10$ Myr) massive stars, similar to that seen in some Wolf-Rayet galaxies. The local HAEs also have properties similar to those of “green pea" galaxies.
In present paper, we extend the study of the properties of the “green peas” by further analysis of about 800 LCGs by @I11, selected from the SDSS DR7. The selection criteria of galaxy sample are briefly described in Section \[s:Sample\]. The correction of LCG fluxes for extinction is discussed in Section \[s:Ext\]. In Section \[s:Dep\] we carry out the statistical investigation of dependence of galaxy luminosities on other LCG characteristics. In Section \[s:SFR\] we discuss star formation rates of LCGs. The SFRs are derived from the extinction-corrected luminosities $L$(H$\alpha$), $L$(FUV) and $L$(NUV) in the H$\alpha$ emission line, the far ultraviolet (FUV) and the near ultrabviolet (NUV) ranges from [*Galaxy Evolution Explorer*]{} ([*GALEX*]{}) observations. The luminosity function for LCGs is discussed in Section \[s:LF\]. We summarise our results in Section \[s:Sum\]. We assume $H_0$ = 75 km s$^{-1}$ Mpc$^{-1}$ for distance estimates.
Sample selection, observational data {#s:Sample}
====================================
It is noted in Section \[s:Introduction\], that LCGs, in contrast to “green pea” galaxies [@C09], are selected on the base of the both their spectroscopic and photometric properties. Selection criteria and LCGs sample properties are described in detail by @I11. Briefly, these criteria are as follows :
– the extinction corrected luminosity of the H$\beta $ emission line is greater than $L$(H$\beta$) = 3$\times$10$^{40}$ erg s$^{-1}$;
– the equivalent width of the H$\beta$ emission line is high, EW(H$\beta$) $\geq $ 50Å. This criterion leads to selection only objects with strong emission lines in their spectra and thus the ones containing young starbursts with ages 3 - 5 Myr;
– only galaxies with well-detected \[O [iii]{}\] $\lambda $4363 [Å]{} emission line in their spectra, with a flux error less than 50 [%]{}, are selected. This criterion allows an accurate abundance determination using the direct method;
– only the star-forming galaxies were selected. Galaxies with obvious evidence of Seyfert 2 features are excluded;
– galaxies on their SDSS images are nearly compact at low redshifts and unresolved at high redshifts. Their typical angular sizes are less than 10.
@I11 used all LCG spectra and Monte Carlo simulations to fit spectral energy distributions in the wavelength range $\lambda$$\lambda$3800 – 9200Å. As for star formation history they assumed a single young burst with the age which is varied in the range $<$ 10 Myr, and a continuous star formation with a constant star formation rate, which started at the lookback time $t_1$ and finished at the lookback time $t_2$ $<$ $t_1$. Parameters $t_1$ and $t_2$ are varied in the range 10 Myr – 13 Gyr. The contribution of gaseous continuum in LCGs is very high, therefore it was fitted first using equivalent width EW(H$\beta$) of the H$\beta$ emission line and subtracted from the observed spectrum prior fitting of the stellar continuum. The masses of the young and old stellar populations, the age of the young burst and parameters for the old stellar population $t_1$ and $t_2$ were parameters of fitting. More details of fitting can be found in @I11. We use the results of modelling obtained by @I11.
We use the [*GALEX*]{} Medium Imaging Survey (MIS) and All-sky Imaging Survey (AIS) data (see http: //galex.stsci.edu/GR4) to estimate the galaxy UV SFR for the LCGs sample. [*GALEX*]{} is a NASA Small Explorer mission performed the all sky ultraviolet survey in two bands: far-UV (FUV, $\lambda_{\rm eff}$= 1528Å), and near-UV (NUV, $\lambda_{\rm eff}$ = 2271Å) [@M05]. MIS and AIS data contain information on fluxes of $\sim $10$^{7}$ galaxies. The prime goal of [*GALEX*]{} is to study star formation in galaxies and its evolution with time. The major science objectives and characteristics of [*GALEX*]{}, and of surveys are described by @Ma05 and @M05.
We matched the [*GALEX*]{} data and the sample of LCGs [@I11] and extracted FUV and NUV fluxes from the [*GALEX*]{} MIS and AIS database. These data combined with the NASA/IPAC Extragalactic Database (NED) data provide the determination of the galaxy UV luminosities.
We excluded the galaxies with the UV flux errors exceeding 50% and the data for the multiple UV sources within the aperture of $\sim$30.
Correction for extinction {#s:Ext}
=========================
Because the radiation of galaxies is reduced by dust extinction, we applied reddening corrections to H$\alpha$ and UV band fluxes using @C89 reddening law. Adopting the $R(V)$-dependent extinction law from @C89 with $R(V)$ = $A(V)/E(B-V)$ = 3.1, we obtain $A$(H$\alpha$) = 2.54$\times$$E(B-V)$ in H$\alpha$, $A$(FUV) = 8.15$\times$$E(B-V)$ in the FUV band and $A$(NUV) = 9.17$\times$$E(B-V)$ in the NUV band. The extinction coefficient $C$(H$\beta$), reddening $E(B-V)_{\rm SDSS}$, and the equivalent width of underlying stellar hydrogen absorption lines were obtained by @I11 from the hydrogen Balmer decrement in the redshift-corrected spectra. All hydrogen line fluxes were corrected for both the reddening and underlying stellar absorption. For comparison, we also use the reddening $E(B-V)_{\rm NED}$ from the NED which was obtained from the Milky Way reddening maps by @S98. The $E(B-V)_{\rm SDSS}$ and $E(B-V)_{\rm NED}$ differ because the former quantity is the total reddening along the line of sight which includes extinction from both the Milky Way and the studied galaxy, while the latter quantity is the reddening in the Milky Way only.
@I11 derived $E(B-V)_{\rm SDSS}$ = 0 for 65 out of 803 galaxies. For some other galaxies they obtained $E(B-V)_{\rm SDSS}$ $<$ $E(B-V)_{\rm NED}$. To correct galaxy fluxes for extinction we use the $E(B-V)_{\rm SDSS}$ for all LCGs. Alternatively, we analyse all the UV data adopting $E(B-V)_{\rm SDSS}$ if $E(B-V)_{\rm SDSS}$ $>$ $E(B-V)_{\rm NED}$ and $E(B-V)_{\rm NED}$ if $E(B-V)_{\rm SDSS}$ $<$ $E(B-V)_{\rm NED}$. The difference between these two approaches is small because of the low extinction in LCGs and does not influence appreciably our results and conclusions (see Sect. \[s:Dep\]).
We derive mean reddenings $E(B-V)_{\rm SDSS}$ of 0.133 and 0.134 for LCGs detected in the FUV and NUV ranges, respectively. We also find that the mean reddening difference $E(B-V)_{\rm SDSS}$ $-$ $E(B-V)_{\rm NED}$ is 0.106. This difference is the rough average internal galaxy reddening which can be used for galaxies without spectroscopic estimates of reddening. In principle, the correction for the Milky Way and intrinsic reddening should be done separately. This is because the Milky Way correction has to be applied to the fluxes at the observed wavelengths, while the correction for the intrinsic reddening should be applied to the fluxes at the redshift-corrected wavelengths. However, ignoring of the separate correction for the Milky Way and intrinsic reddening would introduce very small additional uncertainties in the extinction-corrected fluxes. All LCGs are located at high galactic latitudes where the Milky Way extinction is very low, with the mean $E(B-V)_{\rm NED}$ of $\sim$ 0.03. Most of LCGs are also low-redshift galaxies with $z$ $<$ 0.2. Only few galaxies are at redshifts $z$ $>$ 0.3. Therefore, the difference between the extinction correction of the flux applied with the observed wavelength and the redshift-corrected wavelength for the galaxy with the redshift $z$ = 0.3 and adopting $E(B-V)_{\rm NED}$=0.03 is $\la$ 10% in the FUV and NUV bands, and $\la$3% for the H$\beta$ and H$\alpha$ emission lines. This difference for galaxies with lower redshifts is lower, e.g. it is only $\sim$ 5% in the FUV and NUV bands and $\sim$ 2% for H$\beta$ and H$\alpha$ for the galaxy with $z$ = 0.2. Furthermore, Milky Way extinction maps by @S98 are obtained for large apertures of 6, which are much larger than the angular sizes of LCGs. Therefore, the extinction derived from the @S98 maps may not correspond to the real extinction in the direction on the galaxy if small-scale spatial extinction variations are present. These are reasons why we do not separate correction for the Milky Way and intrinsic reddening and use in a subsequent analysis $E(B-V)_{\rm SDSS}$ derived by @I11 from the hydrogen Balmer decrement.
Accordingly to @I11 we split our sample into two subsamples of 276 “regular” galaxies with the round shape and 519 “irregular” galaxies with some sign of disturbed morphology suggesting the presence of two or more star-forming regions and their interaction. For these two subsamples we obtained $E(B-V)_{\rm SDSS}$ $-$ $E(B-V)_{\rm NED}$ of 0.081 and 0.120, respectively. Using the Student criterion we derived $t$ = 4.98. This value suggests that the probability of the statistically significant difference of $E(B-V)_{\rm SDSS}$ $-$ $E(B-V)_{\rm NED}$ for two subsamples is greater than 99.9%. Therefore, extinction in galaxies with the non-round morphology is higher. On the other hand, we do not find tight correlation between $E(B-V)_{\rm SDSS}$ $-$ $E(B-V)_{\rm NED}$ and heavy element abundances. Apparently, the extinction is determined not only by the dust mass which is expected to be higher in galaxies with higher metallicity, but also by the spatial distribution of dust. This distribution seems to be different in “regular” and “irregular” galaxies.
We correct the galaxy fluxes for extinction according to $I(\lambda)=F(\lambda)\times 2.512^{A(\lambda)}$, where $F(\lambda)$ and $I(\lambda)$ are the observed and the corrected fluxes, respectively. The extinction-corrected [*GALEX*]{} FUV and NUV fluxes from LCGs are nearly three times higher than the observed fluxes. In addition, H$\alpha$ fluxes were corrected for an aperture comparing the total galaxy apparent magnitude $m$ and the magnitude $m$(3) inside the SDSS spectroscopic aperture of 3 in a certain SDSS band depending on the galaxy redshift. We compare SDSS magnitudes $m$ = $r$ and $m$(3) = $r$(3) for galaxies with redshifts $<$ 0.04, $i$ and $i$(3) for galaxies with redshifts in the range 0.04 – 0.26, and $z$ and $z$(3) for galaxies with redshifts $\geq$ 0.26. Then, the aperture H$\alpha$ flux correction is $A$=2.512$^{m-m(3\arcsec)}$, where $m$ = $r,i,z$ depending on the galaxy redshift.
Relations between galaxy luminosities and other global characteristics {#s:Dep}
======================================================================
For each galaxy we calculated its H$\alpha$, FUV and NUV luminosities. We use the regression analysis to study a dependence of the LCG luminosities on other their characteristics. To provide a simple way for comparing $L($H$\alpha)$, $L$(FUV) and $L$(NUV) we use some other parameters being proportional to them. Namely, we use the calibration for SFRs averaged over the reasonable timescale for different SFR tracers and defined by @K98 as $$\label{eqn:1}
{\textrm}{SFR}=k \times L.$$ The conversion factors $k$ between the SFR and the $L$(H$\alpha$), $L$(FUV) and $L$(NUV) in Eq. \[eqn:1\] are derived using the evolutionary synthesis models. The coefficient $k$ depends on the time scale of star formation, initial mass function (IMF) and galaxy metallicity. Adopting the solar metallicity, the IMF with the power-law index 2.35 and mass limits of 0.1 and 100 $M_\odot$ [@S55], SFR in $M_{\odot}$ yr$^{-1}$, $L($H$\alpha)$ in erg s$^{-1}$, $L$(FUV) and $L$(NUV) in erg s$^{-1}$ Hz$^{-1}$ @K98 obtained the coefficient $k$ of $7.9 \times 10^{-42}$ for the H$\alpha$ luminosity and $1.4 \times 10^{-28}$ for the FUV and NUV luminosities.
There are some other modifications of Eq. \[eqn:1\]. In particular, @K2009 proposed a composite SFR calibration based on the luminosities of both the H$\alpha$ emission line and the UV continuum, adopting the IMF by @Kr and obtained $k$ = $5.5 \times 10^{-42}$. For clarity, we will use the values of $k$ derived by @K98 for the solar metallicity and Salpeter IMF. Detailed review and analysis of SFR calibrations based on the H$\alpha$ $\lambda$6563Å and \[O [ii]{}\] $\lambda$3727Å emission lines, far infrared and ultraviolet continua are given by @K98 and @Ca12.
While the calibration defined by Eq. \[eqn:1\] holds for continuous or quasi-continuous star formation which is common in the big galaxies with frequent starbursts, the situation is more complicated in the dwarf galaxies like LCGs with the strong and rare bursts of star formation. In these systems with the instantaneous bursts the time interval of the star formation is not well defined, and the observed H$\alpha$ luminosity strongly decreases on a time scale of a few Myr. Similar conclusions in lesser extent can be drawn for the SFRs derived from the FUV-band and the NUV-band luminosities. However, again for the clarity, we adopt that Eq. \[eqn:1\] can be applied for the determination of SFR in LCGs.
A set of galaxy parameters includes the primary parameters obtained directly from its SDSS spectrum, such as the redshift, the H$\alpha$ flux and the chemical element abundances. For each galaxy we also use secondary parameters obtained by modelling galaxy spectral energy distribution, following @I11, namely the masses of young $M$(young) and old $M$(old) stellar populations, the total mass of a stellar population $M_{*}$, the age of a starburst $t$(young), and the lower ($t_2$) and upper ($t_1$) limits for the age of old stars. All these parameters are distance-independent. Nevertheless, some nearby galaxies from our sample have angular diameters larger than the 3 SDSS aperture. Neglecting an aperture correction for these galaxies would result in the underestimation of stellar population masses. To estimate the proper masses we used the aperture correction similar to the one used for luminosities. The luminosities of galaxies with angular diameters greater than 3 without this correction deviate substantionally from regression relations used for approximation of their SFRs as a function of their luminosities. These deviations vanish after taking into account aperture corrections.
We consider all above-mentioned parameters. However, as one can see below, only two secondary parameters, namely $M$(young) and $t$(young) have the statistically significant impact on the luminosity. Hereafter we will use the brief notations $m$ $\equiv$ $M$(young) and $T$ $\equiv$ $t$(young) for the mass of the young stellar population and its age, respectively.
![Mass of the young stellar population $m$ vs. the starburst age $T$. Dots and open circles correspond to subsamples of “regular” and “irregular” galaxies, respectively[]{data-label="fig:1"}](fig1.eps){width="\columnwidth"}
Consider distributions of some secondary parameters. The values of the starburst age $T$ are distributed rather uniformly over the interval $T<5$ Myr. Galaxies with $T \ga 5.6$ Myr are not included in our sample due to the criterion EW(H$\beta$) $>$ 50Å. This selection is also resulted in a relative decrease of the number of galaxies with $5<T<5.6$ Myr. The distribution of $m$ depends on $T$. This is illustrated in Figure \[fig:1\]. One can see that the mean value of $m$ for $T>3.2$ Myr increases with the increase of $T$ with the best fit $\log(m/10^7 M_\odot)=-0.57 \pm 0.22 +
(0.33 \pm 0.05)\times T$(Myr) for the subsample of “regular” galaxies and $\log(m/10^7 M_\odot)=-0.75 \pm 0.17
+(0.40 \pm 0.04) \times T$(Myr) for the subsample of “irregular” galaxies.
Note that both values $m$ and $T$ were calculated assuming a single star-forming region in the galaxy, while several regions of star formation with different $m$ and $T$ sometimes are observed in the galaxies. In this case we cannot rely on single values of $m$ and especially $T$, the latter value would tend to be larger. To prove that we consider a case of the two star-forming regions with equal stellar masses, one is very young and another is older. The young star-forming region would dominate in the H$\alpha$ luminosity because of the strong dependence of the flux of ionising radiation on a starburst age. On the other hand, the intensity of the optical continuum is less sensitive to the starburst age, therefore both star-forming regions equally contribute to the optical continuum, resulting in lower EW(H$\alpha$) as compared to the case with a single young burst. Consequently, this would result in a larger $T$ because it is determined mainly by EW(H$\alpha$). It is natural to expect that galaxies with larger masses of young stellar population with higher probability consist of several regions of star formation at different evolutionary stages. Therefore they would tend to have larger $T$ as compared to the galaxies with smaller masses of the young stellar population. That is why these galaxies concentrate in the upper right corner in Figure \[fig:1\].
We will show later that the ratio $L/m$ decreases exponentially with increasing $T$ if $T>3.2$ Myr. As a result, the lower right corner in Figure \[fig:1\] is empty because the sample is flux-limited $L$(H$\beta$) $>$ 3$\times$10$^{40}$ erg s$^{-1}$. Galaxies with low $m$ and $T>4$ Myr have luminosities below the threshold and do not enter the sample.
Our goal is to find simple but statistically significant dependences of the galaxy luminosity on other primary and/or secondary galaxy parameters. First, we search for a set of parameters to which the galaxy luminosity is the most sensitive. Later, we will find the best formulae to describe these dependences and analyse them.
At first we do not take into account galaxy metallicities and analyse linear dependences of their luminosities on other parameters and their combinations. Regressors were chosen accordingly to the statistical Fisher test [@ref:F]. We reject the regressors with statistical significance below the threshold of 99.95% and consider the regressions which are good for all six subsamples by selecting 3 wavelengths (H$\alpha$, FUV and NUV) and 2 morphologies - “regular” and “irregular”. These regressions have the form $$\label{eqn:2}
\mathrm{SFR}=C_1+C_2m+C_3mT^2+C_4m^2.$$ The values, the errors and the statistical significances of the coefficients $C_i$ obtained by the least squares method (LSM) are shown in Table \[tbl:1\] for $m$ expressed in solar masses and $T$ in yr. The root mean square (RMS) standard deviation $\sigma$ and the number $N$ of the galaxies in each subsample are also shown in the Table. Significances are characterised by the value $F$ obtained by Fisher’s test. The critical $F$ values, corresponding to the statistical significances of 90, 95, 97.5, 99, 99.5, 99.9 and 99.95% are equal to 2.71, 3.84, 5.02, 6.64, 7.88, 10.83 and 12.10, respectively. One can see from Table \[tbl:1\] that the significances of regressors $C_2$ and $C_3$ in Eq. \[eqn:2\] are higher than 99.95%. For the regressor $C_4$ we choose the threshold value $F>10$. Only for one subsample the threshold attains a higher value. For all other subsamples we assume $C_4=0$ and indicate in parentheses the value of $F$ for the case $C_4\ne 0$.
[lcccccc]{} Subsample&$N$&$\sigma$&$C_1$($F$)&$C_2\times 10^8$($F$)&$C_3\times 10^{22}$($F$)&$C_4\times 10^{18}$($F$)\
\
1. H$\alpha$,“regular” &$276$&$3.4$&$1.15\pm 0.35(10.9)$ &$18.3\pm 0.8( 528)$&$-43.1\pm 2.6( 268)$&$-33.8\pm 8.0(17.8)$\
2. H$\alpha$,“irregular” &$519$&$4.0$&$1.41\pm 0.24(35.3)$ &$17.0\pm 0.4(1602)$&$-45.5\pm 1.6( 766)$&$-(0.7)$\
3. FUV,“regular” &$213$&$4.8$&$0.19\pm 0.45( 0.2)$ &$12.4\pm 1.0( 149)$&$-22.1\pm 4.0( 30)$&$-(3.9)$\
4. FUV,“irregular” &$418$&$4.4$&$1.29\pm 0.31(17.5)$ &$ 8.0\pm 0.6( 159)$&$-14.2\pm 2.4( 34)$&$-(0.4)$\
5. NUV,“regular” &$233$&$5.6$&$0.42\pm 0.50( 0.7)$ &$14.5\pm 1.1( 167)$&$-20.7\pm 4.4( 22)$&$-(2.3)$\
6. NUV,“irregular” &$435$&$7.8$&$0.27\pm 0.53( 0.3)$ &$14.8\pm 1.1( 180)$&$-23.5\pm 4.2( 31)$&$-(8.7)$\
\
1. H$\alpha$,“regular” &$276$&$3.5$&$0$ &$20.2\pm 0.6(1275)$&$-46.3\pm 2.5( 343)$&$-49.2\pm 6.6(55.2)$\
2. H$\alpha$,“irregular” &$519$&$4.1$&$0$ &$18.4\pm 0.4(2592)$&$-49.8\pm 1.5(1062)$&$-(2.0)$\
3. FUV,“regular” &$213$&$4.8$&$0$ &$12.6\pm 0.8( 227)$&$-22.8\pm 3.6( 40)$&$-(1.8)$\
4. FUV,“irregular” &$418$&$4.5$&$0$ &$ 9.6\pm 0.5( 352)$&$-19.3\pm 2.1( 82)$&$-(0.4)$\
5. NUV,“regular” &$233$&$5.6$&$0$ &$15.1\pm 0.9( 264)$&$-22.3\pm 4.0( 32)$&$-(0.5)$\
6. NUV,“irregular” &$435$&$7.8$&$0$ &$15.2\pm 0.9( 303)$&$-24.6\pm 3.6( 47)$&$-(7.1)$\
\[tbl:1\]
Consider an implication of Eq. \[eqn:2\]. One would expect that SFRs vanish at the low-mass limit $m=0$. However, Eq. \[eqn:2\] (case (a) in Table \[tbl:1\]) implies that SFR is equal to non-zero $C_1$ at this limit with a large statistical significance for the three subsamples out of six. We assume that this is due to the uncertainties in the estimation of $m$. We suggest that this statistical effect resembles the well-known Malmquist bias. A similar effect was studied in connection with the large-scale collective galaxy motion [@ref:ParPar08]. To verify this hypothesis we performed some Monte Carlo simulations. For these simulations we need many generated mock catalogues, preferably with the distribution of the parameters similar to that in real subsamples.
First, we adopt the values of $m$ and $T$ from the corresponding real subsamples. Then, we set $C_1 = 0$ and calculate the coefficients $C_2$, $C_3$, and $C_4$ (if the last one is statistically significant) for the regression relation Eq. \[eqn:2\] using the LSM. These coefficients are shown in Table \[tbl:1\], case (b). After that, we derive the SFR values for the case (b) from Eq. \[eqn:2\] with $C_1 = 0$. As a result we obtain the initial “unbiased” set of $m$, $T$ and SFR values. Note that the dependence of SFR on $m$ and $T$ in this sample is functional, not statistical. The real SFR data are taken into account only indirectly via the set of coefficients $C_2$, $C_3$ and $C_4$. Naturally, if we apply the LSM with the regression defined by Eq. \[eqn:2\] to this data we obtain coefficients in the corresponding row of Table \[tbl:1\], case (b), but with $\sigma=0$.
Next, using the Monte Carlo technique we add a noise, i.e. random errors to the unbiased values of SFR or $m$. We find that random normal errors in SFR values result in the nonshifted distributions of $C_i$ values obtained by the LSM. In this case a random value for $C_1$ has the low statistical significance. A completely different situation arises when random errors in $m$ are considered. In this case the distributions of $C_i$ values are shifted relative to the “unbiased" ones and we obtain a non-zero value for $C_1$, sometimes with the large false statistical significance. Other coefficients tend to attain values nearer to zero if the noise increases.
Could the effect of uncertainties in the $m$ determination explain a non-zero value of $C_1$ obtained from the real data? To prove this suggestion we compare the values of $C_i$ in Table \[tbl:1\] (case (a)) with the ones obtained using Monte Carlo simulations. For that, we use the “unbiased” SFR and $T$ values from the sample calculated with the coefficients from case (b) in Table \[tbl:1\] and add a random noise to $m$. We obtain the “biased” value $m_{bias}$. The distribution of $m_{bias}$ is expected to be log-normal, therefore we added the noise using the equation $$\label{eqn:3}
m_{bias}=m(1+s \times \xi),$$ where $\xi$ is a normally distributed random value with zero mean and unity dispersion and $s$ characterises an amplitude of the noise. Then we calculate $C_i$ values applying LSM to the set of “unbiased” SFR and $T$ values and the values of $m_{bias}$. After repeating this procedure 1000 times we obtain the mean values and the distribution of $C_i$ as well as the mean value of $\sigma$. The results for different $s$ are shown in Table \[tbl:2\] for the H$\alpha$ subsample No. 1 in Table \[tbl:1\]. One can see that the calculated values of $C_1$ in Table \[tbl:1\] (case (a)) can be explained by the Monte Carlo model with $s=0.34$ (Table \[tbl:2\]). In addition to the shift of $C_1$ the noise in $m$ also affects the values of other coefficients $C_i$. They are shifted closer to the values from Table \[tbl:1\] (case (a)). The value of modelled $\sigma$ is smaller than that obtained from real data due to the contribution of SFR errors in real data, in addition to errors in $m$.
We used similar Monte Carlo simulations for other two subsamples with statistically significant non-zero $C_1$ values (subsamples No. 2 and 4 in Table \[tbl:1\]) and obtained similar results attained at $s=0.32$ and $s=0.38$, respectively.
Therefore, we adopt that the true preliminary regression for SFR is Eq. \[eqn:2\] with $C_1=0$ and non-zero values of $C_2$, $C_3$ and sometimes $C_4$ (see case (b) in Table \[tbl:1\]). The uncertainties in the estimation of $m$ lead to the appearance of the first regressor in Eq. \[eqn:2\] with a false statistical significance. Hereafter we will use only the models with $C_1=0$. Then we can rewrite Eq. \[eqn:2\] in the form $$\label{eqn:4}
\mathrm{SFR}/m=C_2+C_3T^2+C_4m,$$ introducing a new value SFR/$m$ which gives us a possibility to consider the dependence on $T$ regardless of the dependence on $m$. The first term in Eq. \[eqn:4\] is the main one and its meaning is that SFR of a galaxy is approximately proportional to the mass of the young stellar population. This is quite obvious because radiation in the UV continuum and H$\alpha$ emission line is emitted mainly by young O-stars. We will return later to the consideration of the possible nonlinear dependence of SFR on $m$.
![Ratio $L$(H$\alpha$)/$m$ of the galaxy luminosity in the H$\alpha$ emission line to the mass $m$ of the young stellar population vs. the age of the starburst $T$. Dots and open circles correspond to subsamples of galaxies with “regular” and “irregular” shape. The solid and dashed lines correspond to the best fits for the subsamples 1 and 2 with the mean galaxy metallicity using the regression Eq. \[eqn:9\][]{data-label="fig:2"}](fig2.eps){width="\columnwidth"}
In Figure \[fig:2\] we show the dependence of $L$(H$\alpha$)/$m$ $\propto$ SFR/$m$ on the starburst age $T$. It is seen from the Figure that the ratio $L$(H$\alpha$)/$m$ is practically constant for $T < T_0$=3.2 Myr and decreases practically exponentially for larger $T$. We note that SFR(H$\alpha$), $m$, and $T$ are not directly correlated because they are based on the different features in the spectra: SFR(H$\alpha$) is not modelled and is derived from the H$\alpha$ flux, $m$ and $T$ are modelled but they depend on different features. The mass of young stellar population $m$ is determined by the luminosity and the shape of the stellar continuum in the visible range, while $T$ depends mainly on the equivalent width EW(H$\beta$) of the H$\beta$ emission line. In particular, the equivalent width EW(H$\beta$) is lower by a factor of many times for the burst with the age of 6 Myr as compared to that for the burst with the age of 3 Myr, while the flux of the stellar continuum near H$\beta$ is decreased by only $\sim$ 5% [@L99]. Therefore, SFR(H$\alpha$), $m$, and $T$ can be considered as independent parameters. The most plausible explanation of the $L$(H$\alpha$)/$m$ – $T$ relation is that most massive stars with masses of $\sim$ 100 $M_\odot$, producing most of the ionising radiation, disappear after the starburst age $T$ $\sim$ 3.2 Myr, corresponding to their lifetimes. It also implies that stars in LCGs are formed during very short time periods, otherwise, in the case of extended bursts, $T_0$ would be greater than $\sim$ 3.2 Myr. The best fit for $T > 3.2$ Myr is $\log({\rm SFR}/m) = -5.61\pm 0.07+(-0.316\pm 0.016)\times T({\rm Myr})$ for the “regular” subsample and $\log({\rm SFR}/m) = -5.62\pm 0.05+(-0.326\pm 0.011)\times T({\rm Myr})$ for the “irregular” subsample.
Is this effect statistically significant? The values of the Fisher coefficients for linear terms in regression relations for both subsamples exceed 398. For additional proof we performed the Student test to compare the mean values of the SFR/$m$ for $T<3.5$ Myr and $T>4$ Myr. In all cases the Student test gives $t$ above 3.31, corresponding to the statistical significance of 99.95%. The results are shown in Table \[tbl:3\], where $\sigma$ is the standard deviation. One can see that mean values of SFR/$m$ for $T<3.5$ Myr are essentially larger than those for $T>4$ Myr. This effect is more pronounced for the H$\alpha$ radiation as compared to the FUV and NUV radiation. This is because 1) more massive short-lived stars contribute to the ionising radiation and 2) the luminosity of ionising radiation is much stronger increased with the mass of a star as compared to the UV-radiation. We note, however, that the ratio SFR/$m$ decreases more slowly in comparison with the @SV98 population synthesis models for young stellar populations given the appropriate heavy element abundance.
In accordance with the above discussion we introduce the function $f(T)$ $$\label{eqn:5}
f(T)=\left\{
\begin{array}{rl}
1&\text{if $T<T_0$;}\\
\exp(-p\times(T-T_0))&\text{if $T>T_0$,}\\
\end{array}\right.$$ where $T_0=3.2$ Myr and $p= 0.75$ Myr$^{-1}$. These values are the preliminary ones obtained from Figure \[fig:2\]. They will be improved later. Using $f(T)$ and transforming Eq. \[eqn:4\] we obtain the regression relation $$\label{eqn:6}
SFR= C_5mf(T)+C_4m^2$$ and apply LSM to calculate the values and the errors of the coefficients.
The improved dependence of SFR/$m$ on $T$ (dashed and solid lines in Fig. \[fig:2\] correspond to best fits with the optimal values of $p$) and the existence of the correlation between $T$ and $m$ (Fig. \[fig:1\]) lead us to the following results. For the H$\alpha$ “regular” subsample the second term becomes statistically insignificant and for H$\alpha$ “irregular” subsample the second term becomes positive and statistically significant. The values of RMS standard deviations become smaller than those from Eq. \[eqn:4\] because of the better approximation of SFR/$m$ on $T$.
{width="\columnwidth"} {width="\columnwidth"}
The dependences of $L$(FUV)/$m$ and $L$(NUV)/$m$ on $T$ are more gently sloping and have a larger scattering of $L$(FUV)$/m$ and $L$(NUV)$/m$ as compared to $L$(H$\alpha$)$/m$. The dependences $L$(FUV)$/m$ on $T$ and $L$(NUV)$/m$ on $T$ are shown in Figure \[fig:3\]. The shape and application of $f(T)$ to describe these dependences will be discussed below. First of all we will use the function $f(T)$ for the H$\alpha$ radiation. We now consider the dependence of $L$ on the galaxy metallicity. As a measure of metallicity we choose the oxygen abundance \[O\] $\equiv$ 12 + log(O/H). \[O\] was accurately derived by @I11 for every galaxy in the LCG sample. We adopt these values for the regression determination. To analyse the trend we restrict ourselves to the linear dependence on \[O\] only. Oxygen abundances of the galaxies from the LCG sample vary in the range from 7.52 to 8.47 with the median value of 8.06. We note that the use of nonlinear dependences on \[O\], for example, $10^{\rm [O]}$, does not make any additional progress in reducing of the regression RMS standard deviations.
First, we discuss potentially misleading methods of studying the trends in the dependence of $L$ on \[O\]. The simplest approach is to calculate the mean $L$ for the subsamples with low and high oxygen abundances and to treat the obtained difference as an dependence on the oxygen abundance. Adopting the median value \[O\] = 8.06 as a value dividing galaxies with low and high oxygen abundances, we find that the subsamples with high \[O\] have the mean $L$ (or SFR) values $\sim$ 1.5 times larger than for the subsamples with low \[O\]. However, the mean masses of young stellar population in higher-metallicity subsamples are larger by a factor of $\sim$ 1.5 in comparison with subsamples with lower metallicity. Therefore, the differences in SFRs are mostly due to the differences in masses of the young stellar population and are not directly related to the differences in the metallicity.
The best way to investigate the direct impact of galaxy metallicity on its luminosity for the galaxy with the same values of $m$ and $T$ is to include the metallicity directly in the regression relation. Therefore, we fit the SFR values for H$\alpha$ emission line using the set of regressors $$\label{eqn:7}
SFR= C_5mf(T)+C_4m^2+C_6m({\textrm}{[O]}-\left<{\textrm}{[O]}\right>)$$ with $T_0=3.2$ Myr and $p=0.75$ Myr$^{-1}$. Here $\left<{\textrm}{[O]}\right>$ is the mean oxygen abundance of the sample galaxies. We subtract this value from the galaxy’s oxygen abundance \[O\] to make the last term in Eq. \[eqn:7\] practically orthogonal to the first one and in this way to keep the results of the Fisher test for the first two regressors. Using the LSM we calculate the coefficients, their errors and statistical significances of the regression defined by Eq. \[eqn:7\]. For the H$\alpha$ “regular” subsample we obtain $\sigma=3.42$, $C_5=(2.13 \pm 0.05)\times 10^{-7} (F=1654), C_4=(-3.6 \pm
4.2)\times 10^{-18} (F=0.7), C_6=(-4.7 \pm 1.0)\times 10^{-8}
(F=20)$. For the H$\alpha$ “irregular” subsample the derived values are $\sigma=3.8$, $C_5=(1.94 \pm 0.03)\times10^{-7} (F=4415), C_4=(3.4 \pm
1.6)\times10^{-18} (F=4.9), C_6=(-3.2 \pm 0.7)\times 10^{-8}
(F=23)$. We note the drop of statistical significance of the nonlinear term with $C_4$ below the threshold for both subsamples.
Switching to the UV luminosities, we generalise Eq. \[eqn:4\] by adding the term with the dependence on metallicity: $$\label{eqn:8}
\mathrm{SFR}=C_2m+C_3mT^2+C_4m^2+C_6m({\textrm}{[O]}-\left<{\textrm}{[O]}\right>).$$ Derived coefficients, their errors and statistical significances for FUV and NUV subsamples are shown in Table \[tbl:4\]. All coefficients $C_6$ are negative and have the statistical significanse more than 99.5%. Coefficients $C_4$ are statistically insignificant for all FUV and NUV subsamples, similar to that for H$\alpha$ subsamples.
Summarising, we find that a nonlinear term $m^2$ in the regression relations (Eqs. \[eqn:7\] and \[eqn:8\]) is statistically insignificant for all six subsamples. Speaking more precisely, we conclude that the statistical analysis gives us no reason to justify the existence of such a term. Would it be statistically significant it makes the ratio $L/m$ be depending on the young stellar population masses. This could be treated as an impact of the environment or as a result of some kind of an interaction of several regions of star formation. However, it is difficult to analyse these effects statistically because of the correlation between $m$ and $T$, which would lead to ambiguous conclusions.
[lccccccc]{} Subsample&$N$&$\sigma$&$\bigstrut C_2\times
10^8$($F$)&$C_3\times 10^{22}$($F$)&$C_4\times 10^{18}$($F$)&$C_6\times
10^{8}$($F$)&$\left<[{\rm O}]\right>$\
\
3.FUV,regular &$213$&$4.7$&13.6$\pm$0.9(231)&–31.9$\pm$4.1( 60)&$ 30.9\pm10.5(8.6)$&$ -7.0\pm 1.7(16.2)$&8.05\
4.FUV,irregular &$418$&$4.4$&10.2$\pm$0.5(375)&–20.7$\pm$2.7( 60)&$ -3.4\pm3.5(1.0)$&$ -3.9\pm 1.0(16.7)$&8.13\
5.NUV,regular &$233$&$5.5$&16.0$\pm$1.0(246)&–30.2$\pm$4.7( 42)&$ 23.8\pm12.1(3.9)$&$ -5.8\pm 1.9( 9.1)$&8.05\
6.NUV,irregular &$435$&$7.3$&17.8$\pm$0.9(405)&–39.1$\pm$4.3( 83)&$ 12.9\pm5.6(5.3)$&$-12.0\pm 1.6(57.7)$&8.13\
\
3.FUV,regular &$213$&$4.7$&13.8$\pm$0.9(231)&–26.7$\pm$3.8( 50)&0&$ -4.9\pm 1.6( 9.2)$&8.05\
4.FUV,irregular &$418$&$4.4$&10.3$\pm$0.5(378)&–22.3$\pm$2.2(101)&0&$ -3.8\pm 0.9(16.1)$&8.13\
5.NUV,regular &$233$&$5.5$&16.1$\pm$1.0(248)&–26.2$\pm$4.2( 38)&0&$ -4.2\pm1.7( 5.7)$&8.05\
6.NUV,irregular &$435$&$7.3$&17.3$\pm$0.9(403)&–33.5$\pm$3.6( 88)&0&$-12.2\pm 1.6(59.9)$&8.13\
7.FUV,all &$631$&$4.8$&11.7$\pm$0.5(575)&–25.2$\pm$2.0(157)&0&$ -4.8\pm 0.9(31.7)$&8.10\
8.NUV,all &$668$&$6.8$&17.1$\pm$0.7(647)&–31.6$\pm$2.8(130)&0&$ -9.7\pm1.2(65.4)$&8.10\
\[tbl:4\]
We discard nonlinear term with $m^2$ adopting $C_4$ = 0 in Eq. \[eqn:7\] and obtain a regression relation in the form $$\label{eqn:9}
SFR = C_5 m f(T)+C_6 m ({\textrm}{[O]}-\left<{\textrm}{[O]}\right>).$$ We apply this relation for the H$\alpha$ radiation. For the subsamples No.1 and 2 we use the function $f(T)$ from Eq. \[eqn:5\] with $T_0=3.2$ Myr and $p=0.75$ Myr$^{-1}$. The values, errors and the statistical significances of the coefficients obtained by the LSM are shown in Table \[tbl:5\] (case (a)). Using these coefficients we calculate SFR$_{regr}$(H$\alpha$) and $L_{regr}$(H$\alpha$) = SFR$_{regr}$(H$\alpha$)/$k$ for every galaxy from these subsamples, where $k$ = $7.9 \times 10^{-42}$. The comparison of the calculated values with the observed ones is plotted in Figure \[fig:4\]. It follows from the Figure that Eq. \[eqn:9\] provides a good approximation of the observational H$\alpha$ data in the entire range of SFR(H$\alpha$) = 0.8 – 77 $M_{\odot}$ yr$^{-1}$.
[lcccccc]{} Subsample&$N$&$\sigma$&$p$, Myr$^{-1}$&$\bigstrut C_5\times 10^{7}$($F$) &$C_6\times 10^{8}$($F$)&$\left<[{\rm O}]\right>$\
\
1.H$\alpha$, regular &$276$&$3.4$&0.75&2.10$\pm$0.04(2752)&–4.8$\pm$1.0(21)&8.05\
2.H$\alpha$, irregular &$519$&$3.8$&0.75&1.98$\pm$0.02(7156)&–3.4$\pm$0.7(27)&8.13\
\
1.H$\alpha$, regular &$276$&$3.3$&0.66&1.95$\pm$0.04(2913)&–4.6$\pm$1.0(21)&8.05\
2.H$\alpha$, irregular &$519$&$3.6$&0.65&1.82$\pm$0.02(7849)&–3.2$\pm$0.6(26)&8.13\
9.H$\alpha$, all &$795$&$3.5$&0.65&1.85$\pm$0.02(10851)&–3.6$\pm$0.5(46)&8.10\
7.FUV, all &$631$&$4.7$&0.42&1.11$\pm$0.02(2771)&–5.1$\pm$0.8(39)&8.10\
8.NUV, all &$668$&$6.7$&0.32&1.60$\pm$0.03(3925)&–10.1$\pm$1.1(79)&8.10\
\
7.FUV, all &$631$&$4.7$&0.43&1.12$\pm$0.02(2799)&–5.2$\pm$0.8(41)&8.10\
8.NUV, all &$668$&$6.7$&0.33&1.63$\pm$0.03(3981)&–10.3$\pm$1.1(83)&8.10\
\[tbl:5\]
For the FUV and NUV subsamples we use Eq. \[eqn:8\] with $C_4=0$. We can rewrite it in the form of Eq. \[eqn:9\], introducing $$\label{eqn:10}
f(T)=1-\eta T^2,\ C_5=C_2,\ \eta=-\frac{C_3}{C_2}.$$ From the two last rows of Table \[tbl:4\] (case (b)) we derive $\eta=(21.6\pm 0.9)\times 10^{-3}$ Myr$^{-2}$ for the FUV band and $\eta=(18.5\pm 0.9)\times 10^{-3}$ Myr$^{-2}$ for the NUV band. Certainly, the relation Eq. \[eqn:10\] cannot be used if the starburst age $T$ is greater than 7 Myr because $f(T)$ becomes negative. Note that the errors of $\eta$ were estimated taking into account not only the errors of $C_2$ and $C_3$ in Table \[tbl:4\] but their covariation too by using all elements of the correlation matrix including the nondiagonal ones.
Consider the final fine tuning of the parameters in the Eq. \[eqn:9\]. Using LSM we obtain the optimal values of $T_0$ and $p$ in Eq. \[eqn:5\]. For the H$\alpha$ “regular” subsample we derive $\sigma=3.3$, $T_0=3.3$ Myr, $p=0.69$, while for the H$\alpha$ “irregular” subsample the derived values are $\sigma=3.6$, $T_0=3.3$ Myr, $p=0.68$.
Could Eq. \[eqn:9\] be used for the UV continuum radiation with $f(T)$ in the form of Eq. \[eqn:5\], similar to the H$\alpha$ radiation? Such attempt turns out to be successful. For the subsamples No.7 and 8 we obtain the optimal values $T_0=2.9$ Myr and $T_0=3.1$ Myr, respectively. They are smaller than 3.2 Myr, probably, due to the larger data scatter in comparison with the H$\alpha$ subsamples. With these optimal values of $T_0$ and $p$, the RMS standard deviations for the UV bands are slightly decreased as compared to the case when $f(T)$ is used in the form of Eq. \[eqn:10\]. The minimum of the sum of square residuals $\sum ({\rm SFR}_i-{\rm SFR})^2$, corresponding to the best value of $T_0$, is rather shallow. Therefore, we can adopt a single value $T_0=3.2$ Myr for all subsamples.
In Table \[tbl:5\] (case (b)) we show the final values of the parameters for the subsamples No. 1, 2, 7, 8 and 9 in the form according to Eqs. \[eqn:9\] and \[eqn:5\].
One can see that the approximation Eq. \[eqn:5\] is much better than Eq. \[eqn:10\] for the FUV band and is slightly better for the NUV band. Moreover, the use of Eq. \[eqn:5\] for $f(T)$ in Eq. \[eqn:9\] is more preferable not only because of the decrease of the RMS standard deviation $\sigma$, but also by the same dependence on $T$ as that in the case of H$\alpha$ radiation. Starting from $T_0$, the fading half-times of the H$\alpha$, FUV and NUV emission are 1.1 Myr, 1.6 Myr and 2.1 Myr, respectively.
It is noted in Sect. \[s:Ext\] that for the reddening correction of galaxy fluxes we use $E(B-V)_{\rm SDSS}$ derived from the hydrogen Balmer decrement. Alternatively, we also consider reddening corrections, adopting $E(B-V)_{\rm SDSS}$ if $E(B-V)_{\rm SDSS}$ $>$ $E(B-V)_{\rm NED}$ (for $\sim$ 90% of the sample) and $E(B-V)_{\rm NED}$ if $E(B-V)_{\rm SDSS}$ $<$ $E(B-V)_{\rm NED}$ ($\sim$ 10% of the sample). The coefficients in this case are shown in Table \[tbl:5\], (case (c)). The comparison of case (b) and case (c) coefficients shows that differences are very small, indicating that both approaches can equally be used.
![Luminosities in the ${\textrm}{H}\alpha$ line obtained from the regression relation Eq. \[eqn:9\] with $p= 0.75$ Myr$^{-1}$. vs. the measured ones. Dots and open circles correspond to subsamples of galaxies with “regular” and “irregular” shape[]{data-label="fig:4"}](fig4.eps){width="\columnwidth"}
Star formation rates {#s:SFR}
====================
We already noted in Section \[s:Dep\] that there are different indicators of star formation in a wide range of wavelengths from UV to radio and different calibrations to quantify it. As it is mentioned in numerous papers, each SFR indicator possesses its own strengths and disadvantages. Recently, the hybrid SFR indices were proposed, which are based on the combination of the ultraviolet and infrared tracers, the H$\alpha$ and the infrared or radio continuum tracers, the \[O [ii]{}\] $\lambda$3727Å forbidden-line doublet and the infrared or radio continuum tracers. Studies of star formation rates for different samples of galaxies with different level of star formation activity and with different SF tracers were carried out in many papers [see, e.g. @Bo; @C10; @Gil10; @Hop02; @IP06; @IP08; @K98; @K2009; @K02; @L2009; @Li; @M06; @O09; @SW; @Sch].
In the present paper, the galaxy luminosities in the H$\alpha$ emission line and in the UV non-ionising continuum are used to obtain SFRs (Eq. \[eqn:1\]). The H$\alpha$ emission in the star-forming galaxies is produced by the gas ionised by the most massive short-lived hot O-stars with masses $ \gtrsim$ 17 $M_\odot$ and traces the star formation over the period of a few Myr, corresponding to the lifetime of these stars. The non-ionising UV emission is produced by stars in a wider range of masses and therefore can in principle be used as a SFR tracer on a time scale of up to 100 Myr. However, in the case of strongly star-forming LCGs, similar trends in Figs. \[fig:2\] and \[fig:3\] imply that H$\alpha$, FUV and NUV emission in LCGs are produced by the same young stellar populations. This conclusion is supported by the fact that instantaneous burst with the age of 6 Myr emits $\sim$ 4 times and $\sim$ 3 times less radiation in the FUV and NUV ranges, respectively, as compared to that in the burst with the age of 3 Myr [@L99]. Similar difference is seen in Fig. \[fig:3\]. Calculating SFRs from Eq. \[eqn:1\] we actually use the galaxy luminosities observed at a certain current moment. However, in star-bursting galaxies, the observed H$\alpha$ and UV-luminosities depend on the burst age and may vary on a time scale of several Myr. This effect is most pronounced for the H$\alpha$ luminosity: it is constant over first $\sim$ 3 Myr of a starburst and then quickly declines with time. To take into account the temporal luminosity evolution we introduce the initial value of the H$\alpha$ luminosity $L_0({\rm H}\alpha)\equiv L({\rm H}\alpha)(T=0)$ after the onset of star formation and calculate the value of SFR$_0({\rm H}\alpha)=k\times L_0({\rm H}\alpha)$.
Similarly, we also introduce the initial luminosities $L_0$(FUV) and $L_0$(NUV) in the FUV and NUV ranges and the corresponding values SFR$_0$(FUV) and SFR$_0$(NUV) according to Eq. \[eqn:1\]. Hereafter we consider the temporal evolution in UV ranges in the form of Eq. \[eqn:5\]. To distinguish the functions $f(T)$ for the H$\alpha$ line and the FUV and NUV ranges we will use the corresponding subscripts. These functions differ only in the terms of the coefficient $p$ values, which are presented in the Table \[tbl:5\], case (b). As an illustrative example, we will demonstrate now the certain advantage of using SFR$_0({\rm H}\alpha)$ in comparison with SFR(H$\alpha$).
![Current luminosities and SFRs in the FUV range vs. luminosities and SFRs in the H$\alpha$ emission line. Dots and open circles correspond to subsamples of “regular” and “irregular” galaxies, respectively. Solid line is the line of equal SFRs, dashed lines show the best fits for subsamples[]{data-label="fig:5"}](fig5.eps){width="\columnwidth"}
![Initial luminosities and SFR$_0$ in the FUV range at zero starburst age vs. initial luminosities and SFR$_0$ in the H$\alpha$ emission line. Dots and open circles correspond to subsamples of “regular” and “irregular” galaxies, respectively. Upper solid line is the line of equal initial SFRs, dashed lines show the best fits for subsamples, lower solid line SFR$_0$(FUV)$=0.52\times$SFR$_0$(H$\alpha$) shows their mean ratio[]{data-label="fig:6"}](fig6.eps){width="\columnwidth"}
We have an ample sample of the galaxies with known H$\alpha$ and FUV luminosities. First, we derive SFRs from the observed luminosities. We show in Figure \[fig:5\] the relation between SFR(FUV) and SFR(H$\alpha$) (or equivalently $L$(FUV) and $L$(H$\alpha$)). It is seen that both SFR(H$\alpha$) and SFR(FUV) are of the same orders. The values of the LSM slopes in the dependences log $L$(FUV) on log $L$(H$\alpha$) for subsamples No. 3 and 4 are $1.07\pm 0.05$ and $1.04\pm 0.04$ (see dashed lines in Figure \[fig:5\]). However, the values of the slopes in the inverse dependences log $L$(H$\alpha$) on log $L$(FUV) are much smaller than unity, $0.66\pm 0.03$ and $0.72\pm 0.03$ for subsamples No. 3 and 4, respectively. These values would correspond to the lines with the slopes 1/0.66 = 1.51 and 1/0.72 = 1.40 in Figure \[fig:5\]. The differences in the slopes of the direct and inverse dependences are likely due to the Malmquist bias caused by our selection of only galaxies with high $L($H$\beta$) $\geq$ $3\times 10^{40}$ erg s$^{-1}$.
We reduce $L$s and SFRs to a zero age of a starburst according to $$\label{eqn:11}
L_0=L/f(T),\, \ \mathrm{SFR}_0=\mathrm{SFR}/f(T)$$ with $f_{H\alpha}(T)$ for H$\alpha$ emission line and $f_{\mathrm{FUV}}(T)$ for FUV radiation. Adopting the values of $p$ from Table \[tbl:5\], case b), we obtain the distribution of galaxies shown in Figure \[fig:6\]. The LSM gives the slopes $1.04\pm 0.04$ and $0.95\pm 0.03$ for “regular” and “irregular” galaxies respectively, implying that the initial galaxy FUV luminosity $L_0$(FUV) is proportional to the initial H$\alpha$ luminosity $L_0$(H$\alpha$). Slopes of the inverse linear dependences log SFR$_0$(H$\alpha$) on log SFR$_0$(FUV) for the same subsamples are $0.76\pm 0.03$ and $0.85\pm 0.03$. These values correspond to the lines with the slopes 1/0.76 = 1.31 and 1/0.85 = 1.18 in Figure \[fig:6\]. Thus, though Malmquist bias is also present for the data reduced to the zero starburst age, its effect is much smaller because of smaller differences between the slopes of the direct log SFR$_0$(FUV) - log SFR$_0$(H$\alpha$) and inverse log SFR$_0$(H$\alpha$) - log SFR$_0$(FUV) dependences. We also note that the data point scatter in Figure \[fig:6\] is slightly smaller than that in Figure \[fig:5\].
However, there is a downward shift of the data points relative to the line of equal SFRs in Figure \[fig:6\] indicating that SFRs obtained from different indicators are proportional, but not equal.
@I11 derived a single value of reddening $E(B-V)$ for both the gaseous and stellar emission assuming uniform distribution of dust. However, dust in galaxies is distributed non-uniformly. E.g. @C94 and @CF00 suggested that young massive stars responsible for the H$\alpha$ emission are located in more dusty regions as compared to the stars which produce non-ionising UV radiation, including FUV and NUV ranges. In particular, @C94 assumed that non-ionising UV radiation is produced by older stars which were formed in regions different from those where most massive young stars are present.
Could the non-uniform distribution of dust explain the downward shift in Fig. \[fig:6\]? Apparently, not. Assuming that $E(B-V)$ for FUV and NUV ranges is smaller we obtain lower FUV and NUV luminosities. Therefore, the downward shift would be larger. Furthermore, as it was already noted above, similar trends in Figs. \[fig:2\] and \[fig:3\] imply that H$\alpha$, FUV and NUV emission in LCGs are produced by the same young stellar populations, contrary to assumption by @C94.
In order to equalise the values of SFR$_0$(H$\alpha$) and SFR$_0$(FUV) the coefficients $k$ in Eq. \[eqn:1\] should be changed from their nominal values. We find that the mean ratios of SFR$_0$(H$\alpha$) /SFR$_0$(FUV) and SFR$_0$(H$\alpha$) /SFR$_0$(NUV) are equal to 1.9 and 1.5, respectively. More precisely, we find that $10^{\left<\log\left({\rm SFR}_0{\rm (FUV)/SFR}_0({\rm H}\alpha)\right)\right>}$ = 0.53, $10^{\left<\log\left({\rm SFR}_0{\rm (NUV)/SFR}_0({\rm H}\alpha)\right)\right>}$ = 0.69.
Using these average ratios we obtain statistical relations $$\label{eqn:12}
\begin{array}{l}
L_0({\rm FUV})=3.0\times 10^{-14}\times L_0({\rm H}\alpha),\\
L_0({\rm NUV})=3.9\times 10^{-14}\times L_0({\rm H}\alpha),\\
L_0({\rm NUV})=1.3\times L_0({\rm FUV}).
\end{array}$$ We adopt the factor $k^*=1.4\times 10^{-28}\times \gamma$ in Eq. \[eqn:1\] for the NUV range and derive the values of modified factors $k^*=5.3\times 10^{-42}\times \gamma$ for the H$\alpha$ emission line and $k^*=1.8\times
10^{-28}\times \gamma$ for the FUV range. These values correspond to the rough equality of the SFR$_0$ obtained from the initial H$\alpha$, FUV and NUV luminosities. The multiplier $\gamma$ can be used for overall tuning of the set of factors. It is equal to $\sim$ 1, if the modified coefficient $k^*$(NUV) is set to its nominal value $k$(NUV) by @K98. On the other hand, if $k^*$(H$\alpha$) is set to its nominal value $k$(H$\alpha$) by @K98, then $\gamma$ $\sim$ 1.5. Thus, we obtain estimations of SFR$_0$ $$\label{eqn:13}
\begin{array}{l}
{\rm SFR}_0=5.4\times 10^{-42}\times \gamma \times L({\rm H}\alpha) / f_{H\alpha}(T)\\
\phantom{{\rm SFR}_0}=1.8\times 10^{-28}\times \gamma \times L({\rm FUV}) / f_{FUV}(T)\\
\phantom{{\rm SFR}_0}=1.4\times 10^{-28}\times \gamma \times L({\rm NUV}) / f_{NUV}(T),
\end{array}$$ where $L$(H$\alpha)$, $L$(UV) and SFR are measured in erg s$^{-1}$, erg s$^{-1}$ Hz$^{-1}$ and $M_{\odot}$ yr$^{-1}$, respectively. The relations in Eq. \[eqn:13\] give the approximately equal SFR$_0$(H$\alpha$), SFR$_0$(FUV) and SFR$_0$(NUV) for LCGs from our sample. If the alternative correction for extinction is used, we obtain very similar results. Only one coefficient changes and the first row in Eq. \[eqn:13\] gets the form ${\rm SFR}_0=5.3\times 10^{-42}\times \gamma \times L({\rm H}\alpha) / f_{H\alpha}(T)$.
What value of the parameter $\gamma$ is preferable? We can measure the flux from galaxies in some wavelength ranges but we cannot directly measure their SFRs. These values can be estimated using different indicators of star formation. The values of the coefficients $k$ in Eq. \[eqn:1\] for different wavelength ranges are obtained by modeling and therefore dependant from the parameters and assumptions in these models. Thus they can vary in some intervals. We use the certain values of $k$ for the H$\alpha$ line and the UV range. The values SFR$_0$(H$\alpha$) and SFR$_0$(FUV) are in general proportional, but not equal, as they must be. Thus we have to use some agreed set of the coefficients $k$ for different ranges which matches various estimations of SFR$_0$. We obtain the agreement condition for our sample in Eq. \[eqn:12\]. There is interval of the parameter $\gamma$ in which all coefficients $k$ agree with values obtained from modeling. The increase of the quality of modeling will lead to fine-tuning of the value of $\gamma$. Unfortunately if we use the values of the coefficients in Eq. \[eqn:1\] for different wavelength ranges we get some discrepancy in SFRs obtained. This means that the models used to obtain these values have to be improved. In particular it would be useful to search out the ratio of initial values of the luminosities in different wavelength ranges to SFR after the onset of star formation.
@L2009 studied the consistency between the SFRs derived from the FUV continuum and H$\alpha$ emission for a sample of the dwarf star-forming galaxies. Particularly, authors discuss the dependence of the number ratio of ionising to non-ionising photons in the radiation of dwarf galaxies on its metallicity. We investigated the dependence of both $L$(H$\alpha$)/$L$(FUV) and $L_0$(H$\alpha$)/$L_0$(FUV) on \[O\] and do not find any statistically significant trend. All values of Fisher coefficients do not exceed $F=3.5$. However, we note that the rather small range of \[O\] for the galaxies from our sample makes it difficult to study this dependence.
Using Eq. \[eqn:1\] we derive SFRs and obtain the distributions of SFR(H$\alpha$), SFR(FUV), SFR(NUV) as well as SRF$_0$(H$\alpha$), SFR$_0$(FUV) and SFR$_0$(NUV). SFRs derived from the luminosities in the H$\alpha$ emission line, FUV and NUV continuum vary in the wide ranges $0.8 \div 77$ $M_{\odot}$ yr$^{-1}$, $0.18 \div 86$ $M_{\odot}$ yr$^{-1}$ and $0.24 \div 113$ $M_{\odot}$ yr$^{-1}$, respectively. The corresponding median values of SFRs are 6.7 $M_{\odot}$ yr$^{-1}$, 3.8 $M_{\odot}$ yr$^{-1}$ and 5.2 $M_{\odot}$ yr$^{-1}$. The median values of SFR$_0$(H$\alpha$), SFR$_0$(FUV) and SFR$_0$(NUV) are 8.7 $M_{\odot}$ yr$^{-1}$, 5.1 $M_{\odot}$ yr$^{-1}$ and 6.5 $M_{\odot}$ yr$^{-1}$, respectively. For comparison, the median value of SFR(H$\alpha$) is 0.92 $M_{\odot}$ yr$^{-1}$ for a sample of about 7000 star-forming galaxies from the SDSS DR4 (galaxies being less luminous in H$\beta$ than LCGs) [@IP08]. @C09 derived mean SFR $\sim$10 $M_{\odot}$ yr$^{-1}$ for the sample of “green peas”.
Star formation rates SFRs derived from the H$\alpha$, FUV and NUV luminosities are in better mutual agreement if Eq. \[eqn:13\] is used instead of Eq. \[eqn:1\]. We derive the median values of $5.8\times \gamma \times M_{\odot} {\textrm}{ yr}^{-1}$ for SFR$_0$(H$\alpha$), $6.6\times \gamma \times M_{\odot} {\textrm}{ yr}^{-1}$ for SFR$_0$(FUV), and $6.5\times \gamma \times M_{\odot} {\textrm}{ yr}^{-1}$ for SFR$_0$(NUV).
We already mentioned in Section \[s:Dep\] that in general Eq. \[eqn:1\] for SFR can be applied for the continuous star formation during certain time interval $\Delta t$. Formally, for strongly star-bursting galaxies, we may estimate $\Delta t$ as well, assuming the continuous star formation with the constant SFR and adopting nominal values for the coefficient $k$. Then, it is expressed as $\Delta t$ = 1/$C_5$ with $C_5$ from Table \[tbl:5\] and attains the values in the range $\sim$ 5.1 – 8.9 Myr for different samples of case (b).
The histogram for the H$\alpha$ emission line luminosity $L$(H$\alpha$) is shown in Figure \[fig:7\]. The standard way to study the luminosity function is its approximation by the Schechter function [@ref:Schechter] in the form $$\label{eqn:14a}
\psi(L)dL=const(L/L^*)^\alpha exp(-L/L^*)d(L/L^*)$$ where $\psi$ is the number of galaxies per unit volume in the luminosity interval from $L$ to $L+dL$. Assuming that the volume $V$ of the galaxies with luminosity $L$ entering the sample $V\propto L^{3/2}$ and using the maximum likelihood method we obtain the values $\alpha=-1.04$ and $L^*=8.5\times 10^{41}\,\textrm{erg\,s}^{-1}$. This value of $L^*$ corresponds to $\mathrm{SFR}^*=6.7 M_{\odot} {\textrm}{ yr}^{-1}$ according to Eq. \[eqn:1\]. The value $\alpha=-1.04$ is in agreement with the value obtained from 147,986 galaxy redshifts and fluxes from the SDSS [@Bl].
![Distribution of luminosities $L$(H$\alpha$) and SFR(H$\alpha$). Diagonal crosses show the distribution calculated from the Schechter luminosity function with parameters obtained from the maximum likelihood method. Error bars show the Poisson errors[]{data-label="fig:7"}](fig7.eps){width="\columnwidth"}
The distribution of galaxies calculated with the Schechter function is also shown in Figure \[fig:7\]. One can see that there is some difference between real and modelled distributions of H$\alpha$ luminosities $L$(H$\alpha$). In particular, model underpredicts the number of galaxies with largest luminosities calculated from Eq. \[eqn:14a\]. The distribution of the observed luminosities in the range of low values near $10^{41}$ erg s$^{-1}$ is somewhat above the modelled one (see an inset in Figure \[fig:7\]). According to the Pearson’s $\chi^2$ test this difference has the statistical significance over $99\%$.
The distribution of $L_0$(H$\alpha$) differs from the Schechter function in a larger extent compared to the distribution of $L$(H$\alpha$) because of the larger luminosities. The distributions of the FUV and NUV luminosities are similar to that in Figure \[fig:7\] after the luminosity scaling in accordance with the difference of the median values.
Summary {#s:Sum}
=======
We analyse the properties of H$\alpha$ and UV radiation for the sample of about 800 luminous compact galaxies (LCGs) selected by @I11 from the Data Release 7 (DR7) of the Sloan Digital Sky Survey (SDSS). These low-metallicity galaxies are characterised by an active star formation with the star formation rate SFR(H$\alpha$) in the range $\sim$ 1 – 80 $M_\odot$ yr$^{-1}$ and can be considered as local counterparts of the high-redshift ($z$ $>$ 2) star-forming Lyman-break galaxies (LBGs) and Ly-$\alpha$ emission-line galaxies (LAEs). We use the optical SDSS spectroscopic data for LCGs to derive the luminosity in the H$\alpha$ emission line and SFR(H$\alpha$). [*Galaxy Evolution Explorer*]{} ([*GALEX*]{}) UV fluxes are used for obtaining luminosities and SFRs in the far-UV (FUV, $\lambda_{\rm eff}$ = 1528Å) and in the near-UV (NUV, $\lambda_{\rm eff}$ = 2271Å) ranges. These data are supplemented by other global LCG characteristics derived by @I11 from their SDSS spectra: chemical element abundances of the interstellar medium, masses $m$ and ages $T$ of young starbursts. Our main results are as follows:
1\. We study the extinction in a sample of LCGs. It is found that LCGs are rather unobscured galaxies with a mean reddening about of $E(B-V)=0.136$. The mean internal reddening in the sample is 0.106. For subsamples of “regular” galaxies with round shape and “irrregular” ones having shape with some sign of disturbed morphology we obtain internal reddening of 0.081 and 0.120. The difference of internal reddening for subsamples is statistically significant value.
2\. We find that the ratio $L$(H$\alpha$)/$m$ in starbursts with ages $T$ $<$ 3.2 Myr is constant implying that H$\alpha$ luminosity in young starbursts is proportional to the mass of the young stellar population. At later starburst ages $T$ $\geq$ 3.2 Myr, the ratio $L$(H$\alpha$)/$m$ is declined exponentially with $T$. This temporal dependence of the $L$(H$\alpha$)/$m$ ratio is in general agreement with that from the population synthesis models by @SV98 which predict the decrease of $L$(H$\alpha$) after $\sim$ 3 Myr, the lifetime of the most massive stars.
The dependences of the luminosities per unit mass of the young stellar population $L$(H$\alpha$)/$m$, $L$(FUV)/$m$ and $L$(NUV)/$m$ on $T$ (Eqs. \[eqn:5\], \[eqn:9\]) are similar implying that H$\alpha$, FUV and NUV radiation is produced by the same young populations. However, the dependences of $L$(FUV)/$m$ and $L$(NUV)/$m$ on $T$ are weaker as compared to $L$(H$\alpha$)/$m$. Starting from $T_0\sim 3.2$ Myr, the half-times of the H$\alpha$, FUV and NUV luminosities decline are 1.1 Myr, 1.6 Myr and 2.1 Myr, respectively. The ratios $L$(H$\alpha)/L$(FUV) and $L$(H$\alpha)/L$(NUV) also start to decrease after $\sim 3.2$ Myr. Thus, these ratios can be used for estimation of the starburst age $T$.
With this value of $T$ we can estimate $m$ from $L$(H$\alpha$) in young starbursts without invoking modelling of spectral energy distribution (SED). For that we introduce a function $f_{H\alpha}$($T$) which takes into account the variation of $L$(H$\alpha$)/$m$ with $T$. Then $m$ $\sim$ $L$(H$\alpha$)/$f_{H\alpha}$($T$).
3\. The main impact of galaxy metallicity on its luminosity is the indirect one through the variation of the mass of the young stellar population $m$. Dividing the sample of galaxies into subsamples with high and low metallicities, we obtain that the mean luminosity will be greater for the subsample with high metallicity due to the considerable increase of the mean value of $m$. On the other hand, the direct impact of metallicity has the opposite sign. The ratio $L/m$ slightly decreases with increasing of the galaxy metallicity if the starburst age is constant. Thus, the galaxy luminosity decreases with increasing metallicity at fixed values of $m$ and the starburst age $T$. This direct impact is weaker than the indirect one, but it is statistically significant.
4\. Luminosities in H$\alpha$ and UV decrease rapidly after the starburst age of $\sim 3.2$ Myr. We take into account this temporal evolution and introduce time-independent characteristics of the star formation activity, namely the initial luminosities $L_0$ at the starburst age $T=0$. The initial luminosities in the H$\alpha$ emission line, FUV and NUV ranges can be obtained from the current luminosities and the starburst age $T$ from Eq. \[eqn:11\]. We find that $L_0$(FUV) and $L_0$(NUV) are proportional to $L_0$(H$\alpha$) over a large range of luminosities. We can obtain the approximative equality of the values SFR$_0$ derived from the initial H$\alpha$, FUV and NUV luminosities by tuning the factor $k$ in Eq \[eqn:1\]. The set of factors $k$ for H$\alpha$ emission line and FUV and NUV ranges providing such equality for the sample of LCGs is used in Eq. \[eqn:13\].
5\. We find that SFRs derived from the extinction-corrected H$\alpha$, FUV and NUV luminosities vary in the wide ranges of 0.8$\div$77 $M_\odot$ yr$^{-1}$, 0.18$\div$86 $M_\odot$ yr$^{-1}$ and 0.24$\div$113 $M_\odot$ yr$^{-1}$, respectively. The corresponding median values of SFRs are 6.7 $M_\odot$ yr$^{-1}$, 3.8 $M_\odot$ yr$^{-1}$ and 5.2 $M_\odot$ yr$^{-1}$. The median values of initial SFRs are SFR$_0$(H$\alpha$)=8.7 $M_\odot$ yr$^{-1}$, SFR$_0$(FUV)=5.1 $M_\odot$ yr$^{-1}$ and SFR$_0$(NUV)=6.5 $M_\odot$ yr$^{-1}$. In all cases the nominal coefficients $k$(H$\alpha$), $k$(FUV) and $k$(NUV) by @K98 are adopted. The corresponding equalised median SFR$_0$ values in accordance with Eq. \[eqn:13\] are equal to $5.8\times \gamma \times M_{\odot} {\textrm}{ yr}^{-1}$ for SFR$_0$(H$\alpha$), $6.6\times \gamma \times M_{\odot} {\textrm}{ yr}^{-1}$ for SFR$_0$(FUV), and $6.5\times \gamma \times M_{\odot} {\textrm}{ yr}^{-1}$ for SFR$_0$(NUV).
We thank the anonymous referee for valuable comments which helped to improve the presentation of results.
This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
Funding for the Sloan Digital Sky Survey (SDSS) and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the U.S. Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, and the Max Planck Society, and the Higher Education Funding Council for England.
Abazajian, K. N., Adelman-McCarthy, J. K., Agueros, M. A., et al. 2009, , 182, 543
Amorín, R. O., Pérez-Montero, E., & Vílchez, J. M. 2010, , 715, L128
Amorín, R., Pérez-Montero, E., Vílchez, J. M., & Papaderos, P. 2012, , 749, 185
Blanton, M. R., Hogg, D. W, Bahcall, N. A., et al. 2003, , 592, 819
Boquien, M., Calzetti, D., Kennicutt, R., et al. 2009, , 706, 553
Calzetti, D. 2012, preprint arXiv:1208.2997v1
Calzetti, D., Kinney, A. L., & Storchi-Bergmann, T. 1994, , 429, 582
Calzetti, D., Wu, S.-Y., Hong, S., et al. 2010, , 714, 1256
Cardamone, C., Schawinski, K., Sarzi, M., et al. 2009, , 399, 1191
Cardelli, J.A., Clayton, G. C., & Mathis J. S. 1989, , 345, 245
Chakraborti, S., Yadav, N., Cardamone, C., & Ray, A. 2012, , 746, L6
Charlot, S., & Fall, S. M. 2000, , 539, 718
Fisher, R. A. 1954, “Statistical methods for research workers”, Oliver and Boyd: London
Giavalisco, M. 2002, , 4, 579
Gilbank, D.J., Baldry, I.K., Balogh, M.L., et al. 2010, , 405, 2594
Guseva, N. G., Izotov, Y. I., Fricke, K. J., & Henkel, C. 2011, , 534, A84
Heckman, T. M., Hoopes, C. G., Seibert, M., et al. 2005, , 619, L35
Hopkins, A.M., Schulte-Ladbeck, R.E., Drozdovsky, I.O. 2002, , 124, 862
Hopkins, A. M., Miller, C.J., Nichol, R.C., et al. 2003, , 599, 971
Hudson, D. J. 1964, “Statistics Lectures on Elementary Statistics and Probability”, CERN: Geneva
Iglesias-Paramo, A.J., Buat, V., Takeuchi, T.T., et al. 2006. , 164, 38
Izotov, Y. I., Guseva, N. G., & Thuan, T. X. 2011, , 728, 161
Izotova, I.Yu, & Parnovsky, S.L. 2008, Kinematics and Physics of Celestial Bodies, 24, 183
Kauffmann, G., Heckman, T. M., Tremonti, C., et al. 2003, , 346, 1055
Kennicutt, R. C., Jr. 1998, , 36, 189
Kennicutt, R. C., Jr., Hao, C.-N., Calzetti, D., et al. 2009, , 703, 1672
Kewley, L.J., Geller, M.J., Jansen, R.A., Dopita, M. 2002, , 124, 3135
Kroupa, P. 2001, , 322, 231
Lee, J.C., Gil de Paz, A., Tremonti, C., et al. 2009, , 706, 599
Leitherer, C., Schaerer, D., Goldader, J. D., et al. 1999, , 123, 3
Li, Y., Calzetti, D., Kennicutt, R.C., et al. 2010, , 725, 677
Martin, D. C., Fanson, J., Schiminovich, D., et al. 2005, , 619, L1
Morrissey, P., Schiminovich, D., Barlow, T., et al. 2005, , 619, L7
Moustakas, J., Kennicutt, R., Tremonti, C. 2006, , 642, 775
Overzier, R., Heckman, T.N., Tremonti, C., et al. 2009, , 706, 203
Parnovsky, S. L., & Parnowski, A. S. 2008, Astron. Nachr., 329, 864
Pettini, M., Shapley, A. E., Steidel, C. C., et al. 2001, , 554, 981
Pilyugin, L.S., Vilchez, J.M., Mattsson, L., & Thuan, T. X. 2012, , 421, 1624
Salpeter, E.E. 1955, , 121, 161
Sargsyan, L.A., & Weedman, D.W. 2009, , 701, 1398
Schaerer, D., & Vacca, W. D. 1998, , 497, 618
Schechter, P. 1976, , 203, 297
Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, , 500, 525
Schmitt, H.R., Calzetti, D., Armus, L., et al. 2006, , 643, 173
Shim, H., & Chary, R.-R., 2012, preprint arXiv:1205.0949v1
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents a general method for combining uncertain and paradoxical source of evidences for a wide class of fusion problems. From the foundations of the Dezert-Smarandache Theory (DSmT) we show how the DSm rule of combination can be adapted to take into account all possible integrity constraints (if any) of the problem under consideration due to the true nature of elements/concepts involved into it. We show how the Shafer’s model can be considered as a specific DSm hybrid model and be easily handled by our approach and a new efficient rule of combination different from the Dempster’s rule is obtained. Several simple examples are also provided to show the efficiency and the generality of the approach proposed in this work.'
author:
- |
Jean Dezert\
ONERA\
29 Avenue de la Division Leclerc\
92320 Châtillon, France.\
[email protected]\
- |
Florentin Smarandache\
Department of Mathematics\
University of New Mexico\
Gallup, NM 87301, U.S.A.\
[email protected]
title: |
Combining uncertain and paradoxical\
evidences for DSm hybrid models\
---
DSmT, uncertain and paradoxical reasoning, hybrid-model, data fusion.
[**[MSC 2000]{}**]{}: 68T37, 94A15, 94A17, 68T40.
Introduction
============
A new theory of plausible and paradoxical reasoning (DSmT) has been developed by the authors in the last two years in order to resolve problems that did not work in Dempster-Shafer and other fusion theories. According to each model/problem of fusion occurring, we develop here a DSm hybrid rule which combines two or more masses of independent sources of information and takes care of restraints, i.e. of sets which might become empty at time $t_l$ or new sets/elements that might arise in the frame of discernment at time $t_{l+1}$. DSm hybrid rule is applied in a real time when the hyper-power set $D^\Theta$ changes (i.e. the set of all propositions built from elements of frame $\Theta$ with $\cup$ and $\cap$ operators - see[@Dezert_2003f] for details), either increasing or decreasing its focal elements, or when even $\Theta$ decreases or increases influencing the $D^\Theta$ as well, thus the dynamicity of our DSmT.\
The paper introduces the reader to the independence of sources of evidences, which needs to be deeper studied in the future, then defines the models and the DSm hybrid rule, which is different from other rules of combination such as Dempster’s, Yager’s, Smets’, Dubois-Prade’s and gives seven numerical examples of applying the DSm hybrid rule in various models and several examples of dynamicity of DSmT, then the Bayesian DSm hybrid models mixture.
On the independence of the sources of evidences
===============================================
The notion on independence of sources of evidences plays a major role in the development of efficient data fusion algorithms but is very difficult to formally establish when manipulating uncertain and paradoxical information. Some attempts to define the independence of uncertain sources of evidences have been proposed by P. Smets and al. in the Dempster-Shafer Theory (DST) and Transferable Belief Model in [@Yaghlane_Smets_1999; @Yaghlane_Smets_Mellouli_2002a; @Yaghlane_Smets_Mellouli_2002b] and by other authors in possibility theory [@Dawid_1998; @Dawid_1999; @Fonck_1994; @Shenoy_1994; @Studeny_1993]. In the following we consider that $n$ sources of evidences are independent if the internal mechanism by which each source provides its own basic belief assignment doesn’t depend on the mechanisms of other sources (i.e. there is no internal relationship between all mechanisms) or if the sources don’t share (even partially) same knowledge/experience to establish their own basic belief assignment. This definition doesn’t exclude the possibility for independent sources to provide the same (numerical) basic belief assignments. The fusion of dependent uncertain and paradoxical sources is much more complicated because, one has first to identify precisely the piece of redundant information between sources in order to remove it before applying fusion rules. The problem of combination of dependent sources is under investigation.
DSm rule of combination for free-DSm models
===========================================
Definition of the free-DSm model $\mathcal{M}^f(\Theta)$
--------------------------------------------------------
Let consider a finite frame $\Theta=\{\theta_1,\ldots\theta_n\}$ of the fusion problem under consideration. We abandon the Shafer’s model by assuming here that the fuzzy/vague/relative nature of elements $\theta_i$ $i=1,\ldots,n$ of $\Theta$ can be non-exclusive. We assume also that no refinement of $\Theta$ into a new finer [*[exclusive frame of discernment]{}*]{} $\Theta^{\text{ref}}$ is possible. This is the free-DSm model $\mathcal{M}^f(\Theta)$ which can be viewed as the opposite (if we don’t introduce non-existential constraints - see next section) of Shafer’s model, denoted $\mathcal{M}^0(\Theta)$ where all $\theta_i$ are forced to be exclusive and therefore fully discernable.
Example of a free-DSm model
---------------------------
Let consider the frame of the problem $\Theta=\{\theta_1,\theta_2,\theta_3\}$. The free Dedekind’s lattice $D^\Theta=\{\alpha_0,\ldots,\alpha_{18}\}$ over $\Theta$ owns the following 19 elements [@Dezert_2003; @Dezert_2003f]
$$\begin{array}{|lll|}
\hline
& \text{Elements of } \; D^\Theta \; \text{for} \; \mathcal{M}^f(\Theta)& \\
\hline
\alpha_0\triangleq \emptyset & & \\
\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \neq \emptyset & & \\
\alpha_2\triangleq\theta_1\cap\theta_2 \neq \emptyset & \alpha_3\triangleq\theta_1\cap\theta_3\neq \emptyset & \alpha_4\triangleq\theta_2\cap\theta_3\neq \emptyset \\
\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\neq \emptyset & \alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\neq \emptyset & \alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\neq \emptyset \\
\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\neq \emptyset & & \\
\alpha_9\triangleq\theta_1\neq \emptyset & \alpha_{10}\triangleq\theta_2\neq \emptyset & \alpha_{11}\triangleq\theta_3\neq \emptyset \\
\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3\neq \emptyset & \alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\neq \emptyset & \alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\neq \emptyset \\
\alpha_{15}\triangleq\theta_1\cup\theta_2\neq \emptyset & \alpha_{16}\triangleq\theta_1\cup\theta_3\neq \emptyset & \alpha_{17}\triangleq\theta_2\cup\theta_3\neq \emptyset \\
\alpha_{18}\triangleq\theta_1\cup\theta_2\cup\theta_3\neq \emptyset & & \\
\hline
\end{array}$$
The free-DSm model $\mathcal{M}^f(\Theta)$ assumes that all elements $\alpha_i$, $i>0$, are non-empty. This corresponds to the following Venn diagram where in the Smarandache’s codification “$i$” denotes the part of the diagram which belongs to $\theta_i$ only, “$ij$” denotes the part of the diagram which belongs to $\theta_i$ and $\theta_j$ only, “$ijk$” denotes the part of the diagram which belongs to $\theta_i$ and $\theta_j$ and $\theta_k$ only, etc [@Dezert_2003f]. On such Venn diagram representation of the model, we emphasize the fact that all boundaries of intersections must be seen/interpreted as only vague boundaries just because the nature of elements $\theta_i$ can be, in general, only vague, relative and imprecise.
![Venn Diagram for $\mathcal{M}^f(\Theta)$ []{data-label="fig:2"}](Venn3d.eps){width="4cm"}
We now recall the [*[classical]{}*]{} DSm rule of combination based $\mathcal{M}^f(\Theta)$ over the free Dedekind’s lattice built from elements of $\Theta$ with $\cap$ and $\cup$ operators, i.e. the hyper-power set $D^\Theta$.
Classical DSm rule for 2 sources for free-DSm models
----------------------------------------------------
For two [*[independent ]{}*]{} uncertain and paradoxical sources of information (experts/bodies of evidence) providing generalized basic belief assignment $m_1(.)$ and $m_2(.)$ over $D^\Theta$ (or over any subset of $D^\Theta$), the classical DSm conjunctive rule of combination $m_{\mathcal{M}_f(\Theta)}(.)\triangleq [m_1\oplus m_2](.)$ is given by [@Dezert_2003]
$$\forall A\neq\emptyset\in D^\Theta, \quad m_{\mathcal{M}^f(\Theta)}(A)\triangleq [m_1\oplus m_2](A)=\sum_{\overset{X_1,X_2\in D^\Theta}{(X_1\cap X_2)=A}} m_1(X_1)m_2(X_2)
\label{eq:DSMClassic2}$$
$m_{\mathcal{M}^f(\Theta)}(\emptyset)=0$ by definition, unless otherwise specified in special cases when some source assigns a non-zero value to it (like in the Smets TBM approach [@Smets_1994]). This classic DSm rule of combination working on free-DSm models is commutative and associative. This rule, dealing with both uncertain and paradoxist/conflicting information, requires no normalization process and can always been applied.
Classical DSm rule for $k\geq 2$ sources for free-DSm models
------------------------------------------------------------
The above formula can be easily generalized for the free-DSm model $\mathcal{M}^f(\Theta)$ with $k\geq 2$ independent sources in the following way: $$\forall A\neq\emptyset\in D^\Theta, \quad m_{\mathcal{M}^f(\Theta)}(A)\triangleq [m_1\oplus \ldots m_k](A)=
\sum_{\overset{X_1,\ldots,X_k\in D^\Theta}{(X_1\cap \ldots\cap X_k)=A}} \prod_{i=1}^{k}m_i(X_i)
\label{eq:DSMClassick}$$ $m_{\mathcal{M}^f(\Theta)}(\emptyset)=0$ by definition, unless otherwise specified in special cases when some source assigns a non-zero value to it. This classic DSm rule of combination is still commutative and associative.
Presentation of DSm hybrid models
=================================
Definition
----------
Let $\Theta$ be the general frame of the fusion problem under consideration with $n$ elements $\theta_1$, $\theta_2$, $\ldots$, $\theta_n$. A DSm hybrid model $\mathcal{M}(\Theta)$ is defined from the free-DSm model $\mathcal{M}^f(\Theta)$ by introducing some [*[integrity constraints]{}*]{} on some elements $A$ of $D^\Theta$ if one knows with certainty the exact nature of the model corresponding to the problem under consideration. An integrity constraint on $A$ consists in forcing $A$ to be empty (vacuous element), and we will denote such constraint as $A\overset{\mathcal{M}}{\equiv}\emptyset$ which means that $A$ has been forced to $\emptyset$ through the model $\mathcal{M}(\Theta)$. This can be justified by the knowledge of the true nature of each element $\theta_i$ of $\Theta$. Indeed, in some fusion problems, some elements $\theta_i$ and $\theta_j$ of $\Theta$ can be fully discernable because they are truly exclusive while other elements cannot be refined into finer exclusive elements. Moreover, it is also possible that for some reason with some new knowledge on the problem, an element or several elements $\theta_i$ have to be forced to the empty set (specially if dynamical fusion problems are considered, i.e when $\Theta$ varies with space and time). For example, if we consider a list of three potential suspects into a police investigation, it can occur that, during the investigation, one of the suspect must be stray off the initial frame of the problem if he can prove his gullibility with an ascertainable alibi. The initial basic belief masses provided by sources of information one had on the three suspects, must then be modified by taking into account this new knowledge on the model of the problem.\
There exists several possible kinds of integrity constraints which can be introduced in any free-DSm model $\mathcal{M}^f(\Theta)$ actually. The first kind of integrity constraint concerns [*[exclusivity constraints]{}*]{} by taking into account that some conjunctions of elements $\theta_i,\ldots , \theta_k$ are truly impossible (i.e. $\theta_i\cap\ldots\cap\theta_k\overset{\mathcal{M}}{\equiv}\emptyset$). The second kind of integrity constraint concerns the [*[non-existential constraints]{}*]{} by taking into account that some disjunctions of elements $\theta_i,\ldots,\theta_k$ are also truly impossible (i.e. $\theta_i\cup\ldots\cup\theta_k\overset{\mathcal{M}}{\equiv}\emptyset$). We exclude from our presentation the completely degenerate case corresponding to the constraint $\theta_1\cup\ldots\cup\theta_n\overset{\mathcal{M}}{\equiv}\emptyset$ (total ignorance) because there is no way and interest to treat such vacuous problem. In such degenerate case, we can just set $m(\emptyset)\triangleq 1$ which is useless because the problem remains vacuous and $D^\Theta$ reduces to $\emptyset$. The last kind of possible integrity constraint is a mixture of the two previous ones, like for example $(\theta_i\cap\theta_j)\cup \theta_k$ or any other hybrid proposition/element of $D^\Theta$ involving both $\cap$ and $\cup$ operators such that at least one element $\theta_k$ is subset of the constrained proposition. From any $\mathcal{M}^f(\Theta)$, we can thus build several DSm hybrid models depending on the number of integrity constraints one wants to fully characterize the nature of the problem. The introduction of a given integrity constraint $A\overset{\mathcal{M}}{\equiv}\emptyset \in D^\Theta$ implies necessarily the set of inner constraints $B\overset{\mathcal{M}}{\equiv}\emptyset$ for all $B\subset A$. Moreover the introduction of two integrity constraints, say on $A$ and $B$ in $D^\Theta$ implies also necessarily the constraint on the emptiness of the disjunction $A\cup B$ which belongs also to $D^\Theta$ (because $D^\Theta$ is close under $\cap$ and $\cup$ operators). This implies the emptiness of all $C\in D^\Theta$ such that $C\subset (A\cup B)$. Same remark has to be extended for the case of the introduction of $n$ integrity constraints as well. The Shafer’s model is the unique and most constrained DSm hybrid model including all possible exclusivity constraints [*[without non-existential constraint ]{}*]{} since all $\theta_i\neq\emptyset \in \Theta$ are forced to be mutually exclusive. The Shafer’s model is denoted $\mathcal{M}^{0}(\Theta)$ in the sequel. We denote by $\boldsymbol{\emptyset}_ {\mathcal{M}}$ the set of elements of $D^\Theta$ which have been forced to be empty in the DSm hybrid model $\mathcal{M}$.
Example 1 : DSm hybrid model with an exclusivity constraint
-----------------------------------------------------------
Let $\Theta=\{\theta_1,\theta_2,\theta_3\}$ be the general frame of the problem under consideration and let consider the following DSm hybrid model $\mathcal{M}_1(\Theta)$ built by introducing the following exclusivity constraint $\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_1}{\equiv}\emptyset$. This exclusivity constraint implies however no other constraint because $\alpha_1$ doesn’t contain other elements of $D^\Theta$ but itself. Therefore, one has now the following set of elements for $D^\Theta$ $$\begin{array}{|lll|}
\hline
& \text{Elements of } \; D^\Theta \; \text{for} \; \mathcal{M}_1(\Theta)& \\
\hline
\alpha_0\triangleq \emptyset & & \\
\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_1}{\equiv}\emptyset & & \\
\alpha_2\triangleq\theta_1\cap\theta_2 \neq \emptyset & \alpha_3\triangleq\theta_1\cap\theta_3\neq \emptyset & \alpha_4\triangleq\theta_2\cap\theta_3\neq \emptyset \\
\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\neq \emptyset & \alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\neq \emptyset & \alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\neq \emptyset \\
\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\neq \emptyset & & \\
\alpha_9\triangleq\theta_1\neq \emptyset & \alpha_{10}\triangleq\theta_2\neq \emptyset & \alpha_{11}\triangleq\theta_3\neq \emptyset \\
\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3\neq \emptyset & \alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\neq \emptyset & \alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\neq \emptyset \\
\alpha_{15}\triangleq\theta_1\cup\theta_2\neq \emptyset & \alpha_{16}\triangleq\theta_1\cup\theta_3\neq \emptyset & \alpha_{17}\triangleq\theta_2\cup\theta_3\neq \emptyset \\
\alpha_{18}\triangleq\theta_1\cup\theta_2\cup\theta_3\neq \emptyset & & \\
\hline
\end{array}$$
Hence the initial basic belief mass over $D^\Theta$ has to be transferred over the new constrained hyper-power set $D^\Theta(\mathcal{M}_1(\Theta))$ with the 18 elements defined just above. The mechanism for the transfer of basic belief masses from $D^\Theta$ onto $D^\Theta(\mathcal{M}_1(\Theta))$ will be obtained by the DSm hybrid rule of combination presented in the sequel.
Example 2 : DSm hybrid model with another exclusivity constraint
----------------------------------------------------------------
As second example for DSm hybrid model $\mathcal{M}_2(\Theta)$, let consider $\Theta=\{\theta_1,\theta_2,\theta_3\}$ and the following exclusivity constraint $\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_2}{\equiv}\emptyset$. This constraint implies also $\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3\overset{\mathcal{M}_2}{\equiv}\emptyset$ since $\alpha_1\subset\alpha_2$. Therefore, one has now the following set of elements for $D^\Theta (\mathcal{M}_2(\Theta))$ $$\begin{array}{|lll|}
\hline
& \text{Elements of } \; D^\Theta \; \text{for} \; \mathcal{M}_2(\Theta)& \\
\hline
\alpha_0\triangleq \emptyset & & \\
\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_2}{\equiv}\emptyset & & \\
\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_2}{\equiv}\emptyset & \alpha_3\triangleq\theta_1\cap\theta_3\neq \emptyset & \alpha_4\triangleq\theta_2\cap\theta_3\neq \emptyset \\
\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\neq \emptyset & \alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_2}{\equiv}
\alpha_4\neq \emptyset & \alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_2}{\equiv}
\alpha_3\neq \emptyset \\
\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) \overset{\mathcal{M}_2}{\equiv}\alpha_5 \neq \emptyset & & \\
\alpha_9\triangleq\theta_1\neq \emptyset & \alpha_{10}\triangleq\theta_2\neq \emptyset & \alpha_{11}\triangleq\theta_3\neq \emptyset \\
\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3 \overset{\mathcal{M}_2}{\equiv} \alpha_{11}\neq \emptyset & \alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\neq \emptyset & \alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\neq \emptyset \\
\alpha_{15}\triangleq\theta_1\cup\theta_2\neq \emptyset & \alpha_{16}\triangleq\theta_1\cup\theta_3\neq \emptyset & \alpha_{17}\triangleq\theta_2\cup\theta_3\neq \emptyset \\
\alpha_{18}\triangleq\theta_1\cup\theta_2\cup\theta_3\neq \emptyset & & \\
\hline
\end{array}$$ Note that in this case several non-empty elements of $D^\Theta(\mathcal{M}_2(\Theta))$ coincide because of the constraint ($\alpha_6 \overset{\mathcal{M}_2}{\equiv} \alpha_4$, $\alpha_7 \overset{\mathcal{M}_2}{\equiv} \alpha_3$, $\alpha_8 \overset{\mathcal{M}_2}{\equiv} \alpha_5$, $\alpha_{12} \overset{\mathcal{M}_2}{\equiv} \alpha_{11}$). $D^\Theta(\mathcal{M}_2(\Theta))$ has now only 13 different elements. Note that the introduction of both constraints $\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_2}{\equiv}\emptyset$ and $\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_2}{\equiv}\emptyset$ doesn’t change the construction of $D^\Theta(\mathcal{M}_2(\Theta))$ because $\alpha_1\subset\alpha_2$.
Example 3 : DSm hybrid model with another exclusivity constraint
----------------------------------------------------------------
As third example for DSm hybrid model $\mathcal{M}_3(\Theta)$, let consider $\Theta=\{\theta_1,\theta_2,\theta_3\}$ and the following exclusivity constraint $\alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_3}{\equiv}\emptyset$. This constraint implies now $\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3\overset{\mathcal{M}_3}{\equiv}\emptyset$ since $\alpha_1\subset\alpha_6$, but also $\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_3}{\equiv}\emptyset$ because $\alpha_2\subset\alpha_6$ and $\alpha_4\triangleq\theta_2\cap\theta_3\overset{\mathcal{M}_3}{\equiv}\emptyset$ because $\alpha_4\subset\alpha_6$. Therefore, one has now the following set of elements for $D^\Theta (\mathcal{M}_3(\Theta))$ $$\begin{array}{|lll|}
\hline
& \text{Elements of } \; D^\Theta \; \text{for} \; \mathcal{M}_3(\Theta)& \\
\hline
\alpha_0\triangleq \emptyset & & \\
\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_3}{\equiv}\emptyset & & \\
\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_3}{\equiv}\emptyset & \alpha_3\triangleq\theta_1\cap\theta_3\neq \emptyset & \alpha_4\triangleq\theta_2\cap\theta_3\overset{\mathcal{M}_3}{\equiv}\emptyset \\
\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3 \overset{\mathcal{M}_3}{\equiv}\alpha_{3} \neq \emptyset & \alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_3}{\equiv}\emptyset & \alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_3}{\equiv}\alpha_{3} \neq \emptyset \\
\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_3}{\equiv}\alpha_{5}\neq \emptyset & & \\
\alpha_9\triangleq\theta_1\neq \emptyset & \alpha_{10}\triangleq\theta_2\neq \emptyset & \alpha_{11}\triangleq\theta_3\neq \emptyset \\
\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3\overset{\mathcal{M}_3}{\equiv}\alpha_{11}\neq \emptyset & \alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\neq \emptyset & \alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\overset{\mathcal{M}_3}{\equiv}\alpha_{9}\neq \emptyset \\
\alpha_{15}\triangleq\theta_1\cup\theta_2\neq \emptyset & \alpha_{16}\triangleq\theta_1\cup\theta_3\neq \emptyset & \alpha_{17}\triangleq\theta_2\cup\theta_3\neq \emptyset \\
\alpha_{18}\triangleq\theta_1\cup\theta_2\cup\theta_3\neq \emptyset & & \\
\hline
\end{array}$$
$D^\Theta(\mathcal{M}_3(\Theta))$ has now only 10 different elements.
Example 4 : DSm hybrid model with all exclusivity constraints
-------------------------------------------------------------
As fourth example for DSm hybrid model $\mathcal{M}_4(\Theta)$, let consider $\Theta=\{\theta_1,\theta_2,\theta_3\}$ and the following exclusivity constraint $\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_4}{\equiv}\emptyset$. This model corresponds actually to the Shafer’s model $\mathcal{M}^0(\Theta)$ because this constraint includes all possible exclusivity constraints between elements $\theta_i$, $i=1,2,3$ since $\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \subset \alpha_8$, $\alpha_2\triangleq\theta_1\cap\theta_2 \subset \alpha_8$, $\alpha_3\triangleq\theta_1\cap\theta_3 \subset \alpha_8$ and $\alpha_4\triangleq\theta_2\cap\theta_3 \subset \alpha_8$. Therefore, one has now the following set of elements for $D^\Theta (\mathcal{M}_4(\Theta))$ $$\begin{array}{|lll|}
\hline
& \text{Elements of } \; D^\Theta \; \text{for} \; \mathcal{M}_4(\Theta)& \\
\hline
\alpha_0\triangleq \emptyset & & \\
\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_4}{\equiv}\emptyset & & \\
\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_4}{\equiv}\emptyset & \alpha_3\triangleq\theta_1\cap\theta_3 \overset{\mathcal{M}_4}{\equiv}\emptyset & \alpha_4\triangleq\theta_2\cap\theta_3\overset{\mathcal{M}_4}{\equiv}\emptyset \\
\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\overset{\mathcal{M}_4}{\equiv}\emptyset & \alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_4}{\equiv}\emptyset& \alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_4}{\equiv}\emptyset \\
\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_4}{\equiv}\emptyset & & \\
\alpha_9\triangleq\theta_1\neq \emptyset & \alpha_{10}\triangleq\theta_2\neq \emptyset & \alpha_{11}\triangleq\theta_3\neq \emptyset \\
\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3\overset{\mathcal{M}_4}{\equiv}\alpha_{11}\neq \emptyset & \alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\overset{\mathcal{M}_4}{\equiv}\alpha_{10}\neq \emptyset & \alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\overset{\mathcal{M}_4}{\equiv}\alpha_{9}\neq \emptyset \\
\alpha_{15}\triangleq\theta_1\cup\theta_2\neq \emptyset & \alpha_{16}\triangleq\theta_1\cup\theta_3\neq \emptyset & \alpha_{17}\triangleq\theta_2\cup\theta_3\neq \emptyset \\
\alpha_{18}\triangleq\theta_1\cup\theta_2\cup\theta_3\neq \emptyset & & \\
\hline
\end{array}$$
$D^\Theta (\mathcal{M}_4(\Theta))$ has now $2^{\vert \Theta \vert}=8$ different elements and coincides obviously with the classical power set $2^\Theta$. This corresponds to the Shafer’s model and serves as foundation for the Dempster-Shafer Theory.
Example 5 : DSm hybrid model with a non-existential constraint
--------------------------------------------------------------
As fifth example for DSm hybrid model $\mathcal{M}_5(\Theta)$, let consider $\Theta=\{\theta_1,\theta_2,\theta_3\}$ and the following [*[non-existential]{}*]{} constraint $
\alpha_9\triangleq\theta_1\overset{\mathcal{M}_5}{\equiv}\emptyset$. In other words, we remove $\theta_1$ from the initial frame $\Theta=\{\theta_1,\theta_2,\theta_3\}$. This non-existential constraint implies $\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_5}{\equiv}\emptyset$, $\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_5}{\equiv}\emptyset$, $\alpha_3\triangleq\theta_1\cap\theta_3 \overset{\mathcal{M}_5}{\equiv}\emptyset$ and $\alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_5}{\equiv}\emptyset $. Therefore, one has now the following set of elements for $D^\Theta (\mathcal{M}_5(\Theta))$ $$\begin{array}{|lll|}
\hline
& \text{Elements of } \; D^\Theta \; \text{for} \; \mathcal{M}_5(\Theta)& \\
\hline
\alpha_0\triangleq \emptyset & & \\
\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_5}{\equiv}\emptyset & & \\
\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_5}{\equiv}\emptyset & \alpha_3\triangleq\theta_1\cap\theta_3 \overset{\mathcal{M}_5}{\equiv}\emptyset & \alpha_4\triangleq\theta_2\cap\theta_3\neq\emptyset \\
\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\overset{\mathcal{M}_5}{\equiv}\alpha_4\neq\emptyset & \alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_5}{\equiv}\alpha_4\neq\emptyset& \alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_5}{\equiv}\emptyset \\
\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_5}{\equiv}\alpha_4\neq\emptyset & & \\
\alpha_9\triangleq\theta_1 \overset{\mathcal{M}_5}{\equiv}\emptyset & \alpha_{10}\triangleq\theta_2\neq \emptyset & \alpha_{11}\triangleq\theta_3\neq \emptyset \\
\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3\overset{\mathcal{M}_5}{\equiv}\alpha_{11}\neq \emptyset & \alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\overset{\mathcal{M}_5}{\equiv}\alpha_{10}\neq \emptyset & \alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\overset{\mathcal{M}_5}{\equiv}\alpha_{4}\neq \emptyset \\
\alpha_{15}\triangleq\theta_1\cup\theta_2\overset{\mathcal{M}_5}{\equiv}\alpha_{10}\neq \emptyset & \alpha_{16}\triangleq\theta_1\cup\theta_3\overset{\mathcal{M}_5}{\equiv}\alpha_{11}\neq \emptyset & \alpha_{17}\triangleq\theta_2\cup\theta_3\neq \emptyset \\
\alpha_{18}\triangleq\theta_1\cup\theta_2\cup\theta_3\overset{\mathcal{M}_5}{\equiv}\alpha_{17}\neq \emptyset & & \\
\hline
\end{array}$$
$D^\Theta (\mathcal{M}_5(\Theta))$ has now 5 different elements and coincides obviously with the hyper-power set $D^{\Theta\setminus\theta_1}$.
Example 6 : DSm hybrid model with two non-existential constraints
-----------------------------------------------------------------
As sixth example for DSm hybrid model $\mathcal{M}_6(\Theta)$, let consider $\Theta=\{\theta_1,\theta_2,\theta_3\}$ and the following two [*[non-existential]{}*]{} constraints $
\alpha_9\triangleq\theta_1\overset{\mathcal{M}_6}{\equiv}\emptyset$ and $
\alpha_{10}\triangleq\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset$. Actually, these two constraints are equivalent to choose only the following constraint $\alpha_{15}\triangleq\theta_1\cup\theta_2\overset{\mathcal{M}_5}{\equiv}\emptyset$. In other words, we remove now both $\theta_1$ and $\theta_2$ from the initial frame $\Theta=\{\theta_1,\theta_2,\theta_3\}$. These non-existential constraints implies now $\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_6}{\equiv}\emptyset$, $\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_6}{\equiv}\emptyset$, $\alpha_3\triangleq\theta_1\cap\theta_3 \overset{\mathcal{M}_6}{\equiv}\emptyset$, $\alpha_4\triangleq\theta_2\cap\theta_3 \overset{\mathcal{M}_6}{\equiv}\emptyset$, $\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\overset{\mathcal{M}_6}{\equiv}\emptyset$, $\alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset$, $\alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_6}{\equiv}\emptyset $, $ \alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_6}{\equiv}\emptyset$, $\alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset$, $\alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\overset{\mathcal{M}_6}{\equiv}\emptyset$ . Therefore, one has now the following set of elements for $D^\Theta (\mathcal{M}_6(\Theta))$ $$\begin{array}{|lll|}
\hline
& \text{Elements of } \; D^\Theta \; \text{for} \; \mathcal{M}_6(\Theta)& \\
\hline
\alpha_0\triangleq \emptyset & & \\
\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_6}{\equiv}\emptyset & & \\
\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_6}{\equiv}\emptyset & \alpha_3\triangleq\theta_1\cap\theta_3 \overset{\mathcal{M}_6}{\equiv}\emptyset & \alpha_4\triangleq\theta_2\cap\theta_3\overset{\mathcal{M}_6}{\equiv}\emptyset \\
\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\overset{\mathcal{M}_6}{\equiv}\emptyset & \alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset & \alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_6}{\equiv}\emptyset \\
\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_6}{\equiv}\emptyset & & \\
\alpha_9\triangleq\theta_1 \overset{\mathcal{M}_6}{\equiv}\emptyset & \alpha_{10}\triangleq\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset & \alpha_{11}\triangleq\theta_3\neq \emptyset \\
\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3\overset{\mathcal{M}_6}{\equiv}\alpha_{11}\neq \emptyset & \alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset & \alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\overset{\mathcal{M}_6}{\equiv}\emptyset \\
\alpha_{15}\triangleq\theta_1\cup\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset & \alpha_{16}\triangleq\theta_1\cup\theta_3\overset{\mathcal{M}_6}{\equiv}\alpha_{11}\neq \emptyset & \alpha_{17}\triangleq\theta_2\cup\theta_3\overset{\mathcal{M}_6}{\equiv}\alpha_{11}\neq \emptyset\\
\alpha_{18}\triangleq\theta_1\cup\theta_2\cup\theta_3\overset{\mathcal{M}_6}{\equiv}\alpha_{11}\neq \emptyset & & \\
\hline
\end{array}$$
$D^\Theta (\mathcal{M}_6(\Theta))$ reduces now to only two different elements $\emptyset$ and $\theta_3$. $D^\Theta (\mathcal{M}_6(\Theta))$ coincides obviously with the hyper-power set $D^{\Theta\setminus\{\theta_1,\theta_2\}}$. Because there exists only one possible non empty element in $D^\Theta (\mathcal{M}_6(\Theta))$, such kind of problem is called a trivial problem. If one now introduces all non-existential constraints in free-DSm model, then the initial problem reduces to a [*[vacuous problem]{}*]{} also called [*[impossible problem]{}*]{} corresponding to $m(\emptyset)\equiv1$ (no problem at all since the problem doesn’t not exist now !!!). Such kinds of trivial or vacuous [*[problems]{}*]{} are not considered anymore in the sequel since they present no real interest for engineering data fusion problems.
Example 7 : DSm hybrid model with a mixed constraint
----------------------------------------------------
As seventh example for DSm hybrid model $\mathcal{M}_7(\Theta)$, let consider $\Theta=\{\theta_1,\theta_2,\theta_3\}$ and the following [*[mixed exclusivity and non-existential]{}*]{} constraint $\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset$. This mixed constraint implies $\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset$, $\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_7}{\equiv}\emptyset$, $\alpha_3\triangleq\theta_1\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset$, $\alpha_4\triangleq\theta_2\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset$, $\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset$, $\alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_7}{\equiv}\emptyset$, $\alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_7}{\equiv}\emptyset $, $ \alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_7}{\equiv}\emptyset$ and $\alpha_{11}\triangleq\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset$. Therefore, one has now the following set of elements for $D^\Theta (\mathcal{M}_7(\Theta))$ $$\begin{array}{|lll|}
\hline
& \text{Elements of } \; D^\Theta \; \text{for} \; \mathcal{M}_7(\Theta)& \\
\hline
\alpha_0\triangleq \emptyset & & \\
\alpha_1\triangleq\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset & & \\
\alpha_2\triangleq\theta_1\cap\theta_2 \overset{\mathcal{M}_7}{\equiv}\emptyset & \alpha_3\triangleq\theta_1\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset & \alpha_4\triangleq\theta_2\cap\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset \\
\alpha_5\triangleq(\theta_1\cup\theta_2)\cap\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset & \alpha_6\triangleq(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_7}{\equiv}\emptyset & \alpha_7\triangleq(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_7}{\equiv}\emptyset \\
\alpha_8\triangleq\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_7}{\equiv}\emptyset & & \\
\alpha_9\triangleq\theta_1 \neq \emptyset & \alpha_{10}\triangleq\theta_2\neq\emptyset & \alpha_{11}\triangleq\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset \\
\alpha_{12}\triangleq(\theta_1\cap\theta_2)\cup\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset & \alpha_{13}\triangleq(\theta_1\cap\theta_3)\cup\theta_2\overset{\mathcal{M}_7}{\equiv} \alpha_{10}\neq \emptyset & \alpha_{14}\triangleq(\theta_2\cap\theta_3)\cup\theta_1\overset{\mathcal{M}_7}{\equiv}\alpha_{9}\neq\emptyset \\
\alpha_{15}\triangleq\theta_1\cup\theta_2\neq\emptyset & \alpha_{16}\triangleq\theta_1\cup\theta_3\overset{\mathcal{M}_7}{\equiv}\alpha_{9}\neq \emptyset & \alpha_{17}\triangleq\theta_2\cup\theta_3\overset{\mathcal{M}_7}{\equiv}\alpha_{10}\neq \emptyset\\
\alpha_{18}\triangleq\theta_1\cup\theta_2\cup\theta_3\overset{\mathcal{M}_7}{\equiv}\alpha_{15}\neq \emptyset & & \\
\hline
\end{array}$$
$D^\Theta (\mathcal{M}_7(\Theta))$ reduces now to only four different elements $\emptyset$, $\theta_1$, $\theta_2$, and $\theta_1\cup\theta_2$.
DSm rule of combination for DSm hybrid models
=============================================
We present in this section a general DSm-hybrid rule of combination able to deal with any DSm hybrid models. We will show how this new general rule of combination works with all DSm hybrid models presented in the previous section and we list interesting properties of this new useful and powerful rule of combination.
Notations
---------
Let $\Theta=\{\theta_1,\ldots\theta_n\}$ be a frame of [*[partial discernment]{}*]{} of the constrained fusion problem, and $D^\Theta$ the free distributive lattice (hyper-power set) generated by $\Theta$ and the empty set $\emptyset$ under $\cap$ and $\cup$ operators. We need to distinguish between the empty set $\emptyset$, which belongs to $D^\Theta$, and by $\emptyset$ we understand a set which is empty all the time (we call it [*[absolute emptiness]{}*]{} or [*[absolutely empty]{}*]{}) independent of time, space and model, and all other sets from $D^\Theta$. For example $\theta_1\cap\theta_2$ or $\theta_1\cup\theta_2$ or only $\theta_i$ itself, $1\leq i \leq n$, etc, which could be or become empty at a certain time (if we consider a fusion dynamicity) or in a particular model $\mathcal{M}$ (but could not be empty in other model and/or time) (we call a such element [*[relative emptiness]{}*]{} or [*[relatively empty]{}*]{}). WeÕll denote by $\boldsymbol{\emptyset}_{\mathcal{M}}$ the set of relatively empty such elements of $D^\Theta$ (i.e. which become empty in a particular model $\mathcal{M}$ or at a specific time). $\boldsymbol{\emptyset}_{\mathcal{M}}$ is the set of integrity constraints which depends on the DSm model $\mathcal{M}$ under consideration, and the model $\mathcal{M}$ depends on the structure of its corresponding [*[fuzzy Venn Diagram]{}*]{} (number of elements in $\Theta$, number of non-empty intersections, and time in case of dynamic fusion). Through our convention $\emptyset\notin\boldsymbol{\emptyset}_{\mathcal{M}}$. LetÕs note by $\boldsymbol{\emptyset}\triangleq \{\emptyset,\boldsymbol{\emptyset}_{\mathcal{M}}\}$ the set of all relatively and absolutely empty elements.\
For any $A\in D^\Theta$, let $\phi(A)$ be the [*[characteristic emptiness function]{}*]{} of the set $A$, i.e. $\phi(A)= 1$ if $A\notin \boldsymbol{\emptyset}$ and $\phi(A)= 0$ otherwise. This function helps in assigning the value zero to all relatively or absolutely empty elements of $D^\Theta$ through the choice of DSm hybrid model $\mathcal{M}$. Let’s define the total ignorance on $\Theta = \{\theta_1,\theta_2, \ldots, \theta_n\}$ as $I_t\triangleq\theta_1\cup \theta_2\cup \ldots\cup \theta_n$ and the set of relative ignorances as $I_r\triangleq\{\theta_{i_1}\cup\ldots\cup\theta_{i_k}, \;\text{where}\; i_1, ..., i_k \in \{1,2,...,n\} \; \text{and}\; 2\leq k\leq n-1\}$, then the set of all kind of ignorances as $I = I_t \cup I_r$. For any element $A$ in $D^\Theta$, one considers $u(A)$ as the union of all singletons $\theta_i$ that compose $A$. For example, if $A$ is a singleton then $u(A)=A$; if $A=\theta_1\cap \theta_2$ or $A=\theta_1\cup \theta_2$ then $u(A)=\theta_1\cup \theta_2$; if $A=(\theta_1\cap \theta_2)\cup \theta_3$ then $u(A)=\theta_1\cup \theta_2\cup\theta_3$. ; by convention $u(\emptyset)\triangleq\emptyset$. The second summation of the DSm hybrid rule (see eq. and and denoted $S_2$ in the sequel) transfers the mass of $\emptyset$ \[if any; sometimes, in rare cases, $m(\emptyset) > 0$ (for example in Ph. Smets’ work); we want to catch this particular case as well\] to the total ignorance $I_t=\theta_1\cup\theta_2\cup\ldots\cup\theta_n$. The other part of the mass of relatively empty elements, $\theta_i$ and $\theta_j$ together for example, $i\neq j$, goes to the partial ignorance/uncertainty $m(\theta_i\cup\theta_j)$. $S_2$ multiplies, naturally following the DSm classic network architecture, only the elements of columns of absolutely and relatively empty sets, and then $S_2$ transfers the mass $m_1(X_1)m_2(X_2)\ldots m_k(X_k)$ either to the element $A\in D^\theta$ in the case when $A= u(X_1)\cup u(X_2)\cup \ldots \cup u(X_k)$ is not empty, or if $u(X1)\cup u(X2)\cup \ldots\cup u(X_k)$ is empty then the mass $m_1(X_1)m_2(X_2)É m_k(X_k)$ is transferred to the total ignorance. We include all possible degenerate problems/models in this new DSmT hybrid framework, but the vacuous DSm-hybrid model $\mathcal{M}_{\emptyset}$ defined by the constraint $I_t= \theta_1\cup\theta_2\cup\ldots\cup\theta_n\overset{\mathcal{M}_{\emptyset}}{\equiv}\emptyset $ which is meaningless and useless.\
We provide here the issue for programming the calculation of $u(X)$ from the binary representation of any proposition $X\in D^\Theta$ expressed in the Dezert-Smarandache’s order [@Dezert_2003f; @Dezert_Smarandache_2003]. Let’s consider the Smarandache’s codification of elements $\theta_1,\ldots, \theta_n$. One defines the anti-absorbing relationship as follows: element $i$ anti-absorbs element $ij$ (with $i<j$), and let’s use the notation $i << ij$, and also $j << ij$; similarly $ij << ijk$ (with $i<j<k$), also $jk << ijk$ and $ik << ijk$. This relationship is transitive, therefore $i << ij$ and $ij << ijk$ involve $i << ijk$; one can also write $i << ij << ijk$ as a chain; similarly one gets $j << ijk$ and $k << ijk$. The anti-absorbing relationship can be generalized for parts with any number of digits, i.e. when one uses the Smarandache codification for the corresponding Venn diagram on $\Theta = \{\theta_1,\theta_2,\ldots, \theta_n\}$, with $n\geq 1$. Between elements $ij$ and $ik$, or between $ij$ and $jk$ there is no anti-absorbing relationship, therefore the anti-absorbing relationship makes a partial order on the parts of the Venn diagram for the free DSm model. If a proposition $X$ is formed by a part only, say $i_1 i_2\ldots i_r$, in the Smarandache codification, then $u(X)=\theta_{i_1}\cup \theta_{i_2}\cup\ldots\cup\theta_{i_r}$. If $X$ is formed by two or more parts, the first step is to eliminate all anti-absorbed parts, ie. if $A << B$ then $u({A,B})=u(A)$; generally speaking, a part $B$ is anti-absorbed by part $A$ if all digits of $A$ belong to $B$; for an anti-absorbing chain $A_1 << A_2 << ... << A_s$ one takes $A_1$ only and the others are eliminated; afterwards, when $X$ is anti-absorbingly irreducible, $u(X)$ will be the unions of all singletons whose indices occur in the remaining parts of $X$ - if one digit occurs many times it is taken only once.\
See some examples for the case $n=3$: $12 << 123$, i.e. $12$ anti-absorbs $123$. Between $12$ and $23$ there is no anti-absorbing relationship.
- If $X=123$ then $u(X)=\theta_1\cup\theta_2\cup\theta_3$.
- If $X=\{23,123\}$, then $23 << 123$, thus $u(\{23,123\})=u(23)$, because $123$ has been eliminated, hence $u(X)=u(23)=\theta_2\cup\theta_3$.
- If $X=\{13,123\}$, then $13 << 123$, thus $u(\{13,123\})=u(13)=\theta_1\cup\theta_3$.
- If $X=\{13,23,123\}$, then $13 << 123$, thus $u(\{13,23,123\})=u(\{13,23\})=\theta_1\cup\theta_2\cup\theta_3$ (one takes as theta indices each digit in the $\{13,23\}$) - if one digit is repeated it is taken only once; between $13$ and $23$ there is no relation of anti-absorbing.
- If $X=\{3,13,23,123\}$, then $u(X)=u(\{3,13,23\})$ because $23 << 123$, then $u(\{3,13,23\})=u(\{3,13\})$ because $3 << 23$, then $u(\{3,13\})=u(3)=\theta_3$ because $3 << 13$.
- If $X=\{1,12,13,23,123\}$, then one has the anti-absorbing chain: $1 << 12 << 123$, thus $u(X)=u(\{1,13,23\})=u(\{1,23\})$ because $1 << 13$, and finally $u(X)=\theta_1\cup\theta_2\cup\theta_3$.
- If $X= \{1,2,12,13,23,123\}$, then $1<<12<<123$ and $2 << 23$ thus $u(X)=u(\{1,2,13\})=u(\{1,2\})$ because $1 << 13$, and finally $u(X)=\theta_1\cup \theta_2$.
- If $X=\{2,12,3,13,23,123\}$, then $2 << 23 << 123$ and $3 << 13$ thus $u(X)=u(\{2,12,3\})$, but $2 << 12$ hence $u(X)=u(\{2,3\})=\theta_2\cup\theta_3$.
The DSm hybrid rule of combination for 2 sources
------------------------------------------------
To eliminate the [*[degenerate vacuous fusion problem]{}*]{} from the presentation, we assume from now on that the given DSm hybrid model $\mathcal{M}$ under consideration is always different from the vacuous model $\mathcal{M}_{\emptyset}$ (i.e. $I_t\neq\emptyset$). The DSm hybrid rule of combination, associated to a given DSm hybrid model $\mathcal{M}\neq\mathcal{M}_{\emptyset}$ , for two sources is defined for all $A\in D^\Theta$ as:
$$\begin{gathered}
m_{\mathcal{M}(\Theta)}(A)\triangleq
\phi(A)\Bigl[ \sum_{\overset{X_1,X_2\in D^\Theta}{(X_1\cap X_2)=A}} m_1(X_1)m_2(X_2)\\
+ \sum_{\overset{X_1,X_2\in\boldsymbol{\emptyset}}{[(u(X_1)\cup u(X_2))=A] \vee [(u(X_1)\cup u(X_2)\in\boldsymbol{\emptyset}) \wedge (A=I_t)]}}m_1(X_1)m_2(X_2)\\
+
\sum_{\overset{X_1,X_2\in D^\Theta}{\overset{(X_1\cup X_2)=A}{\overset{X_1\cap X_2\in\boldsymbol{\emptyset}}{}}}} m_1(X_1)m_2(X_2)
\Bigr]
\label{eq:DSmH2}\end{gathered}$$
The first sum entering in the previous formula corresponds to mass $m_{\mathcal{M}^f(\Theta)}(A)$ obtained by the classic DSm rule of combination based on the free-DSm model $\mathcal{M}^f$ (i.e. on the free lattice $D^\Theta$), i.e. $$m_{\mathcal{M}^f(\Theta)}(A)\triangleq \sum_{\overset{X_1,X_2\in D^\Theta}{X_1\cap X_2=A}} m_1(X_1)m_2(X_2)$$
The second sum entering in the formula of the DSm-hybrid rule of combination represents the mass of all relatively and absolutely empty sets which is transferred to the total or relative ignorances.\
The third sum entering in the formula of the DSm-hybrid rule of combination transfers the sum of relatively empty sets to the non-empty sets in the same way as it was calculated following the DSm classic rule.\
The DSm hybrid rule of combination for $k\geq 2$ sources
--------------------------------------------------------
The previous formula of DSm hybrid rule of combination can be generalized in the following way for all $A\in D^\Theta$ : $$\begin{gathered}
m_{\mathcal{M}(\Theta)}(A)\triangleq
\phi(A)\Bigl[ \sum_{\overset{X_1,X_2,\ldots,X_k\in D^\Theta}{(X_1\cap X_2\cap\ldots\cap X_k)=A}} \prod_{i=1}^{k} m_i(X_i)\\
+ \sum_{\overset{X_1,X_2,\ldots,X_k\in\boldsymbol{\emptyset}}{[(u(X_1)\cup u(X_2)\cup \ldots \cup u(X_k))=A]\vee [(u(X_1)\cup u(X_2)\cup \ldots \cup u(X_k)\in\boldsymbol{\emptyset}) \wedge (A=I_t)]}} \prod_{i=1}^{k} m_i(X_i) \\
+ \sum_{\overset{X_1,X_2,\ldots,X_k\in D^\Theta}{\overset{(X_1\cup X_2\cup\ldots\cup X_k)=A}{\overset{X_1\cap X_2\cap \ldots\cap X_k\in\boldsymbol{\emptyset}}{}}}} \prod_{i=1}^{k} m_i(X_i)
\Bigr]
\label{eq:DSmHk}\end{gathered}$$
The first sum entering in the previous formula corresponds to mass $m_{\mathcal{M}^f(\Theta)}(A)$ obtained by the classic DSm rule of combination for $k$ sources of information based on the free-DSm model $\mathcal{M}^f$ (i.e. on the free lattice $D^\Theta$), i.e. $$m_{\mathcal{M}^f(\Theta)}(A)\triangleq \sum_{\overset{X_1,X_2,\ldots,X_k\in D^\Theta}{(X_1\cap X_2\cap\ldots\cap X_k)=A}} \prod_{i=1}^{k} m_i(X_i)
\label{mfk}$$
Remark on the DSm hybrid rule of combination
--------------------------------------------
From and , the previous general formula can be rewritten as
$$m_{\mathcal{M}(\Theta)}(A)\triangleq
\phi(A)\Bigl[ S_1(A) + S_2(A) + S_3(A)\Bigr]
\label{eq:DSmHkBis}$$
where $$S_1(A)\equiv m_{\mathcal{M}^f(\Theta)}(A) \triangleq \sum_{\overset{X_1,X_2,\ldots,X_k\in D^\Theta}{(X_1\cap X_2\cap\ldots\cap X_k)=A}} \prod_{i=1}^{k} m_i(X_i)$$ $$S_2(A)\triangleq \sum_{\overset{X_1,X_2,\ldots,X_k\in\boldsymbol{\emptyset}}{[(u(X_1)\cup u(X_2)\cup \ldots \cup u(X_k))=A]\vee [(u(X_1)\cup u(X_2)\cup \ldots \cup u(X_k)\in\boldsymbol{\emptyset}) \wedge (A=I_t)]}} \prod_{i=1}^{k} m_i(X_i)$$ $$S_3(A)\triangleq\sum_{\overset{X_1,X_2,\ldots,X_k\in D^\Theta}{\overset{(X_1\cup X_2\cup\ldots\cup X_k)=A}{\overset{X_1\cap X_2\cap \ldots\cap X_k\in\boldsymbol{\emptyset}}{}}}} \prod_{i=1}^{k} m_i(X_i)$$
and thus, this combination can be viewed actually as a two steps procedure as follows:
- Step 1: Evaluate the combination of the sources over the free lattice $D^\Theta$ by the classical DSm rule of combination to get for all $A\in D^\Theta$, $S_1(A)=m_{\mathcal{M}^f(\Theta)}(A)$ using . This step preserves the commutativity and associativity properties of the combination.\
- Step 2 : Transfer the masses of the integrity constraints of the DSm hybrid model $\mathcal{M}$ according to formula . Note that this step is necessary only if one has reliable information about the real constraints involved in the fusion problem under consideration.
The second step does not preserve the associativity property but this is not a fundamental requirement in most of fusion systems actually. If one really wants to preserve optimality of the fusion rule, one has first to combine all sources using classical DSm rule (with any clustering of sources) and the ultimate step will consist to adapt basic belief masses according to the integrity constraints of the model $\mathcal{M}$.\
If one first adapts the local basic belief masses $m_1(.),...m_k(.)$ to the hybrid-model $\mathcal{M}$ and afterwards one applies the combination rules, the fusion result becomes only suboptimal because some information is lost during the transfer of masses of integrity constraints. The same remark holds if the transfer of masses of integrity constraints is done at some intermediate steps after the fusion of $m$ sources with $m<k$.\
Let’s note also that this formula of transfer is more general (because we include the possibilities to introduce both exclusivity constraints and non-existential constraints as well) and more precise (because we explicitly consider all different relative emptiness of elements into the general transfer formula ) than the generic transfer formulas used in the DST framework proposed as alternative rules to the Dempster’s rule of combination [@Lefevre_2002] and discussed in section 5.9.
Property of the DSm Hybrid Rule
-------------------------------
$$\sum_{A\in D^\Theta} m_{\mathcal{M}(\Theta)}(A) = \sum_{A\in D^\Theta} \phi(A)\Bigl[ S_1(A) + S_2(A) + S_3(A)\Bigr]=1$$
[**[Proof]{}**]{}: Let’s first prove that $\sum_{A\in D^\Theta} m(A)=1$ where all masses $m(A)$ are obtained by the DSm classic rule. Let’s consider each mass $m_i(.)$ provided by the $i$th source of information, for $1\leq i\leq k$, as a vector of $d={\mid D^\Theta\mid}$ dimension, whose sum of components is equal to one, i.e. $m_i(D^\Theta)=[m_{i1},m_{i2},\ldots,m_{id}]$, and $\sum_{j=1,d} m_{ij}=1$. Thus, for $k\geq 2$ sources of information, the mass matrix becomes $$\mathbf{M}=
\begin{bmatrix}
m_{11} & m_{12} & \ldots & m_{1d}\\
m_{21} & m_{22} & \ldots & m_{2d}\\
\ldots & \ldots & \ldots & \ldots\\
m_{k1} & m_{k2} & \ldots & m_{kd}\\
\end{bmatrix}$$
If one notes the sets in $D^\Theta$ by $A_1$, $A_2$, ..., $A_d$ (doesn’t matter in what order one lists them) then the column $(j)$ in the matrix represents the masses assigned to $A_j$ by each source of information $s_1$, $s_2$, $\ldots$, $s_k$; for example $s_i(A_j)=m_{ij}$, where $1\leq i\leq k$. According to the DSm network architecture [@Dezert_2003f], all the products in this network will have the form $m_{1j_1}m_{2j_2}\ldots m_{kj_k}$, i.e. one element only from each matrix row, and no restriction about the number of elements from each matrix column, $1\leq j_1,j_2,\ldots,j_k\leq d$. Each such product will enter in the fusion mass of one set only from $D^\Theta$. Hence the sum of all components of the fusion mass is equal to the sum of all these products, which is equal to $$\prod_{i=1}^{k} \sum_{j=1}^{d} m_{ij}=\prod_{i=1}^{k} 1 = 1
\label{eq:Formula12}$$ The DSm hybrid rule has three sums $S_1$, $S_2$, and $S_3$. Let’s separate the mass matrix $\mathbf{M}$ into two disjoint sub-matrices $\mathbf{M}_{\emptyset}$ formed by the columns of all absolutely and relatively empty sets, and $\mathbf{M}_N$ formed by the columns of all non-empty sets. According to the DSm network architecture (for $k\geq 2$ rows):
- $S_1$ is the sum of all products resulted from the multiplications of the columns of $\mathbf{M}_N$ following the DSm network architecture such that the intersection of their corresponding sets is non-empty, i.e. the sum of masses of all non-empty sets before any mass of absolutely or relatively empty sets could be transferred to them;
- $S_2$ is the sum of all products resulted from the multiplications of $\mathbf{M}_{\emptyset}$ following the DSm network architecture, i.e. a partial sum of masses of absolutely and relatively empty sets transferred to the ignorances in $I\triangleq I_t\cup I_r$ or to singletons of $\Theta$.
- $S_3$ is the sum of all the products resulted from the multiplications of the columns of $\mathbf{M}_N$ and $\mathbf{M}_{\emptyset}$ together, following the DSm network architecture, but such that at least a column is from each of them, and also the sum of all products of columns of $\mathbf{M}_N$ such that the intersection of their corresponding sets is empty (what did not enter into the previous sum $S_1$), i.e. the remaining sum of masses of absolutely or relatively empty sets transferred to the non-empty sets of the DSm hybrid model $\mathcal{M}$.
If one now considers all the terms (each such term is a product of the form $m_{1j_1}m_{2j_2}\ldots m_{kj_k}$) of these three sums, we get exactly the same terms as in the DSm network architecture for the DSm classic rule, thus the sum of all terms occurring in $S_1$, $S_2$, and $S_3$ is 1 (see formula ) which completes the proof. DSm hybrid rule naturally derives from the DSm classic rule.
Entire masses of relatively and absolutely empty sets in a given DSm hybrid model $\mathcal{M}$ are transferred to non-empty sets according to the above and below formula and thus $$\forall A\in \boldsymbol{\emptyset} \subset D^\Theta, \quad m_{\mathcal{M}(\Theta)}(A)=0$$ The entire mass of a relatively empty set (from $D^\Theta$) which has in its expression $\theta_{j_1}$, $\theta_{j_2}$, $\ldots$, $\theta_{j_r}$, with $1\leq r\leq n$ will generally be distributed among the $\theta_{j_1}$, $\theta_{j_2}$, $\ldots$, $\theta_{j_r}$ or their unions or intersections, and the distribution follows the way of multiplication from the DSm classic rule, explained by the DSm network architecture [@Dezert_2003f]. Thus, because nothing is lost, nothing is gained, the sum of all $m_{\mathcal{M}(\Theta)}(A)$ is equal to 1 as just proved previously, and fortunately no normalization constant is needed which could bring a lost of information in the fusion rule.\
The three summations $S_1(.)$, $S_3(.)$ and $S_3(.)$ are disjoint because:
- $S_1(.)$ multiplies the columns corresponding to non-emptysets only - but such that the intersections of the sets corresponding to these columns are non-empty \[following the definition of DSm classic rule\];
- $S_2(.)$ multiplies the columns corresponding to absolutely and relatively emptysets only;
- $S_3(.)$ multiplies:
- either the columns corresponding to absolutely or relatively emptysets with the columns corresponding to non-emptysets such that at least a column corresponds to an absolutely or relatively emptyset and at least a column corresponds to a non-emptyset,
- or the columns corresponding to non-emptysets - but such that the intersections of the sets corresponding to these columns are empty.
The multiplications are following the DSm network architecture, i.e. any product has the above general form: $m_{1j_1}m_{2j_2}\ldots m_{kj_k}$, i.e. any product contains as factor one element only from each row of the mass matrix $\mathbf{M}$ and the total number of factors in a product is equal to $k$. The function $\phi(A)$ automatically assigns the value zero to the mass of any empty set, and allows the calculation of masses of all non-emptysets.\
On the programming of the DSm hybrid rule
-----------------------------------------
We briefly give here an issue for a fast programming of DSm rule of combination. Let’s consider $\Theta = \{\theta_1, \theta_2, \ldots, \theta_n\}$, the sources $\mathcal{B}_1$, $\mathcal{B}_2$,$\ldots$, $\mathcal{B}_k$, and $p= \min\{n, k\}$. One needs to check only the focal sets, i.e. sets (i.e. propositions) whose masses assigned to them by these sources are not all zero. Thus, if $\mathbf{M}$ is the mass matrix, and we consider a set $A_j$ in $D^\Theta$, then the column $(j)$ corresponding to $A_j$, i.e. $(m_{1j} \; m_{2j}\; \ldots \; m_{kj})$ transposed has not to be identical to the null-vector of $k$-dimension $(0\; 0 \; \ldots \; 0)$ transposed. Let $D^\Theta(\text{step}_1)$ be formed by all focal sets at the beginning (after sources $\mathcal{B}_1$, $\mathcal{B}_2$,$\ldots$, $\mathcal{B}_k$ have assigned massed to the sets in $D^\Theta$). Applying the DSm classic rule, besides the sets in $D^\Theta(\text{step}_1)$ one adds $r$-intersections of sets in $D^\Theta(\text{step}_1)$, thus:
$$D^\Theta(\text{step}_2) = D^\Theta(\text{step}_1) \vee \{A_{i_1}\wedge A_{i_2}\wedge\ldots\wedge A_{i_r}\}$$
where $A_{i_1}$, $A_{i_2}$, …, $A_{i_r}$ belong to $D^\Theta(\text{step}_1)$ and $2\leq r\leq p$.\
Applying the DSm hybrid rule, due to its $S_2$ and $S_3$ summations, besides the sets in $D^\Theta(\text{step}_2)$ one adds $r$-unions of sets and the total ignorance in $D^\Theta(\text{step}_2)$, thus:
$$D^\Theta(\text{step}_3) = D^\Theta(\text{step}_2) \vee I_t \vee \{A_{i_1}\vee A_{i_2}\vee\ldots\vee A_{i_r}\}$$
where $A_{i_1}$, $A_{i_2}$, …, $A_{i_r}$ belong to $D^\Theta(\text{step}_2)$ and $2\leq r\leq p$.\
This means that instead of computing the masses of all sets in $D^\Theta$, one needs to first compute the masses of all focal sets (step 1), second the masses of their $r$-intersections (step 2), and third the masses of $r$-unions of all previous sets and the mass of total ignorance (step 3).
Application of the DSm Hybrid rule on previous examples
-------------------------------------------------------
We present in this section some numerical results of the DSm hybrid rule of combination for 2 independent sources of information. We examine the seven previous examples in order to help the reader to check by himself or herself the validity of our new general formula. Due to space limitation, we will not go in details on all the derivations steps, we will just present main intermediary results (i.e. the value of the three summations) involved into the general formula . The results have been first obtained by hands and then be validated by MatLab programming. We denote $$S_1(A)\equiv m_{\mathcal{M}^f(\Theta)}(A)\triangleq \sum_{\overset{X_1,X_2\in D^\Theta}{X_1\cap X_2=A}} m_1(X_1)m_2(X_2)$$ $$S_2(A)\triangleq \sum_{\overset{X_1,X_2\in\boldsymbol{\emptyset}}{[u(X_1)\cup u(X_2)=A] \vee [(u(X_1)\cup u(X_2)\in\boldsymbol{\emptyset}) \wedge (A=I_t)]}}m_1(X_1)m_2(X_2)$$ $$S_3(A)\triangleq \sum_{\overset{X_1,X_2\in D^\Theta}{\overset{X_1\cup X_2=A}{\overset{X_1\cap X_2\in\boldsymbol{\emptyset}}{}}}} m_1(X_1)m_2(X_2)$$ Now let consider $\Theta=\{\theta_1,\theta_2,\theta_3\}$ and the two following independent bodies of evidence $\mathcal{B}_1$ and $\mathcal{B}_2$ with the generalized basic belief assignments[^1] $m_1(.)$ and $m_2(.)$ given in the following table[^2]. The right column of the table indicates the result of the fusion obtained by the classical DSm rule of combination. $$\begin{array}{llll}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta& m_1(A) & m_2(A) & m_{\mathcal{M}^f(\Theta)}(A) \\
\hline
\emptyset & 0 & 0 & 0 \\
\theta_1\cap\theta_2\cap\theta_3 & 0 & 0 & 0.16 \\
\theta_2\cap\theta_3 & 0 &0.20 & 0.19 \\
\theta_1\cap\theta_3 & 0.10 & 0 & 0.12\\
(\theta_1\cup\theta_2)\cap\theta_3 & 0 & 0 & 0.01\\
\theta_3 & 0.30 & 0.10 & 0.10\\
\theta_1\cap\theta_2 & 0.10 & 0.20 & 0.22\\
(\theta_1\cup\theta_3)\cap\theta_2 & 0 & 0 & 0.05\\
(\theta_2\cup\theta_3)\cap\theta_1& 0 & 0 & 0 \\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) & 0 & 0& 0 \\
(\theta_1\cap\theta_2)\cup\theta_3 & 0 & 0 & 0\\
\theta_2 & 0.20 & 0.10 & 0.03\\
(\theta_1\cap\theta_3)\cup\theta_2 & 0 & 0 & 0\\
\theta_2\cup\theta_3 & 0 & 0 & 0\\
\theta_1 & 0.10 & 0.20 & 0.08\\
(\theta_2\cap\theta_3)\cup\theta_1& 0 & 0 & 0.02\\
\theta_1\cup\theta_3 & 0.10 & 0.20 & 0.02\\
\theta_1\cup\theta_2 & 0.10 & 0 & 0\\
\theta_1\cup\theta_2\cup\theta_3 & 0 & 0 & 0
\end{array}$$ The following subsections present the numerical results obtained by the DSm hybrid rule on the seven previous examples. The tables show all the values of $\phi(A)$, $S_1(A)$, $S_2(A)$ and $S_3(A)$ to help the reader to check by himself or herself the validity of these results. It is important to note that the values of $S_1(A)$, $S_2(A)$ and $S_3(A)$ when $\phi(A)=0$ do not need to be computed in practice but are provided here only for a checking purpose.
### Application of the DSm Hybrid rule on example 1
Here is the numerical result corresponding to example 1 with the hybrid-model $\mathcal{M}_1$ (i.e with the exclusivity constraint $\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_1}{\equiv}\emptyset$). The right column of the table provides the result obtained using the DSm hybrid rule, ie. $\forall A\in D^\Theta$, $$m_{\mathcal{M}_1(\Theta)}(A)=\phi(A)\bigl[ S_1(A) + S_2(A) + S_3(A)\bigr]$$
$$\begin{array}{|l|rrrr|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta& \phi(A) & S_1(A) & S_2(A)& S_3(A) & m_{\mathcal{M}_1(\Theta)}(A)\\
\hline
\emptyset & 0 & 0 & 0 & 0 & 0\\
\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_1}{\equiv}\emptyset & 0 & 0.16 & 0 & 0 & 0\\
\theta_2\cap\theta_3 & 1 & 0.19 & 0 & 0 & 0.19\\
\theta_1\cap\theta_3 & 1 & 0.12 & 0 & 0 & 0.12\\
(\theta_1\cup\theta_2)\cap\theta_3 & 1 & 0.01 & 0 & 0 .02 & 0.03\\
\theta_3 & 1 & 0.10 & 0 & 0 & 0.10\\
\theta_1\cap\theta_2 & 1 & 0.22 & 0 & 0 & 0.22\\
(\theta_1\cup\theta_3)\cap\theta_2 & 1 & 0.05 & 0 & 0.02 & 0.07\\
(\theta_2\cup\theta_3)\cap\theta_1& 1 & 0 & 0 & 0.02 & 0.02\\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) & 1 & 0 & 0 & 0 & 0\\
(\theta_1\cap\theta_2)\cup\theta_3 & 1 & 0 & 0 & 0.07 & 0.07\\
\theta_2 & 1 & 0.03 & 0 & 0 & 0.03\\
(\theta_1\cap\theta_3)\cup\theta_2 & 1 & 0 & 0 & 0.01 & 0.01\\
\theta_2\cup\theta_3 & 1 & 0 & 0 & 0 & 0\\
\theta_1 & 1 & 0.08 & 0 & 0 & 0.08\\
(\theta_2\cap\theta_3)\cup\theta_1& 1 & 0.02 & 0 & 0.02 & 0.04\\
\theta_1\cup\theta_3 & 1 & 0.02 & 0 & 0 & 0.02\\
\theta_1\cup\theta_2 & 1 & 0 & 0 & 0 & 0\\
\theta_1\cup\theta_2\cup\theta_3 & 1 & 0 & 0 & 0 & 0\\
\hline
\end{array}
\quad
\mathbf{D}_{\mathcal{M}_1}=\begin{bmatrix}
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 1 & 1 \\
0 & 0 & 0 & 1 & 1 & 1 \\
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 1 \\
0 & 0 & 1 & 0 & 1 & 0 \\
0 & 0 & 1 & 0 & 1 & 1 \\
0 & 0 & 1 & 1 & 1 & 1 \\
0 & 1 & 1 & 0 & 0 & 1 \\
0 & 1 & 1 & 0 & 1 & 1 \\
0 & 1 & 1 & 1 & 1 & 1 \\
1 & 0 & 1 & 0 & 1 & 0 \\
1 & 0 & 1 & 0 & 1 & 1 \\
1 & 0 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & 0 & 1 & 1 \\
1 & 1 & 1 & 1 & 1 & 1
\end{bmatrix}$$
From the previous table of this first numerical example, we see in column corresponding to $S_3(A)$ how the initial combined mass $m_{\mathcal{M}^f(\Theta)}(\theta_1\cap\theta_2\cap\theta_3)\equiv S_1(\theta_1\cap\theta_2\cap\theta_3)=0.16$ is transferred (due to the constraint of $\mathcal{M}_1$) only onto the elements $(\theta_1\cup\theta_2)\cap\theta_3$, $(\theta_1\cup\theta_3)\cap\theta_2$, $(\theta_2\cup\theta_3)\cap\theta_1$, $(\theta_1\cap\theta_2)\cup\theta_3$, $(\theta_1\cap\theta_3)\cup\theta_2$, and $(\theta_2\cap\theta_3)\cup\theta_1$ of $D^\Theta$. We can easily check that the sum of the elements of the column for $S_3(A)$ is equal to $m_{\mathcal{M}^f(\Theta)}(\theta_1\cap\theta_2\cap\theta_3)=0.16$ as required. Thus after introducing the constraint, the initial hyper-power set $D^\Theta$ reduces to 18 elements as follows $$\begin{split}
D^\Theta_{\mathcal{M}_1}=\{\emptyset,\theta_2\cap\theta_3,\theta_1\cap\theta_3,(\theta_1\cup\theta_2)\cap\theta_3,\theta_3,\theta_1\cap\theta_2,(\theta_1\cup\theta_3)\cap\theta_2,(\theta_2\cup\theta_3)\cap\theta_1,\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2),\\
(\theta_1\cap\theta_2)\cup\theta_3,\theta_2,(\theta_1\cap\theta_3)\cup\theta_2,\theta_2\cup\theta_3,\theta_1,(\theta_2\cap\theta_3)\cup\theta_1,\theta_1\cup\theta_3,\theta_1\cup\theta_2,\theta_1\cup\theta_2\cup\theta_3\}
\end{split}$$
As detailed in [@Dezert_2003f], the elements of $D^\Theta_{\mathcal{M}_1}$ can be described and encoded by the matrix product $\mathbf{D}_{\mathcal{M}_1}\cdot\mathbf{u}_{\mathcal{M}_1}$ with $\mathbf{D}_{\mathcal{M}_1}$ given above and the basis vector $\mathbf{u}_{\mathcal{M}_1}$ defined as $\mathbf{u}_{\mathcal{M}_1}=[<1> <2> <12> <3> <13> <23>]'$. Actually $\mathbf{u}_{\mathcal{M}_1}$ is directly obtained from $\mathbf{u}_{\mathcal{M}^f}$[^3] by removing its component $<123>$ corresponding to the constraint introduced by the model $\mathcal{M}_1$.\
In general, the encoding matrix $\mathbf{D}_{\mathcal{M}}$ for a given DSm hybrid model $\mathcal{M}$ is obtained from $\mathbf{D}_{\mathcal{M}^f}$ by removing all its columns corresponding to the constraints of the chosen model $\mathcal{M}$ and all the rows corresponding to redundant/equivalent propositions. In this particular example with model $\mathcal{M}_1$, we will just have to remove the last column of $\mathbf{D}_{\mathcal{M}^f}$ to get $\mathbf{D}_{\mathcal{M}_1}$ and no row is removed from $\mathbf{D}_{\mathcal{M}^f}$ because there is no redundant/equivalent proposition involved in this example. This suppression of some rows of $\mathbf{D}_{\mathcal{M}^f}$ will however occur in next examples. We encourage the reader to consult the references [@Dezert_2003f; @Dezert_Smarandache_2003] for explanations and details about the generation, the encoding and the partial ordering of hyper-power sets.
### Application of the DSm Hybrid rule on example 2
Here is the numerical result corresponding to example 2 with the hybrid-model $\mathcal{M}_2$ (i.e with the exclusivity constraint $\theta_1\cap\theta_2\overset{\mathcal{M}_2}{\equiv}\emptyset \Rightarrow \theta_1\cap\theta_2\cap\theta_3\overset{\mathcal{M}_2}{\equiv}\emptyset$). One gets now
$$\begin{array}{|l|rrrr|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta& \phi(A) & S_1(A) & S_2(A)& S_3(A) & m_{\mathcal{M}_2(\Theta)}(A)\\
\hline
\emptyset & 0 & 0 & 0 & 0 & 0\\
\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_2}{\equiv}\emptyset & 0 & 0.16 & 0 & 0 & 0\\
\theta_2\cap\theta_3 & 1 & 0.19 & 0 & 0 & 0.19\\
\theta_1\cap\theta_3 & 1 & 0.12 & 0 & 0 & 0.12\\
(\theta_1\cup\theta_2)\cap\theta_3 & 1 & 0.01 & 0 & 0 .02 & 0.03\\
\theta_3 & 1 & 0.10 & 0 & 0 & 0.10\\
\theta_1\cap\theta_2\overset{\mathcal{M}_2}{\equiv}\emptyset & 0 & 0.22 & 0 & 0.02 & 0\\
(\theta_1\cup\theta_3)\cap\theta_2 \overset{\mathcal{M}_2}{\equiv}\theta_2\cap\theta_3 & 1 & 0.05 & 0 & 0.02 & 0.07\\
(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_2}{\equiv}\theta_1\cap\theta_3 & 1 & 0 & 0 & 0.02 & 0.02\\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) \overset{\mathcal{M}_2}{\equiv}(\theta_1\cup\theta_2)\cap\theta_3
& 1 & 0 & 0 & 0 & 0\\
(\theta_1\cap\theta_2)\cup\theta_3 \overset{\mathcal{M}_2}{\equiv} \theta_3 & 1 & 0 & 0 & 0.07 & 0.07\\
\theta_2 & 1 & 0.03 & 0 & 0.05 & 0.08\\
(\theta_1\cap\theta_3)\cup\theta_2 & 1 & 0 & 0 & 0.01 & 0.01\\
\theta_2\cup\theta_3 & 1 & 0 & 0 & 0 & 0\\
\theta_1 & 1 & 0.08 & 0 & 0.04 & 0.12\\
(\theta_2\cap\theta_3)\cup\theta_1& 1 & 0.02 & 0 & 0.02 & 0.04\\
\theta_1\cup\theta_3 & 1 & 0.02 & 0 & 0.04 & 0.06\\
\theta_1\cup\theta_2 & 1 & 0 & 0.02 & 0.07 & 0.09\\
\theta_1\cup\theta_2\cup\theta_3 & 1 & 0 & 0 & 0 & 0\\
\hline
\end{array}$$
From the previous table of this numerical example, we see in column corresponding to $S_3(A)$ how the initial combined masses $m_{\mathcal{M}^f(\Theta)} (\theta_1\cap\theta_2\cap\theta_3) \equiv S_1(\theta_1\cap\theta_2\cap\theta_3)=0.16$ and $m_{\mathcal{M}^f(\Theta)}(\theta_1\cap\theta_2)\equiv S_1(\theta_1\cap\theta_2)=0.22$ are transferred (due to the constraint of $\mathcal{M}_2$) onto some elements of $D^\Theta$. We can easily check that the sum of the elements of the column for $S_3(A)$ is equal to $0.16+0.22=0.38$. Because some elements of $D^\Theta$ are now equivalent due to the constraints of $\mathcal{M}_2$, we have to sum all the masses corresponding to same equivalent propositions/elements (by example $\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) \overset{\mathcal{M}_2}{\equiv}(\theta_1\cup\theta_2)\cap\theta_3$). This can be viewed as the final compression step. One then gets the reduced hyper-power set $D^\Theta_{\mathcal{M}_2}$ having now 13 different elements with the combined belief masses presented in the following table. The basis vector $\mathbf{u}_{\mathcal{M}_2}$ and the encoding matrix $\mathbf{D}_{\mathcal{M}_2}$ for the elements of $D^\Theta_{\mathcal{M}_2}$ are given by $\mathbf{u}_{\mathcal{M}_2}=[<1> <2> <3> <13> <23>]'$ and below. Actually $\mathbf{u}_{\mathcal{M}_2}$ is directly obtained from $\mathbf{u}_{\mathcal{M}^f}$ by removing its components $<12>$ and $<123>$ corresponding to the constraints introduced by the model $\mathcal{M}_2$. $$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_2} & m_{\mathcal{M}_2(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_2\cap\theta_3 & 0.19+0.07=0.26\\
\theta_1\cap\theta_3 & 0.12+0.02=0.14\\
(\theta_1\cup\theta_2)\cap\theta_3 & 0.03+0=0.03\\
\theta_3 & 0.10+0.07=0.17\\
\theta_2 & 0.08\\
(\theta_1\cap\theta_3)\cup\theta_2 & 0.01\\
\theta_2\cup\theta_3 & 0\\
\theta_1 & 0.12\\
(\theta_2\cap\theta_3)\cup\theta_1& 0.04\\
\theta_1\cup\theta_3 & 0.06\\
\theta_1\cup\theta_2 & 0.09\\
\theta_1\cup\theta_2\cup\theta_3 & 0\\
\hline
\end{array}
\qquad\text{and}\qquad
\mathbf{D}_{\mathcal{M}_2}=\begin{bmatrix}
0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 & 1 \\
0 & 0 & 1 & 1 & 1 \\
0 & 1 & 0 & 0 & 1 \\
0 & 1 & 0 & 1 & 1 \\
0 & 1 & 1 & 1 & 1 \\
1 & 0 & 0 & 1 & 0 \\
1 & 0 & 0 & 1 & 1 \\
1 & 0 & 1 & 1 & 1 \\
1 & 1 & 0 & 1 & 1 \\
1 & 1 & 1 & 1 & 1
\end{bmatrix}$$
### Application of the DSm Hybrid rule on example 3
Here is the numerical result corresponding to example 3 with the hybrid-model $\mathcal{M}_3$ (i.e with the exclusivity constraint $(\theta_1\cup\theta_3)\cap\theta_2 \overset{\mathcal{M}_3}{\equiv}\emptyset$). This constraint implies directly $\theta_1\cap\theta_2\cap\theta_3\overset{\mathcal{M}_3}{\equiv}\emptyset$, $\theta_1\cap\theta_2\overset{\mathcal{M}_3}{\equiv}\emptyset$ and $\theta_2\cap\theta_3\overset{\mathcal{M}_3}{\equiv}\emptyset$. One gets now $$\begin{array}{|l|rrrrr|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta& \phi(A) & S_1(A) & S_2(A)& S_3(A) & m_{\mathcal{M}_3(\Theta)}(A)\\
\hline
\emptyset & 0 & 0 & 0 & 0 & 0\\
\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_3}{\equiv}\emptyset & 0 & 0.16 & 0 & 0 & 0\\
\theta_2\cap\theta_3\overset{\mathcal{M}_3}{\equiv}\emptyset & 0 & 0.19 & 0 & 0 & 0\\
\theta_1\cap\theta_3 & 1 & 0.12 & 0 & 0 & 0.12\\
(\theta_1\cup\theta_2)\cap\theta_3 \overset{\mathcal{M}_3}{\equiv} \theta_1\cap\theta_3 & 1 & 0.01 & 0 & 0 .02 & 0.03\\
\theta_3 & 1 & 0.10 & 0 & 0.06 & 0.16\\
\theta_1\cap\theta_2\overset{\mathcal{M}_3}{\equiv}\emptyset & 0 & 0.22 & 0 & 0.02 & 0\\
(\theta_1\cup\theta_3)\cap\theta_2 \overset{\mathcal{M}_3}{\equiv}\emptyset & 0 & 0.05 & 0 & 0.02 & 0\\
(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_3}{\equiv} \theta_1\cap\theta_3 & 1 & 0 & 0 & 0.02 & 0.02\\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) \overset{\mathcal{M}_3}{\equiv} \theta_1\cap\theta_3
& 1 & 0 & 0 & 0 & 0\\
(\theta_1\cap\theta_2)\cup\theta_3 \overset{\mathcal{M}_3}{\equiv} \theta_3 & 1 & 0 & 0 & 0.07 & 0.07\\
\theta_2 & 1 & 0.03 & 0 & 0.09 & 0.12\\
(\theta_1\cap\theta_3)\cup\theta_2 & 1 & 0 & 0 & 0.01 & 0.01\\
\theta_2\cup\theta_3 & 1 & 0 & 0 & 0.05 & 0.05\\
\theta_1 & 1 & 0.08 & 0 & 0.04 & 0.12\\
(\theta_2\cap\theta_3)\cup\theta_1 \overset{\mathcal{M}_3}{\equiv} \theta_1 & 1 & 0.02 & 0 & 0.02 & 0.04\\
\theta_1\cup\theta_3 & 1 & 0.02 & 0 & 0.06 & 0.08\\
\theta_1\cup\theta_2 & 1 & 0 & 0.02 & 0.09 & 0.11\\
\theta_1\cup\theta_2\cup\theta_3 & 1 & 0 & 0.02 & 0.05 & 0.07\\
\hline
\end{array}$$
From the previous table of this numerical example, we see in column corresponding to $S_3(A)$ how the initial combined masses $m_{\mathcal{M}^f(\Theta)}((\theta_1\cup\theta_3)\cap\theta_2)\equiv S_1((\theta_1\cup\theta_3)\cap\theta_2)=0.05$, $m_{\mathcal{M}^f(\Theta)} (\theta_1\cap\theta_2\cap\theta_3) \equiv S_1(\theta_1\cap\theta_2\cap\theta_3)=0.16$, $m_{\mathcal{M}^f(\Theta)}(\theta_2\cap\theta_3)\equiv S_1(\theta_2\cap\theta_3)=0.19$ and $m_{\mathcal{M}^f(\Theta)}(\theta_1\cap\theta_2)\equiv S_1(\theta_1\cap\theta_2)=0.22$ are transferred (due to the constraint of $\mathcal{M}_3$) onto some elements of $D^\Theta$. We can easily check that the sum of the elements of the column for $S_3(A)$ is equal to $0.05+0.16+0.19+0.22=0.62$.\
Because some elements of $D^\Theta$ are now equivalent due to the constraints of $\mathcal{M}_3$, we have to sum all the masses corresponding to same equivalent propositions. Thus after the final compression step, one gets the reduced hyper-power set $D^\Theta_{\mathcal{M}_3}$ having only 10 different elements with the following combined belief masses : $$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_3} & m_{\mathcal{M}_3(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_1\cap\theta_3 & 0.12+0.03+0.02+0=0.17\\
\theta_3 & 0.16+0.07=0.23\\
\theta_2 & 0.12\\
(\theta_1\cap\theta_3)\cup\theta_2 & 0.01\\
\theta_2\cup\theta_3 & 0.05\\
\theta_1 & 0.12+0.04=0.16\\
\theta_1\cup\theta_3 & 0.08\\
\theta_1\cup\theta_2 & 0.11\\
\theta_1\cup\theta_2\cup\theta_3 & 0.07\\
\hline
\end{array}
\qquad \text{and} \qquad
\mathbf{D}_{\mathcal{M}_3}=\begin{bmatrix}
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 \\
0 & 0 & 1 & 1 \\
0 & 1 & 0 & 0 \\
0 & 1 & 0 & 1 \\
0 & 1 & 1 & 1 \\
1 & 0 & 0 & 1 \\
1 & 0 & 1 & 1 \\
1 & 1 & 0 & 1 \\
1 & 1 & 1 & 1
\end{bmatrix}$$ The basis vector $\mathbf{u}_{\mathcal{M}_3}$ is given by $\mathbf{u}_{\mathcal{M}_3}=[<1> <2> <3> <13> ]'$ and the encoding matrix $\mathbf{D}_{\mathcal{M}_3}$ is explicated just above.
### Application of the DSm Hybrid rule on example 4 (Shafer’s model)
Here is the numerical result corresponding to example 4 with the hybrid-model $\mathcal{M}_4$ including all possible exclusivity constraints. This DSm hybrid model corresponds actually to the Shafer’s model. One gets now
$$\begin{array}{|l|rrrrr|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta& \phi(A) & S_1(A) & S_2(A)& S_3(A) & m_{\mathcal{M}_4(\Theta)}(A)\\
\hline
\emptyset & 0 & 0 & 0 & 0 & 0\\
\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_4}{\equiv}\emptyset & 0 & 0.16 & 0 & 0 & 0\\
\theta_2\cap\theta_3\overset{\mathcal{M}_4}{\equiv}\emptyset & 0 & 0.19 & 0 & 0 & 0\\
\theta_1\cap\theta_3\overset{\mathcal{M}_4}{\equiv}\emptyset & 0 & 0.12 & 0 & 0 & 0\\
(\theta_1\cup\theta_2)\cap\theta_3 \overset{\mathcal{M}_4}{\equiv} \emptyset & 0 & 0.01 & 0 & 0 .02 & 0\\
\theta_3 & 1 & 0.10 & 0 & 0.07 & 0.17\\
\theta_1\cap\theta_2\overset{\mathcal{M}_4}{\equiv}\emptyset & 0 & 0.22 & 0 & 0.02 & 0\\
(\theta_1\cup\theta_3)\cap\theta_2 \overset{\mathcal{M}_4}{\equiv}\emptyset & 0 & 0.05 & 0 & 0.02 & 0\\
(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_4}{\equiv} \emptyset & 0 & 0 & 0 & 0.02 & 0\\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) \overset{\mathcal{M}_4}{\equiv} \emptyset
& 0 & 0 & 0 & 0 & 0\\
(\theta_1\cap\theta_2)\cup\theta_3 \overset{\mathcal{M}_4}{\equiv} \theta_3 & 1 & 0 & 0 & 0.07 & 0.07\\
\theta_2 & 1 & 0.03 & 0 & 0.09 & 0.12\\
(\theta_1\cap\theta_3)\cup\theta_2 \overset{\mathcal{M}_4}{\equiv} \theta_2 & 1 & 0 & 0 & 0.01 & 0.01\\
\theta_2\cup\theta_3 & 1 & 0 & 0 & 0.05 & 0.05\\
\theta_1 & 1 & 0.08 & 0 & 0.06 & 0.14\\
(\theta_2\cap\theta_3)\cup\theta_1 \overset{\mathcal{M}_4}{\equiv} \theta_1 & 1 & 0.02 & 0 & 0.02 & 0.04\\
\theta_1\cup\theta_3 & 1 & 0.02 & 0 & 0.15 & 0.17\\
\theta_1\cup\theta_2 & 1 & 0 & 0.02 & 0.09 & 0.11\\
\theta_1\cup\theta_2\cup\theta_3 & 1 & 0 & 0.06 & 0.06 & 0.12\\
\hline
\end{array}$$
From the previous table of this numerical example, we see in column corresponding to $S_3(A)$ how the initial combined masses of the eight elements forced to the empty set by the constraints of the model $\mathcal{M}_4$ are transferred onto some elements of $D^\Theta$. We can easily check that the sum of the elements of the column for $S_3(A)$ is equal to $0.16+0.19+0.12+0.01+0.22+0.05+0=0.75$.\
After the final compression step (i.e. the clustering of all equivalent propositions), one gets the reduced hyper-power set $D^\Theta_{\mathcal{M}_4}$ having only $2^3=8$ (corresponding to the classical power set $2^\Theta$) with the following combined belief masses:
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_4} & m_{\mathcal{M}_4(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_3 & 0.17+0.07=0.24\\
\theta_2 & 0.12+0.01=0.13\\
\theta_2\cup\theta_3 & 0.05\\
\theta_1 & 0.14+0.04=0.18\\
\theta_1\cup\theta_3 & 0.17\\
\theta_1\cup\theta_2 & 0.11\\
\theta_1\cup\theta_2\cup\theta_3 & 0.12\\
\hline
\end{array}
\qquad \text{and} \qquad
\mathbf{D}_{\mathcal{M}_4}=\begin{bmatrix}
0 & 0 & 0 \\
0 & 0 & 1 \\
0 & 1 & 0 \\
0 & 1 & 1 \\
1 & 0 & 0 \\
1 & 0 & 1 \\
1 & 1 & 0 \\
1 & 1 & 1
\end{bmatrix}$$
The basis vector $\mathbf{u}_{\mathcal{M}_4}$ is given by $\mathbf{u}_{\mathcal{M}_4}=[<1> <2> <3> ]'$ and the encoding matrix $\mathbf{D}_{\mathcal{M}_4}$ is explicated just above.
### Application of the DSm Hybrid rule on example 5
Here is the numerical result corresponding to example 5 with the hybrid-model $\mathcal{M}_5$ including the non-existential constraint $\theta_1\overset{\mathcal{M}_5}{\equiv}\emptyset$. This non-existential constraint implies $\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_5}{\equiv}\emptyset$, $\theta_1\cap\theta_2 \overset{\mathcal{M}_5}{\equiv}\emptyset$, $\theta_1\cap\theta_3 \overset{\mathcal{M}_5}{\equiv}\emptyset$ and $(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_5}{\equiv}\emptyset $. One gets now with applying the DSm hybrid rule of combination:
$$\begin{array}{|l|rrrrr|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta& \phi(A) & S_1(A) & S_2(A)& S_3(A) & m_{\mathcal{M}_5(\Theta)}(A)\\
\hline
\emptyset & 0 & 0 & 0 & 0 & 0\\
\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_5}{\equiv}\emptyset & 0 & 0.16 & 0 & 0 & 0\\
\theta_2\cap\theta_3 & 1 & 0.19 & 0 & 0 & 0.19\\
\theta_1\cap\theta_3\overset{\mathcal{M}_5}{\equiv}\emptyset & 0 & 0.12 & 0 & 0 & 0\\
(\theta_1\cup\theta_2)\cap\theta_3 \overset{\mathcal{M}_5}{\equiv} \theta_2\cap\theta_3& 1 & 0.01 & 0 & 0 .02 & 0.03\\
\theta_3 & 1 & 0.10 & 0 & 0.01 & 0.11\\
\theta_1\cap\theta_2\overset{\mathcal{M}_5}{\equiv}\emptyset & 0 & 0.22 & 0 & 0.02 & 0\\
(\theta_1\cup\theta_3)\cap\theta_2 \overset{\mathcal{M}_5}{\equiv}\theta_2\cap\theta_3 & 1 & 0.05 & 0 & 0.02 & 0.07\\
(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_5}{\equiv} \emptyset & 0 & 0 & 0 & 0.02 & 0\\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) \overset{\mathcal{M}_5}{\equiv} \theta_2\cap\theta_3
& 1 & 0 & 0 & 0 & 0\\
(\theta_1\cap\theta_2)\cup\theta_3 \overset{\mathcal{M}_5}{\equiv} \theta_3 & 1 & 0 & 0 & 0.07 & 0.07\\
\theta_2 & 1 & 0.03 & 0 & 0.05 & 0.08\\
(\theta_1\cap\theta_3)\cup\theta_2 \overset{\mathcal{M}_5}{\equiv} \theta_2 & 1 & 0 & 0 & 0.01 & 0.01\\
\theta_2\cup\theta_3 & 1 & 0 & 0 & 0 & 0\\
\theta_1 \overset{\mathcal{M}_5}{\equiv} \emptyset & 0 & 0.08 & 0.02 & 0.08 & 0\\
(\theta_2\cap\theta_3)\cup\theta_1 \overset{\mathcal{M}_5}{\equiv} \theta_2\cap\theta_3 & 1 & 0.02 & 0 & 0.02 & 0.04\\
\theta_1\cup\theta_3 \overset{\mathcal{M}_5}{\equiv} \theta_3 & 1 & 0.02 & 0.02 & 0.17 & 0.21\\
\theta_1\cup\theta_2 \overset{\mathcal{M}_5}{\equiv} \theta_2 & 1 & 0 & 0.06 & 0.09 & 0.15\\
\theta_1\cup\theta_2\cup\theta_3 \overset{\mathcal{M}_5}{\equiv} \theta_2 \cup \theta_3 & 1 & 0 & 0.04 & 0 & 0.04\\
\hline
\end{array}$$
From the previous table of this numerical example, we see in column corresponding to $S_3(A)$ how the initial combined masses of the 5 elements forced to the empty set by the constraints of the model $\mathcal{M}_5$ are transferred onto some elements of $D^\Theta$. We can easily check that the sum of the elements of the column for $S_3(A)$ is equal to $0+0.16+0.12+0.22+0+0.08=0.58$ (sum of S1(A) for which $\phi(A)=0$).\
After the final compression step (i.e. the clustering of all equivalent propositions), one gets the reduced hyper-power set $D^\Theta_{\mathcal{M}_5}$ having only 5 different elements according to: $$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_5} & m_{\mathcal{M}_5(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_2\cap\theta_3 & 0.19+0.03+0.07+0+0.04=0.33\\
\theta_3 & 0.11+0.07+0.21=0.39\\
\theta_2 & 0.08+0.01+0.15=0.24\\
\theta_2\cup\theta_3 & 0+0.04=0.04\\
\hline
\end{array}
\qquad \text{and} \qquad
\mathbf{D}_{\mathcal{M}_5}=\begin{bmatrix}
0 & 0 & 0 \\
0 & 0 & 1 \\
0 & 1 & 1 \\
1 & 0 & 1 \\
1 & 1 & 1
\end{bmatrix}$$
The basis vector $\mathbf{u}_{\mathcal{M}_5}$ is given by $\mathbf{u}_{\mathcal{M}_5}=[<2> <3> <23> ]'$. and the encoding matrix $\mathbf{D}_{\mathcal{M}_5}$ is explicated just above.
### Application of the DSm Hybrid rule on example 6
Here is the numerical result corresponding to example 6 with the hybrid-model $\mathcal{M}_6$ including the two non-existential constraint $\theta_1\overset{\mathcal{M}_6}{\equiv}\emptyset$ and $\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset$. This is a degenerate example actually, since no uncertainty arises in such trivial model. We just want to show here that the DSm hybrid rule still works in this example and provide a legitimist result. By applying the DSm hybrid rule of combination, one now gets: $$\begin{array}{|l|rrrrr|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta& \phi(A) & S_1(A) & S_2(A)& S_3(A) & m_{\mathcal{M}_6(\Theta)}(A)\\
\hline
\emptyset & 0 & 0 & 0 & 0 & 0\\
\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_6}{\equiv}\emptyset & 0 & 0.16 & 0 & 0 & 0\\
\theta_2\cap\theta_3\overset{\mathcal{M}_6}{\equiv}\emptyset & 0 & 0.19 & 0 & 0 & 0\\
\theta_1\cap\theta_3\overset{\mathcal{M}_6}{\equiv}\emptyset & 0 & 0.12 & 0 & 0 & 0\\
(\theta_1\cup\theta_2)\cap\theta_3 \overset{\mathcal{M}_6}{\equiv} \emptyset & 0 & 0.01 & 0 & 0.02 & 0\\
\theta_3 & 1 & 0.10 & 0 & 0.07 & 0.17\\
\theta_1\cap\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset & 0 & 0.22 & 0 & 0.02 & 0\\
(\theta_1\cup\theta_3)\cap\theta_2 \overset{\mathcal{M}_6}{\equiv}\emptyset & 0 & 0.05 & 0 & 0.02 & 0\\
(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_6}{\equiv} \emptyset & 0 & 0 & 0 & 0.02 & 0\\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) \overset{\mathcal{M}_6}{\equiv} \emptyset
& 0 & 0 & 0 & 0 & 0\\
(\theta_1\cap\theta_2)\cup\theta_3 \overset{\mathcal{M}_6}{\equiv} \theta_3 & 1 & 0 & 0 & 0.07 & 0.07\\
\theta_2\overset{\mathcal{M}_6}{\equiv}\emptyset & 0 & 0.03 & 0.02 & 0.11 & 0\\
(\theta_1\cap\theta_3)\cup\theta_2 \overset{\mathcal{M}_6}{\equiv} \emptyset & 0 & 0 & 0 & 0.01 & 0\\
\theta_2\cup\theta_3\overset{\mathcal{M}_6}{\equiv} \theta_3 & 1 & 0 & 0.04 & 0.05 & 0.09\\
\theta_1 \overset{\mathcal{M}_6}{\equiv} \emptyset & 0 & 0.08 & 0 & 0.08 & 0\\
(\theta_2\cap\theta_3)\cup\theta_1 \overset{\mathcal{M}_6}{\equiv} \emptyset & 0 & 0.02 & 0 & 0.02 & 0\\
\theta_1\cup\theta_3 \overset{\mathcal{M}_6}{\equiv} \theta_3 & 1 & 0.02 & 0.02 & 0.19 & 0.23\\
\theta_1\cup\theta_2 \overset{\mathcal{M}_6}{\equiv} \emptyset & 0 & 0 & 0.21 & 0.12 & 0\\
\theta_1\cup\theta_2\cup\theta_3 \overset{\mathcal{M}_6}{\equiv} \theta_3 & 1 & 0 & 0.36 & 0.08 & 0.44\\
\hline
\end{array}$$
After the clustering of all equivalent propositions, one gets the reduced hyper-power set $D^\Theta_{\mathcal{M}_6}$ having only 2 different elements according to: $$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_6} & m_{\mathcal{M}_6(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_3 & 0.17+0.07+0.09+0.23+0.44=1\\
\hline
\end{array}$$ The encoding matrix $\mathbf{D}_{\mathcal{M}_6}$ and the basis vector $\mathbf{u}_{\mathcal{M}_6}$ for the elements of $D^\Theta_{\mathcal{M}_6}$ reduce to $\mathbf{D}_{\mathcal{M}_6}=[0 1]'$ and $\mathbf{u}_{\mathcal{M}_6}=[<3>]$.
### Application of the DSm Hybrid rule on example 7
Here is the numerical result corresponding to example 7 with the hybrid-model $\mathcal{M}_7$ including the [*[mixed exclusivity and non-existential]{}*]{} constraint $( \theta_1\cap \theta_2)\cup\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset$. This mixed constraint implies $\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset$, $\theta_1\cap\theta_2 \overset{\mathcal{M}_7}{\equiv}\emptyset$, $\theta_1\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset$, $\theta_2\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset$, $(\theta_1\cup\theta_2)\cap\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset$, $(\theta_1\cup\theta_3)\cap\theta_2\overset{\mathcal{M}_7}{\equiv}\emptyset$, $(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_7}{\equiv}\emptyset $, $ \{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2)\overset{\mathcal{M}_7}{\equiv}\emptyset$ and $\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset$. By applying the DSm hybrid rule of combination, one now gets: $$\begin{array}{|l|rrrrr|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta& \phi(A) & S_1(A) & S_2(A)& S_3(A) & m_{\mathcal{M}_7(\Theta)}(A)\\
\hline
\emptyset & 0 & 0 & 0 & 0 & 0\\
\theta_1\cap\theta_2\cap\theta_3 \overset{\mathcal{M}_7}{\equiv}\emptyset & 0 & 0.16 & 0 & 0 & 0\\
\theta_2\cap\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset & 0 & 0.19 & 0 & 0 & 0\\
\theta_1\cap\theta_3\overset{\mathcal{M}_7}{\equiv}\emptyset & 0 & 0.12 & 0 & 0 & 0\\
(\theta_1\cup\theta_2)\cap\theta_3 \overset{\mathcal{M}_7}{\equiv} \emptyset & 0 & 0.01 & 0 & 0 .02 & 0\\
\theta_3\overset{\mathcal{M}_7}{\equiv} \emptyset & 0 & 0.10 & 0.03 & 0.10 & 0\\
\theta_1\cap\theta_2\overset{\mathcal{M}_7}{\equiv}\emptyset & 0 & 0.22 & 0 & 0.02 & 0\\
(\theta_1\cup\theta_3)\cap\theta_2 \overset{\mathcal{M}_7}{\equiv}\emptyset & 0 & 0.05 & 0 & 0.02 & 0\\
(\theta_2\cup\theta_3)\cap\theta_1\overset{\mathcal{M}_7}{\equiv} \emptyset & 0 & 0 & 0 & 0.02 & 0\\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) \overset{\mathcal{M}_7}{\equiv} \emptyset
& 0 & 0 & 0 & 0 & 0\\
(\theta_1\cap\theta_2)\cup\theta_3 \overset{\mathcal{M}_7}{\equiv} \emptyset & 0 & 0 & 0 & 0.07 & 0\\
\theta_2 & 1 & 0.03 & 0 & 0.09 & 0.12\\
(\theta_1\cap\theta_3)\cup\theta_2 \overset{\mathcal{M}_7}{\equiv} \theta_2 & 1 & 0 & 0 & 0.01 & 0.01\\
\theta_2\cup\theta_3\overset{\mathcal{M}_7}{\equiv} \theta_2 & 1 & 0 & 0.06 & 0.05 & 0.11\\
\theta_1 & 1 & 0.08 & 0 & 0.06 & 0.14\\
(\theta_2\cap\theta_3)\cup\theta_1 \overset{\mathcal{M}_7}{\equiv} \theta_1 & 1 & 0.02 & 0 & 0.02 & 0.04\\
\theta_1\cup\theta_3 \overset{\mathcal{M}_7}{\equiv} \theta_1 & 1 & 0.02 & 0.01 & 0.22 & 0.25\\
\theta_1\cup\theta_2 & 1 & 0 & 0.02 & 0.09 & 0.11\\
\theta_1\cup\theta_2\cup\theta_3 \overset{\mathcal{M}_7}{\equiv} \theta_1\cup \theta_2 & 1 & 0 & 0.16 & 0.06 & 0.22\\
\hline
\end{array}$$
After the clustering of all equivalent propositions, one gets the reduced hyper-power set $D^\Theta_{\mathcal{M}_6}$ having only 4 different elements according to:
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_7} & m_{\mathcal{M}_7(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_2 & 0.12+0.01+0.11=0.24\\
\theta_1 & 0.14+0.04+0.25=0.43\\
\theta_1\cup \theta_2 & 0.11 + 0.22=0.33\\
\hline
\end{array}$$
The basis vector $\mathbf{u}_{\mathcal{M}_7}$ and the encoding matrix $\mathbf{D}_{\mathcal{M}_7}$ for the elements of $D^\Theta_{\mathcal{M}_7}$ are given by $$\mathbf{u}_{\mathcal{M}_7}=[<1> <2> ]' \qquad \text{and}\qquad
\mathbf{D}_{\mathcal{M}_7}=\begin{bmatrix}
0 & 0 \\
0 & 1 \\
1 & 0 \\
1 & 1
\end{bmatrix}$$
Example with more general basic belief assignments $m_1(.)$ and $m_2(.)$
------------------------------------------------------------------------
We present in this section the numerical results of the DSm hybrid rule of combination applied upon the seven previous models $\mathcal{M}_i$, $i=1,...,7$ with two general basic belief assignments $m_1(.)$ and $m_2(.)$ such that $m_1(A)>0$ and $m_2(A)>0$ for all $A\neq\emptyset\in D^{\Theta=\{\theta_1,\theta_2,\theta_3\}}$. We just provide here results. The verification is left to the reader. The following table presents the numerical values chosen for $m_1(.)$ and $m_2(.)$ and the result of the fusion obtained by the classical DSm rule of combination
$$\begin{array}{|l|ll|l|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta & m_1(A) & m_2(A) & m_{\mathcal{M}^f}(A) \\
\hline
\emptyset & 0 & 0 & 0 \\
\theta_1\cap\theta_2\cap\theta_3 & 0.01 & 0.40 & 0.4389 \\
\theta_2\cap\theta_3 & 0.04 &0.03 & 0.0410 \\
\theta_1\cap\theta_3 & 0.03 & 0.04 & 0.0497 \\
(\theta_1\cup\theta_2)\cap\theta_3 & 0.01 & 0.02 & 0.0257 \\
\theta_3 & 0.03 & 0.04 & 0.0311 \\
\theta_1\cap\theta_2 & 0.02 & 0.20 & 0.1846 \\
(\theta_1\cup\theta_3)\cap\theta_2 & 0.02 & 0.01 & 0.0156 \\
(\theta_2\cup\theta_3)\cap\theta_1 & 0.03 & 0.04 & 0.0459 \\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) & 0.04 & 0.03 & 0.0384 \\
(\theta_1\cap\theta_2)\cup\theta_3 & 0.04 & 0.03 & 0.0296 \\
\theta_2 & 0.02 & 0.01 & 0.0084\\
(\theta_1\cap\theta_3)\cup\theta_2 & 0.01 & 0.02 & 0.0221 \\
\theta_2\cup\theta_3 & 0.20 & 0.02 & 0.0140 \\
\theta_1 & 0.01 & 0.02 & 0.0109 \\
(\theta_2\cap\theta_3)\cup\theta_1 & 0.02 & 0.01 & 0.0090 \\
\theta_1\cup\theta_3 & 0.04 & 0.03 & 0.0136 \\
\theta_1\cup\theta_2 & 0.03 & 0.04 & 0.0175 \\
\theta_1\cup\theta_2\cup\theta_3 & 0.40 & 0.01 & 0.0040 \\
\hline
\end{array}$$
The following table shows the results obtained by the DSm hybrid rule of combination before the final compression step of all redundant propositions for the DSm hybrid models presented in the previous examples.
$$\begin{array}{|l|lllllll|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta & m_{\mathcal{M}_1}(A) & m_{\mathcal{M}_2}(A) & m_{\mathcal{M}_3}(A) & m_{\mathcal{M}_4}(A) & m_{\mathcal{M}_5}(A) & m_{\mathcal{M}_6}(A) & m_{\mathcal{M}_7}(A)\\
\hline
\emptyset & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
\theta_1\cap\theta_2\cap\theta_3 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\
\theta_2\cap\theta_3 & 0.0573 & 0.0573 & 0 & 0 & 0.0573 & 0 & 0\\
\theta_1\cap\theta_3 & 0.0621 & 0.0621 & 0.0621& 0 & 0 & 0 & 0\\
(\theta_1\cup\theta_2)\cap\theta_3 & 0.0324 & 0.0324 & 0.0335 & 0 & 0.0334 & 0 & 0\\
\theta_3 & 0.0435 & 0.0435 & 0.0460 & 0.0494 & 0.0459 & 0.0494 & 0\\
\theta_1\cap\theta_2 & 0.1946 & 0 & 0 & 0 & 0 & 0 & 0\\
(\theta_1\cup\theta_3)\cap\theta_2 & 0.0323 & 0.0365 & 0 & 0 & 0.0365 & 0 & 0\\
(\theta_2\cup\theta_3)\cap\theta_1 & 0.0651 & 0.0719 & 0.0719 & 0 & 0 & 0 & 0\\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) & 0.0607 & 0.0704 & 0.0743 & 0 & 0.0764 & 0 & 0\\
(\theta_1\cap\theta_2)\cup\theta_3 & 0.0527 & 0.0613 & 0.0658 & 0.0792 & 0.0687 & 0.0792 & 0\\
\theta_2 & 0.0165 & 0.0207 & 0.0221 & 0.0221& 0.0207 & 0 & 0.0221\\
(\theta_1\cap\theta_3)\cup\theta_2 & 0.0274 & 0.0309 & 0.0340 & 0.0375 & 0.0329 & 0 & 0.0375\\
\theta_2\cup\theta_3 & 0.0942 & 0.1346 & 0.1471 & 0.1774 & 0.1518 & 0.1850 & 0.1953\\
\theta_1 & 0.0151 & 0.0175 & 0.0175 & 0.0195 & 0 & 0 & 0.0195\\
(\theta_2\cap\theta_3)\cup\theta_1 & 0.0182 & 0.0229 & 0.0243 & 0.0295 & 0.0271 & 0 & 0.0295\\
\theta_1\cup\theta_3 & 0.0299 & 0.0385 & 0.0419 & 0.0558 & 0.0489 & 0.0589 & 0.0631\\
\theta_1\cup\theta_2 & 0.0299 & 0.0412 & 0.0452 & 0.0544 & 0.0498 & 0 & 0.0544\\
\theta_1\cup\theta_2\cup\theta_3 & 0.1681 & 0.2583 & 0.3143 & 0.4752 & 0.3506 & 0.6275 & 0.5786\\
\hline
\end{array}$$
The next tables present the final result of the DSm hybrid rule of combination after the compression step (the merging of all equivalent redundant propositions) presented in previous examples.
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_1} & m_{\mathcal{M}_1(\Theta)}(A)\\
\hline
\emptyset & 0 \\
\theta_2\cap\theta_3 & 0.0573 \\
\theta_1\cap\theta_3 & 0.0621 \\
(\theta_1\cup\theta_2)\cap\theta_3 & 0.0324 \\
\theta_3 & 0.0435 \\
\theta_1\cap\theta_2 & 0.1946 \\
(\theta_1\cup\theta_3)\cap\theta_2 & 0.0323 \\
(\theta_2\cup\theta_3)\cap\theta_1 & 0.0651 \\
\{(\theta_1\cap\theta_2)\cup\theta_3\} \cap(\theta_1\cup\theta_2) & 0.0607 \\
(\theta_1\cap\theta_2)\cup\theta_3 & 0.0527 \\
\theta_2 & 0.0165 \\
(\theta_1\cap\theta_3)\cup\theta_2 & 0.0274 \\
\theta_2\cup\theta_3 & 0.0942 \\
\theta_1 & 0.0151 \\
(\theta_2\cap\theta_3)\cup\theta_1 & 0.0182 \\
\theta_1\cup\theta_3 & 0.0299 \\
\theta_1\cup\theta_2 & 0.0299 \\
\theta_1\cup\theta_2\cup\theta_3 & 0.1681 \\
\hline
\end{array}$$
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_2} & m_{\mathcal{M}_2(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_2\cap\theta_3 & 0.0938\\
\theta_1\cap\theta_3 & 0.1340\\
(\theta_1\cup\theta_2)\cap\theta_3 & 0.1028\\
\theta_3 & 0.1048\\
\theta_2 & 0.0207\\
(\theta_1\cap\theta_3)\cup\theta_2 & 0.0309\\
\theta_2\cup\theta_3 & 0.1346\\
\theta_1 & 0.0175\\
(\theta_2\cap\theta_3)\cup\theta_1& 0.0229\\
\theta_1\cup\theta_3 & 0.0385\\
\theta_1\cup\theta_2 & 0.0412\\
\theta_1\cup\theta_2\cup\theta_3 & 0.2583\\
\hline
\end{array}$$
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_3} & m_{\mathcal{M}_3(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_1\cap\theta_3 & 0.2418\\
\theta_3 & 0.1118\\
\theta_2 & 0.0221\\
(\theta_1\cap\theta_3)\cup\theta_2 & 0.0340\\
\theta_2\cup\theta_3 & 0.1471\\
\theta_1 & 0.0418\\
\theta_1\cup\theta_3 & 0.0419\\
\theta_1\cup\theta_2 & 0.0452\\
\theta_1\cup\theta_2\cup\theta_3 & 0.3143\\
\hline
\end{array}$$
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_4} & m_{\mathcal{M}_4(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_3 & 0.1286\\
\theta_2 & 0.0596\\
\theta_2\cup\theta_3 & 0.1774\\
\theta_1 & 0.0490\\
\theta_1\cup\theta_3 & 0.0558\\
\theta_1\cup\theta_2 & 0.0544\\
\theta_1\cup\theta_2\cup\theta_3 & 0.4752\\
\hline
\end{array}$$
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_5} & m_{\mathcal{M}_5(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_2\cap\theta_3 & 0.2307\\
\theta_3 & 0.1635\\
\theta_2 & 0.1034\\
\theta_2\cup\theta_3 & 0.5024\\
\hline
\end{array}$$
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_6} & m_{\mathcal{M}_6(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_3 & 1\\
\hline
\end{array}$$
$$\begin{array}{|l|r|}
\hline
\text{Element} \: A \: \text{of} \; D^\Theta_{\mathcal{M}_7} & m_{\mathcal{M}_7(\Theta)}(A)\\
\hline
\emptyset & 0\\
\theta_2 & 0.2549\\
\theta_1 & 0.1121\\
\theta_1\cup \theta_2 & 0.6330\\
\hline
\end{array}$$
DSm hybrid rule versus Dempster’s rule of combination
-----------------------------------------------------
We discuss and compare here the DSm hybrid rule of combination with respect to the Dempster’s rule of combination and its alternative proposed in the literature based on the Dempster-Shafer Theory (DST) framework which is frequently adopted in many fusion/expert systems. It is necessary to first recall briefly the basis of the DST [@Shafer_1976].\
### Brief introduction to the DST
The DST starts by assuming an exhaustive and exclusive frame of discernment of the problem under consideration $\Theta=\{\theta_1,\theta_2,\ldots,\theta_n\}$. This corresponds to the Shafer’s model of the problem. The Shafer’s model is nothing more but the DSm model including all possible exclusivity constraints. The Shafer’s model assumes actually that an ultimate refinement of the problem is possible so that $\theta_i$ are well precisely defined/identified in such a way that we are sure that they are exclusive and exhaustive. From this Shafer’s model, a basic belief assignment (bba) $m (.): 2^\Theta \rightarrow [0, 1]$ associated to a given body of evidence $\mathcal{B}$ (also called sometimes corpus of evidence) is defined by $$m(\emptyset)=0 \qquad \text{and}\qquad \sum_{A\in 2^\Theta} m(A) = 1$$ where $2^\Theta$ is called the [*[power set]{}*]{} of $\Theta$, i.e. the set of all subsets of $\Theta$. The set of all propositions $A\in 2^\Theta$ such that $m(A)>0$ is called the core of $m(.)$ and is denoted $\mathcal{K}(m)$. From any bba, one defines the belief and plausibility functions of $A\subseteq\Theta$ as $$\text{Bel}(A) = \sum_{B\in 2^\Theta, B\subseteq A} m(B)
\label{Belg}$$ $$\text{Pl}(A) = \sum_{B\in 2^\Theta, B\cap A\neq\emptyset} m(B)=1- \text{Bel}(\bar{A})
\label{Plg}$$
### The Dempster’s rule of combination
Now let $\text{Bel}_1(.)$ and $\text{Bel}_2(.)$ be two belief functions over the same frame of discernment $\Theta$ and their corresponding bba $m_1(.)$ and $m_2(.)$ provided by two distinct bodies of evidence $\mathcal{B}_1$ and $\mathcal{B}_2$. Then the combined global belief function denoted $\text{Bel}(.)= \text{Bel}_1(.)\oplus \text{Bel}_2(.)$ is obtained by combining the basic belief assignments (called also sometimes [*[information granules]{}*]{} in the literature) $m_1(.)$ and $m_2(.)$ through the following Dempster’s rule of combination $[m_{1}\oplus m_{2}](\emptyset)=0$ and $\forall B\neq\emptyset \in 2^\Theta$, $$[m_{1}\oplus m_{2}](B) =
\frac{\sum_{X\cap Y=B}m_{1}(X)m_{2}(Y)}{1-\sum_{X\cap Y=\emptyset} m_{1}(X) m_{2}(Y)}
\label{eq:DSR}$$
The notation $\sum_{X\cap Y=B}$ represents the sum over all $X, Y \in 2^\Theta$ such that $X\cap Y=B$. The orthogonal sum $m (.)\triangleq [m_{1}\oplus m_{2}](.)$ is considered as a basic belief assignment if and only if the denominator in equation is non-zero. The term $k_{12}\triangleq \sum_{X\cap Y=\emptyset} m_{1}(X) m_{2}(Y)$ is called degree of conflict between the sources $\mathcal{B}_1$ and $\mathcal{B}_2$. When $k_{12}=1$, the orthogonal sum $m (.)$ does not exist and the bodies of evidences $\mathcal{B}_1$ and $\mathcal{B}_2$ are said to be in [*[full contradiction]{}*]{}. Such a case can arise when there exists $A \subset \Theta$ such that $\text{Bel}_1(A) =1$ and $\text{Bel}_2(\bar{A}) = 1$. Same kind of trouble can occur also with the [*[Optimal Bayesian Fusion Rule]{}*]{} (OBFR) [@Dezert_2001a; @Dezert_2001b].\
The DST is attractive for the [*[Data Fusion community]{}*]{} because it gives a nice mathematical model for ignorance and it includes the Bayesian theory as a special case [@Shafer_1976] (p. 4). Although very appealing, the DST presents some weaknesses and limitations because of its model itself, the theoretical justification of the Dempster’s rule of combination but also because of our confidence to trust the result of Dempster’s rule of combination when the conflict becomes important between sources ($k_{12} \nearrow 1$).\
### Alternatives of the Dempster’s rule of combination in the DST framework
The Dempster’s rule of combination has however been [*[a posteriori]{}*]{} justified by the Smet’s axiomatic of the Transferable Belief Model (TBM) in [@Smets_1994]. But we must also emphasize here that an infinite number of possible rules of combinations can be built from the Shafer’s model following ideas initially proposed by Lefèvre, Colot and Vannoorenberghe in [@Lefevre_2002] and corrected here as follows:
- one first has to compute $m(\emptyset)$ by $$m(\emptyset) \triangleq \sum_{A\cap B =\emptyset}m_1(A)m_2(B)$$
- then one redistributes $m(\emptyset)$ on all $(A\neq\emptyset)\subseteq \Theta$ with some given coefficients $w_m(A)\in[0,1]$ such that $\sum_{A\subseteq \Theta} w_m(A)=1$ according to $$\begin{cases}
w_m(\emptyset)m(\emptyset) \rightarrow m(\emptyset)\\
m(A) + w_m(A)m(\emptyset) \rightarrow m(A), \forall A\neq\emptyset
\end{cases}
\label{eq:CEV}$$
The particular choice of the set of coefficients $w_m(.)$ provides a particular rule of combination. Actually there exists an infinite number of possible rules of combination. Some rules can be better justified than others depending on their ability to or not to preserve the associativity and commutativity properties of the combination. It can be easily shown in [@Lefevre_2002] that such general procedure provides all existing rules developed in the literature from the Shafer’s model as alternative to the primeval Dempster’s rule of combination depending on the choice of coefficients $w(A)$. As examples:
- the Dempster’s rule of combination can be obtained from by choosing [@Lefevre_2002] $\forall A\neq\emptyset$ $$w_m(\emptyset)=0\qquad\text{and}\qquad w_m(A)=m(A)/(1-m(\emptyset))$$
- the Yager’s rule of combination is obtained by choosing [@Yager_1987; @Lefevre_2002] $$w_m(\Theta)=1$$
- the Smets’ rule of combination [@Smets_1990; @Lefevre_2002] is obtained by accepting the possibility to deal with bba such that $m(\emptyset)>0$ and thus by choosing $$w_m(\emptyset)=1$$
- with the Lefévre and al. formalism [@Lefevre_2002] and when $m(\emptyset)>0$, the Dubois and Prade’s rule of combination [@Dubois_1998; @Lefevre_2002] is obtained by choosing $$\forall A\subseteq \mathcal{P}, \qquad w_m(A)=\frac{\sum_{\overset{A_1,A_2\mid A_1\cup A_2 = A}{A_1\cap A_2=\emptyset}}m^{\star}}{m(\emptyset}$$ where $m^{\star}\triangleq m_1(A_1)m_2(A_2)$ corresponds to the partial conflicting mass which is assigned to $A_1\cup A_2$ and where $\mathcal{P}$ is the set of all subsets of $2^\Theta$ on which the conflicting mass is distributed defined by
$$\mathcal{P}\triangleq\{A\in 2^\Theta \mid \exists A_1\in \mathcal{K}(m_1), \exists A_2\in \mathcal{K}(m_2), A_1\cup A_2=A \; \text{and} \; A_1\cap A_2=\emptyset \}$$
The computation of the weighting factors $w_m(A)$ of the Dubois and Prade’s rule of combination does not depend only on propositions they are associated with, but also on belief mass functions which have cause the partial conflicts. Thus the belief mass functions leading to the conflict allow to compute that part of conflicting mass which must be assigned to the subsets of $\mathcal{P}$ [@Lefevre_2002]. The Yager’s rule coincides with the Dubois and Prade’s rule of combination when choosing $\mathcal{P}=\{\Theta\}$.
### DSm hybrid rule is not equivalent to the Dempster’s rule of combination
In its essence, the DSm hybrid rule of combination is close to the Dubois and Prade’s rule of combination but more general and precise because it works on $D^\Theta \supset 2^\Theta$ and allows us to include all possible exclusivity and non-existential constraints for the model one has to work with. The advantage of using the DSm hybrid rule is that it does not require the calculation of weighting factors neither the normalization. The DSm hybrid rule of combination is definitely not equivalent to the Dempster’s rule of combination as one can easily prove in the following very simple example:\
Let consider $\Theta=\{\theta_1,\theta_2\}$ and the two sources in full contradiction providing the following basic belief assignments $$m_1(\theta_1)=1 \qquad m_1(\theta_2)=0$$ $$m_2(\theta_1)=0 \qquad m_2(\theta_2)=1$$
Using the classic DSm rule of combination working with the free DSm model $\mathcal{M}^f$, one gets
$$m_{\mathcal{M}^f}(\theta_1)=0 \qquad m_{\mathcal{M}^f}(\theta_2)=0 \qquad m_{\mathcal{M}^f}(\theta_1 \cap \theta_2)=1 \qquad m_{\mathcal{M}^f}(\theta_1 \cup \theta_2)=0$$
If one forces $\theta_1$ and $\theta_2$ to be exclusive to work with the Shafer’s model $\mathcal{M}^0$, then the Dempster’s rule of combination can not be applied in this limit case because of the full contradiction of the two sources of information. One gets the undefined operation 0/0. But the DSm hybrid rule can be applied in such limit case because it transfers the mass of this empty set ($\theta_1 \cap \theta_2\equiv \emptyset$ because of the choice of the model $\mathcal{M}^0$) to non-empty set(s), and one gets:
$$m_{\mathcal{M}^0}(\theta_1)=0 \qquad m_{\mathcal{M}^0}(\theta_2)=0 \qquad m_{\mathcal{M}^0}(\theta_1 \cap \theta_2)=0 \qquad m_{\mathcal{M}^0}(\theta_1 \cup \theta_2)=1$$
This result is coherent in this very simple case with the Yager’s and Dubois-Prade’s rule of combination.\
Now let examine the behavior of the numerical result when introducing a small variation $\epsilon > 0$ on initial basic belief assignments $m_1(.)$ and $m_2(.)$ as follows:
$$m_1(\theta_1)=1-\epsilon \qquad m_1(\theta_2)=\epsilon\qquad \text{and}\qquad m_2(\theta_1)=\epsilon \qquad m_2(\theta_2)=1-\epsilon$$
As shown on figure 1, $\lim_{\epsilon\rightarrow 0} m_{DS}(.)$, where $m_{DS}(.)$ is the result obtained from the Dempster’s rule of combination, is given by
$$m_{DS}(\theta_1)=0.5 \qquad m_{DS}(\theta_2)=0.5\qquad m_{DS}(\theta_1\cap\theta_2)=0 \qquad m_{DS}(\theta_1\cup\theta_2)=0$$
This result is very questionable because it assigns same belief on $\theta_1$ and $\theta_2$ which is more informational than to assign all the belief to the total ignorance. The assignment of the belief to the total ignorance appears to be more justified from our point of view because it properly reflects the almost total contradiction between the two sources and in such cases, it seems legitimist that the information can be drawn from the fusion. When we apply the DSm hybrid rule of combination (using the Shafer’s model $\mathcal{M}^0$), one gets the expected belief assignment on the total ignorance, i.e. $m_{\mathcal{M}^0}(\theta_1 \cup \theta_2)=1$. The figure below shows the evolution of bba on $\theta_1$, $\theta_2$ and $\theta_1\cup\theta_2$ with $\epsilon$ obtained with classical Dempster’s rule and DSm hybrid rule based on Shafer’s model $\mathcal{M}^0$ (i.e. $\theta_1 \cap \theta_2\overset{\mathcal{M}_0}{\equiv} \emptyset$) .
![Comparison of Dempster’s rule with the DSm hybrid rule on $\Theta=\{\theta_1,\theta_2\}$[]{data-label="fig:2"}](Comparison.eps){width="17cm" height="6cm"}
Dynamic fusion
==============
The DSm hybrid rule of combination presented in this paper has been developed for [*[static]{}*]{} problems/models, but is also directly applicable for easily handling [*[dynamic fusion]{}*]{} problems in real time as well, since at each temporal change of the models, one can still apply such hybrid rule. If $D^\Theta$ changes, due to the dynamicity of the frame $\Theta$, from time $t_l$ to time $t_{l+1}$, i.e. some of its elements which at time $t_l$ were not empty become (or are proved) empty at time $t_{l+1}$, or vice versa: if new elements, empty at time $t_l$, arise non-empty at time $t_{l+1}$, this DSm hybrid rule can be applied again at each change. If $\Theta$ tests the same but its set of focal (i.e. non-empty) elements of $D^\Theta$ increases, then again apply the DSm hybrid rule.
Example 1
---------
Let’s consider the testimony fusion problem[^4] with the frame $$\Theta(t_l)\triangleq\{\theta_1\equiv\text{young},\theta_2\equiv\text{old},\theta_3\equiv\text{white hairs}\}$$ with the following two basic belief assignments
$$m_1(\theta_1)=0.5 \qquad m_1(\theta_3)=0.5$$ $$m_2(\theta_2)=0.5 \qquad m_2(\theta_3)=0.5$$
By applying the classical DSm fusion rule, one then gets
$$m_{\mathcal{M}^f(\Theta(t_l))}(\theta_1\cap\theta_2)=0.25 \qquad m_{\mathcal{M}^f(\Theta(t_l))}(\theta_1\cap\theta_3)=0.25 \qquad
m_{\mathcal{M}^f(\Theta(t_l))}(\theta_2\cap\theta_3)=0.25 \qquad
m_{\mathcal{M}^f(\Theta(t_l))}(\theta_3)=0.25$$
Suppose now that at time $t_{l+1}$, one knows that young people don’t have white hairs (i.e $\theta_1\cap\theta_3\equiv \emptyset$). How can we update the previous fusion result with this new information on the model of the problem ? We solve it with the DSm hybrid rule, which transfers the mass of the empty sets (imposed by the constraints on the new model $\mathcal{M}$ available at time $t_{l+1}$) to the non-empty sets of $D^\Theta$, going on the track of the DSm classic rule. Using the DSm hybrid rule with the constraint $\theta_1\cap\theta_3\equiv \emptyset$, one then gets:
$$m_{\mathcal{M}}(\theta_1\cap\theta_2)=0.25 \qquad
m_{\mathcal{M}}(\theta_2\cap\theta_3)=0.25 \qquad
m_{\mathcal{M}}(\theta_3)=0.25\qquad$$ and the mass $m_{\mathcal{M}}(\theta_1\cap\theta_3)=0$, because $\theta_1\cap\theta_3=\{\text{young}\}\cap\{\text{white hairs}\}\overset{\mathcal{M}}{\equiv}\emptyset $ and its previous mass $m_{\mathcal{M}^f(\Theta(t_l))}(\theta_1\cap\theta_3)=0.25$ is transferred to $m_{\mathcal{M}}(\theta_1\cup\theta_3)=0.25$ by the DSm hybrid rule.
Example 2
---------
Let $\Theta(t_l)=\{\theta_1,\theta_2,\ldots,\theta_n\}$ be a list of suspects and let consider two observers who eyewitness the scene of plunder at a museum in Bagdad and who testify to the radio and TV the identities of thieves using the basic beliefs assignments $m_1(.)$ and $m_2(.)$ defined on $D^{\Theta(t_l)}$, where $t_l$ represents the time of the observation. Afterwards, at time $t_{l+1}$, one finds out that one suspect, among this list $\Theta(t_l)$, say $\theta_i$, could not be a suspect because he was on duty in another place, evidence which was certainly confirmed. Therefore he has to be taken off the suspect list $\Theta(t_l)$, and a new frame of discernment is resulting $\Theta(t_{l+1})$. If this one changes again, one applies again the DSm hybrid of combining of evidences, and so on. This is a typically dynamical example where models change with time and where one needs to adapt fusion results with current model over time. In the meantime, one can also take into account new observations/testimonies in the DSm hybrid fusion rule as soon as they become available to the fusion system. If $\Theta$ and $D^\Theta$ diminish (i.e. some of their elements are proven to be empty sets) from time $t_l$ to time $t_{l+1}$, then one applies the DSm hybrid rule in order to transfer the masses of empty sets to the non-empty sets (in the DSm classic rule’s way) getting an updated basic belief assignment $m_{t_{l+1}|t_l}(.)$. Contrarily, if $\Theta$ and $D^\Theta$ increase (i.e. new elements arise in $\Theta$, and/or new elements in $D^\Theta$ are proven different from the empty set and as a consequence a basic belief assignment for them is required), then new masses (from the same or from the other sources of information) are needed to describe these new elements, and again one combines them using the DSm hybrid rule.
Example 3
---------
Let consider a fusion problem at time $t_l$ characterized by the frame $\Theta(t_l)\triangleq \{\theta_1,\theta_2\}$ and two independent sources of information providing the basic belief assignments $m_1(.)$ and $m_2(.)$ over $D^{\Theta(t_l)}$ and assume that at time $t_{l+1}$ a new hypothesis $\theta_3$ is introduced into the previous frame $\Theta(t_l)$ and a third source of evidence available at time $t_{l+1}$ provides its own basic belief assignment $m_3(.)$ over $D^{\Theta(t_{l+1})}$ where $$\Theta(t_{l+1})\triangleq \{ \Theta(t_l),\theta_3\} \equiv \{ \theta_1,\theta_2,\theta_3\}$$
To solve such kind of dynamical fusion problems, we just use the classical DSm fusion rule as follows:
- combine $m_1(.)$ and $m_2(.)$ at time $t_l$ using classical DSm fusion rule to get $m_{12}(.)= [m_1\oplus m_2](.)$ over $D^{\Theta(t_l)}$
- because $D^{\Theta(t_l)} \subset D^{\Theta(t_{l+1})}$, $m_{12}(.)$ assigns the combined basic belief on a subset of $D^{\Theta(t_{l+1})}$, it is still directly possible to combine $m_{12}(.)$ with $m_3(.)$ at time $t_{l+1}$ by the classical DSm fusion rule to get the final result $m_{123}(.)$ over $D^{\Theta(t_{l+1})}$ given by $$m_{t_{l+1}}(.)\triangleq m_{123}(.)= [m_{12}\oplus m_3](.) = [(m_1\oplus m_2)\oplus m_3](.) \equiv [m_1\oplus m_2\oplus m_3](.)$$
- eventually apply DSm hybrid rule if some integrity constraints have to be taken into account in the model $\mathcal{M}$ of the problem
This method can be directly generalized to any number of sources of evidences and, in theory, to any structures/dimension of the frames $\Theta(t_{l})$, $\Theta(t_{l+1})$, ... In practice however, due to the huge number of elements of hyper-power sets, the dimension of the frames $\Theta(t_{l})$, $\Theta(t_{l+1})$, $\dots$ must be not too large. This practical limitation depends on the computer resources available for the real-time processing. Specific suboptimal implementations of DSm rule will have to be developed to deal with fusion problems of large dimension.\
It is also important to point out here that DSmT can easily deal, not only with dynamical fusion problems but with decentralized fusion problems as well working on non exhaustive frames. For example, let consider a set of two independent sources of information providing the basic belief assignments $m_1(.)$ and $m_2(.)$ over $D^{\Theta_{12}(t_l)=\{\theta_1,\theta_2\}}$ and another group of three independent sources of information providing the basic belief assignments $m_3(.)$, $m_4(.)$ and $m_5(.)$ over $D^{\Theta_{345}(t_l)=\{\theta_3,\theta_4,\theta_5,\theta_6\}}$, then it is still possible to combine all information in a decentralized manner as follows:
- combine $m_1(.)$ and $m_2(.)$ at time $t_l$ using classical DSm fusion rule to get $m_{12}(.)= [m_1\oplus m_2](.)$ over $D^{\Theta_{12}(t_l)}$.
- combine $m_3(.)$, $m_4(.)$ and $m_5(.)$ at time $t_l$ using classical DSm fusion rule to get $m_{345}(.)= [m_3\oplus m_4\oplus m_5](.)$ over $D^{\Theta_{345}(t_l)}$.
- consider now the global frame $\Theta(t_l)\triangleq \{\Theta_{12}(t_l) ,\Theta_{345}(t_l)\}$.
- eventually apply DSm hybrid rule if some integrity constraints have to be taken into account in the model $\mathcal{M}$ of the problem.
Note that this [*[static]{}*]{} decentralized fusion can also be extended to decentralized dynamical fusion also by mixing two previous approaches.\
One can even combine all five masses together by extending the vectors $m_i(.)$, $1\leq i \leq 5$, with null components for the new elements arisen from enlarging $\Theta$ to $\{\theta_1,\theta_2, \theta_3,\theta_4,\theta_5\}$ and correspondingly enlarging $D^\Theta$, and using the DSm hybrid rule for $k=5$. And more general combining the masses of any $k \geq 2$ sources.\
We give now several simple numerical examples for such dynamical fusion problems involving non exclusive frames.
### Example 3.1
Let consider $\Theta(t_l)\triangleq \{\theta_1,\theta_2\}$ and the two following basic belief assignments available at time $t_l$:
$$m_1(\theta_1)=0.1 \qquad m_1(\theta_2)=0.2 \qquad m_1(\theta_1\cup\theta_2)=0.3\qquad m_1(\theta_1\cap\theta_2)=0.4$$ $$m_2(\theta_1)=0.5 \qquad m_2(\theta_2)=0.3 \qquad m_2(\theta_1\cup\theta_2)=0.1\qquad m_2(\theta_1\cap\theta_2)=0.1$$
The classical DSm rule of combination gives
$$m_{12}(\theta_1)=0.21 \qquad m_{12}(\theta_2)=0.17 \qquad m_{12}(\theta_1\cup\theta_2)=0.03\qquad m_{12}(\theta_1\cap\theta_2)=0.59$$
Now let consider at time $t_{l+1}$ the frame $\Theta(t_{l+1})\triangleq \{\theta_1,\theta_2,\theta_3\}$ and a third source of evidence with the following basic belief assignment
$$m_3(\theta_3)=0.4 \qquad m_3(\theta_1\cap\theta_3)=0.3\qquad m_3(\theta_2\cup\theta_3)=0.3$$
Then the final result of the fusion is obtained by combining $m_3(.)$ with $m_{12}(.)$ by the classical DSm rule of combination. One thus obtains: $$\begin{gathered}
m_{123}(\theta_1\cap\theta_2\cap\theta_3)=0.464 \quad m_{123}(\theta_2\cap\theta_3)=0.068 \quad m_{123}(\theta_1\cap\theta_3)=0.156\quad m_{123}((\theta_1\cup \theta_2)\cap\theta_3)=0.012\\
m_{123}(\theta_1\cap\theta_2)=0.177 \quad m_{123}(\theta_1\cap(\theta_2\cup\theta_3))=0.063 \quad m_{123}(\theta_2)=0.051 \quad m_{123}((\theta_1\cap \theta_3)\cup\theta_2)=0.009\end{gathered}$$
### Example 3.2
Let consider $\Theta(t_l)\triangleq \{\theta_1,\theta_2\}$ and the two previous following basic belief assignments $m_1(.)$ and $m_2(.)$ available at time $t_l$. The classical DSm fusion rule gives gives as before
$$m_{12}(\theta_1)=0.21 \qquad m_{12}(\theta_2)=0.17 \qquad m_{12}(\theta_1\cup\theta_2)=0.03\qquad m_{12}(\theta_1\cap\theta_2)=0.59$$
Now let consider at time $t_{l+1}$ the frame $\Theta(t_{l+1})\triangleq \{\theta_1,\theta_2,\theta_3\}$ and the third source of evidence as in previous example with the basic belief assignment
$$m_3(\theta_3)=0.4 \qquad m_3(\theta_1\cap\theta_3)=0.3 \qquad m_3(\theta_2\cup\theta_3)=0.3$$
The final result of the fusion obtained by the classical DSm rule of combination corresponds to the result of the previous example, but suppose now one finds out that the integrity constraint $\theta_3= \emptyset$ holds, which implies also constraints $\theta_1\cap\theta_2\cap\theta_3=\emptyset$, $\theta_1\cap\theta_3=\emptyset$, $\theta_2\cap\theta_3=\emptyset$ and $(\theta_1\cup\theta_2)\cap\theta_3=\emptyset$. This is the DSm hybrid model $\mathcal{M}$ under consideration here. We then have to readjust the mass $m_{123}(.)$ of the previous example by the DSm hybrid rule and one finally gets
$$m_{\mathcal{M}}(\theta_1)=0.147$$ $$m_{\mathcal{M}}(\theta_2)=0.060+0.119=0.179$$ $$m_{\mathcal{M}}(\theta_1\cup\theta_2)=0 + 0 + 0.021=0.021$$ $$m_{\mathcal{M}}(\theta_1\cap\theta_2)=0.240+0.413=0.653$$
Therefore, when we restrain back $\theta_3= \emptyset$ and apply the DSm hybrid rule, we don’t get back the same result (i.e. $m_{\mathcal{M}}(.)\neq m_{12}(.)$) because still remains some information from $m_3(.)$ on $\theta_1$, $\theta_2$, $\theta_1\cup\theta_2$, or $\theta_1\cap\theta_2$, i.e. $m_3(\theta_2)=0.3>0$.
### Example 3.3
Let consider $\Theta(t_l)\triangleq \{\theta_1,\theta_2\}$ and two previous following basic belief assignments $m_1(.)$ and $m_2(.)$ available at time $t_l$. The classical DSm fusion rule gives as before $$m_{12}(\theta_1)=0.21 \qquad m_{12}(\theta_2)=0.17 \qquad m_{12}(\theta_1\cup\theta_2)=0.03\qquad m_{12}(\theta_1\cap\theta_2)=0.59$$
Now let consider at time $t_{l+1}$ the frame $\Theta(t_{l+1})\triangleq \{\theta_1,\theta_2,\theta_3,\theta_4\}$ and another third source of evidence with the following basic belief assignment $$m_3(\theta_3)=0.5 \qquad m_3(\theta_4)=0.3 \qquad m_3(\theta_3\cap\theta_4)=0.1 \qquad m_3(\theta_3\cup\theta_4)=0.1$$ Then, the DSm rule applied at time $t_{l+1}$ provides the following combined belief assignment $$\begin{gathered}
m_{123}(\theta_1\cap\theta_3)=0.105 \quad m_{123}(\theta_1\cap\theta_4)=0.063 \quad m_{123}(\theta_1\cap(\theta_3\cup\theta_4))=0.021 \quad m_{123}(\theta_1\cap\theta_3\cap\theta_4)=0.021\\
m_{123}(\theta_2\cap\theta_3)=0.085 \quad m_{123}(\theta_2\cap\theta_4)=0.051 \quad m_{123}(\theta_2\cap(\theta_3\cup\theta_4))=0.017 \qquad m_{123}(\theta_2\cap\theta_3\cap\theta_4)=0.017 \\
m_{123}(\theta_3\cap(\theta_1\cup\theta_2))=0.015 \quad m_{123}(\theta_4\cap(\theta_1\cup\theta_2))=0.009 \quad
m_{123}((\theta_1\cup\theta_2)\cap(\theta_3\cup\theta_4))=0.003\\
m_{123}((\theta_1\cup\theta_2)\cap(\theta_3\cap\theta_4))=0.003\quad
m_{123}(\theta_1\cap\theta_2\cap\theta_3)=0.295 \quad
m_{123}(\theta_1\cap\theta_2\cap\theta_4)=0.177 \\
m_{123}((\theta_1\cap\theta_2)\cap(\theta_3\cup\theta_4))=0.059 \quad
m_{123}(\theta_1\cap\theta_2\cap\theta_3\cap\theta_4)=0.059\end{gathered}$$
Now suppose at time $t_{l+2}$ one finds out that $\theta_3=\theta_4=\emptyset$, then one applies the DSm hybrid rule after re-adjusting the combined belief mass $m_{123}(.)$ by cummulating the masses of all empty sets. Using the DSm hybrid rule, one finally gets:
$$\begin{aligned}
m_{t_{l+2}}(\theta_1)&=m_{123}(\theta_1)+\{m_{12}(\theta_1)m_3(\theta_3) + m_{12}(\theta_1)m_3(\theta_4) + m_{12}(\theta_1)m_3(\theta_3\cup\theta_4) + m_{12}(\theta_1)m_3(\theta_3\cap\theta_4)\}\\
&=0 + \{(0.21\times 0.5) + (0.21\times 0.3) + (0.21\times 0.1) + (0.21\times 0.1)\} = 0.21\end{aligned}$$
$$\begin{aligned}
m_{t_{l+2}}(\theta_2)&=m_{123}(\theta_2)+\{m_{12}(\theta_2)m_3(\theta_3) + m_{12}(\theta_2)m_3(\theta_4) + m_{12}(\theta_2)m_3(\theta_3\cup\theta_4) + m_{12}(\theta_2)m_3(\theta_3\cap\theta_4)\}\\
&=0 + \{(0.17\times 0.5) + (0.17\times 0.3) + (0.17\times 0.1) + (0.17\times 0.1)\} = 0.17\end{aligned}$$
$$\begin{aligned}
m_{t_{l+2}}(\theta_1\cup\theta_2)&=m_{123}(\theta_1\cup\theta_2)+\{m_{12}(\theta_1\cup\theta_2)m_3(\theta_3) + m_{12}(\theta_1\cup\theta_2)m_3(\theta_4) \\
&\qquad + m_{12}(\theta_1\cup\theta_2)m_3(\theta_3\cup\theta_4) + m_{12}(\theta_1\cup\theta_2)m_3(\theta_3\cap\theta_4)\}\\
&\qquad + \sum_{X_1, X_2 \in \{\theta_3, \theta_4, \theta_3\cup\theta_4, \theta_3\cap\theta_4\} } m_{12}(X_1)m_3(X_2)\\
&=0 + \{(0.03\times 0.5) + (0.03\times 0.3) + (0.03\times 0.1) +( 0.03\times 0.1)\} + \{0\} = 0.03\end{aligned}$$
$$\begin{aligned}
m_{t_{l+2}}(\theta_1\cap\theta_2)&=m_{123}(\theta_1\cap\theta_2)+\{m_{12}(\theta_1\cap\theta_2)m_3(\theta_3) + m_{12}(\theta_1\cap\theta_2)m_3(\theta_4) \\
&\qquad + m_{12}(\theta_1\cap\theta_2)m_3(\theta_3\cup\theta_4) + m_{12}(\theta_1\cap\theta_2)m_3(\theta_3\cap\theta_4)\}\\
&=0 + \{(0.59\times 0.5) + (0.59\times 0.3) + (0.59\times 0.1) + (0.59\times 0.1)\} = 0.59\end{aligned}$$
Thus we get the same result as for $m_{12}(.)$ at time $t_l$, which is normal.\
[**[Remark:]{}**]{} note that if the third source of information don’t assign non-null masses to $\theta_1$, or $\theta_2$ (or to their combinations using $\cup$ or $\cap$ operators), then one obtains the same result at time $t_{l+2}$ as at time $t_l$ as in this example 3.3, i.e. $m_{l+2}(.) = m_l(.)$, when imposing back $\theta_3=\theta_4=\emptyset$. But, if the third source of information assigns non-null masses to either $\theta_1$, or $\theta_2$, or to some of their combinations $\theta_1\cup \theta_2$ or $\theta_1\cap \theta_2$, then when one returns from 4 singletons to 2 singletons for $\Theta$, replacing $\theta_3=\theta_4=\emptyset$ and using the DSm hybrid rule, the fusion results at time $t_{l+2}$ is different from that at time $t_l$, and this is normal because some information/mass is left from the third source and is now fusioned with that of the previous sources (as in example 3.2 or in next example 3.4).\
In general, let’s suppose that the fusion of $k \geq 2$ masses provided by the sources $\mathcal{B}_1$, $\mathcal{B}_2$, ..., $\mathcal{B}_k$ has been done at time $t_l$ on $\Theta(t_l) = \{\theta_1, \theta_2, ..., \theta_n\}$. At time $t_{l+1}$ new non-empty elements $\theta_{n+1}$, $\theta_{n+2}$, $\ldots$, $\theta_{n+m}$ appear, $m\geq 1$, thus $$\Theta(t_{l+1}) = \{\theta_1, \theta_2, ..., \theta_n,\theta_{n+1},\theta_{n+2}, \ldots, \theta_{n+m}\}$$ and of course one or more sources (i.e. bodies of evidences) $\mathcal{B}_{k+1}$, $\ldots$, $\mathcal{B}_{k+l}$, where $l\geq 1$, appear to assign masses to these new elements.
- If all these new sources $\mathcal{B}_{k+1}$, $\ldots$, $\mathcal{B}_{k+l}$ assign null masses to all elements from $D^{\Theta(t_{l+1})}$ which contain in their structure/composition at least one of the singletons $\theta_1$, $\theta_2$, $\ldots$, $\theta_n$, then at time $t_{l+2}$ if one sets back the constraints that $\theta_{n+1}=\theta_{n+2}=\ldots = \theta_{n+m}=\emptyset$, then using the DSm hybrid rule, one obtains the same result as at time $t_l$, i.e. $m_{l+2}(.) = m_l(.)$.
- Otherwise, the fusion at time $t_{l+2}$ will be different from the fusion at time $t_l$ because still remains some information/mass from sources $\mathcal{B}_{k+1}$, $\ldots$, $\mathcal{B}_{k+l}$ on singletons $\theta_1$, $\theta_2$, $\ldots$, $\theta_n$ or on some elements from $D^{\Theta(t_l)}$ which contain at least one of such singletons, information/mass which fusions with the previous sources.
### Example 3.4
Let consider $\Theta(t_l)\triangleq \{\theta_1,\theta_2\}$ and the two following basic belief assignments available at time $t_l$: $$m_1(\theta_1)=0.6 \qquad m_1(\theta_2)=0.4 \qquad\text{and}\qquad m_2(\theta_1)=0.7 \qquad m_2(\theta_2)=0.3$$ The classical DSm rule of combination gives $m_{12}(\theta_1)=0.42$, $m_{12}(\theta_2)=0.12$ and $m_{12}(\theta_1\cap\theta_2)=0.46$. Now let consider at time $t_{l+1}$ the frame $\Theta(t_{l+1})\triangleq \{\theta_1,\theta_2,\theta_3\}$ and a third source of evidence with the following basic belief assignment $m_3(\theta_1)=0.5$, $m_3(\theta_2)=0.2$ and $m_3(\theta_3)=0.3$. Then the final result obtained from the classical DSm rule of combination is still as before $$\begin{gathered}
m_{123}(\theta_1)=0.210 \quad m_{123}(\theta_2)=0.024 \quad m_{123}(\theta_1\cap\theta_2)=0.466 \quad m_{123}(\theta_1\cap\theta_3)=0.126\\
m_{123}(\theta_2\cap\theta_3)=0.036 \quad
m_{123}(\theta_1\cap\theta_2\cap\theta_3)=0.138\end{gathered}$$
Suppose now one finds out that the integrity constraint $\theta_1\cap\theta_3= \emptyset$ which also implies $\theta_1\cap\theta_2\cap\theta_3=\emptyset$. This is the DSm hybrid model $\mathcal{M}$ under consideration. By applying the DSm hybrid fusion rule, one forces $m_{\mathcal{M}}(\theta_1\cap\theta_3)=0$ and $m_{\mathcal{M}}(\theta_1\cap\theta_2\cap\theta_3)=0$ and we transfer $m_{123}(\theta_1\cap\theta_2\cap\theta_3)=0.138$ towards $m_{\mathcal{M}}((\theta_1\cap\theta_2)\cup\theta_3)$ and the mass $m_{123}(\theta_1\cap\theta_3)=0.126$ has to be transferred towards $m_{\mathcal{M}}(\theta_1\cup\theta_3)$. One then gets finally $$\begin{gathered}
m_{\mathcal{M}}(\theta_1)=0.210 \quad m_{\mathcal{M}}(\theta_2)=0.024 \quad m_{\mathcal{M}}(\theta_1\cap\theta_2)=0.466 \quad
m_{\mathcal{M}}(\theta_2\cap\theta_3)=0.036 \\
m_{\mathcal{M}}((\theta_1\cap\theta_2)\cup\theta_3)=0.138 \quad
m_{\mathcal{M}}(\theta_1\cup\theta_3)=0.126\end{gathered}$$
### Example 3.5
Let consider $\Theta(t_l)\triangleq \{\theta_1,\theta_2\}$ and the two previous basic belief assignments available at time $t_l$ as in previous example, i.e. $$m_1(\theta_1)=0.6 \qquad m_1(\theta_2)=0.4 \qquad\text{and}\qquad m_2(\theta_1)=0.7 \qquad m_2(\theta_2)=0.3$$ The classical DSm rule of combination gives $$m_{12}(\theta_1)=0.42 \qquad m_{12}(\theta_2)=0.12 \quad m_{12}(\theta_1\cap\theta_2)=0.46$$ Now let consider at time $t_{l+1}$ the frame $\Theta(t_{l+1})\triangleq \{\theta_1,\theta_2,\theta_3\}$ and a third source of evidence with the following basic belief assignment $$m_3(\theta_1)=0.5 \qquad m_3(\theta_2)=0.2 \qquad m_3(\theta_3)=0.3$$
Then the final result of the fusion is obtained by combining $m_3(.)$ with $m_{12}(.)$ by the classical DSm rule of combination. One thus obtains now $$\begin{gathered}
m_{123}(\theta_1)=0.210 \quad m_{123}(\theta_2)=0.024 \quad m_{123}(\theta_1\cap\theta_2)=0.466 \quad m_{123}(\theta_1\cap\theta_3)=0.126\\ m_{123}(\theta_2\cap\theta_3)=0.036 \quad m_{123}(\theta_1\cap\theta_2\cap\theta_3)=0.138\end{gathered}$$
But suppose one finds out that the integrity constraint is now $\theta_3=\emptyset$ which implies necessarily also $\theta_1\cap\theta_3=\theta_2\cap\theta_3=\theta_1\cap\theta_2\cap\theta_3\equiv\emptyset$ and $(\theta_1\cup\theta_2)\cap\theta_3=\emptyset$ (this is our new DSm hybrid model $\mathcal{M}$ under consideration in this example). By applying the DSm hybrid fusion rule, one gets finally the non-null masses $$m_{\mathcal{M}}(\theta_1)=0.336 \qquad m_{\mathcal{M}}(\theta_2)=0.060\qquad
m_{\mathcal{M}}(\theta_1\cap\theta_2)=0.604$$
### Example 3.6
Let consider $\Theta(t_l)\triangleq \{\theta_1,\theta_2,\theta_3,\theta_4\}$ and the following basic belief assignments available at time $t_l$ : $$\begin{cases}
m_1(\theta_1)=0.5 \qquad m_1(\theta_2)=0.4 \qquad m_1(\theta_1\cap\theta_2)=0.1\\
m_2(\theta_1)=0.3 \qquad m_2(\theta_2)=0.2 \qquad m_2(\theta_1\cap\theta_3)=0.1\qquad m_2(\theta_4)=0.4
\end{cases}$$ The classical DSm rule of combination gives $$m_{12}(\theta_1)=0.15 \qquad m_{12}(\theta_2)=0.08 \qquad m_{12}(\theta_1\cap\theta_2)=0.27 \qquad m_{12}(\theta_1\cap\theta_3)=0.05\qquad m_{12}(\theta_1\cap\theta_4)=0.20$$ $$m_{12}(\theta_2\cap\theta_4)=0.16 \qquad m_{12}(\theta_1\cap\theta_2\cap\theta_3)=0.05\qquad m_{12}(\theta_1\cap\theta_2\cap\theta_4)=0.04$$ Now assume that at time $t_{l+1}$ one finds out that $\theta_1\cap\theta_2\overset{\mathcal{M}}{\equiv}\theta_1\cap\theta_3\overset{\mathcal{M}}{\equiv}\emptyset$. Using the DSm hybrid rule, one gets: $$\begin{cases}
m_{\mathcal{M}}(\theta_1\cap\theta_2)=m_{\mathcal{M}}(\theta_1\cap\theta_3)=m_{\mathcal{M}}(\theta_1\cap\theta_2\cap\theta_3)=m_{\mathcal{M}}(\theta_1\cap\theta_2\cap\theta_4)=0\\
m_{\mathcal{M}}(\theta_1)=m_{12}(\theta_1)+m_2(\theta_1)m_1(\theta_1\cap\theta_2)+m_1(\theta_1)m_2(\theta_1\cap\theta_3)=0.15+0.03+0.05=0.23\\
m_{\mathcal{M}}(\theta_2)=m_{12}(\theta_2)+m_2(\theta_2)m_1(\theta_1\cap\theta_2)+m_1(\theta_2)m_2(\theta_1\cap\theta_3)=0.08+0.02+0.04=0.14\\
m_{\mathcal{M}}(\theta_4)=m_{12}(\theta_4)+m_1(\theta_1\cap\theta_2)m_2(\theta_4)=0+0.04=0.04\\
m_{\mathcal{M}}(\theta_1\cap\theta_4)=m_{12}(\theta_1\cap\theta_4)=0.20\\
m_{\mathcal{M}}(\theta_2\cap\theta_4)=m_{12}(\theta_2\cap\theta_4)=0.16\\
m_{\mathcal{M}}(\theta_1\cup\theta_2)=m_{12}(\theta_1\cup\theta_2)+m_1(\theta_1)m_2(\theta_2)+m_2(\theta_1)m_1(\theta_2)+m_1(\theta_1\cap\theta_2)m_2(\theta_1\cap\theta_2)=
0.22\\
m_{\mathcal{M}}(\theta_1\cup\theta_2\cup\theta_3)=m_{12}(\theta_1\cup\theta_2\cup\theta_3)+m_1(\theta_1\cap\theta_2)m_2(\theta_1\cap\theta_3)+m_2(\theta_1\cap\theta_2)m_1(\theta_1\cap\theta_3)\\
\qquad\qquad\qquad\qquad\quad+m_1(\theta_1\cap\theta_2\cap\theta_3)m_2(\theta_1\cap\theta_2\cap\theta_3)=0.01
\end{cases}$$
### Example 3.7
Let consider $\Theta(t_l)\triangleq \{\theta_1,\theta_2,\theta_3,\theta_4\}$ and the following basic belief assignments available at time $t_l$ : $$\begin{cases}
m_1(\theta_1)=0.2 \qquad m_1(\theta_2)=0.4 \qquad m_1(\theta_1\cap\theta_2)=0.1 \qquad m_1(\theta_1\cap\theta_3)=0.2\qquad m_1(\theta_4)=0.1\\
m_2(\theta_1)=0.1 \qquad m_2(\theta_2)=0.3 \qquad m_2(\theta_1\cap\theta_2)=0.2 \qquad m_2(\theta_1\cap\theta_3)=0.1\qquad m_2(\theta_4)=0.3
\end{cases}$$ The classical DSm rule of combination gives $$m_{12}(\theta_1)=0.02 \qquad m_{12}(\theta_2)=0.12 \qquad m_{12}(\theta_1\cap\theta_2)=0.28 \qquad m_{12}(\theta_1\cap\theta_3)=0.06\qquad m_{12}(\theta_4)= 0.03$$ $$m_{12}(\theta_1\cap\theta_4)=0.07\qquad m_{12}(\theta_2\cap\theta_4)=0.15 \qquad m_{12}(\theta_1\cap\theta_2\cap\theta_3)=0.15$$ $$m_{12}(\theta_1\cap\theta_2\cap\theta_4)=0.05\qquad m_{12}(\theta_1\cap\theta_3\cap\theta_4)=0.07$$ Now assume that at time $t_{l+1}$ one finds out that $\theta_1\cap\theta_2\overset{\mathcal{M}}{\equiv}\theta_1\cap\theta_3\overset{\mathcal{M}}{\equiv}\emptyset$. Using the DSm hybrid rule, one gets: $$\begin{cases}
m_{\mathcal{M}}(\theta_1\cap\theta_2)=m_{\mathcal{M}}(\theta_1\cap\theta_3)=m_{\mathcal{M}}(\theta_1\cap\theta_2\cap\theta_3)=m_{\mathcal{M}}(\theta_1\cap\theta_2\cap\theta_4)=0\\
m_{\mathcal{M}}(\theta_1)=m_{12}(\theta_1)+m_1(\theta_1)m_2(\theta_1\cap\theta_2)+m_2(\theta_1)m_1(\theta_1\cap\theta_2)+m_1(\theta_1)m_2(\theta_1\cap\theta_3)+m_2(\theta_1)m_1(\theta_1\cap\theta_3)=
0.11\\
m_{\mathcal{M}}(\theta_2)=m_{12}(\theta_2)+m_1(\theta_2)m_2(\theta_1\cap\theta_2)+m_2(\theta_2)m_1(\theta_1\cap\theta_2)+m_1(\theta_2)m_2(\theta_1\cap\theta_3)+m_2(\theta_2)m_1(\theta_1\cap\theta_3)=
0.33\\
m_{\mathcal{M}}(\theta_4)=m_{12}(\theta_4)+m_1(\theta_4)m_2(\theta_1\cap\theta_2)+m_2(\theta_4)m_1(\theta_1\cap\theta_2)+m_1(\theta_4)m_2(\theta_1\cap\theta_3)+m_2(\theta_4)m_1(\theta_1\cap\theta_3)=
0.15\\
m_{\mathcal{M}}(\theta_1\cap\theta_4)=m_{12}(\theta_1\cap\theta_4)=0.07\\
m_{\mathcal{M}}(\theta_2\cap\theta_4)=m_{12}(\theta_2\cap\theta_4)=0.15\\
m_{\mathcal{M}}(\theta_1\cup\theta_2)=m_{12}(\theta_1\cup\theta_2)+m_1(\theta_1\cap\theta_2)m_2(\theta_1\cap\theta_2)+m_1(\theta_1)m_2(\theta_2)+m_2(\theta_1)m_1(\theta_2)=
0.12\\
m_{\mathcal{M}}(\theta_1\cup\theta_3)=m_{12}(\theta_1\cup\theta_3)+m_1(\theta_1\cap\theta_3)m_2(\theta_1\cap\theta_3)=
0.02\\
m_{\mathcal{M}}(\theta_1\cup\theta_2\cup\theta_3)=m_{12}(\theta_1\cup\theta_2\cup\theta_3)+m_1(\theta_1\cap\theta_2)m_2(\theta_1\cap\theta_3)+m_2(\theta_1\cap\theta_2)m_1(\theta_1\cap\theta_3)=
0.05
\end{cases}$$
Bayesian mixture of DSm hybrid models
=====================================
In the preceding, one has first shown how to combine generalized basic belief assignments provided by $k\geq 2$ independent sources of information with the general DSm hybrid rule of combination which deals with all possible kinds of constraints introduced by the hybrid model of the problem. This approach implicitly assumes that one knows/trusts with certainty that the model $\mathcal{M}$ (usually a DSm hybrid model) of the problem is valid and corresponds to the true model. In some complex fusion problems however (static or dynamic ones), one may have some doubts about the validity of the model $\mathcal{M}$ on which is based the fusion because of the nature and evolution of elements of the frame $\Theta$. In such situations, we propose to consider a set of exclusive and exhaustive models $\{\mathcal{M}_1,\mathcal{M}_2,\ldots,\mathcal{M}_K\}$ with some probabilities $\{P\{\mathcal{M}_1\},P\{\mathcal{M}_2\},\ldots,P\{\mathcal{M}_K\}\}$. We don’t go here deeper on the justification/acquisition of such probabilities because this is highly dependent on the nature of the fusion problem under consideration. We just assume here that such probabilities are available at any given time $t_l$ when the fusion has to be done. We propose then to use the Bayesian mixture of combined masses $m_{\mathcal{M}_i(\Theta)}(.)$ $i=1,\ldots,K$ to obtain the final result : $$\forall A\in D^\Theta, \qquad m_{\mathcal{M}_1,\ldots,\mathcal{M}_K}(A)=\sum_{i=1}^K P\{\mathcal{M}_i\} m_{\mathcal{M}_i(\Theta)}(A)$$
Conclusion
===========
In this paper we have extended the DSmT and the classical DSm rule of combination to the case of hybrid models for the frame of discernment involved in many complex fusion problems. The free-DSm model (which assumes that none of the elements of the frame is refinable) can be interpreted as the opposite of the Shafer’s model (which assumes that all elements of the frame are truly exclusive) on which is based the mathematical theory of evidence (Dempster-Shafer Theory - DST). Between these two extreme models, there exists actually many possible hybrid models for the frames of discernment depending on the real intrinsic nature of elements of the fusion problem under consideration. For real problems, some elements of the frame of discernment can appear to be truly exclusive whereas some others cannot be considered as fully discernable or refinable. This present research work proposes a new DSm hybrid rule of combination for hybrid-models based on the DSmT. The DSm hybrid rule works in any model and is involved in calculation of mass fusion of any number of sources of information, no matter how big is the conflict/paradoxism of sources, and on any frame of discernment (exhaustive or non-exhaustive, with elements which may be exclusive or non-exclusive or both). This is an important rule since does not require the calculation of weighting factors, neither normalization as other rules do, and the transfer of empty-sets’ masses to non-empty sets masses is naturally done following the DSm network architecture which is derived from the DSm classic rule. DSmT together with DSm hybrid rule appears from now on to be a new alternative to classical approaches and to existing combination rules and to be very promising for the development of future complex (uncertain/incomplete/paradoxical/dynamical) information fusion systems.
[99]{} Dawid A.P., *Conditional Independence in statistical theory*, Journal of Royal Statistical Society, Serie B., Vol. 41, pp. 1-31, 1979. Dawid A.P., *Conditional Independence*, 14th Conf. on Uncertainty and Artificial Intelligence, USA, 1998. Dawid A.P., *Conditional Independence*, In Encyclopedia of Statistical Science (Update) Volume 3, Wiley, New York, 1999. Dezert J., *Optimal Bayesian fusion of multiple unreliable classifiers*, Proc. of Fusion 2001 Conf., Montréal, Canada, Aug. 8-11, 2001. Dezert J., *Combination of paradoxical sources of information within the Neutrosophic framework*, Proceedings of the First International Conference on Neutrosophics, pp. 22–46., Univ. of New Mexico, Gallup Campus, NM, USA, 1-3 Dec. 2001, Xiquan, Phoenix, 2002. Dezert J., *Foundations for a new theory of plausible and paradoxical reasoning*, Information & Security, An international Journal, edited by Prof. Tzv. Semerdjiev, CLPP, Bulg. Acad. of Sciences, Vol. 9, 2002. Dezert J., *Fondations pour une nouvelle théorie du raisonnement plausible et paradoxal*, ONERA Tech. Rep. 1/06769 DTIM, Jan. 2003. Dezert J., Smarandache F., *Partial ordering of hyper-power sets and matrix representation of belief functions within DSmT*, Proc. of Fusion 2003 Conf., Cairns, Australia, July 8-11, 2003. Dezert J., Smarandache F., *On the generation of hyper-power sets for the DSmT*, Proceedings of the 6th International Conference on Information Fusion, Cairns, Australia, July 8-11, 2003. Dubois D., Prade H., *Representation and combination of uncertainty with belief functions and possibility measures*, Computational Intelligence, Vol. 4, pp. 244-264, 1998. Fonck P., *Conditional Independence in Possibility Theory*, Uncertainty and Artificial Intelligence, pp. 221-226, 1994. Lefèvre E., Colot O., Vannoorenberghe P. *Belief functions combination and conflict management*, Information Fusion Journal, Elsevier, 2002. Shafer G., *A Mathematical Theory of Evidence*, Princeton Univ. Press, Princeton, NJ, 1976. Shenoy P., *Conditional Independence in Valuation-Based Systems*, International Journal of Approximate reasoning, VoL. 10, pp. 203-234, 1994. Smarandache F. (Editor), *Proceedings of the First International Conference on Neutrosophics*, Univ. of New Mexico, Gallup Campus, NM, USA, 1-3 Dec. 2001, Xiquan, Phoenix, 2002. Smets Ph.,*The combination of evidences in the transferable belief model*, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 12, no. 5, pp. 447-458, 1990. Smets Ph., Kennes R., *The transferable belief model*, Artificial Intelligence, 66(2), pp. 191-234, 1994. Studeny M., *Formal properties of Conditional Independence in Different Calculi of AI*, Proc. of ECSQARU’93, (Clarke K., Kruse R. and Moral S., Eds.), Springer-Verlag, 1993. Yager R.R., *On the Dempster-Shafer framework and new combination rules*, Information Sciences, Vol. 41, pp. 93-138, 1987.. Yaghlane B.B., Smets Ph., Mellouli K., *Independence and Non-Interactivity in the Transferable Belief Model*, Workshop on Conditional Independence Structure and graphical Models, Eds. F. Matus and M. Studeny, Toronto, CA, 1999. Yaghlane B.B., Smets Ph., Mellouli K., *Belief Function Independence: I The marginal case*, International Journal of Approximate Reasoning, Vol. 29, pp. 47-70, 2002. Yaghlane B.B., Smets Ph., Mellouli K., *Belief Function Independence: II conditional case*, International Journal of Approximate Reasoning, Vol. 31, pp. 31-75, 2002.
[^1]: A general numerical example with $m_1(A)>0$ and $m_2(A)>0$ for all $A\neq\emptyset\in D^\Theta$ will be briefly presented in next section.
[^2]: The order of elements of $D^\Theta$ corresponds here to the order obtained from the generation of isotone Boolean functions - see [@Dezert_2003f] for details.
[^3]: $\mathbf{D}_{\mathcal{M}^f}$ was denoted $\mathbf{D}_n$ and $\mathbf{u}_{\mathcal{M}^f}$ as $\mathbf{u}_n$ in reference [@Dezert_2003f].
[^4]: This problem has been proposed to the authors in a private communication by L. Cholvy in 2002.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For a continuous map $T$ of a compact metrizable space $X$ with finite topological entropy, the order of accumulation of entropy of $T$ is a countable ordinal that arises in the context of entropy structure and symbolic extensions. We show that every countable ordinal is realized as the order of accumulation of some dynamical system. Our proof relies on functional analysis of metrizable Choquet simplices and a realization theorem of Downarowicz and Serafin. Further, if $M$ is a metrizable Choquet simplex, we bound the ordinals that appear as the order of accumulation of entropy of a dynamical system whose simplex of invariant measures is affinely homeomorphic to $M$. These bounds are given in terms of the Cantor-Bendixson rank of $\overline{\operatorname{ex}(M)}$, the closure of the extreme points of $M$, and the relative Cantor-Bendixson rank of $\overline{\operatorname{ex}(M)}$ with respect to $\operatorname{ex}(M)$. We also address the optimality of these bounds.'
author:
- David Burguet and Kevin McGoff
title: Orders of Accumulation of Entropy
---
Introduction
============
In this paper, a topological dynamical system is a pair $(X,T)$, where $X$ is a compact metrizable space and $T$ is a continuous mapping of $X$ to itself. For such a system $(X,T)$, the topological entropy $\mathbf{h}_{top}(T)$ provides a well-studied measure of the topological dynamical complexity of the system. We only consider systems with $\mathbf{h}_{top}(T) < \infty$. Let $M(X,T)$ be the space of Borel probability measures on $X$ which are invariant under $T$. The entropy function $h:M(X,T) \rightarrow [0,\infty)$, where $h(\mu)$ is the metric entropy of the measure $\mu$, quantifies the amount of complexity in the system that lies on generic points for $\mu$. In this sense, the entropy function $h$ describes both *where* and *how much* complexity lies in the system. The theory of entropy structures developed by Downarowicz [@D] produces a master entropy invariant in the form of a distinguished class of sequences of functions on $M(X,T)$ whose limit is $h$. The entropy structure of a dynamical system completely determines almost all previously known entropy invariants such as the topological entropy, the entropy function on invariant measures, the tail entropy (or topological conditional entropy [@M]), the symbolic extension entropy, and the symbolic extension entropy function. Entropy structure also produces new entropy invariants, such as the order of accumulation of entropy. Furthermore, the theory of entropy structures and symbolic extensions provides a rigorous description of *how entropy emerges on refining scales*. Entropy structures and the closely related theory of symbolic extensions [@BD] have attracted interest in the dynamical systems literature [@A; @Bur2; @BurNew; @DF; @D; @DM; @DN], especially with the intention of using entropy structure to obtain information about symbolic extensions for various classes of smooth systems. The purpose of the current work is to investigate a new entropy invariant arising from the theory of entropy structures: the order of accumulation of entropy, which is denoted $\al_0(X,T)$.
Given a dynamical system $(X,T)$, one may associate a particular sequence $\H(T)=(h_k)$ to $(X,T)$ with the following properties [@D]:
1. $(h_k)$ is a non-decreasing sequence of harmonic, upper semi-continuous functions from $M(X,T)$ to $[0,\infty)$;
2. $\lim_k h_k = h$;
3. $h_{k+1}-h_k$ is upper semi-continuous for every $k$.
This sequence, or any sequence uniformly equivalent to it (Definition \[uniformEquivDef\]), is called an entropy structure for the system $(X,T)$ [@D]. This distinguished uniform equivalence class of sequences is an invariant of topological conjugacy of the system [@D]. Consequently, we sometimes refer to the entire uniform equivalence class of $\H$ as *the* entropy structure of the system $(X,T)$.
Associated to a non-decreasing sequence $\H = (h_k)$ of functions $h_k : M \to [0,\infty]$, where $M$ is a compact metrizable space, there is a transfinite sequence of functions $u_{\al} : M \rightarrow [0,\infty]$, indexed by the ordinals and defined by transfinite induction as follows. Let $\widetilde{f}$ denote the upper semi-continuous envelope of the function $f$ (Definition \[USCdef\]; by convention $\widetilde{f} \equiv \infty$ if $f$ is unbounded). Let $\tau_k = h - h_k$. Then
- let $u_0 \equiv 0$;
- if $u_{\al}$ has been defined, let $ u_{\al+1} = \lim_{k} \widetilde{u_{\al} + \tau_k}$;
- if $u_{\beta}$ has been defined for all $\beta < \al$ for a limit ordinal $\al$, let $u_{\al} = \widetilde{\sup_{\beta < \al} u_{\beta}}$.
The sequence $(u_{\al})$ is non-decreasing in $\al$ and does not depend on the particular representative of the uniform equivalence class of $\H$. Since $M$ is compact and metrizable, an easy argument (given in [@BD]) implies that there exists a countable ordinal $\al$ such that $u_{\beta} \equiv u_{\al}$ for all $\beta \geq \al$. The least ordinal $\al$ with this property is denoted $\al_0(\H)$ and is called the order of accumulation of $\H$. In the case when $M = M(X,T)$ and $\H$ is an entropy structure for $(X,T)$, the order of accumulation of entropy of $(X,T)$ is defined as $\al_0(\H)$. Because the entropy structure of $(X,T)$ is invariant under topological conjugacy, the sequence $(u_{\al})$ associated to $(X,T)$ and the order of accumulation $\al_0(X,T)$ are invariants of topological conjugacy.
To explain the meaning of $\al_0(X,T)$ and $u_{\al_0(X,T)}$, we discuss symbolic extensions and their relationship to entropy structures. A symbolic extension of $(X,T)$ is a (two-sided) subshift $(Y,S)$ on a finite number of symbols, along with a continuous surjection $\pi : Y \to X$ (the *factor map* of the extension) such that $\pi \circ S = T \circ \pi$. Symbolic extensions have been important tools in the study of some dynamical systems, in particular uniformly hyperbolic systems. A symbolic extension serves as a “lossless finite encoding” of the system $(X,T)$ [@D]. If $\pi$ is the factor map of a symbolic extension $(Y,S)$, we define the extension entropy function $\hext^{\pi} : M(X,T) \rightarrow [0,\infty)$ for $\mu$ in $M(X,T)$ by $$\hext^{\pi}(\mu) = \max \{ h(\nu) : \pi^* \mu = \nu \}.$$ The number $\hext^{\pi}(\mu)$ represents the amount of complexity above the measure $\mu$ in the symbolic extension. The symbolic extension entropy function of a dynamical system $(X,T)$, $\hsex : M(X,T) \rightarrow [0, \medspace \infty]$, is defined for $\mu$ in $M(X,T)$ as $$\hsex(\mu) = \inf \{ \hext^{\pi}(\mu) : \pi \text{ is the factor map of a symbolic extension of } (X,T) \},$$ where the infimum is understood to be $\infty$ if $(X,T)$ admits no symbolic extensions. The symbolic extension entropy function measures the amount of entropy that must be present above each measure in any symbolic extension of the system. Finally, we define the residual entropy function $\hres : M(X,T) \to [0,\infty]$ as $\hres = \hsex - h$. The residual entropy function then measures the amount of entropy that must be added above each measure in any symbolic extension of the system. The functions $\hres$ and $\hsex$ give much finer information about the complexity of the system than the entropy function $h$. These quantities are related to the entropy structure of the system by the following remarkable result of Boyle and Downarowicz.
Let $X$ be a compact metrizable space and $T: X \to X$ a continuous map. Let $\H$ be an entropy structure for $(X,T)$. Then $$\hsex = h + u_{\al_0(X,T)}^{\H}.$$
The conclusion of the theorem may also be stated as $u_{\al_0(X,T)} = \hres$. In this sense, the order of accumulation $\al_0(X,T)$ and the function $u_{\al_0(X,T)}$ each measures a residual complexity in the system that is not detected by the entropy function $h$. The order of accumulation of entropy measures, roughly speaking, *over how many distinct layers residual entropy emerges* in the system [@BD]. It is then natural to ask the following question.
\[realizationQuestion\] Which countable ordinals can be realized as the order of accumulation of entropy of a dynamical system?
It is shown in [@BD] that all finite ordinals can be realized as the order of accumulation of dynamical system. There are constructions in [@Bur2; @DN] (built for other purposes) that show that some infinite ordinals are realized in this way, but these constructions do not allow one to determine exactly which ordinals appear. Moreover, it is stated without proof in [@D] that all countable ordinals are realized.
We prove that all countable ordinals can be realized as the order of accumulation of entropy for a dynamical system (Corollary \[DynRealizationCor\]), answering Question \[realizationQuestion\]. On account of the realization theorem of Downarowicz and Serafin (restated as Theorem \[realization\] in this work), this result reduces to establishing the following result, which is purely functional analytic.
For every countable ordinal $\al$, there exists a metrizable Choquet simplex $M$ and a sequence of functions $\H = (h_k)$ on $M$ such that
- $(h_k)$ is a non-decreasing sequence of harmonic, upper semi-continuous functions from $M$ to $[0,\infty)$;
- $\lim_k h_k$ exists and is bounded;
- $h_{k+1}-h_k$ is upper semi-continuous for every $k$;
- $\al_0(\H) = \al$.
Building on the approach of Downarowicz and Serafin to reduce questions in the theory of entropy structure to the study of functional analysis, we also consider what constraints, if any, the simplex of invariant measures may place on orders of accumulation of entropy.
\[dynQuestion\] Given a metrizable Choquet simplex $M$, which ordinals can be realized as the order of accumulation of a dynamical system $(X,T)$ such that $M(X,T)$ is affinely homeomorphic to $M$?
For a metrizable Choquet simplex $M$, we let $S(M)$ denote the set of all ordinals that can be realized as the order of accumulation of a sequence $\H$ on $M$ satisfying properties (1)-(3). The realization theorem of Downarowicz and Serafin (Theorem \[realization\]) reduces Question \[dynQuestion\] to the following question in functional analysis.
\[introQues\] Given a metrizable Choquet simplex $M$, which ordinals are in $S(M)$?
Theorem \[bauerThm\] answers Question \[introQues\] (and therefore Question \[dynQuestion\]) completely in the event that $M$ is a Bauer simplex by giving a precise description of $S(M)$ in terms of the Cantor-Bendixson rank of the extreme points of $M$. Theorem \[topBounds\] addresses the general case, giving constraints on $S(M)$ in terms of Cantor-Bendixson rank of the closure $\overline{E}$ of the space $E = \operatorname{ex}(M)$ of extreme points of $M$ and the relative Cantor-Bendixson rank of $\overline{E}$ with respect to $E$. Theorems \[threeAlphas\] and \[optimal\] address the optimality of these constraints, and Section \[OpenQuestions\] summarizes our progress on this question and poses some remaining questions.
In the language of dynamical systems, if $M$ is a metrizable Choquet simplex, we have found constraints on the orders of accumulation of entropy that appear within the class of all dynamical systems $(X,T)$ such that $M(X,T)$ is affinely homeomorphic to $M$. These constraints are in terms of the Cantor-Bendixson ranks of the closure $\overline{E}$ of the space $E$ of ergodic measures and the relative Cantor-Bendixson rank of $\overline{E}$ with respect to $E$.
Preliminaries
=============
Ordinals
--------
We assume a basic familiarity with the ordinal numbers, ordinal arithmetic, and transfinite induction. The relevant sections in [@Pin] provide a good introduction. Here we briefly recall some notions that are used in this work.
We view the ordinal $\al$ as the set $\{ \beta : \beta < \al \}$. The symbols $\o$ and $\o_1$ will always be used to denote the first infinite ordinal and the first uncountable ordinal, respectively.
An ordinal $\al$ is **irreducible** if whenever $\al = \al_1 +\al_2$ with $\al_1 \geq \al_2$, it follows that $\al_2=0$.
Recall the well-known Cantor Normal Form of an ordinal.
\[CNF\] For every ordinal $\al > 0$, there exists natural numbers $n_1,\dots, n_k$ and ordinals $\beta_1 > \dots > \beta_k$ such that $\al = \o^{\beta_1}n_1 + \dots + \o^{\beta_k}n_k$. Furthermore, the numbers $n_1, \dots, n_k$ and the ordinals $\beta_1,\dots,\beta_k$ are unique.
The following corollary is an easy consequence of the Cantor Normal Form.
\[irreducible\] An ordinal $\al > 0$ is irreducible if and only if there exists an ordinal $\beta$ such that $\al = \o^{\beta}$.
In light of this corollary, one can view the Cantor Normal Form of $\al$ as a decomposition of $\al$ into a finite sum of irreducible ordinals.
The following corollary is then a simple consequence of Corollary \[irreducible\] and the fact that any non-zero ordinal $\beta$ is either a successor ordinal or a limit ordinal.
\[irrCor\] If $\al > 0$ is countable and irreducible, then either (i) there exists an irreducible ordinal $\widetilde{\al} < \al$ such that $\sup_{n\in \N} \widetilde{\al} n = \al$, or (ii) there exists a strictly increasing sequence of irreducible ordinals $(\al_k)_{k \in \N}$ such that $\sup_{k \in \N} \al_k = \al$.
Any ordinal $\al$ can be viewed as a topological space with the order topology (sets of the form $\{ \gamma \in \al : \gamma < \beta \}$ or $\{ \gamma \in \al : \beta < \gamma \}$ form a subbase for the topology). With this topology, $\al$ is a completely normal, Hausdorff space, and if $\al$ is countable, then it is a Polish space (see below for definition). The space $\al$ is compact if and only if $\al$ is a successor ordinal. The accumulation points in $\al$ are exactly the limit ordinals in $\al$.
For ease of notation, if $\al$ is a successor ordinal, let $\al-1$ denote the unique ordinal $\beta$ such that $\al = \beta+1$. Also, for countable ordinals $\al \leq \beta$, we will write $[\al, \beta]$ to denote the ordinal interval $\{ \gamma : \al \leq \gamma \leq \beta \}$. If $\beta = \o_1$, we make the convention that $[\al, \beta] = \{ \gamma : \al \leq \gamma < \beta \}$. We also make use of the notation $]\al, \beta[ = \{ \gamma : \al < \gamma < \beta \}$, as well as the other possible “half-open” and “half-closed” notations.
Polish Spaces
-------------
A general reference that covers Polish spaces is [@S]. We recall that a topological space $E$ is a Polish space if it is separable and completely metrizable. In particular, any compact metrizable space is Polish. Moreover, any closed subset of a Polish space is itself a Polish space. Some of the definitions and statements below hold for more general topological spaces, but we require them only in the case of Polish spaces.
For any Polish space $E$, let $E'$ denote the set of accumulation points of $E$, $$E' = \{ x \in E : \exists (x_n) \subset E \setminus \{ x \}, \medspace x_n \rightarrow x \}.$$ Note that $E'$ is closed in $E$.
A subset $A$ of a Polish space $E$ is a perfect set if $A$ is a compact subset of $E$ and $A$ contains no isolated points (in the subspace topology). The following result is a special case of the Cantor-Bendixson Theorem.
\[CBThm\] Let $E$ be a Polish space. Then $E = C \cup A$, where $C$ is countable, $A$ is closed and has no isolated points, and $C \cap A = \emptyset$.
We will also use the following fact (see [@S]). Let $\mathcal{C}$ denote the Cantor space.
\[embedCperfect\] Let $A$ be a non-empty Polish space with no isolated points. Then there is an embedding of $\mathcal{C}$ into $A$.
The following statement is an immediate corollary of the previous two theorems.
\[embedCunctble\] Let $E$ be any uncountable Polish space. Then there is an embedding of $\mathcal{C}$ into $E$.
The following corollary is an easy consequence of Corollary \[embedCunctble\].
\[uncountPolish\] Let $E$ be an uncountable Polish space. Then for every countable ordinal $\al$ and every natural number $n$, there exists an embedding $g: \o^{\al}n+1 \rightarrow E$.
Cantor-Bendixson Rank
---------------------
Given a Polish space $E$, we now use transfinite induction to define a transfinite sequence of topological spaces, $\{ \Gamma^{\al}(E) \}$. Let $\Gamma^0(E) = E$. If $\Gamma^{\al}(E)$ has been defined, then let $\Gamma^{\al+1}(E) = (\Gamma^{\al}(E)) ' \subset \Gamma^{\al}(E)$. If $\al$ is a limit ordinal and $\Gamma^{\beta}(E)$ is defined for all $\beta < \al$, then let $\Gamma^{\al}(E) = \cap_{\beta < \al} \Gamma^{\beta}(E)$. Each set $\Gamma^{\al}(E)$ is closed in $E$ and therefore Polish.
Note that $\Gamma^{\al}(E) = \Gamma^{\al+1}(E)$ implies that $\Gamma^{\al}(E)$ has no isolated points (in the subspace topology) and then that $\Gamma^{\beta}(E) = \Gamma^{\al}(E)$ for all $\beta > \al$. For any Polish space $E$, Theorem \[CBThm\] implies that there exists a countable ordinal $\al$ such that $\Gamma^{\al}(E) = \Gamma^{\al+1}(E)$.
With the notation above, the **Cantor-Bendixson rank** of the space $E$, denoted $|E|_{CB}$, is defined to be the least ordinal $\al$ such that $\Gamma^{\al}(E) = \Gamma^{\al+1}(E)$.
When $E$ is compact, $\Gamma^{|E|_{CB}}(E)$ is a perfect set (which may be the empty set). Now we mention a pointwise version of Cantor-Bendixson rank.
\[ptwiseCBDef\] Let $E$ be a Polish space, and let $x$ be in $E$. We define the **topological rank** of $x$, $r(x)$, to be $$r(x) = \left\{
\begin{array}{ll}
\sup \{ \al : x \in \Gamma^{\al}(E) \} & \text{if } x \notin \Gamma^{|E|_{CB}}(E) \\
\o_1 & \text{if } x \in \Gamma^{|E|_{CB}}(E).
\end{array} \right.$$
The following proposition follows directly from the definitions and compactness.
\[accumFacts\] Let $E$ be a countable, compact Polish space. Then
1. $|E|_{CB}$ is a successor ordinal.
2. If $|E|_{CB} = \al+1$, then $\Gamma^{\al}(E)$ is a non-empty, finite set, and $\Gamma^{\al+1}(E) = \emptyset$.
3. $|E|_{CB} = \big( \sup_{x \in E} r(x) \big)+1 = \big(\max_{x \in E} r(x)\big) + 1$.
Now we state a well-known classification of countable, compact Polish spaces, due to Mazurkiewicz and Sierpiński [@MS p. 21]. We denote the cardinality of a set $E$ by $|E|$.
\[classification\] Let $E$ and $F$ be countable, compact Polish spaces, and assume that $|E|_{CB} = \al+1$. Then $E$ and $F$ are homeomorphic if and only if $|E|_{CB}=|F|_{CB}$ and $|\Gamma^{\al}(E)| = |\Gamma^{\al}(F)|$.
\[disjUnion\] Let $\al$ be a countable ordinal. Then $\Gamma^{\al}(\o^{\al}+1) = \{\o^{\al}\}$ and $|\o^{\al}+1|_{CB} = \al+1$. It follows from Theorem \[classification\] that if $\gamma_k$ is any increasing sequence of ordinals such that $\sup_k \gamma_k = \o^{\al}$, then $\o^{\al}+1$ is homeomorphic to the one-point compactification of the disjoint union of the spaces $\gamma_k$, with the point at infinity corresponding to $\o^{\al}$.
Note that for any countable ordinal $\al$, the space $\o^{\al}n+1$ has Cantor-Bendixson rank $\al+1$ and exactly $n$ points of topological rank $\al$ given by $\o^{\al}k$ for $k = 1, \dots, n$. Then by the above classification, the space $\o^{\al}n+1$ provides a representative of the homeomorphism class of countable, compact Polish spaces with Cantor-Bendixson rank $\al+1$ and $n$ points of topological rank $\al$.
Upper-semicontinuity
--------------------
Now we consider functions $f : E \rightarrow \R$, where $E$ is a metrizable space. For such a function $f$, we let $||f|| = \sup_{x \in E} |f(x)|$, where the supremum is taken to be $+\infty$ if $f$ is unbounded.
\[USCdef\] Let $E$ be a compact metrizable space, and let $f : E \rightarrow \R$. Then $f$ is **upper semi-continuous** (u.s.c.) if one of the following equivalent conditions holds:
1. $f = \inf_{\al} g_{\al}$ for some family $\{g_{\al}\}$ of continuous functions;
2. $f = \lim_{n} g_n$ for some nonincreasing sequence $(g_n)_{n\in \N}$ of continuous functions;
3. For each $r \in \R$, the set $\{x : f(x) \geq r \}$ is closed;
4. $\limsup_{y\rightarrow x} f(y) \leq f(x)$, for all $x \in E$.
For any $f : E \rightarrow \R$, the **upper semi-continuous envelope** of $f$, written $\widetilde{f}$, is defined, for all $x$ in $E$, by $$\widetilde{f}(x) = \left\{ \begin{array}{ll} \inf \{ g(x) : g \text{ is continuous, and } g \geq f \}, & \text{ if } f \text{ is bounded} \\ +\infty, & \text{ if } f \text{ is unbounded.}
\end{array}
\right.$$
Note that when $f$ is bounded, $\widetilde{f}$ is the smallest u.s.c. function greater than or equal to $f$ and satisfies $$\widetilde{f}(x) = \max \left( f(x), \limsup_{y \rightarrow x} f(y) \right).$$ It is immediately seen that for any $f, g : E \rightarrow \R$, $\widetilde{f+g} \leq \widetilde{f}+\widetilde{g}$, with equality holding if $f$ or $g$ is continuous.
Let $\pi : E \rightarrow F$ be a continuous map. If $f: F \rightarrow \R$ is any function, we define the **lift** of $f$, denoted $^{\pi} f$, to be the function given by $f \circ \pi$.
If $\pi : E \rightarrow F$ is a surjection and $f: E \rightarrow \R$ is bounded, then the **projection** of $f$, denoted $f^{[F]}$, is the function defined on $F$ by $$f^{[F]}(x) = \sup_{y \in \pi^{-1}(x)} f(y).$$
\[pushDownPullUp\] Let $\pi: E \rightarrow F$ be a continuous surjection.
1. If $f : F \rightarrow \R$, then $(^{\pi} f)^{[F]} = f$.
2. If $f: E \rightarrow \R$, then $^{\pi}(f^{[F]}) \geq f$, and the inequality is strict in general.
3. If $f : E \rightarrow \R$ is u.s.c., then $f^{[F]}$ is also u.s.c. and the supremum is attained.
4. If $f: F \rightarrow \R$ is u.s.c., then $^{\pi} f$ is also u.s.c.
Candidate Sequences {#candSeqs}
-------------------
A **candidate sequence** on a compact, metrizable space $E$ is a non-decreasing sequence $\H = (h_k)$ of non-negative, real-valued functions on $E$ that converges pointwise to a function $h$. We often write $\lim \H = h$. We always assume by convention that $h_0 \equiv 0$.
A candidate sequence $\H$ **has u.s.c. differences** if $h_{k+1} - h_{k}$ is u.s.c. for all $k$. Note that in this case each $h_k$ is u.s.c., since $h_0 \equiv 0$. If $\H$ has u.s.c. differences, we may also refer to $\H$ as a u.s.c.d. candidate sequence, or we may write that $\H$ is u.s.c.d.
Given a candidate sequence $\H$, it is natural to seek a precise description of the manner in which $h_k$ converges to $h$. For example, is this convergence uniform or not? The notion of *uniform equivalence*, as defined by Downarowicz in [@D], captures exactly the manner in which $h_k$ converges to $h$.
\[uniformEquivDef\] Let $\H$ and $\F$ be two candidate sequences on a compact, metrizable space $E$. We say that $\H$ **uniformly dominates** $\F$, written $\H \geq \F$, if for all $\epsilon > 0$, and for each $k$, there exists $\ell$, such that $f_k \leq h_{\ell}+\epsilon$.
The candidate sequences $\H$ and $\F$ are **uniformly equivalent**, written $\H \cong \F$, if $\H \geq \F$ and $\F \geq \H$.
Note that uniform equivalence is in fact an equivalence relation.
Basic Constructions {#basicConstructions}
===================
Order Of Accumulation
---------------------
Let $\H$ be a candidate sequence on $E$. The **transfinite sequence** associated to $\H$, which we write as $(u_{\al}^{\H})$ or $(u_{\al})$, is defined by transfinite induction as follows. Let $\tau_k = h - h_k$. Then
- let $u_0 \equiv 0$;
- if $u_{\al}$ has been defined, let $ u_{\al+1} = \lim_{k} \widetilde{u_{\al} + \tau_k}$;
- if $u_{\beta}$ has been defined for all $\beta < \al$ for a limit ordinal $\al$, let $u_{\al} = \widetilde{\sup_{\beta < \al} u_{\beta}}$.
Note that for each $\al$, either $u_{\al} \equiv +\infty$ or $u_{\al}$ is u.s.c. (since a non-increasing limit of u.s.c. functions is u.s.c.). Furthermore, the sequence $(u_{\al})$ is non-decreasing in $\al$. It is also sub-additive in the following sense.
\[UisZero\] Let $\H$ be a candidate sequence on $E$. Then for any two ordinals $\al$ and $\beta$, $$u_{\al+\beta} \leq u_{\al} + u_{\beta}.$$
Let $\al$ be any ordinal. We prove the statement by transfinite induction on $\beta$. For $\beta=0$, the statement is trivial. Now assume by induction that the statement is true for $\gamma < \beta$. If $\beta$ is a successor ordinal, then by the inductive hypothesis, $$u_{\al+\beta} = \lim_k \widetilde{(u_{\al + (\beta-1)} + \tau_k)} \leq u_{\al} + \lim_k \widetilde{(u_{\beta-1} + \tau_k)} = u_{\al}+u_{\beta}.$$ If $\beta$ is a limit ordinal, then by the inductive hypothesis, $$u_{\al+\beta} = \widetilde{\sup_{\gamma< \beta} u_{\al+\gamma}} \leq u_{\al}+\widetilde{\sup_{\gamma<\beta} u_{\gamma}} \leq u_{\al}+u_{\beta}.$$
If $\H$ is a candidate sequence on $E$, then by Theorem 3.3 in [@BD], there exists a countable ordinal $\al$ such that the associated transfinite sequence satisfies $u_{\al} = u_{\al + 1}$, which then implies that $u_{\beta} = u_{\al}$ for all $\beta > \al$.
In this setting, the least ordinal $\al$ such that $u_{\al} = u_{\al + 1}$ is called the **order of accumulation** of the candidate sequence $\H$, which we write as either $\al_0(\H)$ or $\al_0^{\H}$.
Both the transfinite sequence and the order of accumulation are independent of the choice of representative of uniform equivalence class [@D].
While it is true that $u_{\al}=u_{\al+1}$ implies $u_{\al} = u_{\beta}$ for all $\beta > \al$, it is not true that for a fixed $x$, $u_{\al}(x) = u_{\al+1}(x)$ implies $u_{\beta}(x) = u_{\al}(x)$ for all $\beta > \al$. In fact, in many of the constructions in Section \[examples\] there is a point $\0$ and an ordinal $\al$ such that $u_{\gamma}(\0)=0$ for all $\gamma < \al$ and $u_{\al}(\0)=a>0$. Nonetheless, we make the following definition.
Let $\H$ be a candidate sequence on $E$. Then for each $x$ in $E$, we define the **pointwise order of accumulation** of $\H$ at $x$, $\al_0^{\H}(x)$ or $\al_0(x)$, as $$\al_0^{\H}(x) = \inf \{ \al : u_{\beta}(x) = u_{\al}(x) \text{ for all } \beta > \al \}.$$
\[accumVarPrin\] Note that $\al_0^{\H}(x)$ is always a countable ordinal, and $$\al_0(\H) = \sup_{x \in E} \al_0^{\H}(x).$$
The following proposition relates the pointwise topological rank (Definition \[ptwiseCBDef\]) to the pointwise order of accumulation.
\[pointwiseBound\] Let $\H$ be a candidate sequence on $E$. Then for any $x$ in $E$, $$\al_0(x) \leq \left\{
\begin{array}{ll}
r(x) & \text{ if } r(x) \text{ is finite} \\
r(x) +1 & \text{ if } r(x) \text{ is infinite}.
\end{array} \right.$$
The proof proceeds by transfinite induction on $r(x)$. If $r(x) = 0$ it is easily seen that $u_{\gamma}(x)=0$ for all $\gamma$ and $\al_0(x)=0$.
Suppose the statement is true for all $y$ with $r(y) < \al$, and fix $x$ with $r(x) = \al$. If $\al$ is finite, let $\epsilon = \al$, and if $\al$ is infinite, let $\epsilon = \al+1$. We show that for all $\beta > \al$, $u_{\beta}(x) = u_{\epsilon}(x)$, and here we use transfinite induction on $\beta > \al$. Note that there is an open neighborhood $U$ of $x$ such that for all $y$ in $U$, $r(y) < r(x)$. Thus any real-valued function $f$ on $E$ satisfies $\limsup_{y \rightarrow x} f(y) = \limsup_{y \rightarrow x, \medspace r(y) < r(x)} f(y)$.
Suppose $\beta > \al$ is a successor. Then $$\widetilde{(u_{\beta -1} + \tau_k)}(x) = \max \Bigl( \limsup_{\substack{y \rightarrow x \\ \medspace r(y) < \al}} \, (u_{\beta-1}+\tau_{k})(y), \medspace (u_{\beta-1}+\tau_k)(x) \Bigr).$$ Applying the induction hypotheses to all $y$ with $r(y)<\al$ and $u_{\beta-1}(x)$ gives that $$\widetilde{(u_{\beta -1} + \tau_k)}(x) = \max \Bigl( \limsup_{\substack{y \rightarrow x \\ \medspace r(y) < \al}} \, (u_{\epsilon-1}+\tau_{k})(y), \medspace (u_{\epsilon}+\tau_k)(x) \Bigr).$$ Letting $k$ tend to infinity, we obtain $u_{\beta}(x) = u_{\epsilon}(x)$.
Suppose $\beta$ is a limit ordinal. Then the inductive hypotheses imply $$\begin{aligned}
u_{\beta}(x) & = \max \Bigl( \limsup_{\substack{ y\rightarrow x \\ \medspace r(y) < \al}} \, \sup_{\gamma < \beta} u_{\gamma}(y) , \medspace \sup_{\gamma < \beta} u_{\gamma}(x) \Bigr) \\
& = \max \Bigl( \limsup_{\substack{y\rightarrow x \\ \medspace r(y) < \al}} \, u_{\epsilon-1}(y) , \medspace \sup_{\gamma < \beta} u_{\epsilon}(x) \Bigr) \\
& = u_{\epsilon}(x).\end{aligned}$$
It follows from the proof of Theorem \[bauerThm\] that these pointwise bounds on $\al_0(x)$ are optimal. Also, combining Remark \[accumVarPrin\], Proposition \[pointwiseBound\], and Proposition \[accumFacts\] (3), we obtain the following result.
\[CBrankBound\] Let $\H$ be a candidate sequence on a countable, compact Polish space $E$. Then $$\al_0(\H) \leq \left\{ \begin{array}{ll}
|E|_{CB}-1, & \text{ if } |E|_{CB} \text{ is finite} \\
|E|_{CB}, & \text{ if } |E|_{CB} \text{ is infinite}.
\end{array} \right.$$
Construction of Candidate Sequences
-----------------------------------
Now we discuss various ways of creating candidate sequences. We first begin with elementary constructions that will be studied later in the context of Choquet simplices.
Let $\H$ be a candidate sequence on $E$. If $F$ is a compact subset of $E$, then we define the **restriction candidate sequence**, $\H|_{F}$, on $F$.
Let $\H$ be candidate sequence on $E$, and let $F$ be a compact metrizable space with $\pi: F \rightarrow E$ a continuous surjection. Then the **lifted candidate sequence** of $\H$ to $F$, denoted $^{\pi} \H$, is the candidate sequence on $F$ given by $(^{\pi}h_k) = (h_k \circ \pi)$.
Let $\F = (f_k)$ be a candidate sequence on $F$, and let $g:F\rightarrow E$ be an embedding (continuous injection). The **embedded candidate sequence**, $g\F = (h_k)$, on $E$ is defined to be $$h_k(x) = \left\{
\begin{array}{ll}
f_k \circ g^{-1}(x) & \text{ if } x \in g(F) \\
0 & \text{ if } x \in E \setminus g(F).
\end{array} \right.$$
While all of the constructions in this section will be used, the following two constructions (disjoint union and product candidate sequences) form the basis of the proofs of Theorem \[mainTech\] and Corollary \[realizationCor\].
Let $(\H_n)$ be a countable collection of candidate sequences, where $\H_n = (h_k^n)$ is defined on $E_n$. Then we define the **disjoint union candidate sequence**, $\coprod \H_n$, as follows. Let $E$ be the one-point compactification of the disjoint union of the spaces $E_n$, with the point at infinity denoted $\0$. For each $k$, let $f_k$ be the function on $E$ such that $f_k|_{E_n} = h_k^n$ and $f_k(\0) = 0$. Then the disjoint union candidate sequence, $\coprod \H_n$, is defined to be $(f_k)$.
Recall that $||f||$ denotes the supremum norm of the real-valued function $f$.
\[disUnionLemma\] Let $(\H_n)$ be a sequence of candidate sequences on $E_n$, where $h^n = \lim \H_n$. Let $\H = \coprod \H_n$. If $||h^n|| \rightarrow 0$, then for all $\beta$,
1. $u_{\beta}^{\H}(\0) = \limsup_{n} ||u_{\beta}^{\H_n}||$, and
2. $||u_{\beta}^{\H}|| = \sup_{n} ||u_{\beta}^{\H_n}||$.
For each $n$, $E_n$ is a clopen subset of $E$. It follows that $u_{\gamma}^{\H}(x) = u_{\gamma}^{\H_n}(x)$ for all ordinals $\gamma$, and for all $x$ in $E_n$. Then (2) follows from the definitions and (1). Also, upper semi-continuity of $u_{\beta}^{\H}$ implies that $u_{\beta}^{\H}(\0) \geq \limsup_n ||u_{\beta}^{\H_n}||$. It remains only to show the reverse inequality.
The hypotheses imply that $$u_1^{\H}(\0) \leq \widetilde{h}(\0) \leq \lim_n ||h^n || = 0.$$ Now we use transfinite induction on $\beta$. The case $\beta=0$ is trivial. Suppose $\beta$ is a successor. By sub-additivity of the transfinite sequence (Lemma \[UisZero\]) $u_{\beta}^{\H}(\0) \leq u_{\beta-1}^{\H}(\0) + u_1^{\H}(\0) = u_{\beta-1}^{\H}(\0)$, which, along with induction, implies the desired inequality. Now suppose $\beta$ is a limit ordinal. Monotonicity of the transfinite sequence and induction again imply that $$u_{\beta}^{\H}(\0) = \max \Bigl( \limsup_{y \rightarrow \0} u_{\beta}^{\H}(y), \; \sup_{\gamma < \beta} u_{\gamma}^{\H}(\0) \Bigr) \leq \limsup_n ||u_{\beta}^{\H_n}||.$$
By a marked space $(E, \0)$, we mean a compact, metrizable space $E$ together with a marked point $\0$ in $E$.
Let $\F = (f_k)$ and $\mathcal{G} = (g_k)$ be two candidate sequences defined on the marked spaces $(E_1, \0_1)$ and $(E_2, \0_2)$, respectively. Then we define the **product candidate sequence**, $\H = \F \times \mathcal{G}$, on the marked product space $(E_1 \times E_2, (\0_1,\0_2))$ as the sequence $$h_k(x,y) = \left\{
\begin{array}{ll}
f_k(x) & \text{ if } y = \0_2 \\
g_k(y) & \text{ if } y \neq \0_2
\end{array} \right.$$
Note that this definition is not symmetric under transposition of $\F$ and $\mathcal{G}$. In other words, this product is not commutative, but one may check easily that it is associative.
Let $\H$ be a candidate sequence on the marked space $(E,\0)$. Define $(\H)^{\times p}$ to be the candidate sequence on the product space $(E^p,\0^p)$ given by iterated multiplication: $(\H)^{\times p} = \H^{\times (p-1)} \times \H$.
\[powerLem\] Let $\H$ be a candidate sequence on the marked space $(E,\0)$. Suppose that for some limit ordinal $\al$ and real number $a >0$,
1. $||u_{\gamma}|| \leq a$ for all $\gamma$, and $||u_{\gamma}|| < a$ for $\gamma < \al$;
2. $u_{\gamma}(\0) = 0$, for all $\gamma < \al$, and $u_{\al}(\0)=a$;
3. $\al_0(x) \leq \al$, for all $x$ in $E$.
Then the transfinite sequence associated to $(\H)^{\times p}$ satisfies
1. $||u_{\gamma}^{\H^{\times p}}|| \leq p a$ for all $\gamma$;
2. $||u_{\al k}^{\H^{\times p}}|| \leq k a$ and $||u_{\gamma}^{\H^{\times p}}|| < k a$, for all $\gamma < \al k$ and $k \leq p$;
3. $\al_0^{\H^{\times p}}(x) \leq \al p$, for all $x $ in $E^p$;
4. $u_{\gamma}^{\H^{\times p}}(\0^p) = \ell a$, for all $\al \ell \leq \gamma < \al (\ell+1)$, and $\ell = 0, \dots , p$;
5. $\al_0^{\H^{\times p}}(\0^p) = \al p$.
We argue by induction on $p$. For $p=1$, the claims (1)-(5) follow from (i)-(iii). Assume that (1)-(5) hold for $p$. We prove that (1)-(5) also hold with $p+1$ in place of $p$. Let $(u_{\al}^p)$ be the transfinite sequence for $\H^{\times p} = (h_k^p )$, and let $h^p = \lim \H^{\times p}$. Recall that $E^{p+1} = E^p \times E$. The definition of $\H^{\times (p+1)}$ is that $$h_k^{p+1}(x,y) = \left \{ \begin{array}{ll}
h_k^p(x), & \text{ if } y= \0 \\
h_k(y), & \text{ if } y \neq \0.
\end{array} \right.$$ For all $(x,y)$ in $E^{p+1}$, $(h^{p+1}-h^{p+1}_k)(x,y) \leq (h^p-h^p_k)(x) + (h-h_k)(y)$. It follows from transfinite induction that for all $\gamma$, $u_{\gamma}^{p+1}(x,y) \leq u_{\gamma}^p(x) + u_{\gamma}(y)$. Using the inductive hypotheses, we obtain that $||u_{\gamma}^{p+1}|| \leq ap+a= a(p+1)$ for all $\gamma$, proving (1).
It follows from subadditivity that $||u^{p+1}_{\al k+ \gamma}|| \leq k ||u^{p+1}_{\al}||+ ||u^{p+1}_{\gamma}||$, which means that in order to establish (2) we need only show that for all $\gamma < \al$, $||u^{p+1}_{\gamma}|| < a$. Furthermore, since $u_{\gamma}^{p+1}$ is u.s.c. and therefore attains its supremum, it suffices to show that for all $\gamma < \al$ and all $(x,y)$ in $E^{p+1}$, $u_{\gamma}^{p+1}(x,y) < a$. Let $\gamma < \al$ and let $(x,y)$ be in $E^{p+1}$. If $y \neq \0$, then there exists an open neighborhood $U$ of $(x,y)$ in $E^{p+1}$ such that for all $(s,t)$ in $U$, $t \neq \0$. Then $h^{p+1}_k(s,t) = h_k(t)$ for all $(s,t)$ in $U$. It follows that $u_{\gamma}^{p+1}(x,y) = u_{\gamma}(y) < a$. Now suppose $y = \0$. Let $\epsilon>0$. Since $u_{\gamma}(\0) = 0$ and $u_{\gamma}$ is u.s.c., there exists an open neighborhood $U$ of $\0$ in $E$ such that for all $s$ in $U$, $u_{\gamma}(s) \leq \epsilon$. Then for all $(t,s)$ in the open set $E^p \times U$, $u_{\gamma}^{p+1}(t,s) \leq u_{\gamma}^p(t) + u_{\gamma}(s) \leq u_{\gamma}^{p}(t) + \epsilon$. Since $\epsilon$ was arbitrary, we obtain that $u_{\gamma}^{p+1}(x,\0) \leq u_{\gamma}^{p}(x)$. Using the induction hypothesis for $\H^{\times p}$, we conclude that $u_{\gamma}^{p+1}(x,\0) < a$.
For any point $(x,y)$ in $E^{p+1}$ with $y \neq \0$, we have already shown that $u_{\gamma}^{p+1}(x,y) = u_{\gamma}(y)$ for all $\gamma$. For any point of the form $(x,\0)$, we have shown that $u_{\al}^{p+1}(x,\0) \leq a$. Furthermore, by upper-semicontinuity of $u_{\al}^{p+1}$, we have that $$u_{\al}^{p+1}(x,\0) \geq \limsup_{y \rightarrow \0} u_{\al}^{p+1}(x,y) = \limsup_{y \rightarrow \0} u_{\al}(y) = u_{\al}(\0) = a.$$ Thus $u_{\al}^{p+1}(x,\0)= a$ for all points of the form $(x,\0)$. This fact, in combination with the fact that $u_{\gamma}^{p+1}(x,y) = u_{\gamma}(y) \leq a$ for $y \neq \0$ and all $\gamma$, immediately implies that $u_{\al+\gamma}^{p+1}(x,\0) = u_{\gamma}^p(x)+a$ for all $x$ in $E^p$. Then induction gives statements (3)-(5).
For the rest of this work, we let $\H^p$ denote the **renormalized product** of $\H$ taken $p$ times: if $\H^{\times p} = (h_k^{\times p})$, then let $\H^p = (h_k^p) = (\frac{1}{p} h_k^{\times p})$.
Now we discuss more general products than just powers of the same candidate sequence. We will only consider products of marked spaces. Let $x$ be a point in the product space $(E_N \times \dots \times E_1, \0)$, where $\0 = (\0_N, \dots, \0_1)$. Let $\pi_i$ be projection onto $E_i$. Then define the function $$\operatorname{ind}(x) = \left\{
\begin{array}{ll}
\min\{ i : \pi_i(x) \neq \0_i \} & \text{ if } x\neq \0 \\
N & \text{ if } x = \0.
\end{array} \right.$$ Also, let $\eta_i(x_N, \dots, x_1) = (x_N, \dots, x_i)$. Note that with these notations, if $(h_k) = \H_N \times \dots \times \H_1 $, then $h_k(x) = h_k^{\H_{\operatorname{ind}(x)}}(\pi_{\operatorname{ind}(x)}(x))$ for all $x$.
\[prodLemma\] Let $\al$ be any non-zero countable ordinal, and let $\al = \o^{\beta_1} m_1 + \dots + \o^{\beta_N} m_N$ be the Cantor Normal Form of $\al$. Let $a>0$ be a real number, and suppose $a_1 > \dots > a_N > 0$ such that $$\sum_{i=1}^N a_i = a,$$ and for each $j = 1, \dots, N-1$, $$\label{aiCondition}
\frac{a_j}{m_j} \geq \sum_{i=j+1}^N a_i.$$ (Note that for any $a>0$, such $a_1, \dots, a_N$ exist.) Now suppose that for each $j$ in $\{1, \dots, N\}$, $\F_j$ is a candidate sequence on $(E_j,\0_j)$ such that
1. $||u_{\gamma}^{\F_j}|| \leq a_j$ for all $\gamma$, and $||u_{\gamma}^{\F_j}|| < a_j$ for $\gamma < \o^{\beta_j}$;
2. $u_{\gamma}^{\F_j}(\0_j) = 0$, for all $\gamma < \o^{\beta_j}$;
3. $u_{\o^{\beta_j}}^{\F_j}(\0_j)=a_j$;
4. $\al_0(x) \leq \o^{\beta_j}$, for all $x \neq \0_j$;
5. $\al_0(\0_j) = \o^{\beta_j}$.
Denote $\H_j = \F_j^{m_j}$ and $\al_j = \o^{\beta_j} m_j$. Then the product $\H_N \times \dots \times \H_1$ satisfies
1. $||u_{\gamma}|| \leq a$ for all $\gamma$, and $||u_{\gamma}|| < a$ for $\gamma < \al$;
2. $\al_0(x) \leq \al$, for all $x \neq \0$;
3. $\al_0(\0) = \al$, and $u_{\al_0}(\0) = a$. In particular, $\al_0(\H_N \times \dots \times \H_1) = \al$.
The proof proceeds by induction on $N$. The case $N=1$ follows from (i)-(v). Now we assume that $N>1$ and the statement holds for $N-1$, and we show that it holds for $N$.
Let $\H_N \times \dots \times \H_1 = (h_k)$ be as above, with $h = \lim_k h_k$, and let $\H_{N} \times \dots \times \H_2 = (h'_k)$ with $h'= \lim_k h'_k$. By the definition of the product candidate sequence, we observe that $(h-h_k)(x) \leq (h'-h'_k)(\eta_2(x)) + (h^1 - h^1_k)(\pi_1(x))$. It follows that $u_{\al}(x) \leq u_{\al}^{\H_{N} \times \dots \times \H_2}(\eta_2(x)) + u_{\al}^{\H_1}(\pi_1(x))$ for all $x$ in $E$ and $\al$.
Let $x$ be in $E$. Then there exists an open neighborhood $U$ in $E$ such that for all $y$ in $U$, $\operatorname{ind}(y) \leq \operatorname{ind}(x)$.
If $\operatorname{ind}(x) = 1$, the existence of the neighborhood $U$ implies that $u_{\gamma}^{\H}(x) = u_{\gamma}^{\H_1}(\pi_1(x))$ for all $\gamma$.
Now we prove that for $\gamma < \o^{\beta_1}$ and $x$ such that $\operatorname{ind}(x) > 1$, we have $u_{\gamma}^{\H}(x) \leq u_{\gamma}^{\H_N \times \dots \times \H_2}(\eta_2(x))$. Since $\F_1$ satisfies the hypotheses $(i)-(v)$, we may apply Lemma \[powerLem\] and conclude that $\H_1$ satisfies conclusions (1)-(5) in Lemma \[powerLem\]. Now let $\gamma < \o^{\beta_1}$ and let $x$ be in $E$ with $\operatorname{ind}(x) >1$. By conclusion (4) in Lemma \[powerLem\] applied to $\H_1$, $u_{\gamma}^{\H_1}(\0_1) = 0$. Then for any $\epsilon >0$, using that $u_{\gamma}^{\H_1}$ is u.s.c., there exists an open neighborhood $V$ of $x$ such that for all $y$ in $V$, $u_{\gamma}^{\H}(y) \leq u_{\gamma}^{\H_N \times \dots \times \H_2}(\eta_2(y)) + \epsilon$. Since $\epsilon >0$ was arbitrary, we have the desired inequality.
By the induction hypothesis on $N-1$ applied to $\H_N \times \dots \times \H_2$, we have $\sup_{\gamma < \o^{\beta_1}} u_{\gamma}^{\H_N \times \dots \H_2}(\eta_2(x)) \leq \sum_{j=2}^N a_j$. By conclusion (2) in Lemma \[powerLem\] applied to $\H_1$, $||u_{\o^{\beta_1}}^{\H_1}|| \leq \frac{a_1}{m_1}$. Hence, for $x$ in $E$, $$u_{\o^{\beta_1}}^{\H}(x) = \widetilde{ \bigl(\sup_{\gamma < \o^{\beta_1}} u_{\gamma}^{\H} \bigr)}(x) \leq \max \Bigl( \frac{a_1}{m_1}, \, \sum_{j=2}^{N} a_j \Bigr) \leq \frac{a_1}{m_1}.$$ Then by upper semi-continuity of $u_{\o^{\beta_1}}^{\H}$, we have that for any $x$ with $\operatorname{ind}(x) >1$, $$u_{\o^{\beta_1}}^{\H}(x) \geq \limsup_{\substack{y \rightarrow x \\ \operatorname{ind}(y)=1 }} u_{\o^{\beta_1}}^{\H}(y) = \limsup_{\substack{y \rightarrow x \\ \operatorname{ind}(y)=1 }} u_{\o^{\beta_1}}^{\H_1}(\pi_1(y)) = u_{\o^{\beta_1}}^{\H_1}(\0_1) = \frac{a_1}{m_1}.$$ We conclude that for any $x$ with $\operatorname{ind}(x)>1$, $u_{\o^{\beta_1}}^{\H}(x) = \frac{a_1}{m_1}$. By sub-additivity (Proposition \[UisZero\]), we have that $u_{\o^{\beta_1}m_1}^{\H}(x) \leq a_1$. By upper semi-continuity, for all $x$ with $\operatorname{ind}(x)>1$, $$u_{\o^{\beta_1}m_1}^{\H}(x) \geq \limsup_{\substack{y \rightarrow x \\ \operatorname{ind}(y)=1 }} u_{\o^{\beta}m_1}^{\H_1}(\pi_1(y)) = u_{\o^{\beta}m_1}^{\H_1}(\0_1) = a_1.$$ It follows that $u_{\o^{\beta_1}m_1}^{\H}(x) = a_1$ for all $x$ with $\operatorname{ind}(x) >1$, and then $u_{\o^{\beta_1}m_1+ \gamma}^{\H}(x) = a_1 + u_{\gamma}^{\H_N \times \dots \times \H_2}(\eta_2(x))$ for all $x$ with $\operatorname{ind}(x) >1$ and all $\gamma$. Now with the induction hypothesis on $N-1$ applied to $\H_N \times \dots \times \H_2$, the properties (1)-(3) follow immediately.
We end this section by stating the semi-continuity properties of these new candidate sequences.
1. If $\H_k$ is a sequence of u.s.c.d. candidate sequences and $||h^k|| \rightarrow 0$, then $\H = \coprod \H_k$ is a u.s.c.d. candidate sequence.
2. If $\H_1$ and $\H_2$ are u.s.c.d. candidate sequences and $(\lim\H_2)(\0_2)=0$, then $\H = \H_1 \times \H_2$ is a u.s.c.d. candidate sequence.
3. If $\H$ is a u.s.c.d. candidate sequence on $E$ and $F$ is closed subset of $E$, then $\H|_F$ is a u.s.c.d. candidate sequence.
4. If $\H$ is a u.s.c.d. candidate sequence on $E$ and $^{\pi} \H$ is the lift of $\H$ to $F$, where $\pi : F \rightarrow E$ is a continuous surjection, then $^{\pi} \H$ is a u.s.c.d. candidate sequence.
\(1) The condition $||h^k|| \rightarrow 0$ implies that $\H$ has u.s.c. differences at $\0$ for all $k$.\
(2) Because $\H_1$ is u.s.c.d., the condition $(\lim\H_2)(\0_2)=0$ implies that $\H$ has u.s.c. differences at $(x,\0_2)$ for all $x$ and $k$.\
(3) The restriction of any u.s.c. function to a subset is also u.s.c.\
(4) The lift of any u.s.c. function under a continuous map is also u.s.c.\
Choquet Simplices and Candidate Sequences {#simplices}
-----------------------------------------
The relevant chapters of [@Ph] provide a good reference for most of the basic facts about simplices required in this work.
Let $K$ be a metrizable, compact, convex subset of a locally convex topological vector space. Then the extreme points of $K$, $\operatorname{ex}(K)$, form a non-empty $G_{\delta}$ subset of $K$. We call a function $f: K \rightarrow \R$ **affine** (resp. **convex, concave**) if $f(tx + (1-t)y) = tf(x) + (1-t)f(y)$ (resp. $\leq, \geq$) for all $x$ and $y$ in $K$ and all $t$ in $[0,1]$.
Let $K$ be a compact, convex subset of a locally convex topological vector space. Then $K$ is a **Choquet simplex** if the dual of the continuous affine functions on $K$ is a lattice.
For any Polish space $E$, let $\mathcal{M}(E)$ be the space of all Borel probabilities on $E$ with the weak\* topology. If $E$ is compact, then $\mathcal{M}(E)$ is a Choquet simplex, with the extreme points given by the point measures.
Let $K$ be a Choquet simplex. Then we define the **barycenter map**, $\operatorname{bar}: \mathcal{M}(K) \rightarrow K$, to be the function given for each $\mu$ in $\M(K)$ by $$\operatorname{bar}(\mu) = \int y \medspace d\mu(y),$$ where the integral means that for all continuous, affine functions $f: K \rightarrow \R$, $$f(\operatorname{bar}(\mu)) = \int_K f \medspace d\mu.$$
The barycenter map is well-defined, continuous, affine, and surjective (see [@Ph]).
If $K$ is a metrizable Choquet simplex, then a function $f: K \rightarrow \R$ is called **harmonic** (resp. sub-harmonic, sup-harmonic) if for all $\mu$ in $\M(K)$, $$f(\operatorname{bar}(\mu)) = \int\limits_{\operatorname{ex}(K)} f \; d\mu,$$ (resp. $\leq, \geq$). A harmonic (resp. sub-harmonic, sup-harmonic) function is always affine (resp. convex, concave), but an affine (resp. convex, concave) function need not be harmonic (resp. sub-harmonic, sup-harmonic). On the other hand, a continuous affine (resp. convex, concave) function is always harmonic (resp. sub-harmonic, sup-harmonic). Furthermore, by standard arguments, any u.s.c. affine (resp. concave) function is harmonic (resp. sup-harmonic). It is shown in the proof of Fact \[fonEharIsUSC\] (see Appendix B, Section \[proofAppendix\]) that any u.s.c. convex function is sub-harmonic.
In the metrizable case, Choquet proved the following characterization of Choquet simplices.
\[SimplexThm\] Let $K$ be a metrizable, compact, convex subset of a locally convex topological vector space. Then $K$ is a Choquet simplex if and only if for each point $x$ in $K$, there exists a unique Borel probability measure $\P_x$ on $\operatorname{ex}(K)$ such that for every continuous affine function $f : K \rightarrow \R$, $$f(x) = \int\limits_{\operatorname{ex}(K)} f \; d\P_x.$$
If $K$ is a metrizable Choquet simplex and $f: \operatorname{ex}(K) \rightarrow \R$ is measurable, the **harmonic extension** $f^{har} : K \rightarrow \R$ of $f$ is defined as follows: for $x$ in $K$, let $$f^{har}(x) = \int\limits_{\operatorname{ex}(K)} f \; d\P_x.$$
\[harmRemark\] Using Choquet’s characterization of metrizable Choquet simplices, it is not difficult to show that if $f : K \to \R$ is a measurable function and for each $x$ in $K$, $$f(x) = \int f \; d\P_x,$$ then $f$ is harmonic. It follows that the harmonic extension of a function on $\operatorname{ex}(K)$ is, in fact, harmonic.
In the metrizable case, the following theorem of Choquet characterizes exactly which topological spaces appear as the set of extreme points of a Choquet simplex.
\[choquetWeak\] The topological space $E$ is homeomorphic to the set of extreme points of a metrizable Choquet simplex if and only if E is a Polish space.
The following fact is stated as Fact 2.5 in [@DM], where there is a sketch of the proof. We include a proof as Appendix B (Section \[proofAppendix\]) for the sake of completeness.
\[fonEharIsUSC\] Let $K$ be a metrizable Choquet simplex, and let $f: K \rightarrow [0, \infty)$ be convex and u.s.c. Then $(f|_{\operatorname{ex}(K)})^{har}$ is u.s.c.
If $K$ is a metrizable Choquet simplex, we denote by $\M(\operatorname{ex}(K))$ the set of measures $\mu$ in $\M(K)$ such that $\mu(K \setminus \operatorname{ex}(K)) = 0$. Consider the map $\pi : \M(\operatorname{ex}(K)) \rightarrow K$ given by the restriction of the barycenter map to $\M(\operatorname{ex}(K))$. This restriction inherits the continuity and affinity of the barycenter map. Furthermore, this restriction is always bijective (by Choquet’s characterization of metrizable Choquet simplices, Theorem \[SimplexThm\]), but it may not have a continuous inverse. In fact, $\pi$ has a continuous inverse if and only if $\operatorname{ex}(K)$ is closed in $K$. These considerations lead to the study of Bauer simplices.
A metrizable, compact, convex subset $K$ of a locally convex topological vector space is a **Bauer simplex** if $K$ is a Choquet simplex such that $\operatorname{ex}(K)$ is a closed subset of $K$.
If $E$ is any compact, metrizable space, then $\M(E)$ is a Bauer simplex with $\operatorname{ex}(\M(E))$ homeomorphic to $E$. If $K$ is a Bauer simplex, then the restriction of the barycenter map $\pi : \M(\operatorname{ex}(K)) \rightarrow K$ has a continuous inverse and is therefore an affine homeomorphism from $\mathcal{M}(\operatorname{ex}(K))$ to $K$.
\[uscBauerharm\] If $K$ is a Bauer simplex and $f: K \rightarrow [0,\infty)$ is bounded and harmonic, then $\widetilde{f}$ is harmonic and $\widetilde{f}|_{\operatorname{ex}(K)} = \widetilde{f|_{\operatorname{ex}(K)}}$.
Since $f$ is harmonic, in particular $f$ is affine. Let $x$ and $y$ be in $K$, and let $ax+by$ be a convex combination in $K$. We have $\widetilde{f}(ax+by) \geq f(ax+by) = af(x) + bf(y)$. For fixed $a,b,$ and $y$, the above formula implies that $\widetilde{f}(ax+by) \geq a\widetilde{f}(x) +b f(y)$. Now fixing $a,b,$ and $x$, we obtain $\widetilde{f}(ax+by) \geq a\widetilde{f}(x) +b \widetilde{f}(y)$. Now since $\widetilde{f}$ is u.s.c. and concave, it follows that $\widetilde{f}$ is sup-harmonic.
Let $E = \operatorname{ex}(K)$. It follows from the definitions that $$\label{fIneq}
f(t) = \int_E f|_E \; d \P_t \leq \int_E \widetilde{(f|_E)} \; d\P_t \leq \int_{E} \widetilde{f} \; d\P_t.$$ Now consider the two functions $g_1, g_2 : K \to \R$, given for each $t$ in $E$ by $$\begin{aligned}
g_1(t) & = \left\{ \begin{array}{ll} \widetilde{f|_E}(t), & \text{ if } t \in E, \\
0, & \text{ if } t \notin E,
\end{array} \right. \\
g_2(t) & = \left\{ \begin{array}{ll} \widetilde{f}|_E(t), & \text{ if } t \in E, \\
0, & \text{ if } t \notin E.
\end{array} \right.\end{aligned}$$ Since $E$ is closed, $g_1$ and $g_2$ are u.s.c. They are also obviously convex. Then by Fact \[fonEharIsUSC\], $G_1 = \bigl((g_1)|_E\bigr)^{har}$ and $G_2 = \bigl((g_2)|_E\bigr)^{har}$ are u.s.c. Note that for $t \in K$, $$G_1(t) = \int_E \widetilde{(f|_E)} \; d\P_t, \text{ and } G_2(t) = \int_{E} \widetilde{f} \; d\P_t.$$ Thus, taking the u.s.c. envelope of the expressions in Equation (\[fIneq\]) and using that $G_1$ and $G_2$ are u.s.c., we have that $$\label{inequalities}
\widetilde{f}(t) \leq \int_E \widetilde{(f|_E)} \; d\P_t \leq \int_{E} \widetilde{f} \; d\P_t,$$ which shows that $\widetilde{f}$ is sub-harmonic. Now we have shown that $\widetilde{f}$ is harmonic and the inequalities in Equation (\[inequalities\]) are all equalities.
A candidate sequence $\H = (h_k)$ on a Choquet simplex is said to be harmonic if each $h_k$ is harmonic. The following proposition relates the transfinite sequence of a candidate sequence $\H$ on a Bauer simplex $K$ to the transfinite sequence of $\H|_{\operatorname{ex}(K)}$.
\[Uaffine\] If $\H$ is a harmonic candidate sequence on the Bauer simplex $K$, then for each $\al$, $u_{\al}^{\H}$ is harmonic and $$\label{uBauer}
u_{\al}^{\H} = (u_{\al}^{\H|_{\operatorname{ex}(K)}})^{har}.$$
The proof proceeds by transfinite induction on $\al$. For all $k$, since $h_k$ and $h$ are harmonic, $\tau_k = h - h_k$ is harmonic.
Suppose $u^{\H}_{\al}$ is harmonic and Equation (\[uBauer\]) holds. Then $u^{\H}_{\al}+\tau_k$ is harmonic. By Proposition \[uscBauerharm\], we deduce that $\widetilde{u^{\H}_{\al}+\tau_k}$ is harmonic, and for $t$ in $K$, $$(u^{\H}_{\al}+\tau_k)(t) = \int_E \widetilde{(u_{\al}^{\H}+\tau_k)|_E} d \P_t = \int_E \widetilde{(u_{\al}^{\H|_E}+\tau_k)|_E} d \P_t.$$ Recall that $\{u_{\al}+\tau_k\}_k$ is a non-increasing sequence in $k$. Thus we can take the limit in $k$ and apply the Monotone Convergence Theorem to obtain that $u^{\H}_{\al+1}$ is also harmonic, and for $t$ in $K$, $$u_{\al+1}^{\H}(t) = \int_E u_{\al+1}^{\H|_E} d \P_t,$$ which implies that Equation (\[uBauer\]) holds with $\al+1$ in place of $\al$.
The previous arguments apply in a similar way to the case when $\al$ is a limit ordinal.
\[ConcaveRmk\] Let $K$ be a Choquet simplex which is not necessarily Bauer. Even when the candidate sequence $\H$ on $K$ is harmonic, the functions $u_{\al}^{\H}$ are not in general harmonic. However, we check now that if $\H$ is harmonic, then $u_{\al}^{\H}$ is concave for all $\al$. Assuming by induction that $u_{\al}^{\H}$ is concave, we have that $\widetilde{u_{\al}^{\H}+\tau_k}$ is concave, as it is the u.s.c. envelope of a concave function. Then $u_{\al+1}^{\H}$ is the limit of a sequence of concave functions, and so $u_{\al+1}^{\H}$ is concave. Now for any countable limit ordinal $\al$, there is a strictly increasing sequence $(\al_n)$ of ordinals tending to $\al$. Then $\sup_{\beta < \al} u_{\beta}^{\H} = \lim_n u_{\al_n}^{\H}$ since the sequence $(u_{\beta}^{\H})$ is increasing in $\beta$. Then $\sup_{\beta < \al} u_{\beta}^{\H}$ is concave, as it is the limit of a sequence of concave functions (by induction), and thus $u_{\al}^{\H}$ is concave for any countable limit ordinal as well.
When $\operatorname{ex}(K)$ is not compact, $\mathcal{M}(\operatorname{ex}(K))$ is not a Bauer simplex, and the restriction of the barycenter map to this set is not a homeomorphism. Instead of using this restriction in such cases, we consider the Bauer simplex $\mathcal{M}(\overline{\operatorname{ex}(K)})$ and the continuous surjection $\pi : \mathcal{M}(\overline{\operatorname{ex}(K)}) \rightarrow K$, where $\pi$ is the restriction of the barycenter map to $\mathcal{M}(\overline{\operatorname{ex}(K)})$. In the following two lemmas we consider candidate sequences which may arise as embedded candidate sequences.
\[embedIsHarmUSC\] Let $E$ be a compact, metrizable space, and let $K$ be a metrizable Choquet simplex. Suppose there exists a continuous injection $g : E \rightarrow K$. Let $\F$ be a u.s.c.d. candidate sequence on $E$, let $\H' = (h'_k)$ be the embedded candidate sequence $g\F$, and let $\H$ be the harmonic extension of $\H'|_{\operatorname{ex}(K)}$ to $K$. If $h'_{k+1}-h'_{k}$ is convex for each $k$, then $\H$ is u.s.c.d. In particular, if $g(E) \subset \operatorname{ex}(K)$ then $\H$ is u.s.c.d.
Since $\F$ is u.s.c.d. and $g(E)$ is closed, we have that $h'_{k+1}-h'_k$ is u.s.c. for each $k$. Then $h'_{k+1}-h'_k$ is convex and u.s.c. for each $k$. By applying Fact \[fonEharIsUSC\], we obtain that $h_{k+1}-h_{k}$ is u.s.c. for each $k$. Thus $\H$ is u.s.c.d.
In particular, if $g(E) \subset \operatorname{ex}(K)$, then $h'_{k+1}-h'_k$ takes non-zero values only on $\operatorname{ex}(K)$. Therefore $h'_{k+1}-h'_k$ is convex for each $k$, and by the previous argument, $\H$ is u.s.c.d.
The following lemma is used repeatedly throughout the rest of this work. The utility of this statement lies in the fact that it allows one to compute the transfinite sequence on a (frequently much simpler) subset of the simplex and then write the transfinite sequence on the entire simplex in terms the transfinite sequence on this subset. When $K$ is a Choquet simplex that is not Bauer and $\H$ is a harmonic candidate sequence on $K$, then this statement takes the place of an integral representation of $u_{\al}^{\H}$.
\[embeddingThm\] Let $K$ be a metrizable Choquet simplex with $E = \operatorname{ex}(K)$. Suppose $\H$ is a harmonic candidate sequence on $K$ and there is a set $F \subset E$ such that the sequence $\{(h-h_k)|_{E \setminus F} \}$ converges uniformly to zero. Let $L = \overline{F}$, and let $\pi:\M(\overline{E}) \rightarrow K$ be the restriction of the barycenter map. Then for all ordinals $\al$ and for all $x$ in $K$, $$\label{embeddingEqn}
u_{\al}^{\H}(x) = \max_{\mu \in \pi^{-1}(x)} \int_{L} u_{\al}^{\H|_L} d\mu,$$ and $\al_0(\H) \leq \al_0(\H|_L)$. In particular, if $F$ is compact, then $u_{\al}^{\H}|_F = u_{\al}^{\H|_F}$ for all $\al$ and $\al_0(\H) = \al_0(\H|_F)$.
Note that Equation (\[embeddingEqn\]) implies immediately that $\al_0(\H) \leq \al_0(\H|_L)$. Further, suppose $F$ is compact. Then $L = F \subset \operatorname{ex}(K)$, and if $x$ is in $F$, then $\pi^{-1}(x)=\{\epsilon_x\}$, where $\epsilon_x$ is the point mass at $x$. In this case Equation (\[embeddingEqn\]) implies that $u_{\al}^{\H}|_F = u_{\al}^{\H|_F}$ for all $\al$ and $\al_0(\H) = \al_0(\H|_F)$. We now prove Equation (\[embeddingEqn\]).
Observe that since $L$ is closed and $u_{\al}^{\H|_L}$ is u.s.c., the function $\mathbf{1}_L \cdot u_{\al}^{\H|_L}$ is u.s.c., where $\mathbf{1}_L$ is the characteristic function of the set $L$. Then the function $\mu \mapsto \int_L u_{\al}^{\H|_L} \, d\mu$ is u.s.c., and therefore by Remark \[pushDownPullUp\] (3), for each $x$ in $K$, $$\sup_{\mu \in \pi^{-1}(x)} \int_L u_{\al}^{\H_L} \, d\mu = \max_{\mu \in \pi^{-1}(x)} \int_L u_{\al}^{\H_L} \, d\mu.$$
Let $x$ be in $K$. Since $u_{\al}^{\H}$ is concave (see Remark \[ConcaveRmk\]) and u.s.c., it follows that $u_{\al}^{\H}$ is sup-harmonic. Therefore $$u_{\al}^{\H}(x) \geq \int_{K} u_{\al}^{\H} d\mu, \quad \text{ for all } \mu \in \pi^{-1}(x).$$ Using the fact that $u_{\al}^{\H}|_L \geq u_{\al}^{\H|_L}$, we obtain, for all $\mu \in \pi^{-1}(x)$, $$u_{\al}^{\H}(x) \geq \int_{K} u_{\al}^{\H} d\mu \geq \int_{L} u_{\al}^{\H} d\mu \geq \int_{L} u_{\al}^{\H|_L} d\mu.$$ It follows that for each ordinal $\al$, $$u_{\al}^{\H}(x) \geq \max_{\mu \in \pi^{-1}(x)} \int_{L} u_{\al}^{\H|_L} d\mu.$$
We now prove using transfinite induction on $\al$ that for all $\al$ and $x$ in $K$, $$\label{reverseIneq}
u_{\al}^{\H}(x) \leq \max_{\mu \in \pi^{-1}(x)} \int_{L} u_{\al}^{\H|_L} d\mu,$$ which will complete the proof of the Lemma.
The inequality in Equation (\[reverseIneq\]) is trivial for $\al=0$. Suppose Equation (\[reverseIneq\]) holds for some ordinal $\al$. For the sake of notation, we allow $y=x$ in all expressions involving $\limsup_{y \to x}$ below. First we claim that for any $y$ in $K$, there exists a measure $\mu_y$ supported on $L \cup E$ such that $\mu_y$ is in $\pi^{-1}(y)$ and $$\label{maxExpression}
\max_{\mu \in \pi^{-1}(y)} \int_L u_{\al}^{\H|_L} d\mu = \int_L u_{\al}^{\H|_L} d\mu_y.$$ Indeed, suppose the maximum is obtained by the measure $\nu$. If $\nu(L)=1$, then we are done. Now suppose $\nu(L) < 1$. Then $\nu = \nu(L) \nu_L + (1-\nu(L)) \nu_{\overline{E} \setminus L}$, where $\nu_S$ is the zero measure on $S$ if $\nu(S)=0$ and otherwise $\nu_S(A) = \frac{1}{\nu(S)} \nu(S \cap A)$. Let $z = \operatorname{bar}(\nu_{\overline{E}\setminus L})$, which exists since $\nu_{\overline{E} \setminus L}$ is in $\M(\overline{E})$ (using that $\nu(\overline{E} \setminus L) = 1 - \nu(L) > 0$). Now let $\mu_y = \nu(L) \nu_L + (1-\nu(L)) \P_{z}$. Then $\mu_y$ is supported on $L \cup E$, $\operatorname{bar}(\mu_y) = y$, and $$\int_L u_{\al}^{\H|_L} d\nu \leq \int_L u_{\al}^{\H|_L} d\mu_y.$$ Thus the maximum in Equation (\[maxExpression\]) is obtained by the measure $\mu_y$, which is supported on $L \cup E$ and satisfies $\operatorname{bar}(\mu)=y$.
Now let $\epsilon >0$. Since $\H$ is harmonic, we also have that $\tau_k$ is harmonic. Then for any $y$ in $K$ and $k$ large enough (depending only on $\epsilon$), $$\begin{aligned}
u_{\al}^{\H}(y)+\tau_k(y) & = \max_{\mu \in \pi^{-1}(y)} \int_L u_{\al}^{\H|_L} d\mu + \tau_k(y) \label{induction} \\
& = \int_L u_{\al}^{\H|_L} d\mu_y + \int \tau_k d\mu_y \label{harmonicity} \\
& = \int_L u_{\al}^{\H|_L} d\mu_y + \int_L \tau_k d\mu_y + \int_{E \setminus L} \tau_k d\mu_y \label{integral} \\
& \leq \int_L u_{\al}^{\H|_L} d\mu_y + \int_L \tau_k d\mu_y + \epsilon \label{uniformity} \\
& = \int_L (u_{\al}^{\H|_L} + \tau_k) d\mu_y + \epsilon \label{subset} \\
& \leq \int_L \widetilde{(u_{\al}^{\H|_L} + \tau_k)|_L} d\mu_y + \epsilon \\
& \leq \max_{\mu \in \pi^{-1}(y)} \int_L \widetilde{(u_{\al}^{\H|_L} + \tau_k)|_L} d\mu + \epsilon.\end{aligned}$$ Then we have (allowing $y=x$ in the limit suprema) that $$\begin{aligned}
u_{\al+1}^{\H}(x) & = \lim_k \thinspace \limsup_{y \rightarrow x} \thinspace u_{\al}^{\H}(y)+\tau_k(y) \\
& \leq \lim_k \thinspace \limsup_{y \rightarrow x} \thinspace \max_{\mu \in \pi^{-1}(y)} \int_L \widetilde{(u_{\al}^{\H|_L} +\tau_k)|_L} d\mu + \epsilon \label{usc} \\
& \leq \lim_k \thinspace \max_{\mu \in \pi^{-1}(x)} \int_L \widetilde{(u_{\al}^{\H|_L} +\tau_k)|_L} d\mu + \epsilon \label{limsup} \\
& \leq \max_{\mu \in \pi^{-1}(x)} \int_L u_{\al+1}^{\H|_L} d\mu + \epsilon \label{lim},\end{aligned}$$ where the inequalities in (\[limsup\]) and (\[lim\]) are justified by Lemmas \[limsupLemma\] and \[limLemma\], respectively. Since $\epsilon$ was arbitrary, we have shown the inequality in Equation (\[reverseIneq\]) with the ordinal $\al$ replaced by $\al+1$.
Now suppose the inequality in Equation (\[reverseIneq\]) holds for all $\beta < \al$, where $\al$ is a limit ordinal. Using monotonicity of the sequence $u_{\al}^{\H|_L}$, we see that (allowing $y=x$ in the limit suprema) $$\begin{aligned}
u_{\al}^{\H}(x) & = \widetilde{\sup_{\beta < \al} u_{\beta}^{\H}}(x) \\
& = \limsup_{y \rightarrow x} \thinspace \sup_{\beta< \al} \thinspace \max_{\mu \in \pi^{-1}(y)} \int_L u_{\beta}^{\H|_L} d\mu \\
& \leq \limsup_{y \rightarrow x} \thinspace \max_{\mu \in \pi^{-1}(y)} \int_L u_{\al}^{\H|_L} d\mu \\
& \leq \max_{\mu \in \pi^{-1}(x)} \int_L u_{\al}^{\H|_L} d\mu,\end{aligned}$$ where Lemma \[limsupLemma\] justifies the last inequality. Thus we have shown that the inequality in Equation (\[reverseIneq\]) holds for $\al$, which completes the induction and the proof.
\[embeddingRemark\] Given the assumptions of the Embedding Lemma, if $x$ is in $\operatorname{ex}(K)$, then $\pi^{-1}(x) = \{ \epsilon_x \}$, where $\epsilon_x$ is the point mass at $x$. It follows that, if $x$ is in $L \cap \operatorname{ex}(K)$, then $u_{\al}^{\H}(x) = u_{\al}^{\H|_{L}}(x)$ for all $\al$. Further, if $x$ is in $\operatorname{ex}(K) \setminus L$, then $u_{\al}^{\H}(x)=0$ for all $\al$.
\[embeddingRemarktwo\] With the notation of the Embedding Lemma, Equation (\[embeddingEqn\]) implies that $||u_{\al}^{\H}||=||u_{\al}^{\H|_L}||$ for all $\al$.
\[limsupLemma\] Let $K$ be a metrizable Choquet simplex and $L$ a closed subset of $K$. Let $f: K \rightarrow [0, \infty)$ be u.s.c. Then for all $x$ in $K$, $$\limsup_{y \rightarrow x} \max_{\mu \in \pi^{-1}(y)} \int_L f d\mu \leq \max_{\mu \in \pi^{-1}(x)} \int_L f d\mu,$$ where $\pi$ is the restriction of the barycenter map on $\M(K)$ to $\M(\overline{\operatorname{ex}(K)})$.
Let $T : \M(K) \rightarrow \R$ be defined by $T(\mu) = \int_L f d\mu$. We have that $f \chi_L$ is u.s.c. since $f$ is non-negative and u.s.c. and $L$ is closed. It follows that $T$ is u.s.c. Then the result follows from Remark \[pushDownPullUp\] (3).
\[limLemma\] Let $K$ be a metrizable Choquet simplex and $L$ a closed subset of $K$. Let $\{f_k: K \rightarrow [0, \infty) \}$ be a non-increasing sequence of u.s.c. functions, with $\lim_k f_k = f$. Then for all $x$ in $K$, $$\lim_{k \rightarrow \infty} \max_{\mu \in \pi^{-1}(x)} \int_L f_k d\mu \leq \max_{\mu \in \pi^{-1}(x)} \int_L f d\mu,$$ where $\pi$ is the restriction of the barycenter map on $\M(K)$ to $\M(\overline{\operatorname{ex}(K)})$.
Let $x$ be in $K$. Define $T_k : \M(K) \rightarrow \R$ and $T: \M(K) \rightarrow \R$ by the equations $$T_k(\mu) = \int_L f_k d\mu, \; \text{ and } \; T(\mu) = \int_L f d\mu.$$ Since $f_k \chi_L$ and $f \chi_L$ are u.s.c., $T$ and $T_k$ are u.s.c. Proposition 2.4 of [@BD] states (in slightly greater generality) that $$\label{BDEqn}
\lim_k \max_{\mu \in \pi^{-1}(x)} T_k(\mu) = \max_{\mu \in \pi^{-1}(x)} \lim_k T_k(\mu).$$ By the Monotone Convergence Theorem, $$\label{TequalsLimTk}
T(\mu) = \lim_k T_k(\mu).$$ Combining Equations (\[BDEqn\]) and (\[TequalsLimTk\]) concludes the proof.
Even when the hypotheses of the Embedding Lemma are satisfied, it is possible to have $\al_0(\H) < \al_0(\H|_L)$, as the next example shows.
\[HlessThanPiH\] This example provides a candidate sequence $\H$ satisfying the hypotheses of the Embedding Lemma and $\al_0(\H) < \al_0(\H|_L)$, which proves that the inequality $\al_0(\H) \leq \al_0(\H|_L)$ is not an equality in general. Suppose the set of extreme points of $K$ consists of two points, $b_1$ and $b_2$, sequences $\{ c_n \}$ and $\{d_n \}$ with $c_n \rightarrow b_1$ and $d_n \rightarrow b_2$, and a countable collection $\{ a_n \}$. Let $b = \frac{1}{2}(b_1+b_2)$ in $K$. Suppose further that with the subspace topology inherited from $K$, the set $\{ a_n \} \cup \{b\}$ is homeomorphic to $\o^2 +1$, with the homeomorphism given by $g_1: \o^2 +1 \rightarrow \{ a_n \} \cup \{b\}$ and $g_1(\o^2)=b$. One may construct such a simplex $K$ as the image of $\M(\{a_n\} \cup \{b, b_1, b_2\} \cup \{c_n\} \cup \{d_n\})$ under a continuous affine map (Lemma \[ConstructKLem\]). Let $\F_1 = (f_k^1)$ be u.s.c.d. candidate sequence on $\o^2+1$ such that $\al_0(\F_1)=2$, $u_1^{\F_1}(t) = u_2^{\F_1}(t)$ for $t \neq \o^2$, and $||u_{2}^{\F_1}||=1$. Such a sequence is given by Corollary \[realizationCor\]. Let $\F_2 = (f_k^2)$ be the u.s.c.d. candidate sequence on $\{c_n\} \cup \{d_n\} \cup \{b_1,b_2\}$ given, for $x$ in $\{c_n\} \cup \{d_n\} \cup \{b_1,b_2\}$ and $k \geq 1$, by $$f_k^2(x) = \left \{ \begin{array}{ll} 0 & \text{ if } x = c_n \text{ or } x=d_n, \text{ with } k < n \\
1 & \text{ otherwise}.
\end{array}
\right.$$
Now consider the candidate sequence $\H' = (h'_k)$ on $K$ such that for $x$ in $K$, $$h'_k(x) = \left \{
\begin{array}{ll}
f_k^1(g_1^{-1}(x)) & \text{ if } x = a_n \\
f_k^2(x) & \text{ if } x = c_n, \medspace d_n \\
0 & \text{ otherwise } \\
\end{array} \right.$$ Note that $\H'$ is u.s.c.d., convex, and $h'_{k+1} - h'_k$ is convex. Let $\H$ be the harmonic extension of $\H'|_{\operatorname{ex}(K)}$ on $K$. Then by Lemma \[embedIsHarmUSC\], $\H$ is harmonic and u.s.c.d.
Let $F = \operatorname{ex}(K)$ and $L = \overline{F} = \{a_n\} \cup \{b, b_1, b_2\} \cup \{c_n\} \cup \{d_n\}$. Since $L$ is the disjoint union the two (clopen in $L$) sets $\{a_n\} \cup \{b\}$ and $\{b_1, b_2\} \cup \{c_n\} \cup \{d_n\}$, we see that for $t$ in $L$, $$\begin{aligned}
u_{\al}^{\H|_L} = \left\{ \begin{array}{ll}
u_{\al}^{\F_1}(t), & \text{ if } t \in \{a_n\} \cup \{b\} \\
u_{\al}^{\F_2}(t), & \text{ if } t \in \{b_1, b_2\} \cup \{c_n\} \cup \{d_n\}.
\end{array} \right.\end{aligned}$$ Thus $\al_0(\H|_L) = \max(\al_0(\F_1),\al_0(\F_2)) = \al_0(\F_1) = 2$ and $||u_2^{\H|_L}|| \leq 1$. Also, for all $t \neq b$, $u_1^{\H|_L}(t) = u_2^{\H|_L}(t)$, and for $t \in \{b_1,b_2\}$, $u_1^{\H|_L}(t) = 1$.
Applying the Embedding Lemma, we have that for all $t$ in $K$, $$\label{ETexampleEqn}
u_{\al}^{\H}(t) = \max_{\mu \in \pi^{-1}(t)} \int_L u_{\al}^{\H|_L} \, d\mu.$$ If $\mu \in \pi^{-1}(t)$ and $\mu(\{b\}) >0$, then let $\nu = \frac{1}{2}\mu(\{b\})(\epsilon_{b_1} + \epsilon_{b_2}) + (1-\mu(\{b\})) \mu_{L \setminus \{b\}}$, where $\mu_{L \setminus \{b\}}$ is the measure $\mu$ conditioned on the set $L \setminus \{b\}$. Then $\nu \in \pi^{-1}(t)$, $\nu(\{b\}) = 0$, and $\int_L u_{i}^{\H|_L} \, d\mu \leq \int_L u_{i}^{\H|_L} \, d\nu$ for $i \in \{1, 2\}$. Thus the maximum in Equation (\[ETexampleEqn\]) is obtained by a measure $\mu$ with $\mu(\{b\}) = 0$. Now if $\mu \in \pi^{-1}(t)$ and $\mu(\{b\}) = 0$, then $\int_L u_{1}^{\H|_L} \, d\mu = \int_L u_{2}^{\H|_L} \, d\mu$ since $u_1^{\H|_L}(s) = u_2^{\H|_L}(s)$ for $s \in L \setminus \{b\}$. From these facts we deduce $u_{1}^{\H}(t) = u_2^{\H}(t)$ for all $t$ in $K$, and therefore $\al_0(\H) = 1 < \al_0(\H|_L)$.
Realization of Transfinite Orders of Accumulation {#examples}
=================================================
Recall that for every countable ordinal $\al$, $\o^{\al}+1$ is a countable, compact, Polish space. Then let $K_{\al}$ be the (unique up to affine homeomorphism) Bauer simplex with $\operatorname{ex}(K_{\al}) = \o^{\al}+1$. For notation, let $\0_{\al}$ be the point $\o^{\al}$ in $K_{\al}$, and let $E_{\al} = \operatorname{ex}(K_{\al})$. In this section we construct, for each countable $\al$, a harmonic, u.s.c.d. candidate sequence $\H_{\al}$ on $K_{\al}$ such that $\al_0(\H_{\al})=\al$.
The idea of the following theorem is to construct, for each countable, irreducible ordinal $\al$, a candidate sequence $\H$ such that the transfinite sequence does not converge uniformly at $\al$, in some sense. The main tools of the proof are the disjoint union candidate sequence and the powers candidate sequences.
\[mainTech\] For all real numbers $0< \epsilon < a$, and for all countable, irreducible ordinals $\delta$ and $\al$, with $\delta < \al$, there exists a harmonic, u.s.c.d candidate sequence $\H_{\al}$ on $K_{\al}$ such that
1. $||h|| \leq a$ if $\al$ is finite, and $||h|| \leq \epsilon$ if $\al$ is infinite;
2. $||u_{\delta} || \leq \epsilon$;
3. $||u_{\gamma}|| \leq a$ for all $\gamma$, and $||u_{\gamma}|| < a$ for $\gamma < \al$;
4. $h(\0_{\al})=0, \medspace u_{\gamma}(\0_{\al}) = 0$, for all $ \gamma < \al$, and $u_{\al}(\0_{\al})=a$;
5. $\al_0(\H_{\al}) = \al$.
Suppose that we have constructed an u.s.c.d. candidate sequence $\H'$ on $\o^{\al}+1$ and shown that it possesses properties (1)-(5). Since $K_{\al}$ is Bauer, Proposition \[Uaffine\] implies that we can let $\H_{\al}$ be the harmonic extension of $\H'$ to $K_{\al}$ and properties (1)-(5) carry over exactly. So without loss of generality, we will define $\H_{\al}$ directly on $E_{\al}$ and work exclusively on $E_{\al}$.
The rest of the proof proceeds by transfinite induction on the non-zero irreducible ordinals $\al$ ($\al$ is non-zero because $\delta<\al$). This is equivalent, by Proposition \[irreducible\], to writing $\al = \o^{\beta}$ and using transfinite induction on $\beta$. The base case is when $\beta = 0$.
In this case $E_{\o^0} = E_1 = \o + 1$, the one-point compactification of the natural numbers. Now $\delta$ must be $0$ and by definition $u_0 \equiv 0$. Let $\H= (h_k)$, where $h_k(n) = 0$ if $k \leq n$, $h_k(n) = a$ if $k > n$, and $h_k(\0_1) = 0$. Then $h \leq a$. Since each $n$ is isolated in $E_{1}$, $r(n) = 0$, which implies that $\al_0(n) = 0$ and $u_{\gamma}(n)=0$ for all $\gamma$ (by Proposition \[pointwiseBound\]). The point at infinity, $\0_1$, has topological order of accumulation $1$, which implies that $\al_0(\0_1) \leq 1$ (by Proposition \[pointwiseBound\]). It only remains to check that $u_1(\0_1)=a$. Fix $k$. For any $n > k$, $\tau_k(n) = h(n)-h_k(n) = a$. Thus $\widetilde{\tau_k}(\0_1) \geq a$. Letting $k$ go to infinity gives that $u_1(\0_1) \geq a$. Since $u_1 \leq \widetilde{h} \leq a$, we obtain that $u_1(\0_1) = a$, as desired.
We assume the statement is true for $\o^{\beta}$, and we need to show that it is true for $ \o^{\beta+1} = \sup_n \o^{\beta} n$. In this case $E_{\o^{\beta+1}}$ is homeomorphic to the one-point compactification of the disjoint union of the spaces $(E_{\o^{\beta} n})$ (by Theorem \[classification\]). With this homeomorphism, we may assume without loss of generality that $E_{\o^{\beta+1}}$ is the one-point compactification of the disjoint union of the spaces $E_{\o^{\beta}n}$. Fix $0 < \epsilon < a$, and let $\{a_p\}$ be a sequence of positive real numbers such that $a_p < a$ for all $p$ and $\lim_p a_p = a$. Using the induction hypothesis, for each $p$, we choose a u.s.c.d. candidate sequence $\H_{\o^{\beta}}$ on $E_{\o^{\beta}}$ which satisfies conditions (1)-(5) with parameters $a_p$, $\epsilon$, and $\delta < \o^{\beta}$. For each $p$, let $\H_{\o^{\beta}}^p$ be the $p$-power sequence of this $\H_{\o^{\beta}}$ restricted to $E_{\o^{\beta} p}$ (note that $\o^{\o^{\beta}p}+1 \subset (\o^{\o^{\beta}}+1)^p$). Then $||\lim (\H_{\o^{\beta}}^p)|| \leq \frac{a}{p}$, and $||u_{\o^{\beta}}^{\H_{\o^{\beta}}^p}|| \leq \frac{a_p}{p}$. Let $N$ be such that $\frac{a}{N} \leq \epsilon$, and define $\H_{\o^{\beta+1}} = \coprod_{n \geq N} \H_{\o^{\beta}}^n$. It remains to check (1)-(5) for $\H_{\o^{\beta+1}}$.\
(1) Using that $h(\0_{\o^{\beta+1}})=0$, $$||h|| = \sup_{n \geq N} ||\lim \H_{\o^{\beta}}^n|| = ||\lim \H_{\o^{\beta}}^N|| \leq \frac{a}{N} \leq \epsilon < a.$$\
(2) For irreducible $\delta < \o^{\beta +1}$, we have $\delta \leq \o^{\beta}$. Monotonicity of the transfinite sequence implies $$||u_{\delta}^{\H_{\o^{\beta}}^n}|| \leq ||u_{\o^{\beta}}^{\H_{\o^{\beta}}^n}||,$$ for every $n$. Also, Lemma \[disUnionLemma\] implies $$||u_{\delta}|| = \sup_{n \geq N} ||u_{\delta}^{\H_{\o^{\beta}}^n}||.$$ Putting these inequalities together gives $$||u_{\delta} || = \sup_{n \geq N} ||u_{\delta}^{\H_{\o^{\beta}}^n}|| \leq \sup_{n \geq N} ||u_{\o^{\beta}}^{\H_{\o^{\beta}}^n}|| \leq \frac{a}{N} \leq \epsilon.$$\
(3) For every $\gamma$, Lemma \[disUnionLemma\] and Lemma \[powerLem\] (1) imply $$||u_{\gamma}|| = \sup_{n \geq N} ||u_{\gamma}^{\H_{\o^{\beta}}^n}|| \leq a.$$ Further, for any $\gamma < \al$, there exists $m$ such that $\gamma < \o^{\beta} m$. Using subadditivity (Lemma \[UisZero\]), $||u_{\gamma}^{\H^n_{\o^{\beta}}}|| \leq ||u_{\o^{\beta}m}^{\H^n_{\o^{\beta}}}|| \leq \frac{m}{n} a_n$. Then $$||u_{\gamma}|| = \sup_{n \geq N} ||u_{\gamma}^{\H_{\o^{\beta}}^n}|| \leq \max \Bigl( a_1, \dots, \, a_m, \, \sup_{n > m} \frac{m}{n} a_n \Bigr) < a.$$\
(4) By definition, $h(\0_{\o^{\beta+1}}) = 0$. Let $\gamma < \al$. There exists a $k$ such that $\gamma < \o^{\beta} k$. Then Lemma \[disUnionLemma\], monotonicity, and Lemma \[powerLem\] imply $$u_{\gamma}(\0_{\o^{\beta+1}}) \leq \limsup_{n \rightarrow \infty} ||u_{\gamma}^{\H_{\o^{\beta}}^n}|| \leq \limsup_{n \rightarrow \infty} ||u_{\o^{\beta}k}^{\H_{\o^{\beta}}^n}|| \leq \limsup_{n \rightarrow \infty} \frac{k a}{n} = 0.$$ Also, Lemma \[disUnionLemma\] and Lemma \[powerLem\] imply $$u_{\al}(\0_{\o^{\beta+1}}) \geq \limsup_{n \rightarrow \infty} u_{\al}^{\H_{\o^{\beta}}^n}(\0_{\o^{\beta} n}) = a,$$ which (combining with (3)) implies that $u_{\al}(\0_{\o^{\beta+1}}) = a$.\
(5) For $x \neq \0_{\o^{\beta+1}}$, there exists $n$ such that $x \in E_{\o^{\beta} n}$, which implies that $r(x) \leq \o^{\beta} n$. Then Proposition \[pointwiseBound\] gives that $\al_0(x) \leq \o^{\beta} n +1 < \o^{\beta+1}$. The fact that $\al_0(\0_{\o^{\beta+1}}) = \o^{\beta+1}$ then follows immediately from (3) and (4). Thus $\al_0(\H) = \o^{\beta+1}$.
We assume the statement is true for all $\o^{\xi}$ with $\xi < \beta$, and we need to show that it is true for $\o^{\beta}$. In this case there is a strictly increasing sequence of irreducible ordinals $(\o^{\beta_n})$ with $\sup_n \o^{\beta_n} = \o^{\beta}$, and $E_{\o^{\beta}}$ is homeomorphic to the one-point compactification of the disjoint union of the $E_{\o^{\beta_n}}$ (by Remark \[disjUnion\]). With this homeomorphism, we may assume without loss of generality that $E_{\o^{\beta}}$ is the one-point compactification of the disjoint union of the spaces $E_{\o^{\beta_n}}$. Fix $0 < \epsilon < a$, and let $\{a_n\}$ be a sequence of positive real numbers with $a_n<a$ for all $n$ and $\lim_n a_n =a$. By the induction hypothesis, for each $n>1$, there exists a u.s.c.d. candidate sequence $\H_{\o^{\beta_n}}$ on $E_{\o^{\beta_n}}$ satisfying (1)-(5) with parameters $a_n$, $\frac{\epsilon}{n}$, $\o^{\beta_n}$ and $\delta_n = \o^{\beta_{n-1}}$. Now fix $\delta$ irreducible with $\delta < \o^{\beta}$. Since $\sup_n \o^{\beta_n} = \o^{\beta}$, there exists $N$ such that $\o^{\beta_{N-1}} > \delta$. Let $\H_{\o^{\beta}} = \coprod_{n \geq N} \H_{\o^{\beta_n}}$. All that remains is to verify (1)-(5).\
(1) Using that $h(\0_{\o^{\beta}})=0$, we get $$||h|| = \sup_{n \geq N} ||\lim \H_{\o^{\beta_n}}|| \leq \frac{\epsilon}{N} \leq \epsilon .$$\
(2) Since $\delta < \o^{\beta_{N-1}}$, Lemma \[disUnionLemma\] and monotonicity imply (as in the previous case) $$||u_{\delta}|| \leq \sup_{n \geq N} ||u_{\delta}^{\H_{\o^{\beta_n}}}|| \leq \sup_{n \geq N} ||u_{\o^{\beta_{n-1}}}^{\H_{\o^{\beta_n}}}|| \leq \sup_{n \geq N} \frac{\epsilon}{n} \leq \epsilon.$$\
(3) For any $\gamma$, by construction, $$||u_{\gamma}|| \leq \sup_{n \geq N} ||u_{\gamma}^{\H_{\o^{\beta_n}}}|| \leq a.$$ Further, for $\gamma < \al$, there exists $m$ such that $\gamma < \o^{\beta_m}$. For $n > m$, $||u_{\gamma}^{\H_{\o^{\beta_n}}}|| \leq \frac{\epsilon}{n}$. Then $$||u_{\gamma}|| \leq \sup_{n \geq N} ||u_{\gamma}^{\H_{\o^{\beta_n}}}|| \leq \max \Bigl( a_1, \dots, \, a_m, \, \sup_{n > m} \frac{\epsilon}{n} \Bigr) < a.$$\
(4) By definition, $h(\0_{\o^{\beta}}) = 0$. For any $\gamma < \o^{\beta}$, there exists some $k$ such that for all $n \geq k$, $\o^{\beta_n} > \gamma$. Then $$u_{\gamma}(\0_{\o^{\beta}}) \leq \limsup_{n \rightarrow \infty} ||u_{\gamma}^{\H_{\o^{\beta_n}}}|| \leq \limsup_{n\rightarrow \infty} ||u_{\o^{\beta_{n-1}}}^{\H_{\o^{\beta_n}}}|| \leq \limsup_{n \rightarrow \infty} \frac{\epsilon}{n} = 0.$$\
(5) For any $x \neq \0_{\al}$, there exists $n$ such that $x \in E_{\o^{\beta_n}}$. Then $\al_0(x) \leq r(x) \leq \o^{\beta_n} < \o^{\beta}$. By (3) and (4), $\al_0(\0_{\o^{\beta}}) = \o^{\beta}$. Therefore $\al_0(\H)=\o^{\beta}$.
\[realizationCor\] For all positive real numbers $a$ and non-zero countable ordinals $\al$, there exists a harmonic, u.s.c.d. candidate sequence $\H$ on $K_{\al}$ such that the transfinite sequence corresponding to either $\H$ or $\H|_{\operatorname{ex}(K_{\al})}$ satisfies
1. $||u_{\gamma}|| \leq a$ for all $\gamma$, and $||u_{\gamma}|| < a$ for all $\gamma < \al$;
2. $h(\0_{\al})=0$, and $u_{\al}(\0_{\al})=a$;
3. $\al_0(\H) = \al_0(\H|_{\operatorname{ex}(K_{\al})})= \al$.
Let $\al$ be a non-zero countable ordinal, and suppose the Cantor Normal Form of $\al$ (as in Theorem \[CNF\]) is given by $$\al = \al_1 m_1 + \dots + \al_N m_N.$$ Let $a_1 > \dots > a_N > 0$ be real numbers such that $\sum a_j = a$ and for each $j =1, \dots, N-1$, $$\frac{a_j}{m_j} \geq \sum_{i=j+1}^N a_i.$$ For each $j = 1, \dots, N$, let $\F_j$ be a harmonic, u.s.c.d. candidate sequence given by Theorem \[mainTech\] with parameters $a_j$ and $\al_j$. Define $\H_j$ to be the product sequence $\F_j^{m_j}$ restricted to $K_{\al_j m_j}$, and let $\H = \H_N \times \dots \times \H_1 $ restricted to $K_{\al}$. By definition of $\H$, $h(\0_{\al})=0$. The rest of properties (1)-(3) follow from Lemma \[prodLemma\].
\[alphaPlusOne\] Let $a>0$, and let $\al$ be a countable, infinite ordinal. Then there is a harmonic, u.s.c.d. candidate sequence $\H$ on $K_{\al}$ such that the transfinite sequence corresponding to either $\H$ or $\H|_{\operatorname{ex}(K_{\al})}$ satisfies
1. $||u_{\gamma}|| \leq a$ for all $\gamma$, and $||u_{\gamma}|| < a$ for $\gamma < \al+1$;
2. $h(\0_{\al})=0$ and $u_{\al+1}(\0_{\al}) = a$;
3. $\al_0(\H) = \al_0(\H|_{\operatorname{ex}(K_{\al})}) = \al+1$.
Using Proposition \[Uaffine\], we may deal exclusively with u.s.c.d. candidate sequences on $E_{\al}$ (as opposed to $K_{\al}$), and all properties will carry over to $K_{\al}$.
The proof is executed in two stages. First we prove the statement for the countably infinite, irreducible ordinals. In the second stage, we prove the statement for all countable, infinite ordinals.
**Stage 1.** Let $\al$ be a countably infinite, irreducible ordinal. Let $\al = \o^{\beta}$ (since $\al$ is infinite, $\beta >0$). and let $b = \frac{2}{3} a$. Let $\F$ be given by Theorem \[mainTech\] with parameters $b$, $\al$, $\epsilon$, and $\delta$. Recall from the proof of Theorem \[mainTech\] that we may take $\F = \sqcup \F_n$, where the exact form of the $\F_n$ is as follows. Let $\{a_n\}$ be a sequence of positive real numbers with $a_n < b$ for all $n$ and $\lim_n a_n = b$. If $\beta$ is a successor, then we may take $\F_n = \G^n$, where $\G$ satisfies the conclusions of Theorem \[mainTech\] with parameters $a_n$, $\epsilon$, $\o^{\beta-1}$, and $\delta$. Otherwise, if $\beta$ is a limit with $\beta_n$ increasing to $\beta$, then $\F_n$ satisfies the conclusions of Theorem \[mainTech\] with parameters $a_n$, $\epsilon$, $\o^{\beta_n}$, and $\delta$. Let $\F = (f_k)$, and let $\0_n$ denote the marked point in $E_{\beta_n}$ (so $E_{\beta_n}$ is the domain of $\F_n$). Let $\H = (h_k)$ be defined by the rule $$h_k(x) = \left\{
\begin{array}{ll}
f_k(x) & \text{ if } x \neq \0_n \\
0 & \text{ if } x = \0_n, \medspace k \leq n \\
\frac{b}{2} & \text{ if } x = \0_n, \medspace k > n.
\end{array} \right.$$ By definition, let $h_k(\0_{\al})=0$. Note that $\H$ is again an u.s.c.d. sequence on $E_{\al}$, and $u_{\gamma}^{\H}(x) = u_{\gamma}^{\F}(x)$ for all $\gamma$ and all $x \neq \0_{\al}$. It follows that $u_{\gamma}^{\H}(x) \leq b$ for all $\gamma$ and all $x \neq \0_{\al}$. Computing the transfinite sequence at $\0_{\al}$, we see that $$\begin{aligned}
u_{\ell}^{\H}(\0_{\al}) & = \frac{b}{2}, \text{ for } 1 \leq \ell < \al \\
u_{\al}^{\H}(\0_{\al}) & = b \\
u_{\al+1}^{\H}(\0_{\al}) & = b + \frac{b}{2} = a.\end{aligned}$$ Since $\al_0(\0_{\al}) \leq r(\0_{\al}) + 1 = \al+1$, we conclude that $\al_0(\0_{\al}) = \al+1$. Thus we obtain properties (1)-(3).
**Stage 2.** Let $\al = \o^{\beta_1} m_1 + \dots \o^{\beta_N} m_N$ be the Cantor Normal Form of $\al$.
The construction proceeds by cases. In the first case, suppose $\o^{\beta_N}$ is infinite. Let $a>0$, and select $a_1 > \dots > a_N$ as in Lemma \[prodLemma\]. Let $\F_j$ be given by Lemma \[mainTech\] with parameters $a_j$ and $\o^{\beta_j}$, for $j = 1, \dots , N$. Let $\F'_N$ be given by Stage 1 corresponding to $\frac{a_N}{m_N}$ and $\o^{\beta_N}$. For $j=1, \dots, N-1$, let $\H_j = \F_j^{m_j}$, and for $j=N$, if $m_N >1$, let $\H_j = \F_N^{m_N-1}$. Now let $\H'$ be given by the product (where $\H_N$ is omitted if $m_N = 1$) $$\H' = \F'_N \times (\H_N) \times \dots \times (\H_1),$$ Let $\H$ be the restriction of $\H'$ to $E_{\o^{\al}+1}$. Note that $h(\0_{\al})=0$. Then using Lemmas \[powerLem\] and \[prodLemma\], we conclude that $$\al_0(\H) = \big( \sum_{i=1}^{N-1} \o^{\beta_i} m_i \big) + \o^{\beta_N} (m_N-1) + (\o^{\beta_N}+1) = \al+1.$$
For the second case, we suppose that $\o^{\beta_N}$ is finite, which implies that $\o^{\beta_N} = 1$. Let $a>0$, and select $a_1 > \dots > a_N$ as in Lemma \[prodLemma\], with the additional condition that $\frac{a_{N-1}}{3 m_{N-1}} \geq a_N$. Let $\F_j$ be given by Lemma \[mainTech\] with parameters $a_j$ and $\o^{\beta_j}$, for $j = 1, \dots , N$. Since $\al$ is infinite, it follows that $\o^{\beta_{N-1}}$ is infinite. Let $\F'_{N-1}$ be given by Stage 1 corresponding to $\frac{a_{N-1}}{m_{N-1}}$ and $\o^{\beta_{N-1}}$ (so that the condition $\frac{a_{N-1}}{3 m_{N-1}} \geq a_N$ implies $b/2 \geq a_N$ in the notation of Stage 1). For $j \in \{1, \dots, N-2, N\}$, let $\H_j = \F_j^{m_j}$. If $m_{N-1} >1$, let $\H_{N-1} = \F_{N-1}^{m_{N-1} - 1}$. Now let $\H'$ be given by the product (where $\H_{N-1}$ is omitted if $m_{N-1} = 1$): $$\H' = (\H_N) \times \F'_{N-1} \times (\H_{N-1}) \times \dots \times (\H_1),$$ Let $\H$ be the restriction of $\H'$ to $E_{\o^{\al}+1}$. Note that $h(\0_{\al})=0$. Then the reader may easily adapt the proofs of Lemmas \[powerLem\] and \[prodLemma\] with the additional assumption that $\frac{a_{N-1}}{3 m_{N-1}} \geq a_N$ to check that $$\begin{aligned}
||u_{\o^{\beta_1}m_1 + \dots + \o^{\beta_{N-1}}m_{N-1}}^{\H} || & = \sum_{i=1}^{N-2} a_i + \bigl(\frac{a_{N-1}}{m_{N-1}}(m_{N-1}-1)\bigr) + \frac{a_{N-1}}{m_{N-1}}\bigl(\frac{2}{3}\bigr) \\
||u_{\o^{\beta_1}m_1 + \dots + \o^{\beta_{N-1}}m_{N-1}+1}^{\H} || & = \sum_{i=1}^{N-1} a_i \\
||u_{\o^{\beta_1}m_1 + \dots + \o^{\beta_{N-1}}m_{N-1}+1 + k}^{\H} || & = \sum_{i=1}^{N-1} a_i + \frac{a_N}{m_N} k, \, \text{ for $k = 1, \dots, m_N$,}\end{aligned}$$ and, $$\al_0(\H) = \big( \sum_{i=1}^{N-2} \o^{\beta_i} m_i \big) + \o^{\beta_{N-1}} (m_{N-1}-1) + (\o^{\beta_{N-1}}+1) + m_N = \al+1.$$
\[LimsupToLimitRmk\] In Corollaries \[realizationCor\] and \[alphaPlusOne\], one may further require that $\H|_{\operatorname{ex}(K_{\al})}$ has the following property (P): for any $t$ in $\operatorname{ex}(K_{\al})$, for any sequence $\{s_n\}$ of isolated points in $\operatorname{ex}(K_{\al})$ that converges to $t$, $\limsup_n \tau_k(s_n) = \lim_n \tau_k(s_n)$. Let us prove this fact. In the case $\al = 1$, there is only one sequence of isolated points in $\operatorname{ex}(K_1) \cong \o +1$, and the candidate sequence $\F$ constructed in the proof of Theorem \[mainTech\] satisfies (P). Then we note that if each of the candidate sequences $\F_1, \dots, \F_N$ satisfies this property, then so does the product $\F = \F_1 \times \dots \times \F_N$. To see this fact, note that the projection $\pi_N$ onto the last coordinate of any isolated point $x$ in the product space is not the marked point $\mathbf{0}_N$, and thus $\F(x) = \F_N( \pi_N(x) )$. Hence the product candidate sequence satisfies property (P) because $\F_N$ does. Now suppose there is a sequence $(\F_n)_n$ of candidate sequences such that each $\F_n$ satisfies (P). Let $h^n = \lim \F_n$ and let $I_n$ be the set of isolated points in the domain of $\F_n$. Further suppose that $h^n|_{I_n}$ converges uniformly to $0$. Then $\coprod_n \F_n$ satisfies (P) as well (to see this, note that property (P) is satisfied on the domain of each candidate sequence $\F_n$ separately because $\F_n$ has property (P), and then it is satisfied at the point at infinity because $h^n|_{I_n}$ converges uniformly to $0$). The constructions used in the proofs of Theorem \[mainTech\], Corollary \[realizationCor\] and Corollary \[alphaPlusOne\] only rely on these three types of constructions ($\al = 1$, product sequences, and disjoint union sequences with $h^n|_{I_n}$ tending uniformly to $0$), and thus at each step we may choose candidate sequences satisfying (P). Making these choices yields $\H|_{\operatorname{ex}(K_{\al})}$ with the desired property.
We conclude this section by stating these results in the language of dynamical systems. The following corollary follows from Corollary \[realizationCor\] by appealing to the Downarowicz-Serafin realization theorem (Theorem \[realization\]).
\[DynRealizationCor\] For every countable ordinal $\al$, there is a minimal homeomorphism $T$ of the Cantor set such that $\al$ is the order of accumulation of entropy of $T$.
Characterization of Orders of Accumulation on Bauer Simplices {#Bauer}
=============================================================
For any non-empty countable Polish space $E$, we define $$\rho(E) = \left \{ \begin{array}{ll}
|E|_{CB}-1, & \text{ if } |E|_{CB} \text{ is finite} \\
|E|_{CB}, & \text{ if } |E|_{CB} \text{ is infinite}
\end{array} \right.$$ For any uncountable Polish space $E$, we let $\rho(E) = \o_1$, the first uncountable ordinal.
\[SKdef\] For any metrizable Choquet simplex $K$, we define $$S(K) = \{ \gamma : \text{ there exists a harmonic, u.s.c.d sequence } \H \text{ on } K \text{ with } \al_0(\H) = \gamma \}.$$
Recall our conventions that if $\beta < \o_1$, then $[\al, \beta]$ denotes the ordinal interval $\{ \gamma : \al \leq \gamma \leq \beta \}$, and if $\beta = \o_1$, then $[\al, \beta] = \{ \gamma : \al \leq \gamma < \beta \}$. We also require the use of “open” or “half-open” intervals, which have the usual definitions.
\[bauerThm\] Let $K$ be a Bauer simplex. Then $$S(K) = [0,\rho(\operatorname{ex}(K))].$$
Let $\H$ be a harmonic, u.s.c.d. candidate sequence on $K$. Proposition \[pointwiseBound\] implies that $$\al_0(\H|_{\operatorname{ex}(K)}) \leq \rho(\operatorname{ex}(K)).$$ and it is always true that $\al_0(\H|_{\operatorname{ex}(K)}) < \o_1$. Then since $K$ is Bauer, Proposition \[Uaffine\] implies the same bounds for $\al_0(\H)$. It remains to show that if $\operatorname{ex}(K)$ is countable, then $S(K) \supset [0,\rho(\operatorname{ex}(K))]$, and if $\operatorname{ex}(K)$ is uncountable, then $S(K) \supset [0, \o_1 [$.
Suppose $E = \operatorname{ex}(K)$ is countable. Let $\al < |E|_{CB}$. Then by Proposition \[accumFacts\], there exists $x$ in $E$ such that $r(x) = \al$, which implies that $x$ is isolated in $\Gamma^{\al}(E)$. Let $U$ be a clopen neighborhood of $x$ in $E$ such that $U \cap (\Gamma^{\al}(E)\setminus \{ x \} ) = \emptyset$. Then $|U|_{CB} = \al + 1$ and $|\Gamma^{\al}(U)| = 1$. Then by the classification of countable, compact Polish spaces (Theorem \[classification\]), there is a homeomorphism $g : \o^{\al}+1 \rightarrow U$. Let $\H'$ be the u.s.c.d candidate sequence on $\o^{\al}+1$ given by Corollary \[realizationCor\] with $\al_0(\H') = \al$. Define $\H$ on $K$ to be harmonic extension of the embedded candidate sequence $g\H'$, which is harmonic and u.s.c.d by Lemma \[embedIsHarmUSC\]. Since $\H|_{E \setminus g(\o^{\al}+1)} \equiv 0$, the Embedding Lemma (Lemma \[embeddingThm\]) applies. Since $g(\o^{\al}+1)$ is a compact subset of $\operatorname{ex}(K)$, we obtain that $\al_0(\H) = \al_0(\H') = \al$. Since $\al < |\operatorname{ex}(K)|_{CB}$ was arbitrary, this argument shows that $S(K) \supset [0,|\operatorname{ex}(K)|_{CB}-1]$ (note that since $K$ is Bauer, $\operatorname{ex}(K)$ is compact and $|\operatorname{ex}(K)|_{CB}$ is a successor). If $|\operatorname{ex}(K)|_{CB}$ is infinite, then let $\al = |\operatorname{ex}(K)|_{CB}-1$ and repeat the above argument with $\H'$ given by Corollary \[alphaPlusOne\] so that $\al_0(\H)=\al+1$. In this case we obtain that $S(K) \supset [0,|\operatorname{ex}(K)|_{CB}]$. In any case, we conclude that $S(K) \supset [0,\rho(\operatorname{ex}(K))]$, as desired.
Now suppose $E = \operatorname{ex}(K)$ is uncountable. Fix $\al < \o_1$. Let $g:\o^{\al}+1 \rightarrow E$ be the embedding given by Proposition \[uncountPolish\], and let $\H_{\al}$ be the u.s.c.d. candidate sequence on $\o^{\al}+1$ given by Corollary \[realizationCor\]. Then the harmonic extension $\H$ of the embedded candidate sequence $g\H_{\al}$ on $K$ is harmonic and u.s.c.d. by Lemma \[embedIsHarmUSC\]. Furthermore, $\H$ satisfies $\al_0(\H)=\al_0(\H_{\al}) =\al$, by the Embedding Lemma (as $g(\o^{\al}+1)$ is a compact subset of $\operatorname{ex}(K)$). Since $\al < \o_1$ was arbitrary, $S(K) \supset [0,\o_1 [$.
Orders of Accumulation on Choquet Simplices {#Choquet}
===========================================
In this section we address the extent to which the orders of accumulation that appear on a metrizable Choquet simplex $K$ are constrained by the topological properties of the pair $(\operatorname{ex}(K), \overline{\operatorname{ex}(K)})$.
We will require a relative version of Cantor-Bendixson rank, whose definition we give here.
\[RelCBDerivative\] Given a Polish space $X$ contained in the Polish space $T$, we define the sequence $\{ \Gamma_X^{\al}(T) \}$ of subsets of $T$ using transfinite induction. Let $\Gamma_X^0(T) = T$. If $\Gamma_X^{\al}(T)$ has been defined, then let $\Gamma_X^{\al+1}(T) = \{ t \in T : \exists (t_n) \in \Gamma_X^{\al}(T) \cap X \setminus \{t\} \text{ with } t_n \rightarrow t \}$. If $\Gamma_X^{\beta}(T)$ has been defined for all $\beta < \al$, where $\al$ is a limit ordinal, then we let $\Gamma_X^{\al}(T) = \cap_{\beta < \al} \Gamma_X^{\beta}(T)$.
Note that $\Gamma_X^{\al}(T)$ is closed in $T$ for all $\al$, and $\Gamma_X^{\al}(T) \subset \Gamma_X^{\beta}(T)$ for $\al > \beta$. For $X$ and $T$ Polish, there exists a countable ordinal $\beta$ such that $\Gamma_X^{\al}(T) = \Gamma_X^{\beta}(T)$ for all $\al > \beta$.
\[RelCBRank\] The **Cantor-Bendixson rank of** $T$ **relative to** $X$, denoted $|T|_{CB}^X$, is the least ordinal $\beta$ such that $\Gamma_X^{\al}(T) = \Gamma_X^{\beta}(T)$ for all $\al > \beta$.
If $X$ is countable, then $\Gamma_X^{\al}(T) = \emptyset$ if and only if $\al \geq |T|_{CB}^X$. If $X$ is countable and $T$ is compact, then by the finite intersection property, $|T|_{CB}^X$ is a successor ordinal.
\[ptwiseRelTopRank\] For $t$ in $T$, we also define the pointwise relative topological rank $r_X(t)$ of $t$ with respect to $X$: $$r_X(t) = \left\{
\begin{array}{ll}
\sup \{ \al : t \in \Gamma_X^{\al}(T) \} & \text{if } t \notin \Gamma_X^{|T|^X_{CB}}(T) \\
\o_1 & \text{if } t \in \Gamma_X^{|T|^X_{CB}}(E).
\end{array} \right.$$
It follows that for $X$ countable, for all $t$ in $T$, $r_X(t) \leq |X|_{CB}$, and thus $|T|_{CB}^X \leq |X|_{CB} +1$. Also, $|X|_{CB} \leq |T|_{CB}^X \leq |T|_{CB}$.
For a Polish space $T$, the usual Cantor-Bendixson rank is obtained from the relative version by taking $X = T$ in the above construction. Thus, we have $|T|_{CB}^T = |T|_{CB}$.
Results for Choquet Simplices
-----------------------------
Let $X$ and $T$ be non-empty Polish spaces, with $X \subset T$. If $X$ is countable, let $$\rho_X(T) = \left \{ \begin{array}{ll}
|T|_{CB}^X-1, & \text{ if } |T|_{CB}^X \text{ is finite} \\
|T|_{CB}^X, & \text{ if } |T|_{CB}^X \text{ is infinite}
\end{array} \right.$$ If $X$ is uncountable, let $\rho_X(T) = \o_1$.
Now we present bounds on the set $S(K)$ (see Definition \[SKdef\]) for any metrizable Choquet simplex $K$. Recall our convention that for a countable ordinal $\beta$, $[0,\beta] = \{ \al : 0 \leq \al \leq \beta \}$, but for $\beta= \o_1$, $[0,\beta] = \{ \al : 0 \leq \al < \o_1\} = [0,\o_1 [.$
\[topBounds\] Let $K$ be a metrizable Choquet simplex. Then $$[0, \rho_{\operatorname{ex}(K)}(\overline{\operatorname{ex}(K)}) \, ] \; \subset S(K) \subset [0, \rho(\overline{\operatorname{ex}(K)}) \,].$$
First we prove the lower bound on $S(K)$.
Suppose $\operatorname{ex}(K)$ is uncountable, then by Corollary \[uncountPolish\], for any countable $\al$, there exists a map $g : \o^{\al}+1 \rightarrow \operatorname{ex}(K)$, where $g$ is a homeomorphism onto its image. Let $\F$ be a u.s.c.d. sequence on $\o^{\al}+1$, and let $\H$ be the harmonic extension of the embedded sequence $g\F$ on $K$. $\H$ is a harmonic, u.s.c.d. candidate sequence on $K$ by Lemma \[embedIsHarmUSC\]. Also, $\H|_{\operatorname{ex}(K) \setminus g(\o^{\al}+1)} \equiv 0$. Thus the Embedding Lemma (Lemma \[embeddingThm\]) applies, and then since $g(\o^{\al}+1)$ is a compact subset of $\operatorname{ex}(K)$, we obtain that $\al_0(\H) = \al_0(\F)$. Letting $\F$ vary over all u.s.c.d. candidate sequences on $\o^{\al}+1$, it follows that $S(\M(\o^{\al}+1)) \subseteq S(K)$. By Theorem \[bauerThm\], $S(\M(\o^{\al}+1)) = [0, \rho(\o^{\al}+1)]$. Now $\rho(\o^{\al}+1) = \al$ if $\al$ is finite and $\rho(\o^{\al}+1) = \al+1$ if $\al$ is infinite. In either case, $\rho(\o^{\al}+1) \geq \al$. Hence $S(K) \supset [0,\al]$. Since this inclusion holds for any countable ordinal $\al$, we have that $S(K) \supset [0,\o_1 [$, as desired.
If $\operatorname{ex}(K)$ is countable, then $|\overline{\operatorname{ex}(K)}|_{CB}^{\operatorname{ex}(K)}$ is a successor ordinal. For each ordinal $\al < |\overline{\operatorname{ex}(K)}|_{CB}^{\operatorname{ex}(K)}$, we have $\Gamma_{\operatorname{ex}(K)}^{\al}(\overline{\operatorname{ex}(K)}) \neq \emptyset$. Fix $\al < |\overline{\operatorname{ex}(K)}|_{CB}^{\operatorname{ex}(K)}$, and let $t$ be in $\Gamma_{\operatorname{ex}(K)}^{\al}(\overline{\operatorname{ex}(K)})$. Since $t$ lies in $\Gamma_{\operatorname{ex}(K)}^{\al}(\overline{\operatorname{ex}(K)})$, there exists a map $g : \o^{\al}+1 \rightarrow K$, where $g$ is a homeomorphism onto its image, $g(\o^{\al}+1) \subset \operatorname{ex}(K) \cup \{t\}$ and $g(\0_{\al})=t$, where $\0_{\al}$ is the point $\o^{\al}$ in $\o^{\al}+1$. Given some real number $a>0$, let $\F = (f_k)$ be a u.s.c.d. candidate sequence on $\o^{\al}+1$ with $\al_0(\F)=\al$ and satisfying (1)-(3) of Corollary \[realizationCor\]. Recall that $f_k(\0_{\al}) =0$ for all $k$. Then let $\H' = (h'_k)$ be the embedded candidate sequence $g \F$ on $K$.
Note that for $s$ in $K \setminus \operatorname{ex}(K)$, $(h'_{k+1}-h'_k)(s) = 0$. Also, for $s$ in $\operatorname{ex}(K)$, $(h'_{k+1}-h'_k)(s) \geq 0$. It follows that $h'_{k+1}-h'_k$ is convex on $K$.
Now let $\H = (h_k)$, where $h_k$ is the harmonic extension of $h'_k$ on $K$. By Lemma \[embedIsHarmUSC\], $\H$ is a u.s.c.d. candidate sequence on $K$.
Let $F = g(\o^{\al}+1) \cap \operatorname{ex}(K)$, and note that $\H|_{\operatorname{ex}(K) \setminus F} \equiv 0$. Also $\overline{F} = g(\o^{\al}+1)$ and $\H|_{\overline{F}} = \F \circ g^{-1}$. Applying the Embedding Lemma (Lemma \[embeddingThm\]), we obtain that $\al_0(\H) \leq \al_0(\H|_{\overline{F}}) = \al_0(\F) = \al$. We now show the reverse inequality. Recall that $t = g(\0_{\al})$. For $\gamma < \al$, the Embedding Lemma (Lemma \[embeddingThm\]) implies that $u_{\gamma}^{\H}(t) \leq ||u_{\gamma}^{\F}|| < a$ (where the strict inequality comes from Corollary \[realizationCor\] (1)). Also, $u_{\al}^{\H}(t) \geq u_{\al}^{\F}(\0_{\al}) = a$. From these facts, we have that $\al \leq \al_0^{\H}(t) \leq \al_0(\H)$. Thus $\al_0(\H)=\al$.
Since $\al < |\overline{\operatorname{ex}(K)}|_{CB}^{\operatorname{ex}(K)}$ was arbitrary, we obtain that $S(K) \supset [0,|\overline{\operatorname{ex}(K)}|_{CB}^{\operatorname{ex}(K)} [ $. If $|\overline{\operatorname{ex}(K)}|_{CB}^{\operatorname{ex}(K)}$ is infinite, then we may let $\al = |\overline{\operatorname{ex}(K)}|_{CB}^{\operatorname{ex}(K)}-1$ and repeat the above argument with $\F$ given by Corollary \[alphaPlusOne\] so that $\al_0(\H)=\al+1$. Thus we have that $S(K) \supset [0, \, \rho_{\operatorname{ex}(K)}(\overline{\operatorname{ex}(K)}) \, ]$.
Here we prove the upper bound on $S(K)$. Suppose $\overline{\operatorname{ex}(K)}$ is uncountable. Then $\rho(\overline{\operatorname{ex}(K)}) = \o_1$. Since the order of accumulation of any candidate sequence on $K$ is countable, we have (trivially) that $S(K) \subset [0,\o_1)$. Now suppose $\overline{\operatorname{ex}(K)}$ is countable. If $\H$ is a u.s.c.d., harmonic candidate sequence on $K$, then by Corollary \[CBrankBound\], the restricted sequence $\H|_{\overline{\operatorname{ex}(K)}}$ satisfies $$\al_0(\H|_{\overline{\operatorname{ex}(K)}}) \leq \left\{ \begin{array}{ll}
|\overline{\operatorname{ex}(K)}|_{CB}-1, & \text{ if } |\overline{\operatorname{ex}(K)}|_{CB} \text{ is finite} \\
|\overline{\operatorname{ex}(K)}|_{CB}, & \text{ if } |\overline{\operatorname{ex}(K)}|_{CB} \text{ is infinite},
\end{array} \right.$$ which is exactly the statement that $\al_0(\H|_{\overline{\operatorname{ex}(K)}}) \leq \rho(\overline{\operatorname{ex}(K)})$. Also, the Embedding Lemma (Lemma \[embeddingThm\]) implies that $\al_0(\H) \leq \al_0(\H|_{\overline{\operatorname{ex}(K)}})$. This establishes the upper bound on $S(K)$.
Optimality of Results for Choquet Simplices {#optimality}
-------------------------------------------
In this section we study the optimality of the results in Theorem \[topBounds\].
The following theorem answers a question of Jerome Buzzi, and answers the question of whether the bounds in Theorem \[topBounds\] can be improved using only knowledge of the ordinals $\rho_{\operatorname{ex}(K)}(\overline{\operatorname{ex}(K)})$ and $\rho(\overline{\operatorname{ex}(K)})$.
\[threeAlphas\] Let $\al_1 \leq \al_2 \leq \al_3$ be ordinals such that $\al_1$ and $\al_2$ are countable successors and $\al_3$ is either a countable successor ordinal or $\o_1$. Then there exists a metrizable Choquet simplex $K$ such that $\rho_{\operatorname{ex}(K)}(\overline{\operatorname{ex}(K)}) = \al_1$, $S(K)=[0,\al_2]$, and $\rho(\overline{\operatorname{ex}(K)}) = \al_3$.
We postpone the proof of Theorem \[threeAlphas\] until after the proof of Theorem \[optimal\]. The proofs of these theorems are very similar and we prefer not to repeat the arguments unnecessarily.
Now we address the following question: can the bounds in Theorem \[topBounds\] be improved with knowledge of the homeomorphism class of the compactification $(\operatorname{ex}(K),\overline{\operatorname{ex}(K)})$? We will need some definitions.
If $E$ is a topological space, then a compactification of $E$ is a pair $(\overline{E},g)$, where $\overline{E}$ is a compact, Hausdorff space and $g$ is a homeomorphism of $E$ onto a dense subset of $\overline{E}$.
If $E$ is a topological space and $(\overline{E},g)$ is a compactification of $E$, then we may identify $E$ with $g(E)$ and assume that $E$ is a subset of $\overline{E}$. In such instances, we may refer to $\overline{E}$ as a compactification of $E$, or we may refer to the pair $(E, \overline{E})$ as a compactification. Consider compactifications $(E, \overline{E})$, where $E$ is a topological space and $\overline{E}$ is a compactification of $E$. Suppose there are two such compactifications, $(E_1, \overline{E_1})$ and $(E_2, \overline{E_2})$. We say that the compactifications are homeomorphic, written $(E_1, \overline{E_1}) \simeq (E_2, \overline{E_2})$, if there is a homeomorphism $g : \overline{E_1} \rightarrow \overline{E_2}$ such that $g(E_1)=E_2$. Recall that Theorem \[choquetWeak\] may be strengthened as follows.
\[ChoquetStrong\] Let $E$ be a topological space and $\overline{E}$ a metrizable compactification of $E$. Then there exists a metrizable Choquet simplex $K$ such that $(\operatorname{ex}(K), \overline{\operatorname{ex}(K)}) \simeq (E, \overline{E})$ if and only if $E$ is Polish.
Given a Polish space $E$ and a compactification $\overline{E}$, the proof of Theorem \[optimal\] below involves constructing a metrizable Choquet simplex $K$ such that $(\operatorname{ex}(K), \overline{\operatorname{ex}(K)}) \simeq (E,\overline{E})$ while simultaneously controlling the possible harmonic, u.s.c.d. candidate sequences on $K$. In this sense Theorem \[optimal\] may be viewed as a partial generalization of Theorem \[ChoquetStrong\].
In Theorem \[optimal\], we restrict our attention to metrizable compactifications of Polish spaces. Since we are only interested in studying pairs $(\operatorname{ex}(K),\overline{\operatorname{ex}(K)})$ where $K$ is a metrizable Choquet simplex, Theorem \[ChoquetStrong\] implies that there is no loss of generality in making this restriction.
\[optimal\] Let $E$ be a non-compact, countably infinite Polish space, and let $\overline{E}$ be a metrizable compactification of $E$.
1. If $\overline{E}$ is countable, then for each successor $\beta \in [\rho_E(\overline{E}), \medspace \rho(\overline{E})]$, there exists a Choquet simplex $K$ such that $(\operatorname{ex}(K), \overline{\operatorname{ex}(K)}) \simeq (E, \overline{E})$ and $S(K) = [0, \beta]$.
2. If $E$ is countable and $\overline{E}$ is uncountable, then for each countable ordinal $\beta \geq \rho_E(\overline{E})$, there exists a Choquet simplex $K$ such that $(\operatorname{ex}(K), \overline{\operatorname{ex}(K)}) \simeq (E, \overline{E})$ and $S(K) \supset [0,\beta]$.
Observe that when $E$ is uncountable, Theorem \[topBounds\] gives that for any metrizable Choquet simplex $K$ with $\operatorname{ex}(K)$ homeomorphic to $E$, $S(K) = [0,\o_1[$. The proofs of Theorem \[optimal\] (1) and (2) rely very heavily Lemma \[ConstructKLem\], which in turn relies very heavily on Haydon’s proof (see [@Hay] or [@AE pp. 126-129]) of Theorem \[choquetWeak\].
**Proof of Theorem \[optimal\] (1)**.
### Setup for proof of Theorem \[optimal\] (1)
Let $\beta$ be a successor ordinal with $\rho_{E}(\overline{E}) \leq \beta \leq \rho(\overline{E})$. Let $\beta_0 = \beta$ if $\beta$ is finite, and let $\beta_0 = \beta-1$ if $\beta$ is infinite. For notation, we let $T = \overline{E}$ and $X = E$. Since $T$ is countable and compact, $T \cong \o^{|T|_{CB}-1} n + 1$ for some natural number $n$ (by Theorem \[classification\]). We may assume without loss of generality that $n=1$ (if $n >1$, then $T$ is just the finite disjoint union of the case when $n=1$, and we may repeat the following constructions independently $n$ times). Using this homeomorphism of $T$ and $\o^{|T|_{CB}-1} + 1$, we obtain a well-ordering on $T$ such that the induced order topology coincides with the original topology on $T$. Thus we may assume without loss of generality that $T =\o^{|T|_{CB}-1} + 1$. Also, we fix a complete metric $d( \cdot, \cdot )$ on $T$.
Let $Y \subset T$ be the set $\o^{\beta_0}+1$ in $T = \o^{|T|_{CB}-1} + 1$. Let $Z = Y \setminus X$, which may be empty. There are two cases: either $Y = T$ or $Y \subsetneq T$. The case $Y=T$ occurs if and only if $\beta = \rho(T)$, while the case $Y \subsetneq T$ occurs if and only if $\beta < \rho(T)$. If $Y = T$, then one may ignore the constructions in Sections \[YnoteqTone\], \[YnoteqTtwo\], and \[YnoteqTthree\]. If $Y \subsetneq T$, then $Z$ may be empty. If $Z$ is empty, then one may ignore the construction in Section \[ZnotEmpty\]. We make the convention that an empty sum is zero.
### Definition of the points $z_m, u_m, v_m$ {#ZnotEmpty}
Assuming $Z$ is not empty, we will define distinct points $z_m \in Z$ and $u_m, v_m \in X$. In the simplex $K$, they will satisfy $z_m = \frac{1}{2}(u_m+v_m)$, and it is exactly this formula which allows us to prove that $[0,\beta] \subseteq S(K)$.
Since $T$ is countable, $Z$ is countable, and we may enumerate $Z = \{ z_m \}$ (in the case when $Z$ is finite, this sequence is finite). If $z_m < \o^{|T|_{CB}-1}$ in $T$, then let $u_m = z_m+1$ and $v_m = z_m+2$ (successor ordinals). If $z_m = \o^{|T|_{CB}-1}$ in $T$, we let $u_m = 1$ and $v_m = 2$. Since $X$ is dense in $T$, any isolated point in $T$ must lie in $X$. Therefore any successor ordinals in $T$ must be in $X$. It follows that $u_m, v_m$ are points in $X$.
### Construction of the sets $V_k$ {#YnoteqTone}
Here we will use notations defined previously, such as the relative topological rank, $r_X(x)$, of the point $x$ (Definition \[ptwiseRelTopRank\]) and the relative Cantor-Bendixson derivatives $\Gamma_X^{\al}(Y)$ (Definition \[RelCBDerivative\]). Also, since it is an important hypothesis in this section, we remind the reader that $Y$ is clopen in $T$.
In this section we assume that $T \setminus Y$ is not empty, which occurs exactly when $\beta < \rho(T)$, and we define certain sets $V_k$. The construction of the sets $V_k$ and the points $x_k$ and $y_k$ (see section \[YnoteqTtwo\]) allows one to prove that $S(K) \subseteq [0,\beta]$. In the simplex $K$, all points in the set $V_k$ will lie in the convex hull of $x_k$ and $y_k$, which will imply that the order of accumulation cannot be increased by the points in $V_k \setminus \{ y_k \}$ (see Lemmas \[U\_tLemma\] and \[pointwiseOrderBd\]).
Below, by an interval in a subset $A$ of $T$, we mean the intersection of an interval of $T$ (which may be a singleton) with $A$.
\[VkLemma\] If $T \setminus Y$ is not empty, then there exists a collection $\{V_k\}$ of non-empty subsets of $T$ with the following properties:
1. if $V_k \cap V_j \neq \emptyset$, then $k = j$;
2. for each $V_k$ there exists an ordinal $\al_k \geq 1$ such that $r_X(t) = \al_k$ for all $t$ in $V_k$;
3. each $V_k$ is a clopen interval in $\Gamma_X^{\al_k}(T)$;
4. if $V_k \cap X \neq \emptyset$, then $V_k \cap X = \{ \sup (V_k) \}$;
5. $\Gamma_X^{1}(T) \setminus Y = \cup_k V_k$.
6. $\lim_k \operatorname{diam}(V_k) = 0$.
Suppose $\al \in [1, \rho(T)]$ and the set $A_{\al} = \{ t \in \Gamma_X^{\al}(T) \setminus Y : r_X(t) = \al \}$ is non-empty (which it must be for $\al = 1$ since $Y \neq T$). For $x \in X \cap A_{\al}$, let $a(x) = \min \{ a \in \Gamma_X^1(T) \setminus Y : [a,x] \cap (X \cap A_{\al}) = \{x\} \text{ and } [a,x]\cap \Gamma_X^{\al+1}(T) = \emptyset \}$. Let $U_x = [a(x),x] \cap \Gamma_X^{\al}(T)$ and note that $U_x \subset A_{\al}$. The set $\Gamma_X^{\al+1}(T)$ is closed and does not intersect $A_{\al}$, and the set $X \cap A_{\al}$ has no accumulation points in $A_{\al}$. Thus each $U_x$ is a clopen interval in $\Gamma_X^{\al}(T)$. Now let $U_{\o}^{\al} = A_{\al} \setminus \cup_{x \in X \cap A_{\al}} U_x$, which may be empty.
If $U_{\o}^{\al}$ is non-empty, then we claim that it is also a clopen interval in $\Gamma_X^{\al}(T)$. Let $y_0 = \sup(X \cap A_{\al})$. Note that $Y$ is an initial subinterval of $T$ and $X \cap A_{\al} \subset T \setminus Y$, which implies that $[y_0, \max(T)] \subset T \setminus Y$. We also have that $y_0$ is in $X \cup \Gamma_X^{\al+1}(T)$, which implies that $y_0$ is not in $U_{\o}^{\al}$. We will show that $U_{\o}^{\al} = [y_0+1, \max(T)] \cap \Gamma_X^{\al}(T)$. To see this fact, first note that if $y \leq x$ with $y \in A_{\al}$ and $x \in X \cap A_{\al}$, then $y \in \cup_{x \in X \cap A_{\al}} U_x$. Thus we have that $U_{\o}^{\al} \subset [y_0+1, \max(T)] \cap \Gamma_X^{\al}(T)$. To show the reverse inclusion, we show that $[y_0+1, \max(T)] \cap A_{\al} = [y_0+1, \max(T)] \cap \Gamma_X^{\al}(T)$. We assume for the sake of contradiction that there is a point $t$ in $[y_0+1, \max(T)] \cap \Gamma_X^{\al+1}(T)$. From this assumption and the fact that $[y_0+1, \max(T)]$ is open it follows that $[y_0+1, \max(T)] \cap \Gamma_X^{\al}(T) \cap X$ has $t$ as an accumulation point (and so, in particular, this set is non-empty). If $[y_0+1, \max(T)] \cap \Gamma_X^{\al}(T) \cap X$ contains a single point $s$ with $r_X(s)= \al$, then we see that $s \in X \cap A_{\al}$ and $s > y_0$, which contradicts the definition of $y_0$. Now suppose that for all $s$ in $[y_0+1, \max(T)] \cap \Gamma_X^{\al}(T) \cap X$, $r_X(s) > \al$. Then $[y_0+1, \max(T)] \cap \Gamma_X^{\al}(T) \cap X$ is a non-empty, countable, metrizable space with no isolated points, which implies that it is not Polish. But $[y_0+1, \max(T)] \cap \Gamma_X^{\al}(T)$ is closed in $T$, which implies that it is a $G_{\delta}$ in $T$, and $X$ is Polish in $T$, which implies it is a $G_{\delta}$ in $T$, and the intersection of two $G_{\delta}$ sets is a $G_{\delta}$. Also, any $G_{\delta}$ set in a Polish space is Polish. Thus, $[y_0+1, \max(T)] \cap \Gamma_X^{\al}(T) \cap X$ is Polish, and we arrive at a contradiction.
Let $\{V'_k\}$ be an enumeration of all the non-empty sets $U_x$ and $U_{\o}^{\al}$ constructed above, for any $\al \in [1, \rho(T)]$. The collection $\{ V'_k \}$ satisfies properties (1)-(5) but not necessarily (6). However, given $V'_k$ a clopen interval in $\Gamma_X^{\al}(T)$ contained in $A_{\al}$, we may find a finite collection of pairwise disjoint clopen intervals (in $\Gamma_X^{\al}(T)$) $V'_{k,i}$, contained in $A_{\al}$, whose union is $V'_k$, such that each $V'_{k,i}$ has diameter at most $\frac{1}{k}$. Re-enumerating the collection $\{V'_{k,i} \}$, we obtain the required collection $\{V_k\}$.
Note that since $T \setminus X \subset \Gamma_X^{1}(T)$, we have that $T \setminus (X \cup Y) = \sqcup_k V_k \setminus X$.
### Definition of the points $x_k$ and $y_k$ {#YnoteqTtwo}
The points $x_k$ and $y_k$ are part of the construction that allows one to bound the possible orders of accumulation on $K$ from above.
Assuming $\beta < \rho(T)$, we let $\{V_k\}$ be a collection of non-empty subsets of $T$ given by Lemma \[VkLemma\], and fix a natural number $k$. There are two cases: either $V_k \cap X = \emptyset$ or $V_k \cap X \neq \emptyset$. Suppose $V_k \cap X = \emptyset$. Then choose a point $t_k$ in $V_k$. If $t_k = \sup(T)$, let $x_k=\o^{\beta_0}+3$ and $y_k = \o^{\beta_0}+4$, and otherwise let $x_k = t_k +1$ and $y_k = t_k+2$. If $V_k \cap X \neq \emptyset$, then let $y_k = \sup(V_k)$ (which is in $X$ by conclusion (4) of Lemma \[VkLemma\]). If $y_k = \sup(T)$, let $x_k = \o^{\beta_0}+5$ and otherwise let $x_k = y_k+1$. The fact that the $V_k$ are pairwise disjoint implies that the points $x_k$ and $y_k$ are all distinct. Note that for all $k$, $x_k$ and $y_k$ are in $X$.
Notice that the points $x_k, y_k, z_m, u_m$, and $v_m$ and the sets $V_k$ have been chosen so that (i) the quantities $\operatorname{diam}(V_k)$, $\max_{t \in V_k} \operatorname{dist}(x_k , t)$, and $\max_{t \in V_k} \operatorname{dist}(y_k, t)$ each converge to zero as $k$ tends to infinity, (ii) $d(z_m, u_m)$ and $d(z_m, v_m)$ each converge to zero as $m$ tends to infinity, (iii) the points $x_k, y_k, z_m, u_m$, and $v_m$ are all distinct, (iv) the points $x_k, y_k, u_m, v_m$ are all in $X$, and (v) if $V_k \cap X \neq \emptyset$, then $V_k \cap X = \{y_k\}$ (v) for all $m$ and $k$, $z_m \notin V_k$, and (vi) the sets $V_k$ are pairwise disjoint.
### Definition of $F_k$ and $G_k$ {#YnoteqTthree}
Choose Borel measurable functions $F_k: T \rightarrow [0,1]$ and $G_k : T \rightarrow [0,1]$ with the following properties:
1. $F_k, G_k < 1$ on $T \setminus X$;
2. $F_k$ and $G_k$ are continuous and injective on $V_k$ and $0$ on $T \setminus V_k$;
3. $F_k+G_k = \chi_{V_k}$;
4. $F_k(y_k) = 0$ and $G_k(y_k)=1$.
The existence of such maps follows easily from the fact that $T$ can be order-embedded in $(0,1)$ and $V_k$ is closed.
### Conclusion of the proof of Theorem \[optimal\] (1)
Let $C_n=(\cup_{k=1}^n V_k) \cup \{z_1, \dots, z_n\}$ for each $n$. Consider the collection of points $\{x_k\}\cup\{y_k\}\cup\{u_m\}\cup\{v_m\}$. To each point $x_k$ we associate the function $F_k$. To each point $y_k$ we associate the function $G_k$. To each point $u_m$ or $v_m$, we associate the function $\frac{1}{2}\chi_{z_m}$. Then the hypotheses in Lemma \[ConstructKLem\] are satisfied by the countable collection of closed sets $\{C_n\}\cup\{D_n\}$, the countable collection of points $\{x_k\}\cup\{y_k\}\cup\{u_m\}\cup\{v_m\}$ in $X$, and the associated functions $\{F_k\}\cup\{G_k\}\cup\{\frac{1}{2}\chi_{z_m}\}$. Lemma \[ConstructKLem\] gives a metrizable Choquet simplex $K$ and a homeomorphism $\phi : T \to \overline{\operatorname{ex}(K)}$ such that $\phi(X)=\operatorname{ex}(K)$ and such that for all $t$ in $T \setminus X$, $$\label{phitEqnOne}
\phi(t) = \sum_k F_k(t) \phi(x_k) + G_k(t) \phi(y_k) + \frac{1}{2} \sum_m \chi_{z_m}(t) (\phi(u_m)+\phi(v_m)).$$
\[U\_tLemma\] Let $X,Y,T$, and $K$ be as above. Then for every $t \in T \setminus Y$, there exists an open (in $T$) neighborhood $U_t$ and points $x_t$ and $y_t$ in $X \setminus Y$ such that for all $s$ in $U_t$, either $r_X(s) < r_X(t)$ or else $r_X(s)=r_X(t)$ and $\phi(s) = a_s \phi(x_t) + b_s \phi(y_t)$ in $K$, with $0 \leq a_s, b_s \leq 1$ and $a_s + b_s = 1$.
Let $t \in T \setminus Y$. If $r_X(t)=0$, then $t$ is isolated in $T$ and $t$ is in $X$, since $X$ is dense in $T$. In this case we may choose $U_t = \{t\}$ and the requirement is trivially satisfied.
If $r_X(t) \geq 1$, then $t$ is in $V_k$ for some $k$. Let $U_t$ be any open (in $T$) neighborhood of $t$ with $\Gamma_X^{r_X(t)} \cap U_t \subseteq V_k$ (such a neighborhood exists since $V_k$ is an open interval in in $\Gamma_X^{r_X(t)}(T)$), and let $x_t = x_k$ and $y_t = y_k$. We have that for each $s$ in $U_t$, either $r_X(s) < r_X(t)$ or $s$ is in $V_k$. If $s$ is in $V_k$, then $r_X(s) = r_X(t)$, and it follows from Equation (\[phitEqnOne\]) that $\phi(s) = F_k(s) \phi(x_k) + G_k(s) \phi(y_k)$ in $K$. Also, we have that $F_k(s) + G_k(s) = 1$.
By Lemmas \[pointwiseOrderBd\] and \[SKsubset\], we have that $S(K) \subset [0,\rho(Y)]$. By Lemma \[inclusionLemma\], $S(K) \supset [0, \rho(Y)]$. Thus $S(K)=[0, \rho(Y)] = [0,\beta]$.
*This concludes the proof of Theorem \[optimal\] (1)*.
**Proof of Theorem \[optimal\] (2)**.
### Setup for proof of Theorem \[optimal\] (2)
Let $\beta$ be a successor ordinal with $\beta \geq \rho_E(\overline{E})$. Let $\beta_0 = \beta$ if $\beta$ is finite and let $\beta_0 = \beta-1$ if $\beta$ is infinite. For notation, we let $T = \overline{E}$ and $X = E$. Fix a metric $d$ on $T$ that is compatible with the topology of $T$. Since $T$ is uncountable and compact, $T$ contains an uncountable perfect set $P$. Since $\beta_0$ is countable, $P$ contains a set $Y$ that is homeomorphic to $\o^{\beta_0}+1$. Let $\{a_{\al}\}_{\al=0}^{\o^{\beta_0}}$ be a transfinite sequence of real numbers $a_{\al}$ such that $0 < a_{\al} \leq 1$ and $\sum_{\al \leq \o^{\beta_0}} a_{\al} = 1$ (such a sequence exists since $\o^{\beta_0}$ is countable). Let $Z = Y \setminus X$, and choose an enumeration of $Z = \{z_m\}$. Note that $Z$ may be empty or finite. In the construction to follow, if $Z$ is empty then we do not choose points $u_m$ and $v_m$, and any summation over the index $m$ will be zero by convention.
Let $X_0 = X \sqcup Z = X \cup Y$. Recall that since $X$ is a completely metrizable subset of the the compact metrizable space $T$, $X$ is a $G_{\delta}$ in $T$ (see, for example, [@S]). $Y$ is a $G_{\delta}$ in $T$ because it is compact. Therefore $X_0$ is a $G_{\delta}$ in $T$, since it is the union of two $G_{\delta}$ sets in $T$. Thus we may let $X_0 = \cap_{n \in \N} G_n$, where $G_n$ is open, $G_1 = T$, and $G_{n+1} \subset G_n$. Let $F_n = T \setminus G_n$, which is compact. Fix $n$. Choose a sequence $\epsilon_{\ell}$ strictly decreasing to $0$. Let $D_{\epsilon}(F_n) = \{ t \in T : \operatorname{dist}(t, F_n) \geq \epsilon \}$, which is compact for any $\epsilon$. Then for each $\ell$ there exists a countable collection of open sets $\{U_{\ell}^{j}\}_{j=1}^{\infty}$ such that
- $D_{\epsilon_{\ell}}(F_n) \subset \cup_j U_{\ell}^j \subset D_{\epsilon_{\ell+1}}(F_n)$;
- $\operatorname{diam}(U_{\ell}^j) \leq 2^{-(\ell+n)}$ for all $j$;
- $\operatorname{diam}(U_{\ell}^j)$ tends to $0$ as $j$ tends to infinity;
- the collection $\{U_{\ell}^j\}$ separates the points in $D_{\epsilon_{\ell}}(F_n)$.
Then we may enumerate the collection of all sets $U_{\ell}^j$ to form the sequence $\{V_k^n\}_{k=1}^{\infty}$. Repeating this procedure for all $n$, we obtain a collection of open sets $V_k^n$ such that $\operatorname{diam}(V_k^n) \leq 2^{-n}$ and $\operatorname{diam}(V_k^n)$ converges to $0$ as $k$ tends to infinity with $n$ fixed. The sets $V_k^n$ also satisfy $\cup_k V_k^n=G_n$ and separate points in $G_n$, for each $n$. For each $n$ and $k$, let $g_k^n(t) = \min( \operatorname{dist}(t, T \setminus V_k^n), 1)$ and $f_k^n = 2^{-k} g_k^n$. Then for each $n$, $\sum_k f_k^n$ converges uniformly on $T$. Now let $$h^n_k(t) = \left\{ \begin{array}{ll}
0, & \text{ if } \sum_k f^n_k(t) =0 \\
\frac{f^n_k(t)}{\sum_k f^n_k(t)}, & \text{ if } \sum_k f^n_k(t) > 0
\end{array} \right.$$ The functions $h^n_k$ are all continuous and satisfy $h^n_k(t) > 0$ if and only if $t \in V_k^n$. Furthermore, $\sum_k h^n_k = \chi_{G_n}$. Now we let $p^n_k = h^n_k \cdot \chi_{T \setminus G_{n+1}}$ and notice that $\sum_n \sum_k p^n_k = \chi_{T \setminus (X \cup Y)}$. Also, the collection $p^n_k$ separates points in the sense that if $t \neq s$ with $t$ and $s$ in $T \setminus (X \cup Y)$, then there exists $n$ and $k$ such that $p^n_k(t) >0$ and $p^n_k(s) = 0$.
Using induction (on $m$, $n$, and $k$ simultaneously) and the fact that $X$ is dense in $T$, we choose points $u_m$, $v_m$, $x^n_k$, and $y^n_k$ in $X$ such that (i) $d(z_m, u_m) \leq a_{\al_{z_m}}$ and $d(z_m, v_m) \leq a_{\al_{z_m}}$, (ii) for each $m$, $u_m$ and $v_m$ are not accumulation points of $Y$ (which is possible since the isolated points of $Y$, corresponding to successors of $\o^{\beta_0}+1$, are dense in $Y$ and the set $X \setminus Y$ accumulates at each of the isolated points of $Y$ that is not in $X$) (iii) $x^n_k$ and $y^n_k$ are in $V^n_k$, and (iv) the union of all of these points is a disjoint union.
### Conclusion of the proof of Theorem \[optimal\] (2)
Let $C_n = (T \setminus G_n) \cup \{z_1, \dots, z_n\}$. To each point $x^n_k$ or $y^n_k$, we associate the function $\frac{1}{2}p^n_k$. To each point $u_m$ or $v_m$, we associate the function $\frac{1}{2}\chi_{z_m}$. Then the hypotheses of Lemma \[ConstructKLem\] are satisfied by the countable collection of closed sets $\{C_n\}$, the countable collection of points $\{x^n_k\}\cup\{y^n_k\}\cup\{u_m\}\cup\{v_m\}$ in $X$, and the associated functions$\{\frac{1}{2}p^n_k\}\cup\{\frac{1}{2}\chi_{z_m}\}$. Lemma \[ConstructKLem\] gives a metrizable Choquet simplex $K$ and a homeomorphism $\phi : T \to \overline{\operatorname{ex}(K)}$ such that $\phi(X)=\operatorname{ex}(K)$ and such that for all $t$ in $T \setminus X$, $$\phi(t) = \frac{1}{2} \sum_n \sum_k p^n_k(t) (\phi(x^n_k)+\phi(y^n_k)) + \frac{1}{2} \sum_m \chi_{z_m}(t)( \phi(u_m)+\phi(v_m)).$$ By Lemma \[inclusionLemma\], $S(K) \supset [0,\beta]$.
*This concludes the proof of Theorem \[optimal\] (2)*.
**Proof of Theorem \[threeAlphas\]**.
Fix $\al_1 \leq \al_2 \leq \al_3$ as above. Let $X_1 = \o^{\al_1}+1$, and let $T_1=X_1$. If $\al_2$ is finite, let $T_2 = \o^{\al_2}+1$, and if $\al_2$ is infinite, let $T_2 = \o^{\al_2-1}+1$. In either case, let $X_2$ be all the isolated points (successors) in $T_2$. Let $S$ be a non-empty compact subset of $(0,1) \times \{0\}$ in $\R^2$, chosen so that if $\al_3$ is finite, then $\rho(S) = \al_3-1$, and otherwise $\rho(S) = \al_3$. Let $X_3$ be a bounded, countable subset of $\R^2 \setminus (\R \times \{0\})$ whose set of accumulation points is exactly $S$. Let $T_3 = X_3 \cup S$. Now we let $T = T_1 \sqcup T_2 \sqcup T_3$ and $X = X_1 \sqcup X_2 \sqcup X_3$. Below we will construct a Choquet simplex $K$ such that $(X,T) \simeq (\operatorname{ex}(K),\overline{\operatorname{ex}(K)})$. Let $Y = T_1 \sqcup T_2$, and $Z = Y \setminus X$. Note that $Z$ is actually just the set of accumulation points in $T_2$. We have $$\rho_X(T) = \rho(X_1) = \al_1, \; \rho(T) = \rho(T_3) = \al_3, \text{ and } \rho(Y) = \rho(T_2) = \al_2.$$ Let $Z = \{z_m\}$. If $z_m < \sup(T_2)$, choose $u_m=z_m+1$ and $v_m=z_m+2$. If $z_m = \sup(T_2)$, choose $u_m =1$ and $v_m=2$ in $T_2$. Let $x_0$ and $y_0$ be a choice of two isolated points in $X_3$. Let $F:T \rightarrow [0,1]$ be the function such that, for a point $t$ in $T$, $$F(t) = \left \{ \begin{array}{ll} s, & \text{ if } t = (s,0) \in S \\
0, & \text{ otherwise }.
\end{array}
\right.$$ Let $G: T \rightarrow [0,1]$ be such that for $t$ in $T$, $$G(t) = \left \{ \begin{array}{ll} 1-s, & \text{ if } t = (s,0) \in S \\
0, & \text{ otherwise }.
\end{array}
\right.$$ Let $C_n=S\cup\{z_1, \dots, z_n\}$ for each $n$. To each point $u_m$ or $v_m$, we associate the function $\frac{1}{2}\chi_{z_m}$. To the point $x_0$, we associate the function $F$, and to the point $y_0$ we associate the function $G$. Then the hypotheses in Lemma \[ConstructKLem\] are satisfied by the collection of closed sets $\{C_n\}$, the collection of points $\{x_0,y_0\}\cup\{u_m,v_m\}$, and the associated functions $\{F,G\}\cup\{\frac{1}{2}\chi_{z_m}\}$. Lemma \[ConstructKLem\] gives a Choquet simplex $K$ and a homeomorphism $\phi : T \to \overline{\operatorname{ex}(K)}$ such that $\phi(X)=\operatorname{ex}(K)$ and such that for all $t$ in $T \setminus X$, $$\label{phitEqnTwo}
\phi(t) = F(t) \phi(x_0) + G(t) \phi(y_0) + \frac{1}{2}\sum_m \chi_{z_m}(t) \bigl(\phi(u_m) + \phi(v_m)\bigr).$$
It follows immediately that $\rho_{\operatorname{ex}(K)}(\overline{\operatorname{ex}(K)}) = \rho_X(T) = \al_1$ and $\rho(\overline{\operatorname{ex}(K)}) = \rho(T) = \al_3$. We show that $S(K) = [0,\al_2]$.
\[U\_tLemmaThreeAlphas\] Let $X,Y,T$, and $K$ be as above. Then for every $t \in T \setminus Y$, there exists an open (in $T$) neighborhood $U_t$ and points $x_t$ and $y_t$ in $X \setminus Y$ such that for all $s$ in $U_t$, either $r_X(s) < r_X(t)$ or else $r_X(s)=r_X(t)$ and $\phi(s) = a_s \phi(x_t) + b_s \phi(y_t)$ in $K$, with $0 \leq a_s, b_s \leq 1$ and $a_s + b_s = 1$.
Let $t$ be in $T \setminus Y = T_3$. If $t$ is in $X_3$, then $t$ is isolated in $T$ and we let $U_t = \{t\}$. In this case the requirement on $U_t$ is trivially satisfied.
If $t$ is in $T_3 \setminus X_3$, then $t$ is in $S$ and $r_X(t) = 1$. Let $U_t$ be any open neighborhood of $t$ in $T_3$, and let $x_t = x_0$ and $y_t = y_0$. Let $s$ be in $U_t$. If $s$ is in $X_3$, then $r_X(s) = 0 < r_X(t)$. If $s$ is in $T_3 \setminus X_3$, then $s$ is in $S$, and then we have $r_X(s)=1$ and by Equation (\[phitEqnTwo\]), $\phi(s) = F(s) \phi(x_0)+G(s)\phi(y_0)$, where $F(s) +G(s) = 1$.
Now by Lemmas \[pointwiseOrderBd\] and \[SKsubset\], we have that $S(K) \subset [0,\rho(Y)]$. By Lemma \[inclusionLemma\] we have that $S(K) \supset [0,\rho(Y)]$. Then $S(K)=[0,\rho(Y)] = [0,\al_2]$.
*This concludes the proof of Theorem \[threeAlphas\]*.
### Helpful Lemmas
Recall the following notations. Suppose $T$ is a compact, metrizable space. Let $\mathcal{SM}(T)$ denote the set of all signed, totally finite, Borel measures on $T$. Recall that $\mathcal{SM}(T) = C_{\R}(T)^*$, and therefore $\mathcal{SM}(T)$ inherits the structure of a normed topological vector space over $\R$. For $\mu$ in $\SM(T)$, let $\mu = \mu_1 - \mu_2$ be the Jordan decomposition of $\mu$. Let $|\mu| = \mu_1 + \mu_2$. The norm on $\SM(T)$ is then given by $||\mu|| = |\mu|(T)$. We will use $\mathcal{SM}(T \setminus X)$ to denote the set of measures $\mu$ in $\mathcal{SM}(T)$ such that $|\mu|(X) = 0$. We write $\mathcal{SM}_{prob}(T) = \{ \mu \in \mathcal{SM} : \mu \geq 0, \; ||\mu||=1\}$, and for any subset $\M$ of $\mathcal{SM}(T)$, $\M_1 = \{\mu \in \mathcal{SM} : ||\mu ||\leq 1\}$. Let $\epsilon_{x_k}$ be the point mass at $x_k$.
\[ConstructKLem\] Let $T$ be a compact, metric space, and let $X$ be a dense, Polish subset of $T$. Suppose $\{C_n\}$ is a countable collection of closed subsets of $T$. Suppose $\{w_k\}$ is a countable collection of distinct points in $X$, and to each point $w_k$ there is an associated Borel measurable function $H_k : T \rightarrow [0,1]$. Let $W_k = \operatorname{supp}(H_k)$. Furthermore, suppose the following conditions are satisfied:
1. $C_n \subset C_{n+1}$ for all $n$, $C_0 = \emptyset$, and $\cup_n C_n \setminus X = T \setminus X$;
2. $\sum_k H_k \leq 1$ and $\bigl(\sum_k H_k\bigr)|_{T\setminus X} \equiv 1$;
3. for all $t$ in $T \setminus X$, $H_k(t)<1$;
4. if $H_k(s)=H_k(t)$ for all $k$ with $t,s \in T \setminus X$, then $s=t$;
5. for each $k$, there exists $n_k$ such that $W_k \subset C_{n_k+1} \setminus C_{n_k}$, and with this notation, $H_k$ is continuous on $C_{n_k+1}$;
6. $\max_{t \in W_k} d(t,w_k)$ converges to $0$ as $k$ tends to infinity;
7. if $H_k(x) >0$ for $x$ in $X$, then $x=w_k$ and $H_k(w_k)=1$.
Let $\xi : \mathcal{SM}(T) \rightarrow \mathcal{SM}(T)$, where for $\mu$ in $\mathcal{SM}(T)$, $$\xi(\mu) = \mu - \sum_k \bigl(\int H_k \; d\mu \bigr)\, \epsilon_{w_k}.$$ Let $\M = \{ \xi(\mu) : \mu \in \mathcal{SM}(T \setminus X) \}$, and let $q : \mathcal{SM}(T) \rightarrow \mathcal{SM}(T)/ \M$ be the natural quotient map. Let $\psi : T \rightarrow \mathcal{SM}_{prob}(T)$ be $\psi(t)=\epsilon_t$, and let $\phi = q \circ \psi$. Finally, let $K = q(\mathcal{SM}_{prob}(T))$. Then
1. $\M$ is a closed linear subspace of $\mathcal{SM}(T)$, and thus $\phi$ is continuous;
2. $K$ is a metrizable Choquet simplex;
3. $\phi$ is injective on $T$;
4. $\operatorname{ex}(K) = \phi(X)$;
5. for $t$ in $T \setminus X$, $\phi(t) = \sum_k H_k(t) \phi(w_k)$ in $K$.
This lemma is almost entirely a restatement of Haydon’s proof (see [@Hay] or [@AE pp. 126-129]) of Theorem \[choquetWeak\]. There are two differences. Firstly, we allow $H_k$ to be positive on $X$, while Haydon does not. Secondly, we claim that $\phi$ is injective on all of $T$, whereas Haydon claims injectivity of $\phi$ only on $X$. For the proofs of properties (2), (4) and (5), theses differences do not play any role, and one may repeat Haydon’s proof. For this reason, we will prove only (1) and (3).
\(1) Note that $\mathcal{M}$ is a linear subspace. Recall that $\mathcal{M}$ being closed in the weak\* topology is equivalent to $\mathcal{M}_1$ being closed in the weak\* topology (a proof of this general fact, which follows from the Banach-Dieudonné Theorem, can be found in [@SW]). Let $\sigma_i$ be a sequence of measures in $\mathcal{M}_1$. Since $|| \xi(\mu) || \geq ||\mu||$ for all $\mu$ in $\mathcal{SM}(T \setminus X)$, there exist measures $\mu_i$ in $\mathcal{SM}(T \setminus X)_1$ such that $\sigma_i = \xi(\mu_i)$. Since each $C_{n}$ is compact, each $\mathcal{SM}(C_n)_1$ is compact in the weak\* topology. Therefore a diagonal argument gives a subsequence $\{\nu_i\}$ of $\{\mu_i\}$ such that there exist measures $\hat{\nu}^n \in \mathcal{SM}(C_{n+1})$ such that $\nu_i|_{C_{n+1}}$ converges to $\hat{\nu}^n$ for each $n$. (We note that there may not be a measure $\hat{\nu}$ such that $\hat{\nu}|_{C_{n+1}} = \hat{\nu}^{n}$, since $\hat{\nu}^n|_{C_n}$ may not equal $\hat{\nu}^{n-1}$.)
Let $\nu^n = \hat{\nu}^n|_{C_{n+1} \setminus X}$, and let $\charfun_A$ be the characteristic function of the set $A$. Now define $$\nu = \sum_{n} \nu^n|_{C_{n+1} \setminus C_n} = \sum_{k,n} H_k \charfun_{C_{n+1} \setminus C_n} \nu^n,$$ where the second equality follows from hypotheses (i) and (ii). Let $g^n_k = H_k \cdot \charfun_{C_{n+1} \setminus C_n}$, and note that by hypothesis (v), $g^n_k$ is continuous on $C_{n+1}$ for all $k$ and $n$. Then $g^n_k \nu_i$ weak\* converges to $g^n_k \hat{\nu}^n$ as $i$ tends to infinity. Since $||\nu_i|| \leq 1$ and $g^n_k \nu_i$ weak\* converges to $g^n_k \hat{\nu}^n$, it follows that $||\nu|| \leq 1$ and $\nu$ is in $\mathcal{SM}(T \setminus X)$. Let us show that $\xi(\nu_i)$ converges to $\xi(\nu)$ in the weak\* topology. Let $f \in C_{\R}(T)$. Then for any $\mu$ in $\mathcal{SM}(T \setminus X)$ we have $$\int f d\xi(\mu) = \; \int f d\mu - \sum_{k} \int f(w_k) H_k \; d\mu
= \; \sum_{k} \int (f-f(w_k))H_k \; d\mu
= \; \sum_{n,k} \lambda^n_k(\mu) ,$$ where $$\lambda^n_k(\mu) = \int (f-f(w_k))g^n_k \; d\mu.$$ For each $k$ and $n$, we have that $(f-f(w_k))g^n_k$ is continuous on $C_{n+1}$ by hypothesis (v). Therefore, by the choice of subsequence $\nu_i$, $\lambda^n_k(\nu_i)$ converges to $\lambda^n_k(\hat{\nu}^n)$. Also, using hypothesis (vii), we have that if $H_k(x) \charfun_{C_{n+1} \setminus C_n}(x) >0$ for some $x$ in $X$, then $x = w_k$. It follows that $$\begin{aligned}
\lambda^n_k( \hat{\nu}^n) = & \int (f-f(w_k))H_k \charfun_{C_{n+1}\setminus C_n} \; d\hat{\nu}^n \\
= & \int (f-f(w_k))H_k \charfun_{C_{n+1}\setminus C_n} \; d\nu^n \\
& \quad + (f(w_k)-f(w_k))H_k(w_k) \charfun_{C_{n+1} \setminus C_n}(w_k) \hat{\nu}^n(\{w_k\}) \\
= & \int (f-f(w_k))H_k \charfun_{C_{n+1} \setminus C_n} \; d\nu^n \\
= & \lambda^n_k(\nu^n) = \lambda^n_k(\nu).\end{aligned}$$ This calculation shows that $\lambda^n_k(\nu_i)$ converges to $\lambda^n_k(\nu)$. For fixed $f$ in $C_{\R}(T)$ and $\epsilon >0 $, there exists a $\delta > 0$ such that $|f(t) - f(s)| < \epsilon$ whenever $d(t,s) < \delta$, by uniform continuity. Then since $\max_{t \in W_k} d(w_k,t)$ tends to zero as $k$ tends to infinity, there exists $k_0$ such that for $k \geq k_0$ and $z \in W_k$, $|f(z) - f(w_k)| < \epsilon$. Then for any $\mu$ in $\mathcal{SM}(T \setminus X)$, and $K \geq k_0$ and $N$, $$\begin{aligned}
\Bigl|\int f d\xi(\mu) - \sum_{n=1}^N \sum_{k=1}^{K} \lambda^n_k(\mu) \Bigr| & = \; \Bigl|\sum_{n > N} \sum_{k > K} \lambda^n_k(\mu) \Bigr| \\
& \leq \; \sum_{n > N} \sum_{k > K} \int |f-f(w_k)|g^n_k \; d|\mu|
\leq \; \epsilon ||\mu||,\end{aligned}$$ which implies that $\sum_{n=1}^N \sum_{k=1}^{K} \lambda^n_k(\mu)$ converges uniformly on $\mathcal{SM}(T \setminus X)_1$ to $\int f \; d\xi(\mu)$. Using this uniform convergence and the fact that $\lambda^n_k(\nu_i)$ converges to $\lambda^n_k(\nu)$, we conclude that $\xi(\nu_i)$ converges to $\xi(\nu)$.
\(3) Suppose that $\phi(t) = \phi(s)$, or equivalently, $\epsilon_t - \epsilon_s$ is in $\mathcal{M}$. Thus there exists a measure $\mu$ in $\mathcal{M}(T \setminus X)$ such that $\epsilon_t - \epsilon_s = \xi(\mu)$. We consider three cases.
If $t$ and $s$ are both in $X$, then we notice that $\xi(\mu)=\epsilon_t-\epsilon_s$ has no mass in $T \setminus X$. As $w_k$ are all in $X$, it follows from the definition of $\xi(\mu)$ that we must have $|\mu|(T \setminus X) = 0$, which implies that $\mu$ is the zero measure. Then $\xi(\mu)$ is the zero measure, and we have that $\epsilon_t = \epsilon_s$, which means that $t = s$.
If exactly one of $t$ and $s$ is in $X$, then we may assume without loss of generality that $t \in X$ and $s \in T \setminus X$. In this case, we notice that $- \epsilon_s = (\epsilon_t - \epsilon_s)|_{T \setminus X} = \xi(\mu)|_{T \setminus X} = \mu|_{T \setminus X} = \mu$. Therefore we conclude that $$\epsilon_t = \xi(\mu) + \epsilon_s
= \xi(\mu) - \mu
= \sum_{k} H_k(s) \cdot \epsilon_{w_k} . \label{epsilont}$$ From this equation, we deduce that $t = w_k$ for some $k$. Then $H_k(s) = 1$, which gives a contradiction since $H_k < 1$ on $T \setminus X$ by hypothesis (iii). If $t$ and $s$ are both in $T \setminus X$, then we see that $\xi(\mu) = \epsilon_t - \epsilon_s = \xi(\mu)|_{T\setminus X} = \mu$, which implies that $\int H_k d\mu = 0$ for all $k$. Hence $H_k(t) =H_k(s)$ for all $k$. By hypothesis (iv), we obtain that $t=s$.
\[pointwiseOrderBd\] Let $K$ be a metrizable Choquet simplex. Let $X$ be a Polish subspace of a compact metrizable space $T$, and let $Y$ be clopen in $T$. Let $\phi : T \rightarrow \overline{\operatorname{ex}(K)}$ be a homeomorphism with $\phi(X)=\operatorname{ex}(K)$. Suppose that for every point $t$ in $T \setminus Y$, there exists an open (in $T$) neighborhood $U_t$ and points $x_t$ and $y_t$ in $X \setminus Y$ such that for all $s$ in $U_t$, either $r_X(s) < r_X(t)$ or else $r_X(s)=r_X(t)$ and $\phi(s) = a_s \phi(x_t) + b_s \phi(y_t)$ in $K$, with $0 \leq a_s, b_s \leq 1$ and $a_s + b_s = 1$. Then for each point $t$ in $T \setminus Y$, and any harmonic, u.s.c.d. candidate sequence $\H$ on $K$, $$\label{pointwiseBdEqn}
\al_0^{\H|_{\phi(T)}}(t) \leq \left\{
\begin{array}{ll}
r_X(t) & \text{ if } r_X(t) \text{ is finite} \\
r_X(t) +1 & \text{ if } r_X(t) \text{ is infinite}.
\end{array} \right.$$
For the sake of notation, we identify $X,Y$, and $T$ with their images under $\phi$. Observe that $T \setminus Y$ is clopen in $T$. Thus, for every $t$ in $T \setminus Y$, $u_{\beta}^{\H|_T}(t) = u_{\beta}^{\H|_{T\setminus Y}}(t)$ for all ordinals $\beta$, which implies $\al_0^{\H|_T}(t) = \al_0^{\H|_{T \setminus Y}}(t)$. For the sake of notation, we assume that $\H$ is defined only on $T \setminus Y$ and $u_{\beta}^{\H} = u_{\beta}$.
Now we prove the lemma by transfinite induction on $\al = r_X(t)$. For $\al = 0$, we have that $r_X(t) = 0$, and thus $t$ is isolated in $T$. Then $\al_0^{\H}(t) = 0$.
Suppose the lemma holds for all $t$ in $T \setminus Y$ such that $r_X(t) < \al$. If $\al$ is finite, let $\delta = \al$. If $\al$ is infinite, let $\delta = \al+1$. We now prove that for all $t$ in $T \setminus Y$ with $r_X(t)=\al$ and all $\gamma \geq \delta$, $u_{\gamma}(t)=u_{\delta}(t)$. The proof of this statement is by transfinite induction on $\gamma$.
Let $\gamma > \delta$ be a successor ordinal, and let $t$ be in $T \setminus Y$ with $r_X(t) = \al$. Let $U_t$ be an open neighborhood of $t$, and let $x_t$ and $y_t$ be corresponding to $U_t$ according the hypotheses. Fix $\epsilon > 0$. Choose $k_0$ such that $\max (\tau_k(x_t), \tau_k(y_t)) \leq \epsilon$ for all $k \geq k_0$. Then if $(s_n)$ is a sequence in $U_t$ with $r_X(s_n) < r_X(t)$ for all $n$, then using the inductive hypotheses, we get $$(u_{\gamma-1}+\tau_k)(s_n) = (u_{\delta-1}+\tau_k)(s_n).$$ If $(s_n)$ is a sequence in $U_t$ with $r_X(s_n) \geq r_X(t)$, then by the hypotheses, we have that $r_X(s_n) = r_X(t) = \al$ and $s_n = a_{s_n} x_t +b_{s_n} y_t$. Then by the induction hypothesis on $\gamma$ and the harmonicity of $\tau_k$, we have that $$(u_{\gamma-1}+\tau_k)(s_n) = u_{\delta}(s_n) + a_{s_n} \tau_k(x_t) + b_{s_n} \tau_k(y_t) \leq \epsilon.$$ Thus we may conclude that $$\widetilde{(u_{\gamma-1}+\tau_k)}(t) \leq \max \bigl( \widetilde{(u_{\delta-1}+\tau_k)}(t), u_{\delta}(t) + \epsilon \bigr).$$ Letting $k$ tend to infinity, we obtain that $u_{\gamma}(t) \leq u_{\delta}(t) + \epsilon$. Since $\epsilon$ was arbitrary, we have that $u_{\gamma}(t) = u_{\delta}(t)$.
Now let $\gamma > \delta$ be a limit ordinal, and let $t$ be in $T \setminus Y$ with $r_X(t) = \al$. Fix $U_t$, $x_t$, and $y_t$ as in the hypotheses. Note that by the induction hypotheses, if $s$ is in $U_t$, then $u_{\beta}(s) = u_{\delta}(s)$ for all $\beta < \gamma$. Then $\sup_{\beta < \gamma} u_{\beta}(s) = u_{\delta}(s)$ for all $s$ in $U_t$. Taking upper semi-continuous envelope at $t$, we have that $u_{\gamma}(t) = u_{\delta}(t)$.
We conclude that for all $t$ in $T \setminus Y$ with $r_X(t) = \al$, $\al_0^{\H}(t) \leq \delta$, as desired.
\[SKsubset\] Let $K$ be a metrizable Choquet simplex. Let $X$ be a Polish subspace of a compact metrizable space $T$, and let $Y$ be clopen in $T$. Let $\phi : T \rightarrow \overline{\operatorname{ex}(K)}$ be a homeomorphism with $\phi(X)=\operatorname{ex}(K)$. Suppose that for each point $t$ in $T \setminus Y$ and any harmonic, u.s.c.d. candidate sequence $\H$ on $K$, Equation (\[pointwiseBdEqn\]) holds. Further, suppose that $\rho_X(T) \leq \rho(Y)$. Then $S(K) \subset [0,\rho(Y)]$
Let $\H$ be a harmonic, u.s.c.d. candidate sequence on $K$. For $t$ in $Y$, we have that $\al_0^{\H|_T}(t) = \al_0^{\H|_Y}(t)$ since $Y$ is open in $T$. By Remark \[accumVarPrin\], $\al_0^{\H|_Y}(t) \leq \al_0(\H|_Y)$. By Proposition \[pointwiseBound\], $\al_0(\H|_Y) \leq \rho(Y) $. Putting these facts together, we obtain $\al_0^{\H|_T}(t) \leq \rho(Y)$ for all $t$ in $Y$.
For $t$ in $T \setminus Y$, Equation (\[pointwiseBdEqn\]) gives that if $r_X(t)$ is finite, then $\al_0^{\H|_T}(t) \leq r_X(t)$, and if $r_X(t)$ is infinite, then $\al_0^{\H|_T}(t) \leq r_X(t)+1$. Since $X$ is countable and $T$ is compact, $|T|_{CB}^X$ is a successor, and we have $r_X(t) \leq |T|_{CB}^X -1$. If $|T|_{CB}^X$ is finite, then for all $t$ in $T \setminus Y$ we have $\al_0^{\H|_T}(t) \leq r_X(t) \leq |T|_{CB}^X-1 = \rho_X(T)$. If $|T|_{CB}^X$ is infinite, then for all $t$ in $T \setminus Y$ we have $\al_0^{\H|_T}(t) \leq r_X(t)+1 \leq |T|_{CB}^X = \rho_X(T) \leq \rho(Y)$.
We have shown that for all $t$ in $T$, $\al_0^{\H|_T}(t) \leq \rho(Y)$. Taking supremum over all $t$ in $T$, we have that $\al_0(\H|_T) \leq \rho(Y)$. Now using the Embedding Lemma (Lemma \[embeddingThm\]), we get that $\al_0(\H) \leq \al_0(\H|_T) \leq \rho(Y)$. Hence $S(K) \subset [0,\rho(Y)]$.
\[inclusionLemma\] Let $K$ be a metrizable Choquet simplex. Let $X$ be a Polish subspace of a compact metric space $T$, and let $Y$ be a subset of $T$ with $Y \cong \o^{\beta_0}+1$, where $\beta_0$ is a countable ordinal. Let $\phi : T \rightarrow \overline{\operatorname{ex}(K)}$ be a homeomorphism with $\phi(X)=\operatorname{ex}(K)$. Let $Y \setminus X = \{z_m\}$. Suppose there is countable collection of distinct points $W = \{u_m\}\cup\{v_m\}$ in $X$ such that each point $w$ in $W$ is isolated in $Y \cup W$ and for each $z_m$ in $Y \setminus X$, $\phi(z_m)=\frac{1}{2}(\phi(u_m)+\phi(v_m))$. Further suppose that $d(u_m,z_m)$ and $d(v_m,z_m)$ both tend to $0$ as $m$ tends to infinity. Then $S(K) \supset [0,\rho(Y)]$.
For the sake of notation, we identify $X,Y,W$ and $T$ with their images under $\phi$ and refer to these sets as subsets of $K$. Then let $g : Y \to \o^{\beta_0}+1 $ be a homeomorphism. For any $\gamma$ in $[0,\beta_0]$, there is an u.s.c.d. candidate sequence $\F$ on $\o^{\gamma}+1$ given by Corollary \[realizationCor\] with $\al_0(\F) = \gamma$. Since $\o^{\gamma}+1 \subset \o^{\beta_0}+1$, we may extend $\F$ to a u.s.c.d. candidate sequence on $\o^{\beta_0}+1$ (still denoted $\F$) by letting $\F$ be uniformly $0$ off of $\o^{\gamma}+1$. Note that $\F$ on $\o^{\beta_0}+1$ still has the properties stated in Corollary \[realizationCor\]. We now construct a harmonic, u.s.c.d. sequence $\H$ on $K$ such that $\al_0(\F) = \al_0(\H)$. Let $\F = (f_k)$ be given as above. Then let $\H' = (h'_k)$ be the candidate sequence on $K$ defined as follows. For $t$ in $K$, let $$h'_k(t) = \left \{ \begin{array}{ll}
0, & \text{ if } t \notin Y \cup W \\
f_k(g(t)), & \text{ if } t \in Y \setminus W \\
f_k(g(z_m)), & \text{ if } t = u_m \text{ or } t = v_m.
\end{array} \right.$$ We claim that for each $k$, $h'_{k+1}-h'_k$ is convex and u.s.c. Let $t$ be in $K$. If $t$ is in $X$, then $(h'_{k+1}-h'_k)(t) = \int_X (h'_{k+1}-h'_k) d\P_t$ since $\P_t = \epsilon_t$. If $t$ is in $K \setminus (Y \cup W)$, then $0 =(h'_{k+1}-h'_k)(t) \leq \int_X (h'_{k+1}-h'_k) d\P_t$. If $t$ is in $(Y \cup W) \setminus X = Y \setminus X = Z$, then $t=z_m$ for some $m$, and we have that $\P_{z_m} = \frac{1}{2}(\epsilon_{u_m}+\epsilon_{v_m})$. Then $$\begin{aligned}
(h'_{k+1}-h'_k)(z_m) = (f_{k+1}-f_k)(g(z_m)) & = \frac{1}{2}\Bigl( (h'_{k+1}-h'_k)(u_m) + (h'_{k+1}-h'_k)(v_m) \Bigr) \\ & = \int_X h'_{k+1}-h'_k d\P_t.\end{aligned}$$ We have shown that $h'_{k+1}-h'_k$ is convex.
Let us prove that $h'_{k+1}-h'_{k}$ is u.s.c. Since $\{u_m\}$, $\{v_m\}$ and $\{z_m\}$ each have the same limit points, which are in $Y$ (since $\{z_m\}$ is in $Y$ and $Y$ is closed), we obtain that $Y\cup W$ is compact in $K$. Thus if $t$ is in $K \setminus (Y \cup W)$, then $\widetilde{(h'_{k+1}-h'_k)}(t)=0 = (h'_{k+1}-h'_k)(t)$. For $t$ in $Y \setminus W$, assume $\{t_n\}$ is a sequence in $K \setminus \{t\}$ converging to $t$ in $K$. Since $(h'_{k+1}-h'_k)|_{K \setminus (Y \cup W)} \equiv 0$, we may assume that $t_n$ lies in $Y \cup W$ for all $n$. For each $n$, if $t_n$ is not in $Y$, then there exists a natural number $m_n$ such that $t_n \in \{u_{m_n}, v_{m_n}\}$. If $t_n$ is in $Y$, then $(h'_{k+1}-h'_k)(t_n) = (f_{k+1}-f_k)(g(t_n))$, and if $t_n$ is not $Y$, then there exists a natural number $m_n$ such that $t_n \in \{u_{m_n},v_{m_n}\}$ and $(h'_{k+1}-h'_k)(t_n) = (f_{k+1}- f_k)(g(z_{m_n}))$. By the choice of $\{u_m\}$ and $\{v_m\}$, we have that$\{z_{m_n}\}$ also converges to $t$. Then since $\F$ is u.s.c.d. and $g$ is continuous, we have that $\limsup_n (h'_{k+1}-h'_k)(t_n) \leq (f_{k+1}-f_{k})(g(t)) = (h'_{k+1}-h'_k)(t)$. Thus $\widetilde{(h'_{k+1}-h'_k)}(t) = (h'_{k+1}-h'_k)(t)$. For $t$ in $W$, $t$ is isolated in $Y \cup W$, and we conclude that $\widetilde{(h'_{k+1}-h'_k)}(t) = (h'_{k+1}-h'_k)(t)$. Thus $(h'_{k+1}-h'_k)$ is u.s.c.
Now for $t$ in $K$, let $\H = (h_k)$, where $h_k$ is the harmonic extension of $h'_k$ on $K$. $\H$ is harmonic by definition. Fact \[fonEharIsUSC\] states that the harmonic extension of a non-negative, convex, u.s.c. function on a Choquet simplex $K$ is a harmonic, u.s.c. function on $K$. Applying this fact to each element in the sequence $(h'_{k+1}-h'_k)$, we obtain that $\H$ is a harmonic, u.s.c.d. candidate sequence.
Let $F = (Y \cap X) \cup W$ and $L = \overline{F} = Y \cup W$. Note that $\H|_{X \setminus F} \equiv \H'|_{X \setminus F} \equiv 0$, which implies that we may apply the Embedding Lemma (Lemma \[embeddingThm\]). The Embedding Lemma gives that for all ordinals $\al$ and all $t$ in $K$, $$u^{\H}_{\al}(t) = \max_{\mu \in \pi^{-1}(t)} \int_{L} u_{\al}^{\H|_L} \; d\mu.$$
Let us now show that for all $t$ in $K$, $$\label{computeH}
u^{\H}_{\al}(t) = \max_{\mu \in \pi^{-1}(t)} \int_{Y} u_{\al}^{\H|_L} \; d\mu = \max_{\mu \in \pi^{-1}(t)} \int_{Y} u_{\al}^{\H|_Y} \; d\mu = \max_{\mu \in \pi^{-1}(t)} \int_{Y} u_{\al}^{\F} \circ g \; d\mu.$$ The first equality in Equation (\[computeH\]) has already been justified as an application of the Embedding Lemma. The second equality in (\[computeH\]) will be justified by showing that for all ordinals $\al$, $u_{\al}^{\H|_L}|_{L \setminus Y} \equiv 0$ and $u_{\al}^{\H|_L}|_Y = u_{\al}^{\H|_Y}$. Recall that $\H|_Y = \F' \circ g$, where $\F' = (f'_k)$ is the candidate sequence on $\o^{\beta_0}+1$ defined in terms of $\F = (f_k)$ as follows. If $t$ is in $(\o^{\beta_0}+1) \setminus g(W \cap Y)$, then $f'_k(t) = f_k(t)$, and if $t$ is in $g(W \cap Y)$, then $f'_k(t) = f_k(z_m)$ for $t = g(u_m)$ or $t = g(v_m)$. Since $g$ is a homeomorphism, we have that $u_{\al}^{\H|_Y} = u_{\al}^{\F'} \circ g$ for all ordinals $\al$. Then we will justify the third equality in Equation (\[computeH\]) by proving that $u_{\al}^{\F} = u_{\al}^{\F'}$ for all ordinals $\al$. We proceed with these steps and then conclude the proof of the lemma using Equation (\[computeH\]).
Notice that for all $t$ in $W$, $r_L(t) = 0$ ($t$ is isolated in $L$ by hypothesis). Thus, if $t \in W$, then $u_{\al}^{\H|_L}(t) = 0$ for all $\al$.
For $t$ in $Y$, suppose there is a sequence $s_n \in W$ such that $s_n$ converges to $t$ and $\limsup_{s \to t} \tau_k^{\H|_L}(s) = \lim_n \tau_k^{\H|_L}(s_n)$. Since $s_n$ is in $W$, for each $n$ there exists $m_n$ such that $s_n \in \{u_{m_n}, v_{m_n}\}$. Then $\tau_k^{\H|_L}(s_n) = \tau_k^{\H|_L}(z_{m_n})$, $z_{m_n}$ also converges to $t$, and since $z_{m_n}$ is in $Y$, $\tau_k^{\H|_Y}(z_{m_n}) = \tau_k^{\H|_L}(z_{m_n})$. Thus $\limsup_{s \to t} \tau_k^{\H|_L}(s) = \lim_n \tau_k^{\H|_Y}(z_{m_n})$. By these considerations, we have that for all $t$ in $Y$, $\widetilde{\tau_k^{\H|_L}}(t) = \widetilde{\tau_k^{\H|_Y}}(t)$. Letting $k$ tend to infinity gives that $u_1^{\H|_L}(t) = u_1^{\H|_Y}(t)$, for all $t$ in $Y$.
Now we show by transfinite induction that $u_{\al}^{\H|_Y} = u_{\al}^{\H|_L}|_Y$ for all ordinals $\al$. The equality holds for $\al = 1$ by the previous paragraph. Suppose by induction that it holds for some ordinal $\al$. For the sake of notation, we allow $s = t$ in the following limit suprema. Also, the limit supremum over an empty set is assumed to be $0$ by convention. For $t$ in $Y$, the induction hypothesis implies that $$\begin{aligned}
\widetilde{(u_{\al}^{\H|_L}+\tau_k)}(t) = & \max \Bigl( \limsup_{\substack{ s \to t \\ s \in W}} u_{\al}^{\H|_L}(s)+\tau_k(s), \, \limsup_{\substack{ s \to t \\ s \in Y}} u_{\al}^{\H|_L}(s)+\tau_k(s) \Bigr) \\
= & \max \Bigl( \limsup_{\substack{ s \to t \\ s \in W}} \tau_k(s), \, \limsup_{\substack{ s \to t \\ s \in Y}} u_{\al}^{\H|_Y}(s)+\tau_k(s) \Bigr)\end{aligned}$$ Taking the limit as $k$ tends to infinity gives that $$u_{\al+1}^{\H|_L}(t) = \max \Bigl( u_1^{\H|_L}(t), \, u_{\al+1}^{\H|_Y}(t)\Bigr) = \max \Bigl( u_1^{\H|_Y}(t), \, u_{\al+1}^{\H|_Y}(t)\Bigr) = u_{\al+1}^{\H|_Y}(t).$$ Thus we conclude that $u_{\al+1}^{\H|_Y} = u_{\al+1}^{\H|_L}|_Y$, proving the successor case of our induction. For the limit case, suppose the equality holds for all ordinals $\beta$ less than a limit ordinal $\al$. Then for $t$ in $Y$, we have $$\begin{aligned}
u_{\al}^{\H|_L}(t) = & \max \Bigl( \limsup_{\substack{ s \to t \\ s \in W}} \sup_{\beta < \al} u_{\beta}^{\H|_L}(s), \, \limsup_{\substack{ s \to t \\ s \in Y}} \sup_{\beta < \al} u_{\beta}^{\H|_L}(s) \Bigr) \\
= & \max \Bigl( 0, \limsup_{\substack{ s \to t \\ s \in Y}} \sup_{\beta < \al} u_{\beta}^{\H|_Y}(s) \Bigr) \\
= & u_{\al}^{\H|_Y}(t),\end{aligned}$$ which concludes the limit step of the transfinite induction.
Now we turn our attention towards showing that $u_{\al}^{\F'} = u_{\al}^{\F}$ for all ordinals $\al$. By Remark \[LimsupToLimitRmk\], we assume (without loss of generality) that $\F$ has the property (P) that for $t$ in $\o^{\beta_0}+1$, $$\label{LimsupToLimitEqn}
\limsup_{\substack{s \to t \\ r(s) = 0}} \tau_k^{\F}(s) = \lim_{\substack{s \to t \\ r(s) = 0}} \tau_k^{\F}(s).$$ We also require the following topological fact. For every point $t$ in $Y \setminus I$, there is a sequence in $I \setminus W$ that tends to $t$, where $I$ is the set of isolated points in $Y$. To prove this fact, let $t$ be a point with $r(t) \geq 1$ and let $U$ be an open (in $Y$) neighborhood of $t$. Suppose for the sake of contradiction that $(I \setminus W) \cap U = \emptyset$. Since $Y \cong \o^{\beta_0}+1$ (a countable, compact Polish space), we have that $I$ is dense in $Y$ and $\Gamma^1(Y) \setminus \Gamma^2(Y)$ is dense in $\Gamma^1(Y)$. Since $\Gamma^1(Y) \setminus \Gamma^2(Y)$ is dense in $\Gamma^1(Y)$, there is a point $t'$ in $U$ with $r(t')=1$. Since $I$ is dense in $Y$, there is a sequence $w_n$ in $I \cap U$ tending to $t'$. Since $(I \setminus W) \cap U = \emptyset$, we must have that $w_n$ is in $W$ and then there is a sequence $m_n$ such that $w_n \in \{u_{m_n},v_{m_n}\}$ for all $n$. Then $z_{m_n}$ tends to $t'$. Note that $z_{m_n}$ is not in $W$ by hypothesis, and since $r(t')=1$, we must have that $z_{m_n}$ is isolated in $Y$ for all large $n$. Thus $(I \setminus W) \cap U \neq \emptyset$, a contradiction.
Using that $\F$ satisfies property (P) and the topological fact from the previous paragraph, let us show that for any non-isolated point $t$ in $\o^{\beta_0}+1$, we have $\widetilde{\tau_k^{\F'}}(t) = \widetilde{\tau_k^{\F}}(t)$. First note that for every sequence $s_n$ converging to $t$, there is a sequence $t_n$ converging to $t$ such that $\tau_k^{\F'}(s_n) = \tau_k^{\F}(t_n)$: if $s_n$ is not in $g(W \cap Y)$, then let $t_n = s_n$, and if $s_n$ is in $g(W \cap Y)$, then there exists $m_n$ such that $s_n \in \{g(u_{m_n}),g(v_{m_n})\}$, and one may take $t_n = g(z_{m_n})$. It follows that $\limsup_{s \to t} \tau_k^{\F'}(s) \leq \limsup_{s \to t} \tau_k^{\F}(s)$. Also, since $t$ is not isolated, $t$ is not in $g(W \cap Y)$ and $\tau_k^{\F'}(t) = \tau_k^{\F}(t)$. We deduce that $\widetilde{\tau_k^{\F'}}(t) \leq \widetilde{\tau_k^{\F}}(t)$. Now we show the reverse inequality. If $s_n$ is a sequence converging to $t$ with $r(s_n) >0$, then $s_n$ is not in $g(W \cap Y)$ and thus $\tau_k^{\F'}(s_n) = \tau_k^{\F}(s_n)$. In such a case, we have $\limsup_n \tau_k^{\F'}(s_n) = \limsup_n \tau_k^{\F}(s_n)$. Now let $s_n$ be a sequence converging to $t$ with $r(s_n)=0$. By the topological fact from the previous paragraph, there is a sequence $t_n$ of isolated points in $\o^{\beta_0}+1$ that are not in $g(W \cap Y)$ such that $t_n$ converges to $t$. Using the fact that $\F$ satisfies property (P) (see Equation (\[LimsupToLimitEqn\])), we have $\limsup_n \tau_k^{\F}(s_n) = \limsup_n \tau_k^{\F}(t_n)$. Since the points $t_n$ are not in $g(W \cap Y)$ we also have that $\limsup_n \tau_k^{\F}(t_n) = \limsup_n \tau_k^{\F'}(t_n) \leq \limsup_{s \to t} \tau_k^{\F'}(s)$. We have shown that for every sequence $s_n$ converging to $t$, $\limsup_n \tau_k^{\F}(s_n) \leq \limsup_{s \to t} \tau_k^{\F'}(s)$. It follows that $\widetilde{\tau_k^{\F'}}(t) \geq \widetilde{\tau_k^{\F}}(t)$, and therefore we have shown that $\widetilde{\tau_k^{\F'}}(t) = \widetilde{\tau_k^{\F}}(t)$.
Finally, we show that for all ordinals $\al$, $u_{\al}^{\F'} = u_{\al}^{\F}$ by transfinite induction on $\al$. We make the conventions that we allow $s = t$ in the following limit suprema, and the limit supremum over an empty set is $0$. Note that if $t$ is isolated in $\o^{\beta_0}+1$, then $u_{\al}^{\F}(t) = 0 = u_{\al}^{\F'}(t)$ for all $\al$, and thus we need only show the equality at non-isolated points $t$ in $\o^{\beta_0}+1$. For the sake of induction, suppose the equality holds for an ordinal $\al$. Let $t$ be in $(\o^{\beta_0}+1) \setminus g(I)$. For every sequence $s_n$ converging to $t$, there is a sequence $t_n$ converging to $t$ such that $(u_{\al}^{\F'}+\tau_k^{\F'})(s_n) = (u_{\al}^{\F}+\tau_k^{\F})(t_n)$: if $s_n$ is not in $g(W \cap I)$, then let $t_n = s_n$, and if $s_n$ is in $g(W \cap I)$, then there exists $m_n$ such that $s_n \in \{g(u_{m_n}),g(v_{m_n})\}$, and one may take $t_n = z_{m_n}$. It follows that $\limsup_{s \to t} (u_{\al}^{\F'} + \tau_k^{\F'})(s) \leq \limsup_{s \to t} (u_{\al}^{\F} + \tau_k^{\F})(s)$. Thus we have that $\widetilde{(u_{\al}^{\F'} + \tau_k^{\F'})}(t) \leq \widetilde{(u_{\al}^{\F} + \tau_k^{\F})}(t)$. Now we show the reverse inequality. Let $s_n$ be a sequence in $g(I)$ converging to $t$. Then $(u_{\al}^{\F} + \tau_k^{\F})(s_n) = \tau_k^{\F}(s_n)$ and so $\limsup_n (u_{\al}^{\F} + \tau_k^{\F})(s_n) = \limsup_n \tau_k^{\F}(s_n) \leq \widetilde{\tau_k^{\F}}(t) = \widetilde{\tau_k^{\F'}}(t)$ (recall that we showed the last equality in the previous paragraph). Now let $s_n$ be a sequence in $(\o^{\beta_0}+1) \setminus g(I)$ converging to $t$. Since $s_n$ is not isolated, $s_n$ is not in $g(W \cap Y)$, and we have $(u_{\al}^{\F} + \tau_k^{\F})(s_n) = (u_{\al}^{\F'} + \tau_k^{\F'})(s_n)$. Also, $\limsup_n (u_{\al}^{\F'} + \tau_k^{\F'})(s_n) \leq \widetilde{(u_{\al}^{\F'} + \tau_k^{\F'})}(t)$. Combining these considerations, we have shown that $$\widetilde{(u_{\al}^{\F} + \tau_k^{\F})}(t) \leq \max\Bigl( \widetilde{\tau_k^{\F'}}(t), \widetilde{(u_{\al}^{\F'} + \tau_k^{\F'})}(t) \Bigr) = \widetilde{(u_{\al}^{\F'} + \tau_k^{\F'})}(t).$$ Then we deduce that $\widetilde{(u_{\al}^{\F} + \tau_k^{\F})} = \widetilde{(u_{\al}^{\F'} + \tau_k^{\F'})}$. Taking the limit in $k$ gives that $u_{\al+1}^{\F} = u_{\al+1}^{\F'}$, which concludes the successor step of the transfinite induction. For the limit step, assume that $u_{\beta}^{\F} = u_{\beta}^{\F'}$ for all ordinals $\beta$ less than a limit ordinal $\al$. We show that $u_{\al}^{\F} = u_{\al}^{\F'}$. For $t$ in $\o^{\beta_0}+1$, the induction hypothesis gives that (allowing $s = t$ in the the limit suprema) $$u_{\al}^{\F}(t) = \limsup_{s \to t} \sup_{\beta < \al} u_{\beta}^{\F}(s) = \limsup_{s \to t} \sup_{\beta < \al} u_{\beta}^{\F'}(s) = u_{\al}^{\F'}(t).$$ We conclude that $u_{\al}^{\F} = u_{\al}^{\F'}$ for all ordinals $\al$. This fact completes the verification of Equation (\[computeH\]).
It follows immediately from Equation (\[computeH\]) that $\al_0(\H) \leq \al_0(\F) = \gamma$. We now show the reverse inequality. Let $\0_{\gamma}$ be the marked point in Corollary \[realizationCor\], and let $t=g^{-1}(\0_{\gamma})$. Then $u_{\gamma}^{\H}(t) \geq u_{\gamma}^{\F}(\0_{\gamma}) = a$. For an arbitrary $\al < \gamma$, we also have that $u_{\al}^{\H}(t) \leq ||u_{\al}^{\F}|| < a$ by Equation (\[computeH\]) and Corollary \[realizationCor\] (1). Thus $\gamma = \al_0(t) \leq \al_0(\H)$, and we conclude that $\al_0(\H) = \gamma$.
Since $\gamma \leq \beta_0$ was arbitrary, we deduce that $S(K) \supset [0,\beta_0]$. For $\beta$ finite, $\beta_0 = \beta$ and the proof is finished in this case. On the other hand, if $\beta$ is infinite, then $\beta_0 = \beta-1$ and we may repeat the above argument starting with $\F$ on $\o^{\beta_0}+1$ given by Corollary \[alphaPlusOne\] such that $\al_0(\F)=\beta_0+1$. In this case, we conclude that $S(K) \supset [0,\beta_0+1] = [0, \beta]$, which concludes the proof.
Open Questions {#OpenQuestions}
--------------
In general, our analysis leaves open the following problem.
For a metrizable Choquet simplex $K$, what is $S(K)$?
Theorem \[topBounds\] completely answers this question when $\rho_{\operatorname{ex}(K)}(\overline{\operatorname{ex}(K)}) = \rho(\overline{\operatorname{ex}(K)})$. In particular, when $K$ is Bauer or when $\operatorname{ex}(K)$ is uncountable, Theorem \[topBounds\] gives a complete answer. In general, Theorem \[topBounds\] gives upper and lower bounds on $S(K)$.
Theorem \[threeAlphas\] shows that the bounds in Theorem \[topBounds\] cannot be improved using only knowledge of the ordinals $\rho_{\operatorname{ex}(K)}(\overline{\operatorname{ex}(K)})$ and $\rho(\overline{\operatorname{ex}(K)}))$. Theorem \[optimal\] (1) shows that if $\overline{\operatorname{ex}(K)}$ is countable, then the bounds in Theorem \[topBounds\] cannot be improved using only knowledge of the homeomorphism class of the compactification $(\operatorname{ex}(K),\overline{\operatorname{ex}(K)})$. Theorem \[optimal\] (2) shows that the upper bound in Theorem \[topBounds\] cannot be improved using only knowledge of the homeomorphism class the compactification $(\operatorname{ex}(K),\overline{\operatorname{ex}(K)})$. Thus we have the following question remaining.
\[Qone\] Let $E$ be a countable, non-compact Polish space, and let $\overline{E}$ be an uncountable metrizable compactification of $E$. Let $\beta$ be a successor in $[\rho_E(\overline{E}), \o_1[$. Must there exist a metrizable Choquet simplex $K$ such that $(E,\overline{E}) \simeq (\operatorname{ex}(K),\overline{\operatorname{ex}(K)})$ and $S(K) = [0,\beta]$?
Also, when $E$ is countable and $\overline{E}$ is uncountable, we do not know whether the upper bound on $S(K)$ may be attained. We state this problem as a question as follows.
\[Qtwo\] Let $E$ be a countable, non-compact Polish space, and let $\overline{E}$ be an uncountable metrizable compactification of $E$. Must there exist a metrizable Choquet simplex $K$ such that $(E,\overline{E}) \simeq (\operatorname{ex}(K),\overline{\operatorname{ex}(K)})$ and $S(K) = [0,\o_1 [$?
If the answers to Questions \[Qone\] and \[Qtwo\] are affirmative, then one could conclude that the bounds in \[topBounds\] cannot be improved using knowledge of the homeomorphism class of the compactification $(\operatorname{ex}(K),\overline{\operatorname{ex}(K)})$, and furthermore, one could conclude that these bounds are obtained.
Notice that for every simplex $K$ for which we can compute $S(K)$, $S(K)$ is either $[0,\o_1 [$ or $[0,\beta]$ for a countable successor $\beta$. This observation leads to the following two questions.
\[Qthree\] If $K$ is a metrizable Choquet simplex, must $S(K)$ be an ordinal interval?
\[Qfour\] If $K$ is a metrizable Choquet simplex, must $S(K)$ be either $[0,\o_1 [$ or $[0,\beta]$ for a countable successor $\beta$?
If the answers to Questions \[Qone\], \[Qtwo\], \[Qthree\], and \[Qfour\] are all affirmative, then these results would give a complete description of the constraints imposed on orders of accumulation by the compactification of the ergodic measures for a dynamical system.
**Acknowledgment:** The authors would like to express special thanks to Mike Boyle, who contributed many helpful ideas, conversations, and hours to this work.
Entropy Structures, Symbolic Extensions, and Dynamical Systems {#Rel2DynSys}
==============================================================
For general references on the ergodic theory of topological dynamical systems, see [@Glas; @Pet; @W]. For a topological dynamical system $(X,T)$, we write $M(X,T)$ to denote the space of Borel probability measures on $X$ which are invariant under $T$. We give $M(X,T)$ the weak\* topology. It is well known that in this setting $M(X,T)$ is a metrizable, compact, convex subset of a locally convex topological vector space (see, for example, [@Glas; @Pet]). The set of extreme points of $M(X,T)$ is the set of ergodic measures, $M_{\erg}(X,T)$. Furthermore, the fact that each measure $\mu$ in $M(X,T)$ has a unique ergodic decomposition (see [@Glas; @Pet]) translates to the fact that $M(X,T)$ is a Choquet simplex. Since we are only interested in simplices arising from dynamical systems, we consider only metrizable Choquet simplices. It was shown in [@D3] that every metrizable Choquet simplex $K$ can be obtained as the space of invariant Borel probability measures for a dynamical system.
We write $h : M(X,T) \rightarrow [0, \medspace \infty)$ to denote the function that assigns to each measure $\mu$ in $M(X,T)$ its metric entropy. For any dynamical system $(X,T)$, Boyle and Downarowicz defined a reference candidate sequence $\H_{ref}(X,T)$ on $M(X,T)$ that is u.s.c.d. and harmonic. Further, Downarowicz defined an **entropy structure** on $M(X,T)$ to be any candidate sequence on $M(X,T)$ that is uniformly equivalent to $\H_{ref}$ (see Section \[candSeqs\] for definitions). Almost all known methods of defining or computing entropy can be adapted to form an entropy structure [@D]. The work of Downarowicz and Serafin [@DS] implies the following realization theorem:
\[realization\] Let $\H$ be a candidate sequence on a Choquet simplex $K$ that is uniformly equivalent to a harmonic candidate sequence with u.s.c. differences. Then $\H$ is (up to affine homeomorphism) an entropy structure for a minimal homeomorphism of the Cantor set.
The importance of Theorem \[realization\] lies in the fact that it allows one to translate questions in the theory of entropy structures and dynamical systems into the terms of functional analysis. To understand the theory of entropy structure in dynamical systems, it helps to consider symbolic extensions, and we briefly recall this theory.
A good introduction to symbolic dynamical systems is given in [@LM]. For any finite set $\cA$, we refer to $\cA^{\Z}$ as the full shift on $\cA$. The elements of $\cA$ are referred to as symbols. We give $\cA^{\Z}$ the product topology induced by the discrete topology on $\cA$, which makes $\cA^{\Z}$ a compact metrizable space. Then the left-shift map, $\sigma : \cA^{\Z} \rightarrow \cA^{\Z}$, given by $\sigma(x)_n = x_{n+1}$, is a homeomorphism of $\cA^{\Z}$. If $Y$ is closed subset of $\cA^{\Z}$ satisfying $\sigma(Y) = Y$ and $S = \sigma|_Y$, then we refer to $(Y,S)$ as a symbolic dynamical system, or possibly a subshift of $\cA^{\Z}$.
Let $(X,T)$ be a dynamical system. A **symbolic extension** of $(X,T)$ is a subshift $(Y,S)$ of a full shift on a finite number of symbols, along with a continuous surjection $\pi: Y \rightarrow X$ such that $T \pi = \pi S$.
We think of a symbolic extension as a “lossless finite encoding” of the dynamical system $(X,T)$ [@D].
Downarowicz introduced the study of the entropy of symbolic extensions at the level of measures [@D2].
If $(Y,S)$ is a symbolic extension of $(X,T)$ with factor map $\pi$, then the **extension entropy function**, $h_{ext}^{\pi} : M(X,T) \rightarrow [0,\infty)$, is defined for $\mu$ in $M(X,T)$ by $$h_{ext}^{\pi}(\mu) = \sup \{ h(\nu) : \pi^* \mu = \nu \}.$$ The **symbolic extension entropy function** of a dynamical system $(X,T)$, $h_{sex} : M(X,T) \rightarrow [0, \medspace \infty]$, is defined for $\mu$ in $M(X,T)$, as $$h_{sex}(\mu) = \inf \{ h_{ext}^{\pi}(\mu) : \pi \text{ is the factor map of a symbolic extension of } (X,T) \}.$$ and the **residual entropy function**, $h_{res}: M(X,T) \rightarrow [0,\infty]$ is defined for $\mu$ in $M(X,T)$ as $$h_{res}(\mu) = h_{sex}(\mu)-h(\mu).$$
The study of symbolic extensions is related to entropy structures by the following striking result.
\[sexEntThm\] Let $(X,T)$ be a dynamical system with entropy structure $\H$. Then $$h_{sex} = h + u_{\al_0(\H)}^{\H},$$
This theorem relates the notion of how entropy emerges on refining scales to the symbolic extensions of a system, showing that there is a deep connection between these topics. Using this connection, some progress has been made in understanding the symbolic extensions of certain classes of dynamical systems, with particular interest in smooth dynamical systems. For results of this type, see [@A; @BD; @Bur2; @BurNew; @DF; @D; @DM; @DN]. Note that the functions $u_{\al}$ are, in general, not harmonic, which stands in stark contrast to most objects of study in ergodic theory (in particular, the entropy function $h$ is harmonic [@Glas; @Pet]).
The order of accumulation $\al_0(X,T)$, which is defined as $\al_0(\H)$ for any entropy structure $\H$ of the system $(X,T)$, measures on how many “layers” residual entropy accumulates in system. From Theorem \[sexEntThm\] we see that the complexity in these layers accounts for the extra entropy that must be added to each measure in the system in order to obtain symbolic extensions. Thus the order of accumulation of entropy measures some additional complexity in the system that is not detected by the entropy function $h$.
Proof of Fact \[fonEharIsUSC\] {#proofAppendix}
==============================
The following fact was given as Fact 2.5 in [@DM], where there is a sketch of the proof. In this appendix we fill in some details of this proof for the sake of completeness.
Let $K$ be a metrizable Choquet simplex, and let $f: K \rightarrow [0, \infty)$ be convex and u.s.c. Then $(f|_{\operatorname{ex}(K)})^{har}$ is u.s.c.
Let $f: K \to [0,\infty)$ be convex and u.s.c. Let $g : \M(K) \to [0,\infty)$ be defined for each $\mu$ in $\M(K)$ as $$g(\mu) = \int f d\mu.$$ Now let $G: K \to [0,\infty)$ be given by $G(x) = \sup\{g(\mu) : \operatorname{bar}(\mu)=x\}$ for all $x$ in $K$. We have that $g$ is u.s.c. because $f$ is u.s.c., and $G$ is u.s.c. because $g$ is u.s.c. (Remark \[pushDownPullUp\] (iii)).
Now we claim that $f(x) \leq \int f d \mu$ for any $\mu$ such that $\operatorname{bar}(\mu) = x$. To see this, fix $x$ and $\mu$ such that $\operatorname{bar}(\mu)=x$. Let $f_m$ be a decreasing sequence of continuous functions, $f_m : K \to [0,\infty)$, whose limit is $f$. Let $\delta >0$. Partition the support of $\mu$ into a finite number of sets $S_j$ of diameter smaller than $\delta$. For each $j$, if $\mu(S_j)>0$, let $z_j = \operatorname{bar}( \mu_{S_j} )$, where $\mu_{S_j}$ is the measure $\mu$ conditioned on the set $S_j$. Then let $\nu = \sum_j \mu(S_j) \epsilon_{z_j}$. Note that $\operatorname{bar}(\nu) = \operatorname{bar}(\mu) = x$, and $\nu$ tends to $\mu$ in $\M(K)$ as $\delta$ tends to zero. We have shown that there exists a sequence of measures $\nu_k$ such that each $\nu_k$ is a finite convex combination of point measures, $\nu_k$ converges to $\mu$ in $\M(K)$, and $\operatorname{bar}(\nu_k)=x$ for each $k$. Now choose such a sequence $\nu_k$, and note that for any $m$, any $\epsilon >0$, and any large enough $k$ (depending on $\epsilon$ and $m$), by the convexity of $f$, $$f(x) \leq \int f \; d\nu_k \leq \int f_m \; d \nu_k \leq \int f_m \; d\mu + \epsilon.$$ Letting $m$ tend to infinity, the Dominated Convergence Theorem implies that $f(x) \leq \int f d\mu +\epsilon$. Since $\epsilon$ was arbitrary, we see that $f(x) \leq \int f d\mu$, which implies in particular that $f(x) \leq \int f d\P_x$.
Then for any $\mu$ with $\operatorname{bar}(\mu) = x$, $$\int f d\mu \leq \int \Bigl( \int f d\P_y \Bigr) \; d \mu(y) = \int f \; d\P_x,$$ where the equality of the last two expressions follows from the fact that $x \mapsto \int f \; d\P_x$ defines a harmonic function on $K$ (Remark \[harmRemark\]).
Thus $G(x) = \int f d\P_x$, which shows that $G = (f|_{\operatorname{ex}(K)})^{har}$. Since $G$ is u.s.c., the proof is complete.
[10]{}
M. Asaoka, Hyperbolic sets exhibiting $C\sp 1$-persistent homoclinic tangency for higher dimensions, Proc. Amer. Math. Soc. 136 (2008), no. 2.
L. Asimow and A.J. Ellis, Convexity Theory and its Applications in Functional Analysis, London Mathematical Society Monographs, 16. Academic Press, Inc. \[Harcourt Brace Jovanovich, Publishers\], London-New York, 1980.
M. Boyle and T. Downarowicz, The entropy theory of symbolic extensions, Invent Math 156 2004, no. 1, 119–161.
D. Burguet, Examples of $C^r$ interval maps with large symbolic extension entropy, Discrete Contin. Dyn. Syst. - A, to appear.
D. Burguet, $C^2$ surface diffeomorphisms have symbolic extensions, preprint.
G. Choquet, Lectures on analysis. Vol. II: Representation theory, W. A. Benjamin, Inc., New York-Amsterdam 1969 Vol. II.
L. Díaz, and T. Fisher, Symbolic extensions for partially hyperbolic diffeomorphisms, preprint on arXiv (2009).
T. Downarowicz, The Choquet simplex of invariant measures for minimal flows, Israel J. Math. 74 (1991), no. 2-3, 241–256.
T. Downarowicz, Entropy of a symbolic extension of a dynamical system, Ergodic Theory Dyn. Syst. 21 (2001), no. 4, 1051–1070.
T. Downarowicz, Entropy Structure, Journal d’Analyse Mathematique, 96 (2005), 57–116.
T. Downarowicz and A. Maass, Smooth interval maps have symbolic extensions, Invent. Math. 176 (2009), no. 3.
T. Downarowicz and S. Newhouse, Symbolic extensions and smooth dynamical systems, Invent. Math. 160 (2005), no. 3, 453–499.
T. Downarowicz and J. Serafin, Possible entropy functions, Israel J. Math. 135 (2003), 221–250.
J. Dugundji, Topology, Allyn and Bacon, Inc., Boston, Mass. 1966.
E. Glasner, Ergodic theory via joinings, Mathematical Surveys and Monographs, 101, American Mathematical Society, Providence, RI 2003.
R. Haydon, A new proof that every Polish space is the extreme boundary of a simplex, Bull. London Math. Soc. 7 (1975), 97–100.
D. Lind and B. Marcus, An introduction to symbolic dynamics and coding, Cambridge University Press, Cambridge 1995.
S. Mazurkiewicz and W. Sierpiński, Contribution à la topologie des ensembles dénombrables, Fund. Math. 1, (1920) 17–27.
M. Misiurewicz, Topological conditional entropy, Studia Math. 55 (1976), no. 2, 175–200.
K. Petersen, Ergodic theory, Corrected reprint of the 1983 original. Cambridge Studies in Advanced Mathematics, 2. Cambridge University Press, Cambridge, 1989.
R. Phelps, Lectures on Choquet’s theorem, second edition, Lecture Notes in Mathematics, 1757, Springer-Verlag, Berlin, 2001.
C. Pinter, Set Theory, Addison-Wesley Publishing Co., Reading, Mass.-London-Don Mills, Ont. 1971.
H. Schaefer and M. Wolff, Topological vector spaces, second edition, Graduate Texts in Mathematics, 3. Springer-Verlag, New York, 1999.
S.M. Srivastava, A Course on Borel Sets, Springer-Verlag, New York, 1998.
P. Walters, An introduction to ergodic theory, Graduate Texts in Mathematics, 79. Springer-Verlag, New York-Berlin, 1982.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider quantum clocks in curved spacetimes described by relativistic particles carrying internal degrees of freedom that serve as clocks. The proper time associated with such a clock is defined in terms of an observable $T_C$ (positive operator valued measure) that transforms covariantly with respect to the clock’s free Hamiltonian. This definition ensures that $T_C$ is an unbiased estimator of the elapsed proper time observed by such a clock. The probability that one clock reads a given proper time conditioned on another clock reading a different proper time is derived. From this conditional probability distribution it is shown that when the external degrees of freedom of these clock particles are described by Gaussian wave packets localized in momentum space, the clocks observe classical time dilation in accordance with special relativity. We then illustrate a novel quantum time dilation effect induced by nonclassical states of the clock particles realized by their external degrees of freedom being in a superposition of localized momentum wave packets and give an order of magnitude estimation of such effects. Moreover, we use the Helstrom-Holevo lower bound to derive a time-energy uncertainty relation between the proper time recorded by these clocks and their internal energy.'
author:
- 'Alexander R. H. Smith'
- Mehdi Ahmadi
bibliography:
- 'ProbabilisticTimeDilation.bib'
title: Relativistic quantum clocks observe classical and quantum time dilation
---
Introduction {#Introduction}
============
What allowed Einstein to transcend Newton’s conception of an absolute time was his insistence on an operational notion of time defined in terms of a measurement of a physical system serving as a clock [@einsteinTheoryRelativity1996]:
> ‘*\[Time is\] considered measurable by a clock (ideal periodic process) of negligible spatial extent. The time of an event taking place at a point is then defined as the time shown on the clock simultaneous with the event.*’
However, quantum mechanics has yet to be liberated from the Newtonian conception of time. Standard quantum theory is formulated in terms of an external background time, which is not quantized and with respect to which quantum states are normalized. This fact leads to unitary evolution or, more generally, quantum dynamics defined in terms of completely positive trace preserving maps.
The external time inherent in quantum theory is in tension with the principle of background independence on which general relativity is built. This tension becomes a conceptual hurdle in the construction of a quantum theory of gravity, which leads to the problem of time [@kucharTimeInterpretationsQuantum1992; @Isham1993; @andersonProblemTime2017]. One aspect of this problem is that the Hamiltonian associated with a generally covariant theory is constrained to vanish, leading to the Wheeler-DeWitt equation, and consequently this Hamiltonian does not generate evolution with respect to an external notion of time. One is then tasked with introducing a notion of quantum dynamics that in a particular limit recovers the unitary dynamics of standard quantum theory, a task which has yet to be fully completed in a quantum gravity setting [@kucharTimeInterpretationsQuantum1992; @Isham1993; @rovelliQuantumGravity2004; @kieferQuantumGravity2012; @andersonProblemTime2017].
Given the above, we seek to formulate a relativistic notion of time defined operationally in terms of a measurement of a physical clock. We do so by considering free relativistic particles moving through curved spacetime which carry both external and internal degrees of freedom. The external degrees of freedom of such a particle are associated with its position and momentum, while the internal degrees of freedom (e.g. energy levels of an atom) will be used as a clock which tracks the particle’s proper time. We begin by formulating the classical theory of such relativistic particles in terms of an action principle and derive the associated Hamiltonian, which we show is constrained to vanish. This constraint formulation of the classical theory is analogous to the Hamiltonian formulation of general relativity in the sense that in both cases the constraints are quadratic in the relevant canonical momentum. Therefore the investigation of the former serves as an example that may inform the latter [@ashtekarLecturesNonPerturbativeCanonical1991].
We then canonically quantize this system based on the procedure outlined by Dirac [@diracLecturesQuantumMechanics1964], which results in a Wheeler-DeWitt type equation. Solutions to this equation are projected onto a constant time hypersurface and normalized accordingly. The construction that is followed is based on the Page and Wootters mechanism [@pageEvolutionEvolutionDynamics1983] (i.e. the conditional probability interpretation of time [@woottersTimeReplacedQuantum1984; @pageTimeInaccessibleObservable1989; @pageClockTimeEntropy1994; @gambiniRelationalSolutionProblem2004; @gambiniRealisticClocksUniversial2004; @gambiniConditionalProbabilitiesDirac2009; @corbinSemiClassicalLimitMinimum2009; @dolbyConditionalProbabilityInterpretation2004; @giovannettiQuantumTime2015; @Boette:2016; @smithQuantizingTimeInteracting2017; @marlettoEvolutionEvolutionAmbiguities2017a; @gambiniRealisticClocksUniversial2004; @leonPauliObjection2017; @bryanRealisticClocksUniverse2018; @bryanRequiemIdealClock2018; @rotondoClockTimeQuantum2019]), which allows one to recover a relational dynamics between a physical clock and system of interest from solutions to a Wheeler-DeWitt type equation. We demonstrate that in the context considered here this quantization scheme is equivalent to the standard formulation of quantum theory in terms of a background coordinate time. Given this, the results we present can be viewed as a generalization of the Page and Wootters mechanism to systems described by relativistic quantum mechanics. It is in the spirit of Feynman[^1] that the Page and Wootters mechanism is employed because it offers a coordinate independent description of conditional probabilities associated with measurements of different clocks (see Eq. ).
After quantization we consider the internal degrees of freedom of these relativistic particles to constitute a clock which tracks their proper time. Based on the work of Holevo [@holevoCovariantMeasurementsUncertainty1979; @holevoProbabilisticStatisticalAspects1982] and Busch *et al.* [@buschOperationalQuantumPhysics], we define a proper time observable as a positive operator valued measure (POVM) on these internal clock degrees of freedom that is covariant with respect to their Hamiltonian (as later defined in Eq. ). This covariance property ensures that the proper time observable gives an unbiased estimate of the elapsed proper time experienced by the clock particle, the variance of which is independent of the proper time to be estimated. We then use the Helstrom-Holevo lower bound to state a time-energy uncertainty relation between the proper time measured by these clocks and their internal energy and discuss optimal clocks which saturate this uncertainty relation. Having introduced a proper time observable, we derive the probability distribution associated with one clock reading a particular proper time conditioned on a another clock reading a different proper time. It is shown that this probability distribution depends on the state of the external degrees of freedom of these clock particles.
We then specialize to two such clock particles $A$ and $B$ moving through Minkowski space and evaluate the leading order relativistic correction to the conditional probability that clock $A$ reads the proper time $\tau_A$ given that clock $B$ reads the proper time $\tau_B$. We show that when the external degrees of freedom of clocks $A$ and $B$ are prepared in Gaussian wave packets localized around a particular momentum, their associated clocks agree on average with the classical time dilation effect predicted by special relativity. However, we illustrate a novel quantum time dilation effect stemming from the clock particles being prepared in a Schrödinger cat state realized by a superposition of two Gaussian momentum wave packets. An optimistic order of magnitude estimate is given for this quantum time dilation effect. We conclude that such an effect may be observable in near future experiments, although the feasibility of such experiments remains to be explored.
This paper is organized as follows. In Sec. \[RelativisticClocks\] the classical and quantum theory of $N$ free relativistic particles in a curved spacetime carrying internal degrees of freedom is derived. In Sec. \[ProperTimeObservableSec\] we introduce a covariant proper time observable on these internal degree of freedom, which are taken to constitute a clock. The properties of this proper time observable are discussed and the conditional probability relating the proper times observed by two such clocks is derived. In Sec. \[TimeDilationInMinkowskiSpace\] we specialize to the case of two clocks moving through Minkowski space and recover the classical special relativistic time dilation effect and illustrate a novel quantum time dilation effect. We conclude in Sec. \[Conclusions\] with a summary of our results, a discussion of their relation with other literature, and give an outlook to future questions.
Throughout we will work in units where $\hbar = 1$ and adopt the ($-$ + + +) convention for the metric signature. The greek letters $\mu, \nu$ indicate spacetime indices and the Latin letters $i,j$ indicate spatial indices. For clarity, classical phase space functions and their corresponding quantum operators will be denoted by the same symbols, the meaning of them will be clear from the context in which they appear. Further, $\mathcal{S}(\mathcal{H})$ and $\mathcal{E}(\mathcal{H})$ will denote respectively the space of density operators[^2] and the space of effect operators[^3] acting on the Hilbert space $\mathcal{H}$.
Relativistic quantum clock particles in curved spacetime {#RelativisticClocks}
========================================================
We begin in Sec. \[classicaltheory\] by developing the classical theory of $N$ relativistic particles moving through a curved spacetime each carrying an internal degree of freedom beginning from a local Lorentz invariant action principle. We then quantize this theory via the Dirac prescription in Sec. \[CanonicalQuantization\] and define in Sec. \[RecoveryOfRQM\] the conditional state necessary to implement the Page and Wootters mechanism [@pageEvolutionEvolutionDynamics1983].
Classical theory of free relativistic particles {#classicaltheory}
-----------------------------------------------
Consider a system of $N$ free relativistic particles each carrying a set of internal degrees of freedom, labeled collectively by the configuration variables $q_n$ and their conjugate momentum $P_{q_n}$ ($n=1,\ldots, N$), and suppose these particles are moving through a curved spacetime described by the metric $g_{\mu \nu}$. The action describing such a system is $S = \sum_{n} \int d\tau_n \, L_n (\tau_n)$, where $$\begin{aligned}
L_n(\tau_n) \ce -m_nc^2 + P_{q_n} \frac{dq_n}{d\tau_n} - H^{\rm clock}_n,
\label{Lagrangian1}\end{aligned}$$ is the Lagrangian associated with the $n$th particle, $\tau_n$ and $m_n$ denote respectively this particle’s proper time and rest mass, and $H^{\rm clock}_n = H^{\rm clock}_n \smash{(q_n, P_{q_n})}$ is the Hamiltonian governing its internal degrees of freedom. In the following sections these internal degrees of freedom will be used as a clock which tracks the $n$th particle’s proper time, hence the superscript label. It should be noted that Eq. specifies that $H^{\rm clock}_n$ generates an evolution of the internal degrees of freedom of the $n$th particle with respect to its proper time. Let $x_n^{\mu}$ denote the spacetime position of the $n$th particle; we will refer to $x_n^0$ and $x_n^i$ as the $n$th particle’s temporal and external (spatial/momentum) degrees of freedom, respectively. Figure \[picture1\] depicts the described situation.
The differential proper time $d\tau_n$ along the $n$th particle’s world line $x^{\mu}_n(t_n)$, parametrized in terms of an arbitrary parameter $t_n$, is given by $d\tau_n = \sqrt{-g_{\mu \nu}^{\,}\smash{\dot{x}_n^\mu \dot{x}_n^\nu} \vphantom{X}} dt_n$, where the over dot denotes differentiation with respect to $t_n$. In terms of the parameters $t_n$ the action takes the form $$\begin{aligned}
S &= \sum_{n} \int dt_n \, \sqrt{-g^{\,}_{\mu \nu} \smash{\dot{x}_n^\mu \dot{x}_n^\nu} \vphantom{X}
} \, L_n (t_n). \label{action2}\end{aligned}$$ The action in Eq. is invariant under changes of the world line parameters $t_n$, as long as there is a one-to-one correspondence between $t_n$ and $\tau_n$. This invariance allows for the action to instead be parameterized in terms of a single parameter $t$, which is connected to the $n$th particles proper time through a monotonically increasing function $ f_n(\tau_n) \ce t$. Expressed in terms of the single parameter $t$, the action in Eq. is , where[^4] $$\begin{aligned}
L(t) \ce \sum_n \sqrt{-\dot{x}_{n}^2 } \left( -m_n c^2 + \frac{P_{q_n} \dot{q}_n}{\sqrt{-\dot{x}_{n}^2}} - H^{\rm clock}_n \right), \label{SuperLagrangian}\end{aligned}$$ in which we have used the shorthand $\dot{x}_{n}^2 \ce g_{\mu \nu}\dot{x}_{n}^{\mu} \dot{x}_n^\nu$ and the over dot now denotes differentiation with respect to $t$. This Lagrangian treats the temporal, spatial, and internal degrees of freedom as dynamical variables on equal footing.
The Hamiltonian associated with $L(t)$ is constructed by a Legendre transform of Eq. , which yields $$\begin{aligned}
H &\ce \sum_{n} \left[ g_{\mu \nu} P^\mu_n \dot{x}_{n}^\nu + P_{q_n} \dot{q}_n \right] - L(t) \nn \\
&= \sum_n \left[ g_{\mu \nu} P^\mu_n \dot{x}_{n}^\nu + \sqrt{-\dot{x}_{n}^2 } \left( m_nc^2 + H^{\rm clock}_n \right)\right], \label{Hamiltonian1}\end{aligned}$$ where $P^\mu_n$ is the momentum conjugate to the $n$th particle’s spacetime position $x_n^\mu$ and defined as $$\begin{aligned}
P_n^\mu \ce \frac{\partial L(t)}{\partial \dot{x}_n^\mu } = \frac{\dot{x}_n^\mu}{\sqrt{-\dot{x}_{n}^2}}\left( m_n c^2 + H^{\rm clock}_n \right) . \label{conjugateP}\end{aligned}$$ Upon substituting Eq. into the above Hamiltonian $H$, we see that each term in Eq. is constrained to vanish $$\begin{aligned}
H_n &\ce g_{\mu \nu} P_n^\mu \dot{x}_n^\nu + \sqrt{-\dot{x}_{n}^2 } \left( m_n + H^{\rm clock}_n\right) \nn \\
& = \frac{g_{\mu \nu} \dot{x}_n^\mu \dot{x}_n^\nu}{\sqrt{-\dot{x}_{n}^2}} \left( m_n c^2 + H^{\rm clock}_n\right) \nn \\
&\qquad + \sqrt{-\dot{x}_{n}^2 } \left( m_n c^2 + H^{\rm clock}_n\right) \nn \\
& = \frac{\dot{x}_{n}^2}{\sqrt{-\dot{x}_{n}^2}}\left [ \left( m_n c^2 + H^{\rm clock}_n \right) - \left( m_n c^2 + H^{\rm clock}_n\right) \right] \nn \\
&\approx 0, \label{constraint1}\end{aligned}$$ where $\approx$ means $H_n$ vanishes as a constraint [@diracLecturesQuantumMechanics1964]. Furthermore, the velocity appearing in the definition of $H_n$ above may be replaced with the momentum $P_n^\mu$ using Eq. , which allows the $N$ constraints in Eq. to be expressed in the form $$\begin{aligned}
C_{H_n} \ce g_{\mu\nu}P_n^\mu P_n^\nu + \left(m_nc^2 + H^{\rm clock}_n\right)^2 \approx 0, \quad \forall \, n, \label{constraint2}\end{aligned}$$ which we will refer to as the Hamiltonian constraints. These are a collection of primary first class constraints which are quadratic in the particles’ momentum. One might heuristically interpret Eq. as requiring the total energy shared between each particles external and internal degrees of freedom to be conserved.
[Each of these constraints may be factorized[^5] as $C_{H_n} = C_n^+ C_n^-$, where $C_n^\pm$ is defined as $$\begin{aligned}
C_n^\pm \ce P_n^0 \pm h_n,
\label{constraint3}\end{aligned}$$ and $$\begin{aligned}
h_n \ce \frac{ 1}{\sqrt{-g_{00}}} \sqrt{ g_{ij} P_n^i P_n^j + \left(m_nc^2 + H^{\rm clock}_n \right)^2} \label{defofh}\end{aligned}$$ In Eq. we have assumed the time-space components of the metric vanish, $g_{0i}=0$. Such an assumption is not necessary, although in the sections that follow this will be the case and thus we make this assumption now for simplicity. ]{}
By construction, the momentum conjugate to the $n$th particle’s spacetime coordinates satisfy the canonical Poisson relations $\{ x_m^\mu, P_n^\nu \} = \delta^{\mu\nu} \delta_{mn}$. This implies that the canonical momentum $P_n^\mu$ generates translations in the spacetime coordinate $x_n^\mu$. Moreover, if it is the case that $C_n^\pm \approx 0 $, it follows that $P_n^0 = \pm h_n$, which is the generator of translations in the $n$th particle’s time coordinate. Said another way, $\pm h_n$ is the Hamiltonian for both the external and internal degrees of freedom of the $n$th particle, generating an evolution of these degrees of freedom with respect to the particle’s temporal coordinate $x_n^0$.
Canonical quantization {#CanonicalQuantization}
----------------------
In what follows we will make use of the canonical quantization scheme outlined by Dirac [@diracLecturesQuantumMechanics1964; @ashtekarLecturesNonPerturbativeCanonical1991; @marolfGroupAveragingRefined2002], which is applicable to the quantization of the theory of $N$ relativistic particles described above. We begin by promoting the phase space variables associated with the $n$th particle to operators acting on appropriate Hilbert spaces: $x_n^0$ and $P_n^0$ become canonically conjugate self-adjoint operators acting on the Hilbert space $\mathcal{H}_n^0 \simeq L^2 \left( \mathbb{R} \right)$ associated with the $n$th particle’s temporal degree of freedom; $x_n^i$ and $P_n^i$ become canonically conjugate operators acting on the Hilbert space $\mathcal{H}_n^{\rm ext} \simeq L^2 \left( \smash{\mathbb{R}^3} \right)$ associated with the $n$th particle’s external degrees of freedom; and $q_n$ and $P_{q_n}$ become canonically conjugate operators acting on the Hilbert space $\mathcal{H}_n^{\rm clock}$ associated with the $n$th particle’s internal degrees of freedom. The Hilbert space describing the $n$th particle is thus $\mathcal{H}_n \simeq \mathcal{H}_n^{0} \otimes \mathcal{H}_n^{\rm ext} \otimes \mathcal{H}_n^{\rm clock}$.
[The constraint functions in Eqs. and become operators $C_{H_n}$ and $C_n^\pm$ acting on $\mathcal{H}_n$. The quantum analogue of the constraints is to demand that physical states of the theory are annihilated by these constraint operators $$\begin{aligned}
C_{H_n} \kket{\Psi} = C_n^+ C_n^- \kket{\Psi} = 0 , \quad \forall \, n, \label{quantumconstraints}\end{aligned}$$ where $\kket{\Psi} \in \mathcal{H}$ is a physical state that is an element of the physical Hilbert space $\mathcal{H}$ [@rovelliQuantumGravity2004; @kieferQuantumGravity2012]. The reason for introducing the physical Hilbert space is because the spectrum of $C_{H_n}$ is continuous around zero, which implies solutions to Eq. are not normalizable in the kinematical Hilbert space $\mathcal{K} \simeq \bigotimes_n \mathcal{H}_n$. To fully specify $\mathcal{H}$ a physical inner product must be defined, which is done below in Eq. . Note that because $[ C_n^+ ,C_n^- ] = 0$, it follows $C_{H_n} \kket{\Psi} = 0$ if either $C_n^+ \kket{\Psi} =0$ or $C_n^- \kket{\Psi} =0$.]{}
The conditional state and the recovery of relativistic quantum mechanics {#RecoveryOfRQM}
------------------------------------------------------------------------
In this subsection we recover the standard formulation of relativistic quantum mechanics with respect to an external (coordinate) time by implementing the Page and Wootters mechanism. To do so, the physical state $\kket{\Psi}$ describing all $N$ particles is normalized on a spatial hypersurface. This normalization is realized by projecting a physical state $\kket{\Psi}$ onto a subspace in which the temporal degree of freedom of each particle is in an eigenstate state $\ket{t_n}$ of the operator $x_n^0$ associated with the eigenvalue $t \in \mathbb{R}$ in the spectrum of $x_n^0$, that is, $x_n^0\ket{t_n} = t \ket{t_n}$. Explicitly, this projection takes the form $$\begin{aligned}
\Pi_t \otimes I_S \kket{\Psi} = \ket{t} \ket{\psi_S(t)},
\label{SpatialHypersurfaceProjection}\end{aligned}$$ where $I_S$ denotes the identity on $\mathcal{H}_S \simeq \bigotimes_n\mathcal{H}_n^{\rm ext} \otimes \mathcal{H}_n^{\rm clock}$, $\Pi_t \ce \ket{t}\!\bra{t}$ is a projector onto the subspace of $\mathcal{H}$ in which the temporal degree of freedom of each particle is in a definite temporal state $\ket{t}\ce \bigotimes_n \ket{t_n}$ associated with the eigenvalue $t$. Equation defines the conditional state $$\begin{aligned}
\ket{\psi_S(t)} &\ce \big( \bra{t} \otimes I_S \big) \kket{\Psi} \in \mathcal{H}_S,\label{defconditinalstate}\end{aligned}$$ which describes the state of the external and internal degrees of freedom of all $N$ particles conditioned on their temporal degree of freedom being in the state $\ket{t_n}$. We will interpret the conditional state as describing the external and internal degrees of freedom of the particles on a spatial hypersurface defined by their temporal degrees of freedom being in $\ket{t_n}$ and demand that the conditional state is normalized on this hypersurface $\braket{\psi_S(t) | \psi_S(t)}=1$ for all $t\in\mathbb{R}$. The normalization of the conditional state implies that the physical states are normalized with respect to the inner product [@smithQuantizingTimeInteracting2017] $$\begin{aligned}
\braket{\braket{\Psi | \Psi }}_{\rm PW} &\ce \braket{\braket{\Psi | \Pi_t \otimes I_S | \Psi }} \nn \\
& = \braket{\psi_S(t) | \psi_S(t) } \nn \\
&= 1,
\label{PhysicalInnerProduct}\end{aligned}$$ for all choices of $t \in \mathbb{R}$.
Note that the set $\left\{ \Pi_t \, |\, t \in \mathbb{R} \right\}$ constitutes a projective valued measure (PVM) on the Hilbert space $\bigotimes_n \mathcal{H}_n^0$, since $\Pi_t \Pi_{t'} = \delta(t-t')$ and $\int dt \, \Pi_t = I^0$, where $I^0$ is the identity on $ \bigotimes_n \mathcal{H}_n^0$. Given this observation and the definition of the conditional state in Eq. , it is seen that the physical state $\kket{\Psi}$ is an entangled state between the temporal degrees of freedom of the particles described by $\mathcal{H}^0_n$ and their external and internal degrees of freedom described by $\mathcal{H}_n^{\rm ext} \otimes \mathcal{H}_n^{\rm clock}$ $$\begin{aligned}
\kket{\Psi} &= \left(\int dt \, \Pi_t \otimes I_S \right)\kket{\Psi} = \int dt\, \ket{t}\ket{\psi_S(t)}.
\label{entangledPhysicalState}\end{aligned}$$ [In what follows we consider physical states that satisfy $$\begin{aligned}
C_n^+ \kket{\Psi} = \left(P_n^0 + h_n \right) \kket{\Psi} = 0,\end{aligned}$$ for all values of $n$, where $h_n$ is now the operator equivalent of Eq. . This amounts to demanding that the conditional state of the system has positive energy as measured by $h_n$.]{}
We now show that the conditional state $\ket{\psi_S(t)}$, as defined in Eq. , satisfies the Schrödinger equation in the parameter $t$. Recall that $x_n^0$ and $P_n^0$ are canonically conjugate variables satisfying $[x_n^0,P_n^0] = i $, from which it follows that the operators $P_n^0$ generate translations in $x_n^0$, $$\begin{aligned}
\ket{t'_n} = e^{-i (t' - t) P_n^0} \ket{t_n}, \label{generator}\end{aligned}$$ where $\ket{t_n}$ and $\ket{t_n'}$ are eigenkets of the operator $x_n^0$ with respective eigenvalues $t$ and $t'$. Now consider how $\ket{\psi_S(t)}$ changes with the parameter $t$: $$\begin{aligned}
i \frac{d}{dt} \ket{\psi_S(t)} &= i \frac{d}{dt} \left( \bigotimes_n \bra{t}_n \otimes I_S\right) \kket{\Psi} \nn \\
% &= (i)^2 \bigotimes_n \bra{t}_n \otimes I \sum_m P_m^0 \otimes I_{n \neq m } \otimes I_S \kket{\Psi}, \label{derivation1}
&= - \left( \sum_m \bra{t} P_m^0 \otimes I_{n \neq m }^0 \otimes I_S \right) \kket{\Psi}, \label{derivation1}\end{aligned}$$ where $I_{n \neq m }^0$ denotes the identity operator on all of the Hilbert spaces $\mathcal{H}_n^0$ for which $n\neq m$ and the derivative with respect to $t$ was evaluated using Eq. . Recall that $\kket{\Psi}$ satisfies the constraint $C_n^+ \kket{\Psi} =0$, from which it follows that $$\begin{aligned}
P_m^0 \otimes I_{n \neq m }^0 \otimes I_S \kket{\Psi}
= - I^0 \otimes h_m \otimes I_{S_{m\neq n}} \kket{\Psi}, \label{constraintrearanged}\end{aligned}$$ for all $m$ where $h_m$ is now the operator equivalent of Eq. acting on $ \mathcal{H}_n^{\rm ext} \otimes \mathcal{H}_n^{\rm clock}$ and $I_{S_{m\neq n}}$ is the identity on . Substituting Eq. into Eq. yields $$\begin{aligned}
i \frac{d}{dt} \ket{\psi_S(t)}
&= \left(\sum_m \bra{t} \otimes h_m \otimes I_{S_{m\neq n}} \right) \kket{\Psi} \nn \\
%&= \left( \sum_m \bra{t} \otimes h_m^\pm \otimes I_{S_{m\neq n}} \right) \nn \\
%& \quad\times \left( \int dt' \, \Pi_{t'} \otimes I_S\right) \kket{\Psi} \nn \\
&= \sum_m \int dt' \, \braket{t | t' } h_m \otimes I_{S_{m\neq n}} \ket{\psi_S(t')} \nn \\
&= \sum_m h_m \otimes I_{S_{m\neq n}} \ket{\psi_S(t)},
\label{derivation2}\end{aligned}$$ where the second equality is obtained from the first using Eq. . Equation asserts that the conditional state $\ket{\psi_S(t)}$ satisfies the Schrödinger equation $$\begin{aligned}
i \frac{d}{dt} \ket{\psi_S(t)} &= H_S \ket{\psi_S(t)}, \label{SchrodingerEq}\end{aligned}$$ where $H_S \ce \sum_m h_m \otimes I_{S_{m\neq n}}$ is the relativistic Hamiltonian describing the evolution of all $N$ particles. The solution to Eq. may be expressed in terms of a unitary operator $\ket{\psi_S(t)} = U(t-t_0) \ket{\psi_S(t_0)}$, where $U(t-t_0) \ce e^{-iH_S(t-t_0)}$.
Given Eq. , we are justified in identifying the conditional state $\ket{\psi_S(t)}$ as the usual time-dependent state of the $N$ particles in the standard formulation of relativistic quantum mechanics with respect to an external coordinate time as described for example in Ref. [@zychQuantumSystemsGraviational2017]. It is thus seen that in this context the implementation of the Page and Wootters mechanism above is equivalent to the standard formulation of relativistic quantum mechanics. However, the former does not require a priori the notion of a background coordinate time and, as will be shown in the following section, provides a natural framework in which conditional probabilities between measurement outcomes of different particles may be computed from the physical state $\kket{\Psi}$ using the Born rule.
[Had we instead considered physical states satisfying the constraints $C_n^- \kket{\Psi} = 0$ for all $n$, the conditional state would still satisfy the Schrödinger equation in Eq. under the replacement $H_S \to - H_S$. However, a general solution to the constraints in Eq. will not result in a conditional state that satisfies a Schrödinger equation. As shown in Appendix \[KGequation\], if the physical state was chosen to satisfy the constraints given in Eq. and considering the background spacetime to be Minkowski space, $g_{\mu \nu} = \eta_{\mu \nu}$, the conditional state would satisfy the Klein-Gordon equation $$\begin{aligned}
\left[ \Box + \left(mc^2 + H^{\rm clock}\right)^2 \right] \ket{\psi_S(t, \mathbf{x})} = 0,
\label{KleinGordonEq}\end{aligned}$$ where $\ket{\psi_S(t, \mathbf{x})} \ce \left( \bra{\mathbf{x}} \otimes I_C \right) \ket{\psi_S(t)}$ is a further conditioning of the physical sate, $\ket{\mathbf{x}} \ce \ket{x^1}\ket{x^2}\ket{x^3}$ with $\ket{x^i}$ denoting an eigenstate of the operator $x^i$, and $\Box \ce \eta^{\mu \nu} \partial_\mu \partial_\nu$ is the d’Alembertian operator.]{}
Proper time observables {#ProperTimeObservableSec}
=======================
Defining a proper time observable
---------------------------------
We define a clock to be the quadruple $\{\mathcal{H}^{\rm clock}, \rho_C, H^{\rm clock}, T_C\}$, where the clock is described by the Hilbert space $\mathcal{H}^{\rm clock}$, the fiducial state $\rho_C \in \mathcal{S}(\mathcal{H}^{\rm clock})$, a Hamiltonian $H^{\rm clock}$, and time observable $T_C$ with measurement outcomes $\tau$ contained in the set $G \subseteq \mathbb{R}$, which correspond to the time read by the clock. The time observable is defined as a POVM $$\begin{aligned}
T_C : \, G &\to \mathcal{E}(\mathcal{H}^{\rm clock}), \nn \\
\tau &\mapsto E(\tau) \ \mbox{ s.t. } \int_{G} d\tau \, E(\tau) = I_C,
\label{ProperTimeObservable}\end{aligned}$$ where each measurement outcome $\tau \in G $ is associated with an effect operator $E(\tau) \in \mathcal{E}(\mathcal{H}^{\rm clock})$. The defining property of a time observable is that it is covariant with respect the group of time translations generated by the clock Hamiltonian $H^{\rm clock}$ [@holevoCovariantMeasurementsUncertainty1979; @holevoProbabilisticStatisticalAspects1982; @buschOperationalQuantumPhysics], which amounts to the following condition on the effect operators $$\begin{aligned}
E(\tau+\tau') = U_{C}(\tau) \, E(\tau') \, U_{C}(\tau)^\dagger,
\label{covariant}\end{aligned}$$ where $U_{C}(\tau) \ce e^{-i H^{\rm clock} \tau}$ is the unitary representation of the group element parametrized by $\tau \in G$.
The physical significance of the covariance condition in Eq. is that it implies that a covariant time observable satisfies the following two physical properties commonly associated with a clock, which we state in the following theorem.
Let $T_C$ be a covariant time observable satisfying Eq. , the fiducial state $\rho_C$ be such that on average a measurement of $T_C$ yields $\braket{T_C}_{\rho_C} = 0$, and define $\rho_C(\tau) \ce U_C(\tau) \rho_C U_C(\tau)^\dagger$. The following two physical properties of such a time observable follow:
1. $T_C$ is an unbiased estimator of the parameter $\tau$ so that on average $T_C$ yields the value $\braket{T_C}_{\rho_C(\tau)} = \tau$ on the state $\rho_C(\tau)$.
2. The variance $\braket{\Delta T_C^2}_{\rho_C(\tau)}$ of the time observable $T_C$ is independent of the parameter $\tau$ on the state $\rho_C(\tau)$.
The first statement follows from a direct computation of the average of $T_C$ on the state $\rho_C(\tau)$ $$\begin{aligned}
\braket{T_C}_{\rho_C(\tau)} &= \int_{G} d\tau' \, \tr \left[ E(\tau') \rho_C(\tau) \right] \tau' \nn \\
&= \int_{G} d\tau' \, \tr \left[ U(\tau)^\dagger E(\tau') U(\tau) \rho_C \right] \tau' \nn \\
&= \int_{G} d\tau' \, \tr \left[ E(\tau' - \tau) \rho_C \right] \tau' \nn \\
&= \int_{G} d\tau' \, \tr \left[ E(\tau') \rho_C \right] \left(\tau' + \tau\right) \nn \\
&= \tau,\end{aligned}$$ where the third equality follows from Eq. , the fourth equality follows from a change of variables $\tau' \to \tau'-\tau$, and in arriving at the last equality we used the fact that by construction $\braket{T_C}_{\rho_C} = 0$.
The second statement follows in a similar manner $$\begin{aligned}
\braket{\Delta T_C^2}_{\rho_C(\tau)} &= \int_{G} d\tau' \, \tr \left[ E(\tau') \rho_C(\tau) \right] \tau'^2 - \braket{T_C}_{\rho_C(\tau)}^2 \nn \\
&= \int_{G} d\tau' \, \tr \left[ E(\tau'-\tau) \rho_C\right] \tau'^2 - \tau^2 \nn \\
&= \int_{G} d\tau' \, \tr \left[ E(\tau') \rho_C\right] \left(\tau'+\tau\right)^2 - \tau^2 \nn \\
&= \braket{\Delta T_C^2}_{\rho_C}.
\label{VarainceInTime}\end{aligned}$$
It is this theorem that justifies interpreting $T_C$ as a measurement of proper time — we expect a system functioning as a clock to estimate its proper time on average when measured (i.e. the clock should be unbiased) and that the variance of this measurement is independent of the proper time being estimated.
The uncertainty relation between proper time and clock energy
--------------------------------------------------------------
For an unbiased estimator, the Helstrom-Holevo lower bound [@helstromQuantumDetectionEstimation1976; @holevoProbabilisticStatisticalAspects1982] places the fundamental limit on the variance of the proper time measured by the clock $$\begin{aligned}
\braket{\Delta T_C^2}_{\rho_C} \geq \frac{1}{4\braket{(\Delta H^{\rm clock} )^2}_{\rho_C}},
\label{TimeEnergyUncertanity}\end{aligned}$$ where $\braket{(\Delta H^{\rm clock} )^2}_{\rho_C}$ is the variance of $H^{\rm clock}$ on the fiducial state $\rho_C$. Equation is a time-energy uncertainty relation between the proper time estimated by $T_C$ and a measurement of the clock’s energy $H^{\rm clock}$.
This inequality can be saturated and the optimal proper time observable constructed provided we make the additional assumption that the effect operators $E(\tau)$ defining $T_C$ in Eq. are proportional to ‘projection’ operators $$\begin{aligned}
E(\tau) = \mu \ket{\tau}\!\bra{\tau} \in \mathcal{E}\left( \mathcal{H}^{\rm clock} \right),
\label{projectionCondition}\end{aligned}$$ for $\mu\in \mathbb{R}$ such that Eq. is satisfied. We will refer to $\ket{\tau}$ as the clock state corresponding to the proper time $\tau \in G$, which may in general be unnormalizable. The motivation for this assumption is that measurements not described by one-dimensional projectors have less resolution [@braunsteinStatisticalDistanceGeometry1994]. Note that the clock states $\ket{\tau}$ are not necessarily orthogonal, $\braket{\tau | \tau '} \neq 0$, and form a possibly overcomplete basis for $\mathcal{H}^{\rm clock}$.
Braunstein and Caves [@braunsteinStatisticalDistanceGeometry1994] proved that covariant observables that satisfy Eq. , like the proper time observable $T_C$ considered here, constitute an optimal measurement to estimate the parameter $\tau$ unitarily encoded in the state $\rho_C(\tau) \ce U_C(\tau) \rho_C U_C(\tau)^\dagger$, provided that the fiducially state is pure $\rho_C = \ket{\psi^{\rm clock}}\! \bra{\psi^{\rm clock}}$ and $$\begin{aligned}
\ket{\psi^{\rm clock}} = \int_G d\tau \, \abs{\psi^{\rm clock}(\tau)} e^{i \tau \braket{H^{\rm clock}}_{ \rho_C}} \ket{\tau}
%\in \mathcal{H}^{\rm clock}
,\end{aligned}$$ where $\psi^{\rm clock}(\tau) \ce \sqrt{\mu} \braket{\tau | \psi^{\rm clock}} $ is the wave function of the fiducial state in the basis furnished by the clock states $\{ \ket{\tau}, \ \forall \,\tau \in G \}$ and $\abs{\psi^{\rm clock}(\tau)}$ is an arbitrary real function of $\tau$ such that . Such a proper time observable $T_C$ is optimal in the sense that it maximizes the Fisher information[^6] [over all estimates of $\tau$]{}; the Fisher information in this context is given by $$\begin{aligned}
F(\tau; \rho_C(\tau)) = 4 \braket{(\Delta H^{\rm clock} )^2}_{\rho_C}.\end{aligned}$$ The covariance condition in Eq. ensures that the Fisher information is independent of $\tau$.
Time dilation between quantum clocks in curved spacetime
--------------------------------------------------------
Let us now consider two relativistic particles $A$ and $B$, each with an internal degree of freedom serving as a clock, $\{\mathcal{H}^{\rm clock}_A, \rho_A, H^{\rm clock}_A, T_A\}$ and $ \{\mathcal{H}^{\rm clock}_B, \rho_B, H^{\rm clock}_B, T_B\}$. To probe time dilation effects between these clocks we consider the probability that clock $A$ reads the proper time $\tau_A$ conditioned on clock $B$ reading the proper time $\tau_B$. This conditional probability is evaluated using the physical state $\kket{\Psi}$ and the Born rule as follows $$\begin{aligned}
&\prob \left[T_A = \tau_A \ | \ T_B = \tau_B \right] \nn \\
&\hspace{.9em }= \frac{\prob \left[T_A = \tau_A \ \& \ T_B = \tau_B \right] }{\prob \left[ T_B = \tau_B \right]} \nn \\
&\hspace{.9em }= \frac{\bbra{\Psi} E_{A}(\tau_A) E_{B}(\tau_B) \kket{\Psi} }{\bbra{\Psi} E_{B}(\tau_B) \kket{\Psi}} \nn \\
%&\hspace{.9em }= \int dt \, \bra{\psi_S} U(t)^\dagger E_{T_A}(\tau_A) E_{T_B}(\tau_B) U(t) \ket{\psi_S}\nn \\
&\hspace{.9em }= \frac{\int dt \, \tr \left [ E_{A}(\tau_A) \rho_{A}(t) \right] \tr \left [ E_{B}(\tau_B) \rho_{B}(t) \right]}{\int dt \, \tr \left [ E_{A}(\tau_B) \rho_{B}(t) \right]},
\label{ConditionalProbability} \end{aligned}$$ where we have made use of Eq. in arriving at the last equality, assumed the state of the clock particles $A$ and $B$ is unentangled $\ket{\psi_S} = \ket{\psi_{S_A}}\ket{\psi_{S_B}}$, and defined the reduced state of the internal clock degrees of freedom as $$\begin{aligned}
\rho_n(t) \ce \tr_{\mathcal{H}_S \backslash \mathcal{H}_n^{\rm clock} } \left[U(t) \ket{\psi_{S_n}}\! \bra{\psi_{S_n}} U(t)^\dagger \right], \label{nInt}\end{aligned}$$ where $U(t-t_0) \ce e^{-iH_S(t-t_0)}$, as defined below Eq. , and the partial trace is over the complement of the clock Hilbert space $\mathcal{H}_S \backslash \mathcal{H}_n^{\rm clock}$.
Probabilistic time dilation in Minkowski space {#TimeDilationInMinkowskiSpace}
==============================================
We now apply the formalism developed in Secs. \[RelativisticClocks\] and \[ProperTimeObservableSec\] to investigate time dilation effects between two quantum clocks in relative motion moving through Minkowski space. It is demonstrated that when the external degrees of freedom of these particles are localized in momentum space these clocks observe on average time dilation described by special relativity. Then a quantum time dilation effect is exhibited by considering nonclassical states of the external degrees of freedom of these clock particles realized by a superposition of localized momentum wave packets.
Quantum clocks and time dilation
--------------------------------
Consider the case where two clock particles $A$ and $B$ are moving through Minkowski space and have equal mass $m$. From Eq. it is seen that the physical state takes the form $$\begin{aligned}
\kket{\Psi} &= \int dt\, \ket{t}\ket{\psi_S(t)} \nn \\
&= \int dt\, \ket{t} \bigotimes_{n \in \{A,B\}} U_n(t) \ket{\psi^{\rm ext}_{n}} \ket{\psi^{\rm clock}_n} ,
\label{PhysicalState2}\end{aligned}$$ where $U_n(t) \ce e^{-ih_n t}$ and in writing Eq. we have supposed that the conditional state at $t=0$ is unentangled, $\ket{\psi_S(0)} = \ket{\psi_{S_A}} \ket{\psi_{S_B}}\in \mathcal{H}_A \otimes \mathcal{H}_B$, and that the external and internal degrees of freedom of both particles are unentangled, $\ket{\psi_{S_n}} = \ket{\psi^{\rm ext}_{n}} \ket{\psi^{\rm clock}_n}$, where $\ket{\psi^{\rm ext}_{n}} \in \mathcal{H}^{\rm ext}_{n} $ is the state of the external degrees of freedom and $\ket{\psi^{\rm clock}_{n}} \in \mathcal{H}^{\rm clock}_{n} $ is the state of the internal clock.
Suppose that the Hilbert space of both clocks is $\mathcal{H}^{\rm clock}_A \simeq L_2(\mathbb{R}) \simeq \mathcal{H}^{\rm clock}_B$ and their Hamiltonians are equal to the momentum operator acting on these spaces, ${H}^{\rm clock}_A = p_{A}$ and $H^{\rm clock}_B = p_{B}$. Equation and the covariance condition in Eq. fix the time observables $T_A$ and $T_B$ to be equal to the position operator on $L_2(\mathbb{R})$ so that $G=\mathbb{R}$. Such clocks represent a commonly used idealization in which the time observable is sharp, that is, the clock states are orthogonal $\braket{\tau|\tau'} = \delta(\tau-\tau')$ and so outcomes of different clock measurements are perfectly distinguishable. It follows that the time observable is a PVM that can be associated with a self-adjoint operator canonically conjugate to the clock Hamiltonian, $\left[ T_{A} , H^{\rm clock}_{A} \right] = \left[ T_{B} , H^{\rm clock}_{B} \right] = i$. We employ such clocks for their mathematical simplicity in illustrating the quantum time dilation effect, however we stress that for any covariant time observable, on account of Eq. , a quantum time dilation effect is expected.
Finally, suppose that the fiducial state of the internal clock degrees of freedom of each particle is prepared in the state $$\begin{aligned}
\ket{\psi^{\rm clock}_n} := \frac{1}{\pi ^{1/4} \sqrt{\sigma}} \int d\tau \, e^{-\frac{\tau^2}{2 \sigma^2}} \ket{\tau_n},
\label{intialClockState}\end{aligned}$$ corresponding to a Gaussian wave packet with a spread in the basis formed by the clock states $\{ \ket{\tau_n}, \, \tau \in \mathbb{R} \}$. The above suppositions define clock $A$ by the quadruple $\{\mathcal{H}^{\rm clock}_A, \ket{\psi^{\rm clock}_A}, H^{\rm clock}_A, T_A \}$ and clock $B$ by the quadruple $ \{\mathcal{H}^{\rm clock}_B, \ket{\psi^{\rm clock}_B}, H^{\rm clock}_B, T_B\}$.
To compute the conditional probability in Eq. we require the reduced states of both clocks, $\rho_{A}(t)$ and $\rho_{B}(t)$, as defined in Eq. with $U(t) = e^{-i h_A t } \otimes e^{-i h_B t }$. Specializing to Minkowski space in Eq. , the $n$th particle’s Hamiltonian is $$\begin{aligned}
h_n % \sqrt{\mathbf{P}^2_n c^4 + (mc^2 + H_{n}^{\rm clock})^2} \nn \\
&= mc^2 \sqrt{ \frac{\mathbf{P}^2_n}{m^2c^4} + \left(1 + \frac{H_{n}^{\rm clock}}{mc^2}\right)^2} \nn \\
&=H_n^{\rm clock} + H_n^{\rm ext } + H_n^{\rm int},
\label{MinkowskiHamiltonian}\end{aligned}$$ where $\mathbf{P}_n^2 \ce \eta_{ij} P^i_n P^j_n$ and $\eta_{ij}$ is the Minkowski metric; in the last equality we have dropped an overall constant $mc^2 $ and defined the external Hamiltonian $H_n^{\rm ext} \ce \mathbf{P}^2_n/2m$ and the interaction Hamiltonian $$\begin{aligned}
H_n^{\rm int} &\ce
- \frac{1}{mc^2} \left( H_n^{\rm ext}\otimes H_{n}^{\rm clock} + \left(H_n^{\rm ext}\right)^2 \right) \nn \\
&\quad +\mathcal{O}\!\left( \left( \tfrac{ H_n^{\rm ext } }{mc^2} \right)^2 , \left( \tfrac{ H_n^{\rm clock} }{mc^2} \right)^2 \right),\label{interactionHamiltonian}\end{aligned}$$ which is arrived at by expanding the square root in Eq. in both $\mathbf{P}_n/mc $ and $H_{n}^{\rm clock}/mc^2$ and retaining only the leading-order relativistic correction. Using Eqs. - the conditional probability in Eq. may be evaluated perturbatively. This is done explicitly in Appendix \[DerivationOfCondtionalProb\] with the result that the probability that clock $A$ reads the proper time $T_A = \tau_A$ conditioned on clock $B$ reading the proper time $T_B = \tau_B$ is given by $$\begin{aligned}
&\prob \left[T_A = \tau_A \ | \ T_B = \tau_B \right] \nn \\
&= \frac{e^{- \frac{(\tau_A-\tau_B)^2}{2 \sigma^2}}}{\sqrt{2 \pi} \sigma} \left( 1 + \frac{\braket{H^{\rm ext}_{A}}- \braket{H^{\rm ext}_{B}}}{2 mc^2} \left(1 - \frac{ \tau_A^2 - \tau_B^2}{ \sigma^2} \right)\right) \nn \\
&\quad + \mathcal{O}\!\left( \tfrac{1}{\sigma}\left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}} \sigma}{mc^2} \right)^2 \right),
\label{ConditionalProbabilityDistirbutionMinkowski}\end{aligned}$$ where $\braket{H_n^{\rm ext }} \ce \bra{\psi^{\rm ext}_{n}} H_n^{\rm ext } \ket{\psi^{\rm ext}_{n}}$. As described by this probability distribution, the average proper time read by clock $A$ conditioned on clock $B$ indicating the time $\tau_B$ is $$\begin{aligned}
\avg{T_A} &= \int d\tau \, \prob \left[T_A = \tau \ | \ T_B = \tau_B \right] \tau \nn \\
&= \left( 1 - \frac{\braket{H^{\rm ext}_{A}} - \braket{H^{\rm ext}_{B}}}{mc^2} \right)\tau_B \nn \\
&\quad + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}} \sigma}{mc^2} \right)^2 \tau_B \right),
\label{averageEffect}\end{aligned}$$ and the variance in such a measurement is $$\begin{aligned}
\braket{\Delta T_A^2} &= \int d\tau \, \prob \left[T_A = \tau \ | \ T_B = \tau_B \right] \tau^2 - \avg{T_A}^2 \nn \\
&= \left( 1 - \frac{\braket{H^{\rm ext}_{A}} - \braket{H^{\rm ext}_{B}}}{mc^2} \right) \sigma^2 \nn \\
&\quad + \mathcal{O}\! \left( \left( \tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}} \sigma}{mc^2} \right)^2 \right).
\label{varianceEffect}\end{aligned}$$ As might have been anticipated, the variance in a measurement of $T_A$ is proportional to $\sigma^2$, which quantifies the spread in the fiducially clock state given in Eq. . Further, we note that this variance is modified at leading-order in $\braket{H_n^{\rm ext }} /mc^2$.
Recovering special relativistic time dilation for classical clocks {#ClassicalTimeDilation}
------------------------------------------------------------------
Suppose that the external degrees of freedom of both clock particles $A$ and $B$ are prepared in a Gaussian state localized around an average momentum $\bar{\mathbf{p}}_n$ with spread $\Delta_n >0$, $$\begin{aligned}
\ket{\psi^{\rm ext}_{n}} &= \frac{1}{\pi ^{1/4} \sqrt{\Delta_n}} \int d\mathbf{p} \, e^{-\frac{(\mathbf{p}-\bar{\mathbf{p}}_n)^2}{ 2\Delta_n^2}} \ket{\mathbf{p}_n}
=: \ket{\bar{\mathbf{p}}_n},
\label{GaussExtState}\end{aligned}$$ for which $
\braket{H_n^{\rm ext }} =\frac{\bar{\mathbf{p}}_n^2}{2m} + \frac{\Delta_n^2}{4m}$. It follows that the observed average time dilation between two such clocks is $$\begin{aligned}
\braket{T_A}
&= \left[1 - \frac{ \bar{\mathbf{p}}_A^2 - \bar{\mathbf{p}}_B^2 + \tfrac{1}{2}(\Delta_A^2 - \Delta_B^2) }{2m^2c^2} \right] \tau_B \nn \\
&\quad + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}} \sigma}{mc^2} \right)^2 \tau_B \right). \label{averageTime}\end{aligned}$$
If instead the two clocks were classical moving at velocities $\mathbf{v}_A=\bar{\mathbf{p}}_A/m$ and $\mathbf{v}_B=\bar{\mathbf{p}}_B/m$ corresponding to the average velocity of the momentum wave packets of the clocks just considered, then the proper time $\tau_A$ read by clock $A$ given that clock $B$ reads the proper time $\tau_B$ is $$\begin{aligned}
\tau_A &= \frac{\gamma_B}{\gamma_A} \tau_B \nn \\
& = \left[ 1 - \frac{ \bar{\mathbf{p}}_A^2 - \bar{\mathbf{p}}_B^2} {2m^2c^2}\right] \tau_B + \mathcal{O}\! \left( \left( \tfrac{\braket{H^{\rm ext}_{n}}}{mc^2} \right)^2 \right), \label{classicaltimedilation}\end{aligned}$$ where $\gamma_n \ce 1/\sqrt{1-v_n^2/c^2}$. Therefore, upon comparison with Eq. and supposing that $\Delta_A = \Delta_B$, quantum clocks whose external degrees of freedom are prepared in Gaussian wave packets localized around a particular momentum agree on average with the time dilation effect described by special relativity.
Nonclassical states of clocks and quantum time dilation effects {#QuantumTimeDilationEffect}
---------------------------------------------------------------
In the previous subsection we showed that quantum clocks prepared in momentum wave packets behave classically in the sense that they observe on average special relativistic time dilation. It is natural to ask: Do novel quantum time dilation effects arise if instead the external degrees of freedom of the clock particles are prepared in nonclassical states?
To answer this question we consider two clocks, $A$ and $B$, and suppose that the state of the external degree of freedom of clock $A$ begins in a nonclassical Schrödinger cat state realized by a superposition of two Gaussian wave packets with average momenta $\bar{\mathbf{p}}_{A}$ and $\bar{\mathbf{p}}_{A}'$, $$\begin{aligned}
\ket{\psi^{\rm ext}_A} \ce \frac{1}{\sqrt{N}} \Big( \cos \theta \ket{\bar{\mathbf{p}}_{A}} + \sin \theta e^{i\phi} \ket{\bar{\mathbf{p}}_{A}'} \Big) \in \mathcal{H}_A^{\rm ext},
\label{nonclassical}\end{aligned}$$ where $\theta \in [0,2\pi)$, $\phi \in [0,\pi]$, $\ket{\bar{\mathbf{p}}_{A}}$ and $\ket{\bar{\mathbf{p}}'_{A}}$ are defined by Eq. , and $$\begin{aligned}
N \ce 1 + \cos (\phi) \sin (2 \theta ) e^{- \frac{\left( \bar{\mathbf{p}}_A - \bar{\mathbf{p}}_A' \right)^2}{4 \Delta_A^2}} ,\end{aligned}$$ is a normalization constant. Further, suppose that the external degrees of freedom of clock $B$ are prepared in a Gaussian wave packet with average momentum $\bar{\mathbf{p}}_B$, like in Eq. , and that the internal degrees of freedom of both clocks are again given by Eq. .
In such a setup, the probability that clock $A$ reads the proper time $\tau_A$ conditioned on clock $B$ reading the proper time $\tau_B$ is given by Eq. . The difference between this case and the case considered in Sec. \[ClassicalTimeDilation\] is that the relativistic correction is now proportional to $$\begin{aligned}
\frac{\braket{H^{\rm ext}_{A}}- \braket{H^{\rm ext}_{B}}}{mc^2} & =K_{\rm classical} + K_{\rm quantum},
% & = \frac{ \bar{\mathbf{p}}_A^2 + \bar{\mathbf{p}}_A'^2 + \Delta_A^2 }{ 4 m^2c^2 N } - \frac{ 2 \bar{\mathbf{p}}_B^2 + \Delta_B^2 }{4m^2c^2}
% \nn \\ &\quad + \frac{ \left( \bar{\mathbf{p}}_A^2 - \bar{\mathbf{p}}_A'^2\right) \cos (2 \theta)
% }{ 4 m^2c^2 N }
% + \frac{ \tfrac{1}{2}\left( \bar{\mathbf{p}}_A + \bar{\mathbf{p}}_A'\right)^2 + \Delta_A^2}{ 4 m^2c^2 N }
% e^{- \frac{\left( \bar{\mathbf{p}}_A - \bar{\mathbf{p}}_A' \right)^2}{4 \Delta_A^2}} \cos ( \phi) \sin (2 \theta )
% \nn \\
% = \frac{\abs{\alpha}^2 \left(\bar{\mathbf{p}}_A^2 + \tfrac{1}{2}\Delta^2\right) + \abs{\beta}^2 \left(\bar{\mathbf{p}}_A'^2 + \tfrac{1}{2}\Delta^2 \right) }{ 2 m^2c^2 N } \nn \\
% + \frac{\cos \phi e^{- \frac{\left( \bar{\mathbf{p}}_A - \bar{\mathbf{p}}_A' \right)^2}{4 \Delta_A^2}} \left( \tfrac{1}{2}\left( \bar{\mathbf{p}}_A + \bar{\mathbf{p}}_A' \right)^2 + \Delta^2 \right)} { 2 m^2c^2 N }
%\nn \\ - \frac{ 2 \bar{\mathbf{p}}_B^2 + \Delta_B^2 }{4m^2c^2}
\label{DeviationFromClassicality} \end{aligned}$$ where $K_{\rm classical}$ is equal to the contribution corresponding to particle $A$ being prepared in a classical (incoherent) mixture of the momentum wave packets with average momentum $\bar{\mathbf{p}}_A$ with probability $\cos^2 \theta$ and $\bar{\mathbf{p}}_A'$ with probability $\sin^2 \theta$, $$\begin{aligned}
K_{\rm classical}
%&\ce \cos^2 \theta \frac{\bar{\mathbf{p}}_A^2}{2m^2 c^2 } +
%\sin^2 \theta \frac{\bar{\mathbf{p}}_A'^2 }{2m^2 c^2 } - \frac{\bar{\mathbf{p}}_B^2}{2m^2 c^2 } \nn \\
%& \qquad + \frac{\Delta_A^2- \Delta_B^2}{4m^2 c^2 } , \nn \\
&\ce \frac{ \bar{\mathbf{p}}_A^2 \cos^2 \theta + \bar{\mathbf{p}}_A'^2 \sin^2 \theta - \bar{\mathbf{p}}_B^2}{2m^2 c^2 } + \frac{\Delta_A^2- \Delta_B^2}{4m^2 c^2 },
%\frac{ \abs{\alpha}^2 \bar{\mathbf{p}}_A^2 + \abs{\beta}^2 \bar{\mathbf{p}}_A'^2 - \bar{\mathbf{p}}_B^2}{2m^2c^2} + \frac{\Delta^2_A - \Delta_B^2 }{4m^2c^2} ,
\label{Kclass} \end{aligned}$$ and $K_{\rm quantum}$ quantifies the contribution from the nonclassical aspect of the state of the external degrees of freedom of clock $A$, $$\begin{aligned}
K_{\rm quantum} &\ce \frac{ \sin 2 \theta \cos \phi }{8 m^2 c^2N} e^{- \frac{\left( \bar{\mathbf{p}}_A' - \bar{\mathbf{p}}_A \right) ^2}{4\Delta_A^2}} \nn \\
& \quad \times \!
\Big[ 2\left( \bar{\mathbf{p}}_A'^2 - \bar{\mathbf{p}}_A^2 \right) \cos 2 \theta - \left(\bar{\mathbf{p}}_A' - \bar{\mathbf{p}}_A \right)^2 \Big].
\label{Kquant} \end{aligned}$$ It is clear from Eq. that if either or , the quantum contribution vanishes, $K_{\rm quantum} = 0$. This is expected given that in these cases the external state of the clock particle is no longer a superposition of momentum wave packets; see Eq. . From Eq. it is clear that choosing $\Delta_A = \Delta_B$ and $\bar{\mathbf{p}}_B^2 = \bar{\mathbf{p}}_{A}^2 \cos^2 \theta + \bar{\mathbf{p}}_{A}'^2 \sin^2 \theta$ results in $K_{ \rm classical} =0$ and any observed time dilation between clock $A$ and $B$ in such a case would result from the quantum contribution $K_{\rm quantum}$. Moreover, note that the average and variance of the observed time dilation of clock $A$ with respect to clock $B$ will be given by Eqs. and using Eq. .
\
![The strength of the quantum time dilation effect $K_{\rm quantum}$ is plotted as a function of the difference $\abs{\bar{\mathbf{p}}_A'-\bar{\mathbf{p}}_A}/mc = (\bar{p}_A'-\bar{p}_A)/mc$, where $\bar{\mathbf{p}}_A = (\bar{p}_A,0,0)$ and $\bar{\mathbf{p}}_A' = (\bar{p}_A',0,0)$ denote the average momentum of the wave packets comprising the superposition state in Eq. . Different values of average total momentum $(\bar{p}_A'+\bar{p}_A)/mc$ are shown, in all cases $\theta = \pi/8$ and $\Delta_A/mc = 0.01$. The thin black line traces the trajectory of the optimal momentum difference $\bar{p}_{\rm opt}$ of for different total momentum $\left( \bar{p}_A'+\bar{p}_A\right)/mc$. []{data-label="NonClassicalStates1"}](plot1.pdf "fig:"){width="\linewidth"}\
![The strength of the quantum time dilation effect $K_{\rm quantum}$ is plotted as a function of the difference $\abs{\bar{\mathbf{p}}_A'-\bar{\mathbf{p}}_A}/mc = (\bar{p}_A'-\bar{p}_A)/mc$, where $\bar{\mathbf{p}}_A = (\bar{p}_A,0,0)$ and $\bar{\mathbf{p}}_A' = (\bar{p}_A',0,0)$ denote the average momentum of the wave packets comprising the superposition state in Eq. . Different values of average total momentum $(\bar{p}_A'+\bar{p}_A)/mc$ are shown, in all cases $\theta = \pi/8$ and $\Delta_A/mc = 0.01$. The thin black line traces the trajectory of the optimal momentum difference $\bar{p}_{\rm opt}$ of for different total momentum $\left( \bar{p}_A'+\bar{p}_A\right)/mc$. []{data-label="NonClassicalStates1"}](Legend1a.pdf "fig:"){width="\linewidth"}\
![The strength of the quantum time dilation effect $K_{\rm quantum}$ is plotted as a function of the parameter $\theta$, which quantifies the weight of the momentum wave packets with average momentum $\bar{\mathbf{p}}_A = (\bar{p}_A,0,0)$ and $\bar{\mathbf{p}}_A' = (\bar{p}_A',0,0)$ comprising the superposition state in Eq. . Different values of the sum of these average momentum $(\bar{p}_A' + \bar{p}_A)/mc$ are shown, in all cases $(\bar{p}_A' - \bar{p}_A)/mc = 0.017$ and $\Delta_A/mc = 0.01$.[]{data-label="NonClassicalStates2"}](plot2.pdf "fig:"){width="\linewidth"} ![The strength of the quantum time dilation effect $K_{\rm quantum}$ is plotted as a function of the parameter $\theta$, which quantifies the weight of the momentum wave packets with average momentum $\bar{\mathbf{p}}_A = (\bar{p}_A,0,0)$ and $\bar{\mathbf{p}}_A' = (\bar{p}_A',0,0)$ comprising the superposition state in Eq. . Different values of the sum of these average momentum $(\bar{p}_A' + \bar{p}_A)/mc$ are shown, in all cases $(\bar{p}_A' - \bar{p}_A)/mc = 0.017$ and $\Delta_A/mc = 0.01$.[]{data-label="NonClassicalStates2"}](Legend1a.pdf "fig:"){width="\linewidth"}\
To illustrate the behaviour of the quantum time dilation effect stemming from the nonclassicality of the external state of clock particle $A$, the quantity $K_{ \rm quantum}$ is plotted in Figs. \[NonClassicalStates1\] and \[NonClassicalStates2\]; for simplicity the one-dimensional case is exhibited by supposing $\bar{\mathbf{p}}_A = (\bar{p}_A ,0,0)$ and $\bar{\mathbf{p}}_A' = (\bar{p}_A' ,0,0)$ with $p_A' > p_A$.
Figure \[NonClassicalStates1\] shows the behaviour of $K_{\rm quantum}$ as a function of the difference $\left( \bar{p}_A' - \bar{p}_A \right)/mc$ in the average momentum of each wave packet comprising the momentum superposition in Eq. for different values of their total momentum $\left( \bar{p}_A' + \bar{p}_A\right)/mc$. It is seen that there is an optimal difference in the average momentum of the two wave packets $p_{\rm opt}$, which increases as the total average momentum of the wave packets $\left( \bar{p}_A' + \bar{p}_A\right)/mc$ increases; this is depicted by the thin black line in Fig. \[NonClassicalStates1\]. Further, from Fig. \[NonClassicalStates1\] it is seen that $K_{\rm quantum}$ increases as the total average momentum of the wave packets $\left( \bar{p}_A' + \bar{p}_A\right)/mc$ increases.
Figure \[NonClassicalStates2\] is a plot of $K_{\rm quantum}$ as a function of the parameter $\theta$ quantifying the weight of each momentum wave packet comprising the superposition in Eq. for a fixed value of the difference in average momentum of each wave packet $\left( \bar{p}_A' - \bar{p}_A \right)/mc$. It is observed that when $\left( \bar{p}_A' - \bar{p}_A \right)/mc = 0$, $K_{\rm quantum}$ is negative for all values of $\theta$ and reaches its most negative value at $\theta = \pi/4$. As the total average momentum increases, $K_{\rm quantum}$ increases for $0< \theta < \pi/4$ and decreases for $\pi/4 < \theta < \pi/2$ with the largest positive value at $\theta \approx\pi/8$ and largest negative value at $\theta \approx 3\pi/8$. Recall that the plot in Fig. \[NonClassicalStates1\] is for $p_A' \geq p_A$, so it is not unexpected that the largest (negative) value occurs at $\theta \approx 3\pi/8$ corresponding to more weight being attributed to the wave packet with larger momentum in Eq. .
Quantum time dilation in experiment?
------------------------------------
[Consider the two clock particles to be ${^{87}}$Rb atoms, which have a mass of $m = 1.4 \times 10^{-25}$kg and atomic radius of $r = \hbar/\Delta_A = 2.5 \times 10^{-10}$m. Suppose these clock particles can be prepared in a superposition of momentum wave packets such that $\left( \bar{p}_A' + \bar{p}_A \right)/mc = \left( \bar{p}_A' - \bar{p}_A \right)/mc= 6.7 \times 10^{-8}$, corresponding each branch of the momentum superposition moving at average velocities of $\bar{v}_A=5$ m/s and $\bar{v}_A' =15$ m/s. We note that (classical) special relativistic time dilation has been observed with atomic clocks moving at these velocities [@reinhardtTestRelativisticTime2007; @chouOpticalClocksRelativity2010], and perhaps the momentum superposition can be prepared by a momentum beam splitter realized using coherent momentum exchange between atoms and light [@Berman:1997; @cladeLargeMomentumBeamsplitter2009]. Supposing that $\theta = 3 \pi/4$ and $\phi =0$ results in $K_{\rm quantum} \approx 10^{-15}$. Assuming that the resolution of the clock formed by the internal degrees of freedom of the ${^{87}}$Rb atoms is $10^{-14}$s, corresponding to the resolution of ${^{87}}$Rb atomic clocks [@camparoRubidiumAtomicClock2007], it is seen from Eq. that the coherence time of the momentum superposition must be on the order of 10s to observe a quantum time dilation effect. We note that the required coherence time is comparable to coherence times of the superpositions created in the experiments of Kasevich *et al.* [@kovachyQuantumSuperpositionHalfmetre2015], which are on the order of seconds; however, in these experiments spatial superpositions are constructed as opposed to the momentum superpositions considered here.]{}
Alternatively, Bushev et al. [@bushevSingleElectronRelativistic2016a] have proposed to use the spin precession of a single electron in a Penning trap as a clock to observe relativistic time dilation, and it is conceivable that such a clock might be able to witness the quantum time dilation effect.
Conclusions and Outlook {#Conclusions}
=======================
We began by formulating the classical theory of $N$ free relativistic particles with internal degrees of freedom moving through a curved spacetime in terms of a Lorentz invariant action principle. The importance of formulating the theory in terms of an action principle, as remarked by Dirac [@diracLecturesQuantumMechanics1964], is that it is easy to implement the conditions for the theory to be relativistic: one simply has to require that the action integral be invariant. With an action comes a Lagrangian from which we formulated the theory in terms of a Hamiltonian constraint as given in Eqs. and . We then passed over to the quantum theory via the canonical quantization procedure and implemented the constraints in Eq. as a Wheeler-DeWitt type equation.
We then considered the internal degrees of freedom of these relativistic particles to function as clocks that track their proper time. In doing so we constructed an optimal covariant proper time observable which gives an unbiased estimate of the clock’s proper time. It was shown that the Helstrom-Holevo lower bound [@helstromQuantumDetectionEstimation1976; @holevoProbabilisticStatisticalAspects1982] implies a time-energy uncertainty relation between the proper time read by such a clock and a measurement of its energy. From the Born rule the conditional probability distribution that one such clock reads the proper time $\tau_A$ conditioned on another clock reading the proper time $\tau_B$ was derived in Eq. .
We then specialized to two such clock particles moving through Minkowski space and evaluated the leading-order relativistic correction to this conditional probability distribution. It was shown that on average these quantum clocks measure a time dilation consistent with special relativity when the state of their external degrees of freedom is localized in momentum space. However, when the state of their external degrees of freedom is in a superposition of such localized momentum states, we demonstrated that a quantum time dilation effect occurs. We exhibited how this quantum time dilation effect depends on the parameters defining the momentum superposition and gave an order of magnitude estimate for the size of this effect. We conclude that such a quantum time dilation effect may be observable with present day technology, but note that the experimental feasibility of observing this effect remains to be explored.
It should be noted that the conditional probability distribution in Eq. associated with clocks reading different times was a nonperturbative expression for clocks in arbitrary nonclassical states in a curved spacetime. It thus remains to investigate the effect of other nonclassical features of the clock particles such as shared entanglement among the clocks and spatial superpositions. In regard to the latter, it will be interesting to recover previous relativistic time dilation effects in quantum systems related to particles prepared in spatial superpositions and each branch in the superposition experiencing a different proper time due to gravitational time dilation [@zychGeneralRelativisticEffects2012a; @pikovskiUniversalDecoherenceDue2015; @zychQuantumInterferometricVisibility2011a; @khandelwalGeneralRelativisticTime2019]; this work has largely focused on gravitational decoherence effects. We emphasize that the quantum time dilation effect described in Sec. \[QuantumTimeDilationEffect\] differs from these results in that it is a consequence of a momentum superposition rather than gravitational time dilation. Nonetheless, it will be interesting to examine such gravitational time dilation effects in the framework developed above and make connections with previous literature on quantum aspects of the equivalence principle [@violaTestingEquivalencePrinciple1997; @zychQuantumFormulationEinstein2018; @hardyImplementationQuantumEquivalence2019]. We also note that while we exhibited the quantum time dilation effect for a specific clock model in Sec. \[TimeDilationInMinkowskiSpace\], based on the preceding analysis in terms covariant time observables it is expected that any clock will witness quantum time dilation. Given this, it will be fruitful to examine our results in relation to other models of quantum clocks that have been considered [@saleckerQuantumLimitationsMeasurement1958; @peresMeasurementTimeQuantum1980; @buzekOptimalQuantumClocks1999; @erkerAutonomousQuantumClocks2017; @paigeQuantumClocksNot2018; @woodsAutonomousQuantumMachines2018].
Another avenue of exploration is the construction of relativistic quantum reference frames from the relativistic clock particles considered here [@bartlettRelativisticallyInvariantQuantum2005; @bartlettReferenceFramesSuperselection2007; @bartlettQuantumCommunicationUsing2009; @ahmadiCommunicationInertialObservers2015; @loveridgeSymmetryReferenceFrames2018]. In particular, one might define relational coordinates with respect to a reference particle and examine the corresponding relational quantum theory and the possibility of changing between different reference frames [@poulinToyModelRelational2006; @angeloPhysicsQuantumReference2011; @palmerChangingQuantumReference2014; @Safranek2015; @smithQuantumReferenceFrames2016; @smithCommunicatingSharedReference2018; @giacominiRelativisticQuantumReference2018; @giacominiQuantumMechanicsCovariance2019]. Related is the perspective-neutral interpretation of the Hamiltonian constraint in terms of which a formalism for changing clock reference systems has recently been developed [@vanrietveldeChangePerspectiveSwitching2018; @Hoehn:2019; @hoehnHowSwitchRelational2018; @vanrietveldeSwitchingQuantumReference2018]. Further, one might also wish to consider interactions among the clock particles which has been shown to lead to interesting effects in related contexts [@Hoehn:2012; @smithQuantizingTimeInteracting2017; @ruizEntanglementQuantumClocks2017].
We note that the results presented here complement those of Greenberger [@greenbergerConceptualProblemsRelated2010] and his desire for proper time and rest mass to be treated as quantum observables. The approach adopted here differs in that we construct a proper time observable in terms of a covariant POVM rather than a self-adjoint operator. This has the advantage of overcoming Pauli’s objection to the construction of a canonically conjugate self-adjoint time observable [@pauliGeneralPrinciplesQuantum1980]. Like Greenberger [@greenbergerConceptualProblemsRelated2010] we obtain in Eq. an uncertainty relation between proper time and rest mass, provided we interpret $M \ce m I_C + H^{\rm clock}/c^2$ as a mass operator [@zychQuantumFormulationEinstein2018].
As remarked in the introduction, the results presented here constitute the generalization of the Page and Wootters mechanism into the regime of relativistic quantum mechanics. The advantage of employing this formalism was three fold: it a quantization scheme independent of a background coordinate time as described by the Wheeler-DeWitt equation in Eq. , the recovery of the standard formulation of relativistic quantum mechanics as described by the Schrödinger equation in Eq. , and provided a framework in which to compute the conditional probability that different clocks read different proper times that is independent of a background coordinate time (see Eq. ). We view these results as contributing to the program of extending the Page and Wootters mechanism into the relativistic regime, including relativistic quantum field theory and quantum gravity [@gambiniRealisticClocksUniversial2004; @gambiniRelationalSolutionProblem2004; @gambiniConditionalProbabilitiesDirac2009; @rotondoClockTimeQuantum2019]. In this regard, once a clear field-theoretic formulation of the Page and Wootters mechanism is available, it will be interesting to make connections with other studies of physical clocks in such settings [@lorekIdealClocksConvenient2015; @lockRelativisticQuantumClocks2017; @lockQuantumClassicalEffects2019]. Pushing the Page and Wootters mechanism into the relativistic regime is an important program to pursue because it has the potential to address aspects of the problem of time by providing a way in which a relational dynamics can emerge from solutions to the Wheeler-DeWitt equation.
This work was supported by the Natural Sciences and Engineering Research Council of Canada and the Dartmouth College Society of Fellows. We’d like thank and Philipp A. Höhn and Maximilian P. E. Lock for useful discussions.
Derivation of the Klein-Gordon equation in the Page and Wootters formalism {#KGequation}
==========================================================================
For simplicity, let us consider a single particle situated in Minkowski space so that the constraint in Eq. becomes $$\begin{aligned}
C_{H} \kket{\Psi} = \left( \eta_{\mu\nu}P^\mu P^\nu \otimes I_C + I^0 \otimes I_{\rm ext} \otimes \left(mc^2 + H^{\rm clock}\right)^2 \right) \kket{\Psi}= 0, \label{constraintAppendix}\end{aligned}$$ where $\eta_{\mu \nu}$ denotes the Minkowski metric and we have suppressed the subscript $n$. Given a physical state satisfying this constraint, Eq. defines the conditional state of the external and internal degrees of freedom of the particle $$\begin{aligned}
\ket{\psi_S(t, \mathbf{x})} \ce \big( \bra{t} \bra{\mathbf{x}} \otimes I_{C} \big) \kket{\Psi} ,\end{aligned}$$ where $x^0 \ket{t} = t \ket{t}$ and $\ket{\mathbf{x}} \ce \ket{x^1}\ket{x^2}\ket{x^3}$ with $\ket{x^i}$ denoting an eigenstate of the operator $x^i$. Now consider the action of the d’Alambertian operator $\Box \ce \eta^{\mu \nu} \partial_\mu \partial_\nu$ on the conditional state $$\begin{aligned}
\Box\ket{\psi_S(t, \mathbf{x})} &= \eta^{\mu \nu}\partial_\mu \partial_\nu \big( \bra{t} \bra{\mathbf{x}} \otimes I_{C} \big) \kket{\Psi} \nn \\
&= \eta^{\mu \nu}\partial_\mu \partial_\nu \left( \bra{t_0} e^{iP^0 t}\bra{\mathbf{x}_0} e^{i \mathbf{P} \cdot \mathbf{x} } \otimes I_{C} \right) \kket{\Psi} \nn \\
&= \big( \bra{t} \bra{\mathbf{x}} \otimes I_{C} \big) \big( \eta_{\mu \nu} P^\mu P^\nu \otimes I_{C} \big) \kket{\Psi} \nn \\
&= - \big( \bra{t} \bra{\mathbf{x}} \otimes I_{C} \big) \left( I^0 \otimes I_{\rm ext} \otimes \left(mc^2 + H^{\rm clock}\right)^2 \right) \kket{\Psi} \nn \\
&= - \left( I^0 \otimes I_{\rm ext} \otimes \left(mc^2 + H^{\rm clock}\right)^2 \right) \ket{\psi_S(t, \mathbf{x})},\label{stepsToKG}\end{aligned}$$ where the third equality is obtained using Eq. . Upon rearranging Eq. we find that the conditional state satisfies $$\begin{aligned}
\left[ \Box + \left(mc^2 + H^{\rm clock}\right)^2 \right] \ket{\psi_S(t, \mathbf{x})} = 0,
\label{KleinGordonEq}\end{aligned}$$ where we have suppressed the identity operators $I^0$, $I_C$, and $I_{\rm ext}$. If one suppose $H^{\rm clock}$ vanishes, then Eq. reduces to the usual Klein-Gordon equation.
Derivation of the conditional probability distribution given in Eq. . {#DerivationOfCondtionalProb}
=====================================================================
Let us define the free evolution of the external and internal clock degrees of freedom respectively as $$\begin{aligned}
\bar{\rho}^{\rm ext}_n(t) \ce e^{-iH_n^{\rm ext} t} \rho_n^{\rm ext} e^{iH_n^{\rm ext} t} \in \mathcal{S}(\mathcal{H}^{\rm ext}_n)\ \mbox{ and } \ \bar{\rho}_n(t) \ce e^{-iH_n^{\rm clock} t} \rho_n e^{iH_n^{\rm clock} t} \in \mathcal{S}(\mathcal{H}^{\rm clock}_n),\end{aligned}$$ for $n \in \{A,B\}$. Then the reduced state of of the clock is given by $$\begin{aligned}
\rho_n(t) &= \tr_{\rm ext} \left(e^{-i H_n^{\rm int} t} \bar{\rho}_n^{\rm ext}(t) \otimes \bar{\rho}_n(t) e^{i H_n^{\rm int} t} \right) \nn \\
&= \tr_{\rm ext} \Big(\bar{\rho}_n(t) \otimes \bar{\rho}_n^{\rm ext}(t) - it \left[ H_n^{\rm int} ,\bar{\rho}_n^{\rm ext}(t) \otimes \bar{\rho}_n(t) \right] + \mathcal{O}\!\left( (H_n^{\rm int} t)^2 \right) % \mathcal{O}\!\left( mc^2 t^2 \right)
\Big) \nn \\
&= \bar{\rho}_n(t) -it \tr_{\rm ext} \Big( [H_n^{\rm int} , \bar{\rho}_n^{\rm ext}(t) \otimes \bar{\rho}_n(t)] \Big) + \mathcal{O}\!\left( (H_n^{\rm int} t)^2 \right) % \mathcal{O}\!\left( mc^2 t^2 \right)
\nn \\
&= \bar{\rho}_n(t) + it \frac{\braket{H_n^{\rm ext }}}{mc^2} \left[ H_n^{\rm clock} , \bar{\rho}_n(t) \right] + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} H_n^{\rm clock}}{mc^2} t \right)^2 \right) .\label{expansion}\end{aligned}$$ Using Eq. the integrands defining the conditional probability distribution in Eq. may be evaluated perturbatively $$\begin{aligned}
\tr \big( E_{n}(\tau_n) \rho_{n}(t) \big) &= \tr \big( E_{n}(\tau_n) \bar{\rho}_n(t)\big)
+ it \frac{\braket{H_n^{\rm ext }}}{m c^2} \tr \big( E_{n}(\tau_n) \left[ H_{n}^{\rm clock} , \bar{\rho}_n(t) \right]\big) + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}}}{mc^2} t \right)^2 \right).
\label{integrand1}\end{aligned}$$
Suppose that the fiducial state $\ket{\psi^{\rm clock}_n} \in \mathcal{H}_n^{\rm clock}$ of the clock is Gaussian with a spread $\sigma$ as given by Eq. , then the first term in Eq. is given by $$\begin{aligned}
\tr \left [ E_{n}(\tau_n) \bar{\rho}_n(t)\right] =
\abs{ \bra{\tau_n} e^{- i H_n^{\rm clock} t } \ket{\psi^{\rm clock}_n} }^2 = \abs{\int_{\mathbb{R}} d \tau' \, \braket{\tau_n |\tau_n'} \frac{e^{-\frac{(\tau'-t)^2}{2 \sigma^2}} }{\pi^{\frac{1}{4}} \sqrt{\sigma}} }^2 = \frac{e^{-\frac{(\tau-t)^2}{\sigma^2}}}{\sqrt{\pi} \sigma} .
\label{FirstTerm}\end{aligned}$$ Defining $\ket{\psi^{\rm clock}_n(t)} \ce e^{- i H_n^{\rm clock} t } \ket{\psi^{\rm clock}_n}$, the trace in the second term is given by $$\begin{aligned}
\tr \left [ E_{T_n}(\tau_n) \left[ H_{n}^{\rm clock} , \bar{\rho}_n(t) \right]\right] &= \bra{\tau_n} \left[ H_{n}^{\rm clock} , \bar{\rho}_n(t) \right]\ket{\tau_n} \nn \\
&= \bra{\tau_n } H_n^{\rm clock} \ket{\psi^{\rm clock}_n(t)} \braket{\psi^{\rm clock}_n(t) | \tau_n}
-\braket{\tau_n | \psi^{\rm clock}_n(t)} \bra{\psi^{\rm clock}_n(t)} H_n^{\rm clock} \ket{\tau_n} \nn \\
&= \frac{ e^{-\frac{(\tau-t)^2}{2 \sigma^2}}}{\pi ^{\frac{1}{4}} \sqrt{\sigma}} \big( \bra{\tau_n } H_n^{\rm clock} \ket{\psi^{\rm clock}_n(t)} - \bra{\psi^{\rm clock}_n(t)} H_n^{\rm clock} \ket{\tau_n}\big).
\label{trace2ndTerm}\end{aligned}$$ It follows from the covariance relation in Eq. that the clock states satisfy $$\begin{aligned}
\ket{\tau + \tau'} = e^{-iH^{\rm clock}_n \tau'}\ket{\tau},\end{aligned}$$ which implies that $H^{\rm clock}_n\equiv -i \partial/ \partial\tau $ is the displacement operator in the $\ket{\tau}$ representation [@wisemanQuantumMeasurementControl2010]. This observation allows us to evaluate the probability amplitudes in Eq. $$\begin{aligned}
\bra{\tau_n } H_n^{\rm clock} \ket{\psi^{\rm clock}_n(t)} &= - i \frac{\partial}{\partial \tau} \frac{e^{-\frac{(\tau-t)^2}{2 \sigma^2}} }{\pi^{\frac{1}{4}} \sqrt{\sigma}}
= i\frac{e^{-\frac{(\tau-t)^2}{ 2 \sigma^2}}}{\pi^{\frac{1}{4}} \sqrt{\sigma}} \frac{\tau-t}{\sigma^2}.
\label{appendixSimp1}\end{aligned}$$ This allows for the simplification of Eq. $$\begin{aligned}
\tr \left [ E_{T_n}(\tau_n) \left[ H_{n}^{\rm clock} , \bar{\rho}_n(t) \right]\right]
&=
2i\frac{e^{-\frac{(\tau-t)^2}{ \sigma^2}} }{\sqrt{\pi} \sigma} \frac{\tau-t}{\sigma^2},\end{aligned}$$ and together with Eq. , this implies Eq. simplifies to $$\begin{aligned}
\tr \left [ E_{T_n}(\tau_n) \rho_{n}(t) \right] &= \tr \left [ E_{T_n}(\tau_n) \bar{\rho}_n(t)\right]
+ it \frac{\braket{H_n^{\rm ext }}}{mc^2} \tr \left [ E_{T_n}(\tau_n) \left[ H_{n}^{\rm clock} , \bar{\rho}_n(t) \right]\right] + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}}}{mc^2} t \right)^2 \right) \nn \\
&= \frac{e^{-\frac{(\tau-t)^2}{\sigma^2}}}{\sqrt{\pi} \sigma}
+ it \frac{ \braket{H_n^{\rm ext }} }{mc^2}\left(2i\frac{e^{-\frac{(\tau-t)^2}{ \sigma^2}} }{\sqrt{\pi} \sigma} \frac{\tau-t}{\sigma^2}\right) + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}}}{mc^2} t \right)^2 \right) \nn \\
&= \frac{ e^{-\frac{(\tau-t)^2}{\sigma^2}}}{\sqrt{\pi} \sigma^2} \left[1
- 2 \frac{ \braket{H_n^{\rm ext }} }{mc^2} \frac{t ( \tau - t) }{ \sigma^2} + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}}}{mc^2} t \right)^2 \right) \right].
\label{integrand2}\end{aligned}$$
Finally, with Eq. the conditional probability defined in Eq. maybe evaluated, yielding the result stated in Eq. $$\begin{aligned}
\prob \left[T_A = \tau_A \ | \ T_B = \tau_B \right] &= \frac{\int dt \, \tr \left [ E_A(\tau_A) \rho_{A}(t) \right] \tr \left [ E_B(\tau_B) \rho_{B}(t) \right]}{\int dt \, \tr \left [ E_B(\tau_B) \rho_{B}(t) \right]} \nn \\
&= \frac{\frac{e^{- \frac{(\tau_A-\tau_B)^2}{2 \sigma^2}}}{\sqrt{2 \pi} \sigma} \left[1 + \frac{ \braket{H^{\rm ext}_A} + \braket{H^{\rm ext}_B}}{2 mc^2} - \frac{\braket{H^{\rm ext}_A} - \braket{H^{\rm ext}_B} }{2 mc^2} \frac{\tau_A^2 - \tau_B^2}{\sigma^2} + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}}}{mc^2} \sigma \right)^2 \right) \right] }{1 + \frac{\braket{H^{\rm ext}_B}}{mc^2} + \mathcal{O}\!\left( \left( \frac{\braket{H^{\rm ext}_B}}{mc^2} \right)^2\right) } \nn \\
%&= \frac{e^{- \frac{(\tau_A-\tau_B)^2}{2 \Sigma^2}}}{\sqrt{2 \pi} \Sigma} \left( 1 + \frac{\Sigma^2 - \tau_A^2 + \tau_B^2}{2 \Sigma^2} \frac{\braket{H^{\rm ext}_{B}}}{mc^2} - \frac{\Sigma^2 - \tau_A^2 + \tau_B^2}{2 \Sigma^2} \frac{\braket{H^{\rm ext}_{A}}}{mc^2} \right) + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}}}{mc^2} \Sigma \right)^2 \right) \nn \\
&= \frac{e^{- \frac{(\tau_A-\tau_B)^2}{2 \sigma^2}}}{\sqrt{2 \pi} \sigma} \left[ 1 + \frac{\braket{H^{\rm ext}_{A}}- \braket{H^{\rm ext}_{B}}}{mc^2} \frac{\sigma^2 - \tau_A^2 + \tau_B^2}{2 \sigma^2} + \mathcal{O}\!\left( \left(\tfrac{\braket{H_n^{\rm ext }} \braket{H_n^{\rm clock}}}{mc^2} \sigma \right)^2 \right)\right].
\label{endResult}\end{aligned}$$
[^1]: The following text is an excerpt that has been edited for clarity from Feynman’s 1964 Messenger lectures delivered at Cornell University [@feynmanLectureRelationMathematics1964]:
> Consider two identical theories A and B, which look completely different psychologically and have different ideas in them, but all their consequences are exactly the same. A thing that people often say is how are we going to decide which one is right?
>
> No way! Not by science because both theory A and theory B agree with experiment to the same extent so there is no way to distinguish one from the other. So if two theories, though they may have deeply very different ideas behind them, can be shown to be mathematically equivalent then people usually say in science that the theories can not be distinguished.
>
> However, theories A and B for psychological reasons, in order to guess new theories, are very far from equivalent because one gives the scientist very different ideas than the other. By putting a theory in a given framework you get an idea of what to change. It may be the case that a simple change in theory A may be a very complicated change in theory B. In other words, although theories A and B are identical before they’re changed, there are certain ways of changing one that look natural which don’t look natural in the other.
>
> Therefore, psychologically we must keep all the theories in our head and every theoretical physicist that is any good knows six or seven different theoretical representations for exactly the same physics, and knows that they are all equivalent, and that nobody is every going to be able to decide which one is right at that level. But they keep these representations in their head hoping they will give them different ideas for guessing.
[^2]: $\mathcal{S}(\mathcal{H}) \ce \{ \rho \in \mathcal{T}(\mathcal{H}) \ | \ \rho \succeq 0 \mbox{ and } \tr \rho = 1\}$, where $\mathcal{T}(\mathcal{H})$ is the space of trace class operators acting on $\mathcal{H}$.
[^3]: $\mathcal{E}(\mathcal{H}) \ce \{ E \in \mathcal{B}(\mathcal{H}) \ | \ 0 \preceq E \preceq I \mbox{ and } E = E^\dagger \}$, where $\mathcal{B}(\mathcal{H})$ is the space of bounded operators acting on $\mathcal{H}$.
[^4]: A similar trick was employed in Ref. [@marneliusLagrangianHamiltonianFormulation1974] to formulate the theory of $N$ relativistic particles which do not carry internal degrees of freedom.
[^5]: Similar constraint factorization is described in more detail in Refs. [@hoehnHowSwitchRelational2018; @vanrietveldeChangePerspectiveSwitching2018].
[^6]: The Fisher information is defined as $$F(x; \rho(x))\ce \int d\xi \,p(\xi |x)\left(\frac{\mathrm{d} \ln p(\xi |x)}{dx}\right)^2, \nn$$ where $p(\xi |x) \ce \tr\left(E(\xi)\rho(x)\right)$ is the probability of getting the measurement result $\xi$ given that the parameter $x$ is encoded state of the quantum system $\rho(x)$. For an unbiased estimator, the Fisher information places a lower bound on the variance of the estimator, which is known as the Cramér-Rao bound [@wisemanQuantumMeasurementControl2010].
| {
"pile_set_name": "ArXiv"
} |
Dense liquids have highly nontrivial density correlations arising because the harshly repulsive molecular cores cannot overlap [@widomsci; @wca; @hansenmac]. Because of the constantly changing molecular arrangements, such correlations play a much more fundamental role in liquids than they do in other condensed phases such as glasses and solids, which sample only a few basic configurations. Indeed, a model with only repulsive intermolecular forces [@wca] can give a surprisingly accurate description of the full density correlations seen in a uniform dense simple liquid like Ar because the vector sum of the longer ranged attractive forces on a given particle essentially [*cancels*]{} [@widomsci] in most typical configurations.
[*Nonuniform*]{} liquids present a greater and qualitatively different challenge, since even the averaged effects of attractive forces clearly do not cancel [@wsb]. We discuss here an example where [*both*]{} attractive and repulsive forces can greatly influence the liquid’s structure: a Lennard-Jones (LJ) fluid next to a hard wall. We obtain accurate numerical results using a physically suggestive generalized mean field description of the attractive forces [@ourotherpapers]. We consider first the effects of these slowly varying forces on the liquid’s structure before taking account of the response to the rapidly varying (hard core like) part of the external field. This treatment of attractive interactions is quite different from that used in conventional integral equation and density functional methods [@denfungeneral], and we believe it offers important conceptual and computational advantages.
Fluid particles interact with a known external (wall) field $\phi ({\bf r})$ and through the LJ pair potential $w(r_{ij})\equiv u_0(r_{ij})+u_1(r_{ij})$, divided as usual [@wca] so that all the repulsive intermolecular forces arise from $u_0$ and all the attractive forces from $u_1.$ We assume that the external field $\phi ({\bf r})\equiv $ $\phi _0({\bf r})+\phi _1({\bf r}%
) $ can be separated in a similar way, where the subscript $0$ denotes in all that follows a harshly repulsive interaction and the subscript $1$ a much more slowly varying interaction usually associated with attractive forces. Here we take $\phi ({\bf r})$ to be a hard wall potential, setting $%
\phi _1({\bf r})=0$ and $\phi _0({\bf r})=$ $\phi _{HW}(z),$ where $\phi
_{HW}=\infty $ for $z\leq 1$ (in reduced units) and $0$ otherwise, and we let $\rho _B$ be the bulk density far from the wall.
We relate the structure of the nonuniform LJ system to that of a simpler [*nonuniform*]{} [*reference fluid* ]{}[@wsb; @wvk], with only repulsive intermolecular pair interactions $u_0(r_{ij})$ (equal to the LJ repulsions) in a different [*effective reference field*]{} (ERF) $\phi _{\mbox{\tiny R}}(%
{\bf r})$. While the replacement of attractive pair interactions by an approximate “molecular field” is an essential step in mean field theory, we can think of other more general choices. Here we determine $\phi _{%
\mbox{\tiny R}}({\bf r})$ formally by the requirement that it has a functional form such that the [*local*]{} (singlet) density at every point $%
{\bf r}$ in the reference fluid equals that of the full LJ fluid [@sullstell]: $$\rho _0({\bf r;[}\phi _{\mbox{\tiny R}}])=\rho ({\bf r;[}\phi ])\,.
\label{singletden}$$ The subscript $0$ reminds us that the reference system pair interactions arise only from $u_0$ and the notation ${\bf [}\phi _{\mbox{\tiny R}}]$ indicates that all distribution functions are functionals of the appropriate external field.
To find $\phi _{\mbox{\tiny R}}$ explicitly, we subtract the first equations of the YBG hierarchy [@hansenmac] for the full and reference systems with $\phi _{\mbox{\tiny
R}}$ chosen so that Eq. (\[singletden\]) is satisfied [@wsb; @wvk]. The result can be written exactly as $$\begin{aligned}
&-&\nabla _1[\phi _{\mbox{\tiny R}}({\bf r}_1)-\phi ({\bf r}_1)]=-\int d{\bf %
r}_2\rho _0({\bf r}_2|{\bf r}_1;[\phi _{\mbox{\tiny R}}])\nabla _1u_1(r_{12})
\nonumber \\
&&-\int d{\bf r}_2\{\rho ({\bf r}_2|{\bf r}_1;[\phi ])-\rho _0({\bf r}_2|%
{\bf r}_1;[\phi _{\mbox{\tiny R}}])\}\nabla _1w(r_{12})\,. \label{exactybg}\end{aligned}$$ Here $\rho _0({\bf r}_2|{\bf r}_1;{\bf [}\phi _{\mbox{\tiny R}}])\equiv \rho
_0^{(2)}({\bf r}_1,{\bf r}_2{\bf ;[}\phi _{\mbox{\tiny
R}}])/\rho _0({\bf r}_1{\bf ;[}\phi _{\mbox{\tiny R}}])$ is the [*conditional*]{} singlet density, i.e., the density at ${\bf r}_2$ given that a particle is fixed at ${\bf r}_1$.
If we assume that Eq. (\[singletden\]) produces similar local environments for the (identical) repulsive cores in the two fluids, which mainly determine density correlations through excluded volume effects, then the conditional singlet densities in the two fluids should also be very similar. This key structural assumption introduces a generalized mean field theory in which the reference fluid still has nontrivial pair and higher order correlations induced by the repulsive forces. This suggests that the last term on the R.H.S. in Eq. (\[exactybg\]) is often very small. If we ignore it completely [@wsbaccuracy] we obtain the approximate equation for the field $\phi _{\mbox{\tiny R}}$ suggested by Weeks, Selinger, and Broughton [@wsb]: $$\begin{aligned}
\nabla _1[\phi _{\mbox{\tiny R}}({\bf r}_1)-\phi ({\bf r}_1)]=\int d{\bf r}%
_2\rho _0({\bf r}_2|{\bf r}_1;{\bf [}\phi _{\mbox{\tiny
R}}])\nabla _1u_1(r_{12})\,. \label{wsb}\end{aligned}$$
Eq. (\[wsb\]) incorporates mean field ideas, but it appropriately focuses on [*forces* ]{}[@widomsci; @wca]. The relation to ordinary mean field theory becomes clearer [@wvk] if we replace $\rho _0({\bf r}_2|{\bf r}_1;%
{\bf [}\phi _{\mbox{\tiny R}}])$ by $\rho _0({\bf r}_2;{\bf [}\phi _{%
\mbox{\tiny R}}])$ in Eq.(\[wsb\]). This approximation is much better than one might at first suppose, since the main difference in these two functions occurs when ${\bf r}_2$ is close to ${\bf r}_1,$ but then for small $r_{12}$ the multiplicative factor $-\nabla _1u_1(r_{12})$ (the [*attractive*]{} part of the LJ force) vanishes identically. The gradient $\nabla
_1$ can then be taken outside the integral and Eq.(\[wsb\]) can be integrated. Choosing the constant of integration so that the density far from the wall equals $\rho _B$, we obtain the simplified mean field equation [@wvk]:
$$\begin{aligned}
\phi _{\mbox{\tiny R}}({\bf r}_{1})-\phi ({\bf r}_{1}) &\equiv & \nonumber
\\
\phi _{s}({\bf r}_{1}) &=&\int d{\bf r}_{2}\,[\rho _{0}({\bf r}_{2};{\bf [}%
\phi _{\mbox{\tiny R}}])-\rho _{B}]\,u_{1}(r_{12})\,. \label{mfint}\end{aligned}$$
Because of the integration over the slowly varying attractive potential “weighting function” $u_1(r_{12})$, $\phi _s({\bf r})$ in Eq. (\[mfint\]) extends smoothly into the repulsive core region where $\rho _0({\bf r};%
{\bf [}\phi _{\mbox{\tiny R}}])$ vanishes. Outside the wall it is smooth and relatively slowly varying even when $\rho _0({\bf r};{\bf [}\phi _{%
\mbox{\tiny R}}])$ itself has pronounced oscillations. Physically $\phi _s(%
{\bf r})$ mimics the effects of the unbalanced attractive forces in the LJ system, giving a soft [*repulsive*]{} interaction [@wsb] that tends to push the reference particles away from the wall.
In order to solve equations like (\[wsb\]) or (\[mfint\]) to obtain the self-consistent ERF $\phi _{\mbox{\tiny R}}({\bf r})$, we must determine the required reference fluid distribution functions arising from a given external field. In previous work [@wsb; @wvk], computer simulations were used for this purpose. We now introduce a simple and accurate numerical method for calculating these distribution functions and illustrate it here by solving (\[mfint\]) for the case of the LJ fluid near the hard wall.
We note that the ERF $\phi _{\mbox{\tiny
R}}({\bf r})\equiv \phi _{\mbox{\tiny
R}0}({\bf r})+\phi _{\mbox{\tiny
R}1}({\bf r})$ in Eq. (\[mfint\]) (and other related equations) can be naturally separated into the sum of a harshly repulsive part, $\phi _{%
\mbox{\tiny
R}0}({\bf r}),$ and a much more slowly varying “smooth” part $\phi _{%
\mbox{\tiny
R}1}({\bf r}),$ arising physically mainly from the attractive interactions in the original system. Eq. (\[mfint\]) suggests the identification $\phi
_{\mbox{\tiny
R}0}({\bf r})=$ $\phi _0({\bf r})$ and $\phi _{\mbox{\tiny
R}1}({\bf r})=$ $\phi _s({\bf r})+\phi _1({\bf r}).$ More generally, we can define $\phi _{\mbox{\tiny R}0}({\bf r})=$ $\phi _0({\bf r})-\phi _{0s}({\bf %
r})$ and $\phi _{\mbox{\tiny R}1}({\bf r})=$ $\phi _s({\bf r})+\phi _{0s}(%
{\bf r})+\phi _1({\bf r}),$ where $\phi _{0s}({\bf r})$ is an essentially arbitrary smooth function that is nonzero only in the repulsive core region but with $\phi _{0s}({\bf r})<<\phi _0({\bf r})$, so that $\phi _{%
\mbox{\tiny
R}0}$ remains a harshly repulsive interaction. In the present case it is sufficient to take the separation suggested by Eq. (\[mfint\]), with $\phi
_{\mbox{\tiny
R}0}=$ $\phi _{HW}$ and $\phi _{\mbox{\tiny
R}1}=\phi _s$.
Our task is now to determine the local density $\rho _0({\bf r};{\bf [}\phi
_{\mbox{\tiny
R}}])\equiv $ $\rho _{0,\mbox{\tiny
R}}({\bf r})$ produced by a given ERF $\phi _{\mbox{\tiny
R}}$. We provide a new way to solve this basic problem, quite independent of its origins in the mean field equation (\[mfint\]). Initially we treat the LJ repulsive potential $u_0$ as a hard core interaction, but then use standard methods [@hansenmac] to correct for its finite softness in our final numerical results. We expect that there will be very different responses of the reference fluid density to the rapidly and slowly varying parts of the ERF $\phi _{\mbox{\tiny R}}\equiv \phi _{\mbox{\tiny R}0}+\phi
_{\mbox{\tiny
R}1}$ and anticipate that any large oscillations arise mainly from the harshly repulsive part $\phi _{\mbox{\tiny R}0}$. These oscillations cause problems in density functional methods, which use a variety of somewhat arbitrary “weighting functions” to arrive at some underlying “smooth density” for use in a free energy functional [@denfungeneral].
=2.8in
Instead, we first determine the response to the [*slowly varying part of the ERF alone*]{}, followed by the response to the harshly repulsive part, using different methods in each step appropriate for the different density responses. In the first step, we determine the associated “smooth interface” $\rho _0({\bf r;[}\phi _{\mbox{\tiny
R}1}])\equiv $ $\rho _{0,\mbox{\tiny
R}1}({\bf r})$ that arises naturally from the slowly varying part $\phi _{%
\mbox{\tiny
R}1}=\phi _s$ of the ERF [*alone.*]{} Physically, this takes account of the effects of the attractive interactions modeled by $\phi _{\mbox{\tiny
R}1}$. We start from the basic equation relating small changes in the potential and density [@hansenmac]: $$-\beta \delta \phi _{\mbox{\tiny
R}1}({\bf r}_1)\,{\bf =}\int d{\bf r}_2\chi _0^{-1}({\bf r}_1,{\bf r}_2;{\bf %
[}\rho _{0,\mbox{\tiny
R}1}])\delta \rho _{0,\mbox{\tiny
R}1}({\bf r}_2)\,, \label{linresp}$$ through the generalized linear response function $\chi _0^{-1}({\bf r}_1,%
{\bf r}_2;{\bf [}\rho _{0,\mbox{\tiny
R}1}])\equiv \delta ({\bf r}_1\!-\!{\bf r}_2)/\rho _{0,\mbox{\tiny
R}1}({\bf r}_1)\!-\!c_0({\bf r}_1,{\bf r}_2;{\bf [}\rho _{0,\mbox{\tiny
R}1}])$. Here $c_0({\bf r}_1,{\bf r}_2;{\bf [}\rho _{0,\mbox{\tiny
R}1}])$ is the direct correlation function of the reference fluid with density $\rho _{0,\mbox{\tiny R}1}$ and $\beta =1/k_BT.$ Specializing to the case when the change is a small displacement of the field, we find the exact equation [@lmb]: $$\begin{aligned}
\nabla _1\rho _{0,\mbox{\tiny
R}1}({\bf r}_1) &/&\rho _{0,\mbox{\tiny
R}1}({\bf r}_1)=-\beta \nabla _1\phi _{\mbox{\tiny R}1}({\bf r}_1) \nonumber
\\
&+&\int d{\bf r}_2c_0({\bf r}_1,{\bf r}_2;{\bf [}\rho _{0,\mbox{\tiny R}%
1}])\nabla _2\rho _{0,\mbox{\tiny R}1}({\bf r}_2)\,. \label{LMB}\end{aligned}$$
If $\rho _{0,\mbox{\tiny
R}1}$ is relatively slowly varying, we can accurately approximate $c_0({\bf r%
}_1,{\bf r}_2;{\bf [}\rho _{0,\mbox{\tiny
R}1}])$ under the integral in Eq. (\[LMB\]) by the [*uniform fluid*]{} function $c_0(r_{12};\bar{\rho}_{12}),$ where $\bar{\rho}_{12}$ is some intermediate density associated with the two points [@frishleb]. A natural choice that gives very good results when $\rho _{0,\mbox{\tiny
R}1}$ is reasonably smooth is $\bar{\rho}_{12}=[\rho _{0,\mbox{\tiny
R}1}({\bf r}_1)+\rho _{0,\mbox{\tiny
R}1}({\bf r}_2)]/2$. Starting with a given $\phi _{\mbox{\tiny
R}1}$, we can then solve Eq. (\[LMB\]) for the associated $\rho _{0,%
\mbox{\tiny
R}1}$ by iteration, making use of the analytic and accurate Percus-Yevick (PY) expressions for the direct correlation function of the uniform hard sphere fluid [@frishleb; @hansenmac]. If necessary, we can choose $\phi _{%
\mbox{\tiny
R}1}$ inside the repulsive core region to help ensure that $\rho _{0,%
\mbox{\tiny
R}1}$ is smooth enough for the expansion method to be accurate; this procedure is important in some other applications [@ourotherpapers].
A special case where this step can be carried out analytically arises when $%
\rho _{0,\mbox{\tiny R}1}$ varies so slowly that it is accurate to keep only the first term in the expansion of $\nabla _2\rho _{0,\mbox{\tiny
R}1}({\bf r}_2)\,$ in Eq. (\[LMB\]) about ${\bf r}_1$. After integrating, we arrive at the simple [*local hydrostatic relation*]{} [@frishleb] between $\rho _{0,\mbox{\tiny
R}1}$ and $\phi _{\mbox{\tiny
R}1}$: $$\mu _0(\rho _{0,\mbox{\tiny
R}1}({\bf r}))+\phi _{\mbox{\tiny
R}1}({\bf r})=\mu _{0B}\,, \label{hydrostatic}$$ where $\mu _0(\rho )$ is the chemical potential of the uniform (hard sphere) reference fluid at density $\rho $ and $\mu _{0B}=$ $\mu _0(\rho _B).$
The smooth profile $\rho _{0,\mbox{\tiny R}1}$ is analogous to one that could be calculated using a single occupancy lattice gas (Ising) model, where correlations arise [*only*]{} from attractive interactions [@widomsci]. A realistic fluid has additional short wavelength correlations due to the repulsive intermolecular forces. These show up primarily in the second step of our method, where we take account of the response to $\phi _{%
\mbox{\tiny R}0}$, the remaining harshly repulsive part of the ERF.
Consider first a [*small*]{} perturbing potential $\delta \phi _{%
\mbox{\tiny
R}0}$ that is nonzero only inside the wall region with $z<1$. Evaluating Eq. (\[linresp\]) for $z_1>1$ gives an exact relation between the small induced density changes inside and outside the wall region. However, it has been shown that even large density fluctuations in a hard sphere fluid are accurately described by gaussian fluctuation theory [@crooks]. This suggests that if we could somehow [*impose*]{} the proper values on the wall density field for $z<1$ arising from the [*full*]{} $\phi _{\mbox{\tiny R}0}$, we could then still use the linear response relation to determine the large density change $\Delta \rho _{0,\mbox{\tiny
R}}({\bf r})\equiv \rho _{0,\mbox{\tiny
R}}({\bf r})-\rho _{0,\mbox{\tiny
R}1}({\bf r})$ induced for $z>1$. Imposing accurate density values in general is very difficult [@walloz], but for the hard wall potential $%
\phi _{\mbox{\tiny
R}0}=\phi _{HW}$ we have the exact result $\rho _{0,\mbox{\tiny
R}}({\bf r})=0$ for all $z\leq 1$. Thus replacing $\delta \rho _{0,%
\mbox{\tiny
R}1}$ by $\Delta \rho _{0,\mbox{\tiny
R}}$ in (\[linresp\]) and setting $\rho _{0,\mbox{\tiny
R}}=0$ for all $z\leq 1$, we find for $z_1>1$: $$\Delta \rho _{0,\mbox{\tiny
R}}({\bf r}_1)/\!\rho _{0,\mbox{\tiny
R}1}({\bf r}_1)=\int \!d{\bf r}_2\,c_0({\bf r}_1,{\bf r}_2;{\bf [}\rho _{0,%
\mbox{\tiny
R}1}])\Delta \rho _{0,\mbox{\tiny R}}({\bf r}_2). \label{wallpy}$$
Eq. (\[wallpy\]) is a linear equation for $\Delta \rho _{0,\mbox{\tiny
R}}({\bf r}_1),$ which we can directly solve by iteration or other means, approximating $c_0({\bf r}_1,{\bf r}_2;{\bf [}\rho _{0,\mbox{\tiny
R}1}])$ by that of an appropriate uniform system, just as we did before. When $\rho _{0,\mbox{\tiny R}1}({\bf r)}=\rho _B,$ Eq. (\[wallpy\]) is equivalent to the usual hard-wall, hard-particle PY equation, which has an analytic solution [@inteqns]. Eq. (\[wallpy\]) is quite adequate for our purposes here, though small errors can be seen at the highest densities. If still more accuracy is required, we could use modified GMSA type equations related to the PY equation [@inteqns]. It may also be possible to use new and very accurate density functional methods for hard core fluids in this step of our method [@rosenfeld].
The net result of this two step process is the desired $\rho _{0,%
\mbox{\tiny
R}}$ arising from a given $\phi _{\mbox{\tiny
R}}.$ This can be substituted into Eq. (\[mfint\]), which can then be iterated to determine the final self-consistent $\phi _{\mbox{\tiny R}}.$ In Fig. (1) we give the self-consistent potentials $\phi _{\mbox{\tiny R}%
1}=\phi _s$ that satisfy Eqs. (\[mfint\]), (\[LMB\]), and (\[wallpy\]) for two different states along the near critical isotherm $T=1.35$. We see that $\phi _{\mbox{\tiny R}1}$ is indeed a slowly varying repulsive interaction in both cases. In Fig. (2) we give the associated smooth density profiles $\rho _{0,\mbox{\tiny R}1}$ from (\[LMB\]) for each state, as well as the full profiles $\rho _{0,\mbox{\tiny R}}$ determined from Eq. (\[wallpy\]). These are compared to Monte Carlo simulations we carried out [@ourotherpapers] of the reference system in the ERFs of Fig. (1). This directly tests the accuracy of our two step procedure for calculating the effects of $\phi _{\mbox{\tiny R}}$ on the reference system. The agreement is excellent.
In Fig. (3) we test the simplified mean field treatment of the attractive interactions in Eq. (\[mfint\]) by comparing the reference profiles $\rho
_{0,\mbox{\tiny R}}$ to those of the full LJ fluid in the presence of the hard wall, as determined by MC calculations. There is good agreement, though small quantitative differences can be seen. Thus even the simplest mean field treatment of attractive interactions is capable of capturing the major changes in the density profile as the density is decreased, and at lower temperature at coexistence we find complete drying states where a stable vapor-liquid interface can exist arbitrarily far from the wall [@ourotherpapers].
Our emphasis thus far has been on quantitative numerical calculations. However, the qualitative features of our method are equally important. A long-standing problem of liquid state theory, well illustrated by the nonuniform fluid example studied here, is how to treat consistently the oscillating molecular scale “excluded volume” correlations arising from repulsive intermolecular forces and the more slowly varying and longer ranged correlations arising from attractive forces and often associated with the formation of interfaces. In principle these issues are addressed by modern density functional and integral equation methods, but in practice a number of uncontrolled and often mathematically motivated approximation are made. It is often difficult to assess their physical implications, and to determine where the major sources of error lie. Here, we have divided this problem into several distinct parts, whose accuracy can be examined separately, and where the physical content and limitations of the methods used are more clear. We used here the simplest mean field equation to determine the ERF $\phi _{\mbox{\tiny R}}$ but more accurate (though more complicated) equations derived from (\[exactybg\]) are available [@ourotherpapers].
To determine the structure of the reference fluid in the presence of a given ERF, we first calculated the response to $\phi _{\mbox{\tiny
R}1}$, the slowly varying part of the ERF, by expanding about a uniform system. Next we used a gaussian field model [@crooks] (equivalent to a modified wall-particle PY equation) to calculate the response to the remaining harshly repulsive part $\phi _{\mbox{\tiny R}0}$ of the ERF. More accurate methods could be used in both steps if necessary, and for qualitative purposes both steps can be simplified considerably. For example, Lum, Chandler, and Weeks [@lcw] have developed very simple approximations for use with continuum Landau-Ginsburgh type equations that give good qualitative results in a number of different cases, including hydrophobic interactions in water. Application of these ideas to a variety of different problems is underway [@ourotherpapers; @lcw].
This work was supported by NSF Grant No. CHE9528915. KV acknowledges support from the Deutsche Forshungsgemeinschaft. We thank D. Chandler, K. Lum, and J. Broughton for helpful discussions.
B. Widom, Science [**157**]{}, 375 (1967).
J. D. Weeks, D. Chandler and H. C. Andersen, J. Chem. Phys. [**54**]{}, 5237 (1971); D. Chandler, J. D. Weeks, and H. C. Andersen, Science [**220**]{}, 787 (1983).
See, e.g., J.P. Hansen and I.R. McDonald, [*Theory of Simple Liquids*]{}, (Academic, London, 1986).
J. D. Weeks, R .L .B. Selinger and J. Q. Broughton, Phys. Rev. Lett. [**75**]{}, 2694 (1995)
A more complete discussion of this problem and the related application to the liquid-vapor interface is given by K. Katsov, K. Vollmayr, J. D. Weeks, and J. Broughton (to be published). Studies of the uniform fluid, and confined fluids in slits and cylinders are also being carried out \[K. Vollmayr, K. Katsov, and J. D. Weeks (to be published)\].
For a general review, see R. Evans, in [*Fundamentals of Inhomogeneous Fluids*]{}, edited by D. Henderson (Dekker, New York, 1992).
J. D. Weeks, K. Vollmayr and K. Katsov, Physica A [**244**]{}, 461(1997).
D. E. Sullivan and G. Stell, J. Chem. Phys. [**69**]{}, 5450 (1978); D. E. Sullivan, D. Levesque, and J. J. Weis, Ibid. [**72**]{}, 1170 (1980).
We expect that this approximation will be most accurate at high density, as is the case for the uniform fluid. However, it also gives [*exact*]{} results as the density tends to zero, since then the last term vanishes. But the next order term in a density expansion can be evaluated exactly and does not vanish. As discussed in [@ourotherpapers], we can use this result in Eq. (\[exactybg\]) to derive modified equations that improve on the predictions of Eqs. (\[wsb\]) and (\[mfint\]) at very low but nonzero densities, and in some cases such as the liquid-vapor interface, these corrections can be quantitatively important.
R. A. Lovett, C. Y. Mou and F. P. Buff, J. Chem. Phys. [**65**]{}, 570 (1976); M. Wertheim, Ibid. [**65**]{}, 2377 (1976).
J. K. Percus, in [*The Equilibrium Theory of Classical Fluids*]{}, edited by H. L. Frisch and J. L. Lebowitz (W. A. Benjamin, Inc., New York, 1964), p. II-33.
D. Chandler, Phys Rev. E [**48**]{}, 2989 (1993); G. E. Crooks and D. Chandler, Phys. Rev. E [**56**]{}, 4217 (1997). See also G. Hummer, et al., Proc. Natl. Acad. Sci. [**93**]{}, 8951 (1996).
By definition [@frishleb], the [*exact*]{} density change $\Delta \rho _{0,\mbox{\tiny
R}}({\bf r})\equiv \rho _{0,\mbox{\tiny
R}}({\bf r})-\rho _{0,\mbox{\tiny
R}1}({\bf r})$ that arises when $\phi _{\mbox{\tiny
R}0}$ is taken into account satisfies Eq. (\[linresp\]) when the factor $%
-\beta \delta \phi _{\mbox{\tiny
R}1}({\bf r)}$ is replaced by the (wall-particle) [*direct correlation function*]{} $C_{\mbox{\tiny
R}0}({\bf r;[}\rho _{0,\mbox{\tiny
R}1}]).$ Accurately determining this for a general potential is very difficult. However, for the [*hard wall*]{}, if we make the the PY assumption that $C_{\mbox{\tiny
R}0}$ has the same range as $\phi _{\mbox{\tiny
R}0}$ and hence vanishes for $z>1,$ we again arrive at Eq. (\[wallpy\]) for $z>1.$
J. L. Lebowitz, Phys. Rev. [**133**]{}, A895 (1964); E. Waisman, Mol. Phys. [**25**]{}, 45 (1973).
Y. Rosenfeld, et al., Phys. Rev. E[**55**]{}, 4245 (1997).
K. Lum, D. Chandler, and J. D. Weeks (to be published).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the problem of reconstructing a wide sense stationary band-limited process from its local averages taken either at the Nyquist rate or above. As a result, we obtain a sufficient condition under which average sampling expansions hold in mean square and for almost all sample functions. Truncation and aliasing errors of the expansion are also discussed.'
address: 'Laboratory of Mathematics Applied to Systems, Ecole Centrale Paris, Grande voie de vignes, 92290 Chtâtenay Malabry, France'
author:
- Gilles Fay
- Sinuk Kang
title: 'Average sampling of band-limited stochastic processes'
---
average sampling ,wide sense stationary stochastic process ,sampling theorem 42C15 ,94A12
Introdution
===========
The classical Whittaker-Shannon-Kotel’nikov (WSK) sampling theorem [@Shannon:1949uo; @Whittaker:1915wb; @Kotelnikov:1933vx] says that a signal $f(t)$ in $PW_{\pi\omega}$, the Paley-Wiener space of functions band-limited to $[-\pi\omega,\pi\omega]$ (to be specified in Section \[notationdefinition\]), is uniquely determined by its discrete samples $f({n}/{\omega})$’s, $n\in\mathbb Z$, and can be reconstructed via $$f(t) = \sum_{n \in \mathbb Z } f(\frac{n}{\omega}) {\textnormal{sinc}}(\omega t -n)$$ which converges in $L^2(\mathbb R)$ and uniformly and absolutely on $\mathbb R$, where ${\textnormal{sinc}}t := {\sin \pi t}/({\pi t})$.
It is well known that the WSK sampling theorem has its counterpart for stochastic processes. The counterpart for wide sense stationary band-limited processes, definition of which is to be specified in Section \[notationdefinition\], is first introduced by Balakrishnan [@Balakrishnan:1957gx], and developed further by many more authors, among them Lloyd [@Lloyd:1959vh] and Beutler [@Beutler:1961wk]. Note that the sampling expansion in [@Lloyd:1959vh; @Beutler:1961wk] is shown to converge both in mean square and with probability 1, while the expansion in [@Balakrishnan:1957gx] is shown to converge in mean square only. There also has been an effort to extend the class of stochastic processes to which the WSK sampling theorem applies. Zakai [@ZAKAI:1965ha] extended the notion of band-limited stochastic processes and proved the WSK sampling theorem hold over this extended notion. Using Zakai’s technique, Lee [@Lee:1976uw] extended it to the class of second order measurable mean square continuous processes whose covariances are polynomially bounded. Belyaev [@Belyaev:1959iv] and Piranashvili [@Piranashvili:1967bi] defined classes of the analytic processes, for which almost all trajectories can be analytically continued. Then they derived the WSK sampling expansion which holds either with probability 1 or for almost all sample functions in their classes.
Recently, an average sampling expansions (ASE’s) for stochastic processes have been investigated in [@Song:2007fu; @He:2011ep; @Olenko:2011db]. Since acquisition devices do not produce signal values at the exact instances, in practice it seems more reasonable to use local averages instead of point evaluations. We call this sampling procedure the average sampling. The ASE on band-limited functions was first presented by Gröchenig [@Grochenig:1992ht], and then [extended]{} in [@Aldroubi:2002cz; @Feichtinger:1994td; @Sun:2002eq]. In [@Grochenig:1992ht], the author treated the problem of reconstructing a band-limited function from its local averages $\langle f , u_n \rangle = \int f(t)u_n(t) dt$ around $t_n$, where $\{u_n(t) : n \in \mathbb Z\}$ is a sequence of average functions satisfying $$\label{v1.0condaverage}
{\textnormal{supp}\,}u_n(t) \subset [t_n - \frac{\delta}{2}, t_n + \frac{\delta}{2}],~ 0 \leq u_n(t),~\textnormal{and } \int_{-\infty}^{\infty} u_n (t)dt = 1.$$ More precisely, he proved that for any $f(t)$ in $PW_{\pi\omega}$, if $0< t_{n+1} - t_n \leq \delta < {1}/({\sqrt{2}\pi\omega})$ then $f(t)$ is uniquely determined by the local averages $\langle f , u_n \rangle$’s, $n \in \mathbb Z$, and can be reconstructed by some iteration scheme. Applying the iteration scheme with $\{ \langle f , u_n \rangle: n \in \mathbb Z\}$ is equivalent to finding a frame expansion of $f$ on $PW_{\pi\omega}$, coefficients of which correspond to $\{ \langle f , u_n \rangle: n \in \mathbb Z\}$. That is, if the aforementioned condition is satisfied then there exists a frame $\{ r_n (t) : n \in \mathbb Z\}$ of $PW_{\pi\omega}$ such that $$\label{v1.0eq0.1}
f(t) = \sum_{n \in \mathbb Z} \langle f , u_n \rangle r_n (t), ~ f \in PW_{\pi\omega}$$ which converges in $L^2 (\mathbb R)$ and pointwise on $\mathbb R$. [Note here that $\{ \mathcal{P}u_n : n \in \mathbb Z\}$ and $\{r_n : n \in \mathbb Z\}$ are dual frame pairs of $PW_{\pi\omega}$ where $\mathcal{P}$ denotes the orthogonal projection of $L^2(\mathbb R)$ onto $PW_{\pi\omega}$.]{}
Using Gröchenig’s result, Song et al. [@Song:2007fu] addressed an ASE for band-limited stochastic processes. To be precise, let $X(t)$, $-\infty < t < \infty$, be a wide sense stationary stochastic process band-limited to $[-\pi\omega, \pi \omega]$ and $R_X(t)$ be its autocovariance function. Under the same notation and the assumption as in Gröchenig’s, they proved that if $$\label{v2.3eq2}
\{ t_n : n \in \mathbb Z\} \textnormal{ {is} relatively separable,}$$ i.e., there is some constant $N$ such that $[k,k+1] \cap \{ t_n : n \in \mathbb Z\}$ contains at most $N$ elements for all $k \in \mathbb Z$, and $$\label{v1.0eq1.2}
|R_X (t)| \leq R_X(0) (1+|t|)^{-\eta}~ \textnormal{ for some } \eta >1,$$ then $$\label{v2.1averageexp}
X(t) = \sum_{n \in \mathbb Z} \langle X, u_n \rangle r_n (t)$$ which converges in mean square, i.e., $$\lim_{N \rightarrow \infty} E \Big| X(t) - \sum_{n =-N}^{N} \langle X,u_n\rangle r_n (t) \Big|^2 = 0,$$ for any $t \in \mathbb R$. It is assumed here that $\{r_n (t) : n \in \mathbb Z\}$ is a frame of $PW_{\pi\omega}$ for which holds. Later, He et al. [@He:2011ep] provided an ASE to approximate wide sense stationary band-limited processes and, more generally, Olenko et al. [@Olenko:2011db] presented an ASE to approximate Piranashvili’s processes. But their ASE’s are asymptotically equivalent to the WSK sampling expansion, i.e., a point sampling.
It should be noted that the condition seems too strong. For instance, consider class of linear combination[s]{} of $\{ {\textnormal{sinc}}(\omega t-n) : n \in \mathbb Z\}$, denoted by $PW_{\pi\omega}^o$. Then $PW_{\pi\omega}^o$ is dense in $PW_{\pi\omega}$, i.e., any function in $PW_{\pi\omega}$ can be approximated [in $L^2(\mathbb R)$]{} by a limit of a sequence of functions in $PW_{\pi\omega}^o$, and for any $f(t) \in PW_{\pi\omega}^o$, $|f(t)|(1+|t|)^{\eta} \rightarrow \infty$ as $|t| \rightarrow \infty$ whenever $\eta >1$. [In fact, one can prove Theorem 2.2 of [@Song:2007fu] without assuming the condition and , based on the observation that converges unconditionally on $\mathbb R$ (see Theorem \[v1.0thm2.2\]).]{}
In this paper, we show that the average sampling theorem given in [@Song:2007fu] remains true without the condition and . We also provide an average sampling theorem with local averages taken at the Nyquist rate, while only oversampled local averages were considered in [@Song:2007fu]. Success of a perfect reconstruction from local averages via the resulting expansion depends only on the length of support of average functions $u_n$, which improves the results of [@Song:2007fu; @He:2011ep]. The latter ASE of ours not only converges in mean square but also converges for almost all sample functions. Under band-guard condition (see e.g. [@Helms:1962wo]) we derive explicit upper bounds on the truncation error of the ASE. Aliasing error is also discussed.
This paper is organized as follows. In Section \[notationdefinition\] we introduce notations and definitions needed throughout the paper. In Section \[secASO\] we present an average sampling theorem in which local averages are taken above the Nyquist rate (oversampling). In Section \[Nyquist\] we show that band-limited stochastic processes can also be reconstructed by its local averages taken at the Nyquist rate. It is shown that the resulting ASE converges both in mean square and for almost all sample functions. Finally in Section \[error\] truncation and aliasing errors of the expansion are discussed.
Notations and definitions {#notationdefinition}
=========================
The Paley-Wiener space of signals band-limited to $[-\pi\omega, \pi\omega]$ is defined by $$PW_{\pi\omega} := \{ f \in L^2(\mathbb R) \cap C(\mathbb R) : \textnormal{supp} \hat{f} \subset [-\pi\omega, \pi\omega] \}$$ where we define the Fourier transform as $\mathcal{F}[f](\xi) = \hat{f}(\xi) := \int_{-\infty}^{\infty} f(t) e^{-it\xi} dt$, $f \in L^1(\mathbb R)$, and extend it to an isomorphism from $L^2(\mathbb R)$ to $L^2(\mathbb R)$.
A sequence $\{\phi_n:n\in \mathbb{Z}\}$ of vectors in a separable Hilbert space $\mathcal{H}$ equipped with the norm $\| \cdot \|_\mathcal{H}$ is
- a frame of $\mathcal{H}$ with bounds $(A,B)$ if there are constants $B \geq A >0$ such that $$A\Vert f \Vert^{2}_\mathcal{H}\leq \sum_{n\in \mathbb{Z}}|\langle f ,\phi _{n}\rangle_\mathcal{H}|^{2}\leq B\Vert f\Vert^{2}_\mathcal{H},~f \in \mathcal{H};$$
- a Riesz (or stable) basis of $\mathcal{H}$ with bounds $(A,B)$ if $\{\phi_n:n\in \mathbb{Z}\}$ is complete in $\mathcal H$ and there are constants $B \geq A >0$ such that $$A\|\mathbf{c}\|^{2}\leq \Big\| \sum_{n\in \mathbb{Z}}c(n)\phi _{n}\Big\|_\mathcal{H}^{2}\leq B\| \mathbf{c} \|^{2},~\mathbf c := \{ c(n)\}_n \in \ell^2 (\mathbb Z)$$ where $\|\mathbf{c}\|^{2} = \sum_{n \in \mathbb Z} |c(n)|^2$.
A stochastic process $\{ X(t) : t \in \mathbb R \}$ is wide sense stationary if $E(X(t)) = 0$ and $E|X(t)|^2 < \infty$ for $t \in \mathbb R$ and the autocovariance function $R_X (t,s) :=E(X(t) \overline{X(s)})$ depends only on the difference $t-s$. By the spectral representation theorem [@Doob:1990us], a wide sense stationary process $\{ X(t) : t \in \mathbb R \}$ has a spectral representation: $X(t) = \int_{-\infty}^{\infty} e^{it\lambda} dy(\lambda)$, $ t \in \mathbb R$, where the process $y$ has orthogonal increments and $F$ is the spectral distribution function of $\{ X(t) : t \in \mathbb R \}$ such that $E|dy(\lambda)|^2 = dF(\lambda)$.
$\{ X(t) : t \in \mathbb R \}$ is said to be band-limited to $[-\pi\omega, \pi\omega]$ if its spectrum [has the support]{} $[-\pi\omega, \pi\omega]$, i.e., $X(t) = \int_{-\pi\omega}^{\pi\omega}e^{it\lambda} dy(\lambda)$, $t \in \mathbb R$. [Since $X(t) = \int_{-\pi\omega}^{\pi\omega}e^{it\lambda} dy(\lambda)$ if and only if $R_X (t,s) = \int_{-\pi\omega}^{\pi\omega}e^{i(t-s)\lambda} dF(\lambda) $ [@Cramer:1940vr], $\{ X(t) : t \in \mathbb R \}$ is also said to be band-limited to $[-\pi\omega,\pi\omega]$ if $R_X(t)$ is band-limited to $[-\pi\omega,\pi\omega]$, i.e., $R_X (t) \in PW_{\pi\omega}$.]{}
We denote by $\langle \cdot, \cdot \rangle$ the usual inner product in $L^2(\mathbb R)$ unless otherwise specified.
Oversampled local averages {#secASO}
==========================
We extend Theorem 2.2 of [@Song:2007fu] into the following theorem by removing the aforementioned condition and the constraint for $\{ t_n : n \in \mathbb Z\}$ being relatively separable.
\[v1.0thm2.2\] Let $\{u_n (t) : n \in \mathbb R\}$ be a sequence of average functions satisfying . For a wide sense stationary process $X(t)$ band-limited to $[-\pi\omega,\pi\omega]$, if $t_{n+1} - t_n \leq \delta < {1}/({\sqrt{2}\pi\omega})$ then $$\label{v1.0thm2.1eq4}
X(t) = \sum_{n \in \mathbb Z} \langle X,u_n \rangle r_n (t)$$ which converges in mean square, [uniformly on any compact subset of $\mathbb R$,]{} where $\{ r_n (t): n \in \mathbb Z\}$ is a frame of $PW_{\pi\omega}$ for which holds.
Since $E|X(t)|^2 <\infty$ for $t \in \mathbb R$, we have by Fubini’s theorem that for any $t \in \mathbb R$ and any positive integer $N$, $$\begin{aligned}
\label{v1.0eq2}
&& E \Big| X(t) - \sum_{n =-N}^{N} \langle X,u_n\rangle r_n (t) \Big|^2 \nonumber \\
&& = R_X (0) - \sum_{n =-N}^{N} \langle R_X(t-\cdot) ,u_n(\cdot) \rangle r_n (t)
- \sum_{n =-N}^{N} \overline{ \langle R_X(t-\cdot) ,u_n(\cdot) \rangle r_n (t)} \nonumber \\
&& + \sum_{k,n =-N}^{N} \int_{t_k-\frac{\delta}{2}}^{t_k+\frac{\delta}{2}} \int_{t_n-\frac{\delta}{2}}^{t_n+\frac{\delta}{2}} {R_X(x-y)}u_n (x) u_k (y) dx dy \,r_k(t) \overline{r_n (t)}.
\end{aligned}$$ We first show that the double summation converges unconditionally for any $t \in \mathbb R$ as $N$ goes to infinity. Since $R_X(t'-\cdot) \in PW_{\pi\omega}$ for $t' \in \mathbb R$, we obtain from that [for a given $t' \in \mathbb R$]{} $$\label{v1.0eq5.1}
R_X(t'-t) = \sum_{k \in \mathbb Z} \int_{t_k-\frac{\delta}{2}}^{t_k+\frac{\delta}{2}} R_X(t'-y)u_k(y)dy \,r_k (t).$$ It should be noticed that converges unconditionally [for each $t\in\mathbb R$ and uniformly on $\mathbb R$]{} since is a frame expansion in $PW_{\pi\omega}$ so that it converges unconditionally both in $L^2 (\mathbb R)$ and pointwise on $\mathbb R$ [and moreover the pointwise convergence is uniform]{}[: this follows by Corollary 3.1.5 of [@Christensen:2008wy] together with $PW_{\pi\omega}$ being a reproducing kernel Hilbert space [with bounded reproducing kernel]{} [@Higgins:1996uoa].]{} Setting $t = t'$, we have $$\label{v3.1sec3eq1}
R_X(0) = \sum_{k \in \mathbb Z} \int_{t_k-\frac{\delta}{2}}^{t_k+\frac{\delta}{2}} R_X(t-y)u_k(y)dy \,r_k (t)$$ which converges unconditionally for any $t\in \mathbb R$. [Furthermore, converges uniformly on any compact subset of $\mathbb R$ (Theorem 7.13 of [@Rudin:1964wf]).]{}
Now, for a given $x \in \mathbb R$, let $a_{k,x}(t) := \int_{t_k-{\delta}/{2}}^{t_k+{\delta}/{2}}{R_X(x-y)} u_k (y) dy \,r_k (t)$, $t \in \mathbb R$. Then $\sum_{k \in \mathbb Z} a_{k,x}(t)$ converges unconditionally to $R_X (x-t)$ for any $t \in \mathbb R$ and uniformly on $\mathbb R$ (with respect to t). Thus $$\begin{aligned}
\lim_{N \rightarrow \infty} \sum_{n=-N}^{N} \int_{t_n-\frac{\delta}{2}}^{t_n+\frac{\delta}{2}} \Big( \sum_{k \in \mathbb Z} a_{k,x}(t) \Big)u_n (x) dx\, \overline{r_n (t)}
= \sum_{n\in\mathbb Z} \int_{t_n-\frac{\delta}{2}}^{t_n+\frac{\delta}{2}} R_X(x-t) u_n (x) dx\, \overline{r_n (t)}\end{aligned}$$ converges unconditionally to $\overline{R_X(0)}$ for any $t \in \mathbb R$ and uniformly on any compact subset of $\mathbb R$, so that the double summation converges unconditionally for any $t \in \mathbb R$ and uniformly on any compact subset of $\mathbb R$ as $N$ goes to infinity.
[Since the absolute convergence is equivalent to the unconditional convergence for real- or complex-valued series (see e.g. Lemma 3.3 of [@Heil:2011jr]),]{} $$\label{v1.7theproof}
\lim_{N \rightarrow \infty} \sum_{k,n =-N}^{N} \int_{t_k-\frac{\delta}{2}}^{t_k+\frac{\delta}{2}} \int_{t_n-\frac{\delta}{2}}^{t_n+\frac{\delta}{2}} {R_X(x-y)}u_n (x) u_k (y) dx dy\, {r}_k(t) \overline{r_n (t)}$$ converges absolutely for any $t \in \mathbb R$.
Thus it follows by that for any $t \in \mathbb R$ $$\begin{aligned}
\eqref{v1.7theproof}
& = & \sum_{n \in \mathbb Z} \int_{t_n-\frac{\delta}{2}}^{t_n+\frac{\delta}{2}} \Big( \sum_{k \in \mathbb Z} \int_{t_k-\frac{\delta}{2}}^{t_k+\frac{\delta}{2}} {R_X(x-y)} u_k (y) dy \,{r}_k(t)\Big) \, u_n (x) ds\, \overline{r_n (t)} \\
& = & \sum_{n \in \mathbb Z} \int_{t_n-\frac{\delta}{2}}^{t_n+\frac{\delta}{2}}R_X (x-t) \, u_n (x) ds \, \overline{r_n (t)} \\
& = & \overline{R_X(0)}\end{aligned}$$ from which we have $\lim_{N \rightarrow \infty} E \Big| X(t) - \sum_{n =-N}^{N} \langle X,u_n\rangle r_n (t) \Big|^2 = 0$, [uniformly on any compact subset of $ \mathbb R$.]{}
The globally uniform convergence of is not guaranteed in general. It is worth mentioning that the necessary and sufficient condition for the WSK sampling expansion of so-called I-process, i.e., a band-limited stochastic process possessing an absolutely continuous spectral distribution function, to be globally uniformly convergent is addressed in [@Boche:2010dv].
The WSK sampling theorem for band-limited stochastic processes [@Balakrishnan:1957gx; @Lloyd:1959vh; @ZAKAI:1965ha] states that any $X(t)$ band-limited to $[-\pi\omega, \pi\omega]$ can be reconstructed by $$X(t) = \sum_{n \in \mathbb Z} X(\frac{n}{\omega}){\textnormal{sinc}}(\omega t - n)$$ which converges in mean square or with probability 1 for any $t \in \mathbb R$. In this case, the samples $\{X({n}/{\omega}):n \in \mathbb Z\}$ are taken at the Nyquist rate $\omega$. However, Theorem \[v1.0thm2.2\] does not cover the case of local averages taken at the Nyquist rate, $\{ \langle X, u_n \rangle : n \in \mathbb Z \}$, since if $t_n = {n}/{\omega}$ then $t_{n+1} - t_n = {1}/{\omega} > {1}/({\sqrt{2}\omega\pi})$, $n \in \mathbb Z$, where ${\textnormal{supp}\,}u_n \subset [t_n -{\delta}/{2}, t_n + {\delta}/{2}]$.
Local averages taken at the Nyquist rate {#Nyquist}
========================================
In this section we consider the case of local averages $\{ \langle X, v_n \rangle : n \in \mathbb Z \}$ taken at the Nyquist rate. In what follows, we assume $\omega = 1$. The aim of Section \[Nyquist\] is to derive an ASE of the form: $$\label{v3.3sec3eq1}
X(t) = \sum_{n \in \mathbb Z} \langle X,v_n \rangle s_n(t)$$ which converges in a proper sense. Here, $\{ s_n (t) : n \in \mathbb Z \} $ is a frame of $PW_\pi$ and $\{ v_n(t) : n \in \mathbb Z\}$ is a sequence of average functions satisfying $$\begin{aligned}
\label{v1.0eq7.1}
&& \textnormal{supp}\, v_n \subset [n-a,n+b] ~\textnormal{ for } a,b \geq 0 \textnormal{ and } a+b > 0, \\
&& 0 \leq v_n \in L^2 (\mathbb R), \textnormal{ and } \int_{-\infty}^{\infty} v_n (t) dt = \int_{n-a}^{n+b} v_n (t) dt = 1,~n \in \mathbb Z. \nonumber\end{aligned}$$
Our main results, Theorem \[v1.0thm2.4\] and \[v1.0thm2.5\], are based on the following ASE in $PW_\pi$.
\[v1.0prop2.3\] Let $\{v_n (t) : n \in \mathbb R\}$ be a sequence of average functions satisfying and let $\delta := \max\{a,b\}$. If $\sqrt{\delta(a+b)} < {1}/{\pi}$, then there is a frame $\{ s_n (t) : n \in \mathbb Z\} $ of $PW_{\pi}$ such that $$\label{v1.4propeq8}
f(t)= \sum_{n \in \mathbb Z} \langle f, v_n \rangle s_n (t), ~f \in PW_{\pi}$$ which converges in $L^2(\mathbb R)$ and uniformly and absolutely on $\mathbb R$. [In this case, $\{ s_n (t) : n \in \mathbb Z\} $ and $\{ \mathcal{P} v_n (t) : n \in \mathbb Z \} $ are dual frame pairs of $PW_\pi$ where $\mathcal{P}$ is the orthogonal projection of $L^2(\mathbb R)$ onto $PW_\pi$. ]{}
Let $\phi(t) := {\textnormal{sinc}}(t)$. Note that $\phi$ is differentiable, $\phi' \in L^2 (\mathbb R)$, $|Z_\phi (0,\xi)| = 1$ for $\xi \in \mathbb R$, and $\| Z_{\phi'}(t,\xi) \|_{L^\infty(\mathbb R^2)} = \pi$, where $Z_f (t,\xi):= \sum_{n \in \mathbb Z} f (t-n) e^{in\xi}$ denotes the Zak transform [@JANSSEN:1988ul] of $f(t) \in L^2 (\mathbb R)$. Then Proposition \[v1.0prop2.3\] is an immediate consequence of Theorem 3.2 of [@Kang:2011bf] with $\phi(t) = {\textnormal{sinc}}(t)$.
Since $R_X(t'-\cdot) \in PW_\pi$ for a given $t' \in \mathbb R$, assuming $\sqrt{\delta(a+b)} < {1}/{\pi}$, we have by Proposition \[v1.0prop2.3\] $$\label{v1.0sec4lem3.2}
R_X (t'-t) = \sum_{n \in \mathbb Z } \langle R_X (t'-\cdot), v_n (\cdot) \rangle s_n (t),~ t \in \mathbb R$$ which converges in $L^2 (\mathbb R)$ and absolutely and uniformly on $\mathbb R$. As already mentioned in the proof of Theorem \[v1.0thm2.2\], the absolute convergence is equivalent to the unconditional convergence for real- or complex-valued series, so also converges unconditionally on $\mathbb R$.
With the same definition as in Proposition \[v1.0prop2.3\], we have a counterpart statement for stochastic processes:
\[v1.0thm2.4\] Let $\{ X(t) : t \in \mathbb R \}$ be a wide sense stationary process band-limited to $[-\pi, \pi]$. If $\sqrt{\delta(a+b)} < {1}/{\pi}$, then converges in mean square, [uniformly on any compact subset of $\mathbb R$,]{} where $\{ s_n (t) : n \in \mathbb Z \}$ is a frame of $PW_{\pi}$ for which holds.
The proof is essentially the same as the proof of Theorem \[v1.0thm2.2\].
Furthermore, we prove the ASE converges for almost all sample functions.
\[v1.0thm2.5\] Let the notation and the assumption be the same as in Theorem \[v1.0thm2.4\]. Then holds for almost all sample functions.
Let $$\label{v1.9thmproofeq}
\tilde{X}(t) := \lim_{N \to \infty} \sum_{|n| < N } \langle X,v_n \rangle s_n (t)$$ in mean square sense. We have shown in the proof of Theorem \[v1.0thm2.2\] (with a proper modification of the notation) that for any $t \in \mathbb R$ $$\begin{aligned}
\label{v1.0eqepsilon}
E|X(t) - \tilde{X}(t) |^2 = \lim_{N \rightarrow \infty} E|X(t) - \sum_{n =-N}^N \langle X,v_n \rangle s_n (t) |^2.\end{aligned}$$ Thus, $X(t) = \tilde{X}(t)$ with probability 1 by Theorem \[v1.0thm2.4\]. The right of converges uniformly [on any compact subset of $\mathbb R$]{} and $X(t)$ is continuous for almost all sample functions: in fact, almost all sample functions are entire functions [@Belyaev:1959iv]. Then the theorem follows.
Error estimation {#error}
================
The aim of this section is to estimate truncation and aliasing errors of the ASE . To this end we always assume $\sqrt{\delta(a+b)} < {1}/{\pi}$ so that, by Theorem \[v1.0thm2.5\], $\{ \mathcal{P}v_n (t) : n \in \mathbb Z\}$ and $\{ s_n (t): n \in \mathbb Z\}$ are dual frame pairs of $PW_\pi$ for which holds for almost all sample functions of a given wide sense stationary process $X(t)$ band-limited to $[-\pi,\pi]$. Here, $\mathcal{P}$ is the orthogonal projection of $L^2(\mathbb R)$ onto $PW_\pi$. Note that $\mathcal{P}$ is a shift-invariant operator, i.e., $\mathcal{P}[f(\cdot-n)](t) = \mathcal{P}f (t-n),~n \in \mathbb Z$, for any $f \in L^2(\mathbb R)$.
In the following we assume further that $u_n (t) = u(t-n)$ for $n \in \mathbb Z$. Then $\{ \mathcal{P}u_n(t) = \mathcal{P}u(t-n) : n \in \mathbb Z\}$ is a frame of $PW_\pi$ if and only if there exist constants $B \geq A >0$ such that $$\label{v2.3sec4eq0.1}
A \leq | \hat{u}(\xi) | \leq B \textnormal{ a.e. on } [-\pi,\pi]$$ (Theorem 2 of [@Garcia:2005fz]). As a matter of fact, the condition is also a sufficient and necessary condition for $\{ \mathcal{P}u(t-n) : n \in \mathbb Z\}$ to be a Riesz basis of $PW_\pi$.
Since $\mathcal{P}u(t) \in PW_\pi$, we have by $$\mathcal{P}u(t) = \sum_{n \in \mathbb Z} \langle \mathcal{P}u(\cdot) , u(\cdot -n) \rangle s(t-n)$$ so that, via the Fourier transform, $$\begin{aligned}
\hat{u}(\xi)\chi_{[-\pi,\pi]}(\xi)
& = & \Big( \sum_{n \in \mathbb Z} \langle \hat{u}(\xi)\overline{\hat{u}(\xi)} , e^{-in\xi} \rangle_{L^2[-\pi,\pi]} e^{-in\xi} \Big) \hat{s}(\xi) \\
& = & {| \tilde{\hat{u}}(\xi)|^2 \hat{s}(\xi)}\end{aligned}$$ which holds in $L^2(\mathbb R)$ [where $\tilde{\hat{u}}(\xi)$ is $2\pi$-periodic extension of $\hat{u}(\xi)\big|_{[-\pi,\pi]}$, the restriction of $\hat{u}(\xi)$ on $[-\pi,\pi]$.]{} Thus we have $$\label{v2.3sec4eq1}
\overline{\hat{s}(\xi)} = \frac{1}{{\hat{u}(\xi)}} \chi_{[-\pi,\pi]}(\xi) \textnormal{ a.e. on } \mathbb R.$$
In summary we consider the ASE of the form $$\label{v2.6sec4eq2}
{X}(t) = \sum_{n \in \mathbb Z} \langle X(\cdot),u(\cdot-n) \rangle s(t-n)$$ where $u(t)$ and $s(t)$ satisfy and , respectively.
It is, by definition, unavoidable that $\hat{s}(\xi)$ has discontinuities at $\pm \pi$. Thus, for a given $t \in \mathbb R$, $s(t-n)$ decays slowly as $n$ goes to infinity so that convergence speed of is also slow. To overcome this, we adapt so-called oversampling technique by the guard-band assumption. This is introduced in [@Yao:1966hf; @Helms:1962wo; @BrownJr:1969wr] to estimate truncation error bound of the WSK sampling expansion of band-limited functions. The same method is also applied to truncation error estimation of the WSK sampling expansion of band-limited stochastic processes [@BrownJr:1968uh].
Consider $f(t) \in PW_\omega$ where $0< \omega <\pi$, i.e., ${\textnormal{supp}\,}\hat{f} \subset [-\omega,\omega]\subsetneq [-\pi,\pi]$. Since $PW_\omega \subseteq PW_\pi$, we have by Proposition \[v1.0prop2.3\] $$f(t) = \sum_{n \in \mathbb Z} \langle f( \cdot), u(\cdot-n) \rangle s(t-n)$$ which is, via the Fourier transform, equivalent to $$\label{v2.8sec4eq1}
\hat{f}(\xi) = \hat{f}(\xi)\theta(\xi) = \sum_{n \in \mathbb Z} \langle f( \cdot), u(\cdot-n) \rangle \hat{s}(\xi) \theta(\xi)e^{-in\xi}$$ where $\theta(\xi)$ is an arbitrary smooth function satisfying $\theta(\xi) = 1$ on $[-\omega,\omega]$ and $\theta(\xi)=0$ on $\mathbb R \backslash [-\pi,\pi]$. Applying the inverse Fourier transform on gives $$\label{v2.8sec4eq2}
f(t) = \sum_{n \in \mathbb Z} \langle f( \cdot), u(\cdot-n) \rangle \tilde{s}(t-n)$$ where $$\label{v2.8sec4eq3}
\hat{\tilde s} (\xi) = \hat s (\xi)\theta(\xi) = \frac{1}{\overline{\hat u (\xi)}} \theta (\xi)$$ in $L^2(\mathbb R)$. Note that $\tilde s (t) \in PW_\pi$.
\[v2.3sec4lemma1\] For $\tilde s(t)\in PW_\pi$ satisfying , if $\theta(\xi)$ is $p$-times continuously differentiable for some integer $p>1$, i.e., $\theta(\xi) \in C^p (\mathbb R)$, then for any given $t \in \mathbb R$ $$\label{v2.4sec4lemeq1}
|\tilde s(t-n)| \leq \frac{C_p(t)}{|n|^p},~ n \in \mathbb Z$$ where $$\label{v3.4sec4eq2}
C_p(t) := \frac{1}{2\pi} \int_{-\pi}^{\pi} \Big| (\frac{1}{\hat{u}(\xi)} \theta(\xi)e^{-it\xi} )^{(p)} \Big| d\xi < \infty .$$
Note first that $u(t)$ is compactly supported so that $\hat{u}(\xi)$ is infinitely many differentiable on $\mathbb R$. Since $$\tilde s(t-n) = \frac{1}{2\pi} \int_{-\pi}^{\pi}\hat{s}(\xi)\theta(\xi) e^{it\xi} e^{-in\xi} d\xi = \frac{1}{2\pi} \int_{-\pi}^{\pi} \frac{1}{\overline{\hat{u}(\xi)}} \theta(\xi) e^{it\xi} e^{-in\xi} d\xi$$ is the $n$-th coefficient of the Fourier series of $p$-times continuously differentiable function ${ \theta(\xi) e^{it\xi} }/{\overline{\hat{u}(\xi)}}$ vanishing at $\pm \pi$, on $[-\pi,\pi]$, we have $$\begin{aligned}
\int_{-\pi}^{\pi} (\frac{1}{\overline{\hat{u}(\xi)}}\theta(\xi)e^{it\xi} )^{(p)} e^{-in\xi} d\xi
= - (-in)^p \int_{-\pi}^{\pi}\frac{1}{\overline{\hat{u}(\xi)}}\theta(\xi)e^{it\xi} e^{-in\xi} d\xi
= - 2\pi(-in)^p \tilde s(t-n)\end{aligned}$$ which implies .
Let $X(t)$ be a wide sense stationary process band-limited to $[-\omega, \omega] \subsetneq [-\pi,\pi]$ and $R_X(t)$ the autocovariance function of $X(t)$. Assume that ASE holds. Then we have for any $t \in \mathbb R$ $$\label{v2.8sec4eq5}
X(t) = \sum_{n \in \mathbb Z} \langle X( \cdot), u(\cdot-n) \rangle \tilde{s}(t-n)$$ which converges in mean square where $\tilde s(t)$ is given by . Morever, if $\theta(\xi) $ in belongs to $ C^p (\mathbb R)$ for some integer $p>1$ then $$E|X(t) - X_N (t) |^2 \leq \frac{4 R_X(0) C_p(t)^2 }{(p-1)^2 N^{2(p-1)}}$$ where $$X_N(t):= \sum_{|n| \leq N} \langle X(\cdot), u(\cdot -n) \rangle \tilde s(t-n)$$ and $C_p (t)$ is given by .
Since $R_X(t'-\cdot) \in PW_\omega$ for a given $t' \in \mathbb R$, it follows by that $$\label{v2.8sec4eq4}
R_X (t'-t) = \sum_{n \in \mathbb Z } \langle R_X (t'-\cdot), u (\cdot-n) \rangle \tilde s(t-n), ~t \in \mathbb R$$ which converges in $L^2 (\mathbb R)$ and absolutely and uniformly on $\mathbb R$. Using we can obtain by the same argument as in the proof of Theorem \[v1.0thm2.2\].
For any $t \in \mathbb R$ it follows that $$\begin{aligned}
E|X(t) - X_N (t) |^2
&=& E\Big| \sum_{|n|>N} \langle X, u_n \rangle \tilde s_n (t) \Big|^2 \\
&=& \sum_{|n|>N} \int_{n-a}^{n+b} \Big( \sum_{|k|>N} \int_{k-a}^{k+b} R_X(x-y)u_k(y)dy \tilde s_k (t) \Big) u_n (x)dx \, \overline{\tilde s_n (t)} \\
& \leq & \sum_{|n|>N} \int_{n-a}^{n+b} \left( \sum_{|k|>N} \Big| \int_{k-a}^{k+b} R_X(x-y)u_k(y)dy\Big| \Big| \tilde s_k (t)\Big| \right) u_n (x)dx \, \overline{\tilde s_n (t)}\\
& \leq & \sup_{-(a+b) \leq t \leq a+b} |R_X (t)| \sum_{|n|>N} \int_{n-a}^{n+b} \Big( \sum_{|k|>N} | \tilde s_k (t)| \Big) u_n (x)dx \, \overline{\tilde s_n (t)} \\
& \leq &R_X (0) \Big( \sum_{|k|>N} | \tilde s_k (t)| \Big)^2\end{aligned}$$ in which $\tilde s_k(t) = \tilde s(t-k)$ and $u_k (t) = u(t-k)$ for $k \in \mathbb Z$. By Lemma \[v2.3sec4lemma1\] combined with the integral test, $$\begin{aligned}
\sum_{|k|>N} | \tilde s(t-k)| \leq
\sum_{ k > N} \frac{2 C_p(t)}{k^p}
\leq \frac{2 C_p(t)}{(p-1)N^{p-1}}\end{aligned}$$ which proves the theorem.
As already mentioned, $\{\mathcal{P}u(t-n) : n \in \mathbb Z\}$ and $\{ s(t-n) : n \in \mathbb Z\}$ are dual Riesz basis pairs of $PW_\pi$. Thus the orthogonal projection $\mathcal{P}$ of $L^2(\mathbb R)$ onto $PW_\pi$ can be written as $$\mathcal{P}f (t) = \sum_{n \in \mathbb Z} \langle f(\cdot), \mathcal{P}u(\cdot-n) \rangle s(t-n).$$ To consider an aliasing error of ASE , we extend $\mathcal{P}$ for stochastic processes as $$\label{v2.6sec4proj}
{\mathcal{P}}X(t) := \sum_{n \in \mathbb Z} \langle X(\cdot), \mathcal{P}u(\cdot-n) \rangle s(t-n).$$ Assuming that ASE holds for wide sense stationary processes band-limited to $[-\pi,\pi]$, one can easily see that for any $t \in \mathbb R$ $$E|X(t) - {\mathcal{P}}X(t) |^2 = \int_{|\lambda| > \pi} F(d\lambda)$$ where $F$ is the spectral distribution function of $X(t)$.
Acknowledgement {#acknowledgement .unnumbered}
===============
Sinuk Kang is partially supported by Erasmus Mundus BEAM program funded by European Commission.
[31]{} natexlab\#1[\#1]{}\[2\][\#2]{} , , () . , , () . , , (). , , () . , , () . , , () . , , () . , , () . , , () . , , () . , , , , , () . , , , () . , , , () . , , () . , , () . , , , in: , , , , pp. . , , , () . , , , () . , , Wiley Classics Library, , , . , , () . , , an introductory course, , . , , Volume 1: Foundations, , . , , Second edition, , , . , , Applied and Numerical Harmonic Analysis, , edition, . , , , () . , , , () . , , () . , , , , () . , , , () . , , () . , , () .
| {
"pile_set_name": "ArXiv"
} |
---
author:
- ZEUS Collaboration
date: July 2007
title: ' Three- and four-jet final states in photoproduction at HERA '
---
[ The ZEUS Collaboration ]{}
S. Chekanov$^{ 1}$, M. Derrick, S. Magill, B. Musgrave, D. Nicholass$^{ 2}$, , R. Yoshida\
[*Argonne National Laboratory, Argonne, Illinois 60439-4815*]{}, USA $^{n}$
M.C.K. Mattingly\
[*Andrews University, Berrien Springs, Michigan 49104-0380*]{}, USA
M. Jechow, N. Pavel $^{\dagger}$, A.G. Yagües Molina\
[*Institut für Physik der Humboldt-Universität zu Berlin, Berlin, Germany*]{}
S. Antonelli, P. Antonioli, G. Bari, M. Basile, L. Bellagamba, M. Bindi, D. Boscherini, A. Bruni, G. Bruni, , F. Cindolo, A. Contin, M. Corradi, S. De Pasquale, G. Iacobucci, , R. Nania, A. Polini, G. Sartorelli, A. Zichichi\
[*University and INFN Bologna, Bologna, Italy*]{} $^{e}$
D. Bartsch, I. Brock, H. Hartmann, E. Hilger, H.-P. Jakob, M. Jüngst, O.M. Kind$^{ 3}$, , E. Paul$^{ 4}$, R. Renner$^{ 5}$, U. Samson, V. Schönberg, R. Shehzadi, M. Wlasenko\
[*Physikalisches Institut der Universität Bonn, Bonn, Germany*]{} $^{b}$
N.H. Brook, G.P. Heath, J.D. Morris\
[*H.H. Wills Physics Laboratory, University of Bristol, Bristol, United Kingdom*]{} $^{m}$
M. Capua, S. Fazio, A. Mastroberardino, M. Schioppa, G. Susinno, E. Tassi\
[*Calabria University, Physics Department and INFN, Cosenza, Italy*]{} $^{e}$
J.Y. Kim$^{ 6}$, K.J. Ma$^{ 7}$\
[*Chonnam National University, Kwangju, South Korea*]{} $^{g}$
Z.A. Ibrahim, B. Kamaluddin, W.A.T. Wan Abdullah\
[*Jabatan Fizik, Universiti Malaya, 50603 Kuala Lumpur, Malaysia*]{} $^{r}$
Y. Ning, Z. Ren, F. Sciulli\
[*Nevis Laboratories, Columbia University, Irvington on Hudson, New York 10027*]{} $^{o}$
J. Chwastowski, A. Eskreys, J. Figiel, A. Galas, M. Gil, K. Olkiewicz, P. Stopa, L. Zawiejski\
[*The Henryk Niewodniczanski Institute of Nuclear Physics, Polish Academy of Sciences, Cracow, Poland*]{} $^{i}$
L. Adamczyk, T. Bołd, I. Grabowska-Bołd, D. Kisielewska, J. Łukasik, , L. Suszycki\
[*Faculty of Physics and Applied Computer Science, AGH-University of Science and Technology, Cracow, Poland*]{} $^{p}$
A. Kotański$^{ 8}$, W. S[ł]{}omiński$^{ 9}$\
[*Department of Physics, Jagellonian University, Cracow, Poland*]{}
V. Adler$^{ 10}$, U. Behrens, I. Bloch, C. Blohm, A. Bonato, K. Borras, R. Ciesielski, N. Coppola, , V. Drugakov, J. Fourletova, A. Geiser, D. Gladkov, P. Göttlicher$^{ 11}$, J. Grebenyuk, I. Gregor, T. Haas, W. Hain, C. Horn$^{ 12}$, A. Hüttmann, B. Kahle, I.I. Katkov, U. Klein$^{ 13}$, U. Kötz, H. Kowalski, , B. Löhr, R. Mankel, I.-A. Melzer-Pellmann, S. Miglioranzi, A. Montanari, T. Namsoo, D. Notz, L. Rinaldi, P. Roloff, I. Rubinsky, R. Santamarta, , A. Spiridonov$^{ 14}$, H. Stadie, D. Szuba$^{ 15}$, J. Szuba$^{ 16}$, T. Theedt, G. Wolf, K. Wrona, C. Youngman,\
[*Deutsches Elektronen-Synchrotron DESY, Hamburg, Germany*]{}
W. Lohmann,\
[*Deutsches Elektronen-Synchrotron DESY, Zeuthen, Germany*]{}
G. Barbagli, E. Gallo, P. G. Pelfer\
[*University and INFN Florence, Florence, Italy*]{} $^{e}$
A. Bamberger, D. Dobur, F. Karstens, N.N. Vlasov$^{ 17}$\
[*Fakultät für Physik der Universität Freiburg i.Br., Freiburg i.Br., Germany*]{} $^{b}$
P.J. Bussey, A.T. Doyle, W. Dunne, M. Forrest, D.H. Saxon, I.O. Skillicorn\
[*Department of Physics and Astronomy, University of Glasgow, Glasgow, United Kingdom*]{} $^{m}$
I. Gialas$^{ 18}$, K. Papageorgiu\
[*Department of Engineering in Management and Finance, Univ. of Aegean, Greece*]{}
T. Gosau, U. Holm, R. Klanner, E. Lohrmann, H. Salehi, P. Schleper, , J. Sztuk, K. Wichmann, K. Wick\
[*Hamburg University, Institute of Exp. Physics, Hamburg, Germany*]{} $^{b}$
C. Foudas, C. Fry, K.R. Long, A.D. Tapper\
[*Imperial College London, High Energy Nuclear Physics Group, London, United Kingdom*]{} $^{m}$
M. Kataoka$^{ 19}$, T. Matsumoto, K. Nagano, K. Tokushuku$^{ 20}$, S. Yamada, Y. Yamazaki$^{ 21}$\
[*Institute of Particle and Nuclear Studies, KEK, Tsukuba, Japan*]{} $^{f}$
A.N. Barakbaev, E.G. Boos, N.S. Pokrovskiy, B.O. Zhautykov\
[*Institute of Physics and Technology of Ministry of Education and Science of Kazakhstan, Almaty,* ]{}
V. Aushev$^{ 1}$, M. Borodin, A. Kozulia, M. Lisovyi\
[*Institute for Nuclear Research, National Academy of Sciences, Kiev and Kiev National University, Kiev, Ukraine*]{}
D. Son\
[*Kyungpook National University, Center for High Energy Physics, Daegu, South Korea*]{} $^{g}$
J. de Favereau, K. Piotrzkowski\
[*Institut de Physique Nucléaire, Université Catholique de Louvain, Louvain-la-Neuve, Belgium*]{} $^{q}$
F. Barreiro, C. Glasman$^{ 22}$, M. Jimenez, L. Labarga, J. del Peso, E. Ron, M. Soares, J. Terrón,\
[*Departamento de Física Teórica, Universidad Autónoma de Madrid, Madrid, Spain*]{} $^{l}$
F. Corriveau, C. Liu, R. Walsh, C. Zhou\
[*Department of Physics, McGill University, Montréal, Québec, Canada H3A 2T8*]{} $^{a}$
T. Tsurugai\
[*Meiji Gakuin University, Faculty of General Education, Yokohama, Japan*]{} $^{f}$
A. Antonov, B.A. Dolgoshein, V. Sosnovtsev, A. Stifutkin, S. Suchkov\
[*Moscow Engineering Physics Institute, Moscow, Russia*]{} $^{j}$
R.K. Dementiev, P.F. Ermolov, L.K. Gladilin, L.A. Khein, I.A. Korzhavina, V.A. Kuzmin, B.B. Levchenko$^{ 23}$, O.Yu. Lukina, A.S. Proskuryakov, L.M. Shcheglova, D.S. Zotkin, S.A. Zotkin\
[*Moscow State University, Institute of Nuclear Physics, Moscow, Russia*]{} $^{k}$
I. Abt, C. Büttner, A. Caldwell, D. Kollar, W.B. Schmidke, J. Sutiak\
[*Max-Planck-Institut für Physik, München, Germany*]{}
G. Grigorescu, A. Keramidas, E. Koffeman, P. Kooijman, A. Pellegrino, H. Tiecke, M. Vázquez$^{ 19}$,\
[*NIKHEF and University of Amsterdam, Amsterdam, Netherlands*]{} $^{h}$
N. Brümmer, B. Bylsma, L.S. Durkin, A. Lee, T.Y. Ling\
[*Physics Department, Ohio State University, Columbus, Ohio 43210*]{} $^{n}$
P.D. Allfrey, M.A. Bell, A.M. Cooper-Sarkar, R.C.E. Devenish, J. Ferrando, B. Foster, K. Korcsak-Gorzo, K. Oliver, S. Patel, V. Roberfroid$^{ 24}$, A. Robertson, P.B. Straub, C. Uribe-Estrada, R. Walczak\
[*Department of Physics, University of Oxford, Oxford United Kingdom*]{} $^{m}$
P. Bellan, A. Bertolin, R. Brugnera, R. Carlin, F. Dal Corso, S. Dusini, A. Garfagnini, S. Limentani, A. Longhin, L. Stanco, M. Turcato\
[*Dipartimento di Fisica dell’ Università and INFN, Padova, Italy*]{} $^{e}$
B.Y. Oh, A. Raval, J. Ukleja$^{ 25}$, J.J. Whitmore$^{ 26}$\
[*Department of Physics, Pennsylvania State University, University Park, Pennsylvania 16802*]{} $^{o}$
Y. Iga\
[*Polytechnic University, Sagamihara, Japan*]{} $^{f}$
G. D’Agostini, G. Marini, A. Nigro\
[*Dipartimento di Fisica, Università ’La Sapienza’ and INFN, Rome, Italy*]{} $^{e}~$
J.E. Cole, J.C. Hart\
[*Rutherford Appleton Laboratory, Chilton, Didcot, Oxon, United Kingdom*]{} $^{m}$
H. Abramowicz$^{ 27}$, A. Gabareen, R. Ingbir, S. Kananov, A. Levy\
[*Raymond and Beverly Sackler Faculty of Exact Sciences, School of Physics, Tel-Aviv University, Tel-Aviv, Israel*]{} $^{d}$
M. Kuze, J. Maeda\
[*Department of Physics, Tokyo Institute of Technology, Tokyo, Japan*]{} $^{f}$
R. Hori, S. Kagawa$^{ 28}$, N. Okazaki, S. Shimizu, T. Tawara\
[*Department of Physics, University of Tokyo, Tokyo, Japan*]{} $^{f}$
R. Hamatsu, H. Kaji$^{ 29}$, S. Kitamura$^{ 30}$, O. Ota, Y.D. Ri\
[*Tokyo Metropolitan University, Department of Physics, Tokyo, Japan*]{} $^{f}$
M.I. Ferrero, V. Monaco, R. Sacchi, A. Solano\
[*Università di Torino and INFN, Torino, Italy*]{} $^{e}$
M. Arneodo, M. Ruspa\
[*Università del Piemonte Orientale, Novara, and INFN, Torino, Italy*]{} $^{e}$
S. Fourletov, J.F. Martin\
[*Department of Physics, University of Toronto, Toronto, Ontario, Canada M5S 1A7*]{} $^{a}$
S.K. Boutle$^{ 18}$, J.M. Butterworth, C. Gwenlan$^{ 31}$, T.W. Jones, J.H. Loizides, M.R. Sutton$^{ 31}$, M. Wing\
[*Physics and Astronomy Department, University College London, London, United Kingdom*]{} $^{m}$
B. Brzozowska, J. Ciborowski$^{ 32}$, G. Grzelak, P. Kulinski, P. [Ł]{}użniak$^{ 33}$, J. Malka$^{ 33}$, R.J. Nowak, J.M. Pawlak, A. Ukleja, A.F. Żarnecki\
[*Warsaw University, Institute of Experimental Physics, Warsaw, Poland*]{}
M. Adamus, P. Plucinski$^{ 34}$\
[*Institute for Nuclear Studies, Warsaw, Poland*]{}
Y. Eisenberg, I. Giller, D. Hochman, U. Karshon, M. Rosin\
[*Department of Particle Physics, Weizmann Institute, Rehovot, Israel*]{} $^{c}$
E. Brownson, T. Danielson, A. Everett, D. Kçira, D.D. Reeder$^{ 4}$, P. Ryan, A.A. Savin, W.H. Smith, H. Wolfe\
[*Department of Physics, University of Wisconsin, Madison, Wisconsin 53706*]{}, USA $^{n}$
S. Bhadra, C.D. Catterall, Y. Cui, G. Hartner, S. Menary, U. Noor, J. Standage, J. Whyte\
[*Department of Physics, York University, Ontario, Canada M3J 1P3*]{} $^{a}$
$^{\ 1}$ supported by DESY, Germany\
$^{\ 2}$ also affiliated with University College London, UK\
$^{\ 3}$ now at Humboldt University, Berlin, Germany\
$^{\ 4}$ retired\
$^{\ 5}$ self-employed\
$^{\ 6}$ supported by Chonnam National University in 2005\
$^{\ 7}$ supported by a scholarship of the World Laboratory Björn Wiik Research Project\
$^{\ 8}$ supported by the research grant no. 1 P03B 04529 (2005-2008)\
$^{\ 9}$ This work was supported in part by the Marie Curie Actions Transfer of Knowledge project COCOS (contract MTKD-CT-2004-517186)\
$^{ 10}$ now at Univ. Libre de Bruxelles, Belgium\
$^{ 11}$ now at DESY group FEB, Hamburg, Germany\
$^{ 12}$ now at Stanford Linear Accelerator Center, Stanford, USA\
$^{ 13}$ now at University of Liverpool, UK\
$^{ 14}$ also at Institut of Theoretical and Experimental Physics, Moscow, Russia\
$^{ 15}$ also at INP, Cracow, Poland\
$^{ 16}$ on leave of absence from FPACS, AGH-UST, Cracow, Poland\
$^{ 17}$ partly supported by Moscow State University, Russia\
$^{ 18}$ also affiliated with DESY\
$^{ 19}$ now at CERN, Geneva, Switzerland\
$^{ 20}$ also at University of Tokyo, Japan\
$^{ 21}$ now at Kobe University, Japan\
$^{ 22}$ Ram[ó]{}n y Cajal Fellow\
$^{ 23}$ partly supported by Russian Foundation for Basic Research grant no. 05-02-39028-NSFC-a\
$^{ 24}$ EU Marie Curie Fellow\
$^{ 25}$ partially supported by Warsaw University, Poland\
$^{ 26}$ This material was based on work supported by the National Science Foundation, while working at the Foundation.\
$^{ 27}$ also at Max Planck Institute, Munich, Germany, Alexander von Humboldt Research Award\
$^{ 28}$ now at KEK, Tsukuba, Japan\
$^{ 29}$ now at Nagoya University, Japan\
$^{ 30}$ Department of Radiological Science\
$^{ 31}$ PPARC Advanced fellow\
$^{ 32}$ also at Łódź University, Poland\
$^{ 33}$ Łódź University, Poland\
$^{ 34}$ supported by the Polish Ministry for Education and Science grant no. 1 P03B 14129\
$^{\dagger}$ deceased\
-------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$^{a}$ supported by the Natural Sciences and Engineering Research Council of Canada (NSERC)
$^{b}$ supported by the German Federal Ministry for Education and Research (BMBF), under contract numbers HZ1GUA 2, HZ1GUB 0, HZ1PDA 5, HZ1VFA 5
$^{c}$ supported in part by the MINERVA Gesellschaft für Forschung GmbH, the Israel Science Foundation (grant no. 293/02-11.2) and the U.S.-Israel Binational Science Foundation
$^{d}$ supported by the German-Israeli Foundation and the Israel Science Foundation
$^{e}$ supported by the Italian National Institute for Nuclear Physics (INFN)
$^{f}$ supported by the Japanese Ministry of Education, Culture, Sports, Science and Technology (MEXT) and its grants for Scientific Research
$^{g}$ supported by the Korean Ministry of Education and Korea Science and Engineering Foundation
$^{h}$ supported by the Netherlands Foundation for Research on Matter (FOM)
$^{i}$ supported by the Polish State Committee for Scientific Research, grant no. 620/E-77/SPB/DESY/P-03/DZ 117/2003-2005 and grant no. 1P03B07427/2004-2006
$^{j}$ partially supported by the German Federal Ministry for Education and Research (BMBF)
$^{k}$ supported by RF Presidential grant N 8122.2006.2 for the leading scientific schools and by the Russian Ministry of Education and Science through its grant Research on High Energy Physics
$^{l}$ supported by the Spanish Ministry of Education and Science through funds provided by CICYT
$^{m}$ supported by the Particle Physics and Astronomy Research Council, UK
$^{n}$ supported by the US Department of Energy
$^{o}$ supported by the US National Science Foundation. Any opinion, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.
$^{p}$ supported by the Polish Ministry of Science and Higher Education as a scientific project (2006-2008)
$^{q}$ supported by FNRS and its associated funds (IISN and FRIA) and by an Inter-University Attraction Poles Programme subsidised by the Belgian Federal Science Policy Office
$^{r}$ supported by the Malaysian Ministry of Science, Technology and Innovation/Akademi Sains Malaysia grant SAGA 66-02-03-0048
-------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Introduction
============
In photoproduction at HERA, a quasi-real photon, emitted by the incoming electron[^1], interacts with the proton. Hard photoproduction [@pl:b79:83; @np:b166:413; @pr:d21:54; @zfp:c6:241; @proc:hera:1987:331; @prl:61:275; @prl:61:682; @pr:d39:169; @zfp:c42:657; @pr:d40:2844] may be categorised at leading order (LO) as being either direct, if the photon interacts as a point-like particle, or resolved, if it fluctuates into a partonic system, and subsequently transfers only a fraction of its momentum to the hard interaction.
Photoproduction collisions at HERA can lead to final states with multiple jets. Multi-jet events are of particular interest as they are produced by processes that are manifestly beyond LO in the strong coupling constant, $\alpha_s$. Presently, predictions from perturbative quantum chromodynamics (pQCD) pertaining to multi-jet final states in photoproduction are only available up to $\mathcal{O}(\alpha\alpha_s^{2})$, where $\alpha$ is the fine structure constant.
The hadron-like structure of the photon in resolved processes gives rise to the possibility of multi-parton interactions (MPIs) at HERA. In the MPI picture, more than one pair of partons takes part in the hard interaction. A schematic of an MPI event is shown in Fig. \[fig:MPI\]. The secondary scatters generate additional hadronic energy flow in the event, the topology and magnitude of which is poorly understood theoretically. Potentially, this energy flow may lead to the formation of jets and so MPIs may constitute a source of multi-jets in the final state. Multi-parton interactions have been studied before at the Tevatron [@MPIsTEV4; @MPIsTEV1; @MPIsTEV2; @MPIsTEV5], in dijet photoproduction events at HERA [@Art:zeusxg; @MPIsHERA1] and are expected to be prevalent at the LHC [@MPIsLHC1; @MPIsLHC2].
Three-jet photoproduction events have been studied before at HERA in the $M_{3j}\ge50$ GeV [@Art:zeusPP1] region of three-jet invariant mass. Multi-jet final states have also been studied in deep inelastic scattering (DIS) at HERA [@Art:h1mjdis; @Art:h13jdis; @Art:zeus3jalpha; @Art:zeus3jdiff]. In this paper, differential cross sections are presented for the three-jet photoproduction final state, in a wider $M_{3j}$ region and with over seven times the luminosity of the previous ZEUS publication [@Art:zeusPP1]. The four-jet photoproduction cross sections are presented here for the first time. Also examined is the description of the data by two Monte Carlo (MC) models and whether or not it is improved by the introduction of simulated MPIs. In addition, an $\mathcal{O}(\alpha\alpha_s^2)$ prediction has been compared to the three-jet data.
Experimental conditions
=======================
The data were collected using the ZEUS detector during the 1996 to 2000 running periods. In 1996 and 1997, HERA collided positrons with protons, with energies of $E_e=27.5$ GeV and $E_p=820$ GeV, respectively, corresponding to a beam centre-of-mass energy, $\sqrt{s}$, of 300 GeV. From 1998 onwards the proton beam energy was raised to $E_p=920$ GeV ($\sqrt{s}=318$ GeV). Furthermore, a subsample of these later data consists of electron-proton collisions. The total sample corresponds to an integrated luminosity of $(121\pm2)~{\rm~pb^{\rm -1}}$, of which $82~{\rm~pb^{\rm -1}}$ were collected at $\sqrt{s}=318$ GeV. A detailed description of the ZEUS detector can be found elsewhere [@pl:b297:404; @zeus:1993:bluebook]. A brief outline of the components that are of most relevance to this analysis is given below.
The high-resolution uranium–scintillator calorimeter (CAL) [@nim:a309:77; @*nim:a309:101; @*nim:a336:23; @*nim:a321:356] consists of three parts: the forward, the barrel and the rear calorimeters. Each part is subdivided transversely into towers and longitudinally into one electromagnetic and either one (in the rear) or two (in the barrel and forward) hadronic sections. The smallest subdivision of the calorimeter is called a cell. The CAL relative energy resolutions, as measured under test-beam conditions, are $0.18/\sqrt{E}$ for electrons and $0.35/\sqrt{E}$ for hadrons, with $E$ in GeV.
Charged particles are measured in the central tracking detector (CTD) [@nim:a279:290; @*npps:b32:181; @*nim:a338:254], which operates in a magnetic field of 1.43 T provided by a thin superconducting coil. The CTD covers the polar-angle[^2] region . The relative transverse-momentum resolution for full-length tracks is $\sigma_{p_T}/p_T=0.0058p_T \oplus 0.0065 \oplus 0.0014/p_T$, with $p_T$ in GeV. Both tracking and calorimetry were used to reconstruct the transverse energy and direction of jets as described in Section \[eventrecon\].
The luminosity was measured from the rate of the bremsstrahlung process $ep~\rightarrow~e\gamma p$, where the photon is detected in a lead–scintillator calorimeter [@desy-92-066-tim; @*zfp:c63:391; @*desy-01-141] placed in the HERA tunnel at $Z=-107$ m.
Event reconstruction and selection {#eventrecon}
==================================
A three-level trigger system was used to select events online [@zeus:1993:bluebook; @epj:c1:109]. At the first two levels, general characteristics of photoproduction collisions were required and background from beam-gas events was rejected. At the third level, jets were reconstructed by applying a cone jet algorithm to the CAL cells. Events with at least two jets satisfying $E_T^{\rm jet}\ge4.5$ GeV and $\eta^{\rm jet}\le2.5$ were accepted, where $E_T^{\rm jet}$ and $\eta^{\rm jet}$ are the transverse energy and pseudorapidity of the jet evaluated in the laboratory frame.
In the offline analysis, the hadronic final state was reconstructed using energy-flow objects [@Art:efo1; @thesis:hefa] (EFOs), which are formed from a combination of track and calorimeter information. This approach optimises the energy resolution and the one-to-one correspondence between the detector-level objects and the hadrons. The EFOs were corrected [@Art:efocorr; @Art:ZEUSEFO] to account for energy losses in the dead material and were forced to be massless by setting the energy component equal to the magnitude of the three-momentum.
The jets were reconstructed using the $k_T$ cluster algorithm [@Art:ktclus] in the longitudinally invariant inclusive mode [@Art:ktclusinc] using the $p_T^2$ recombination scheme. At the detector level, jets were formed from the EFOs. The energy and angular reconstruction of the detector-level jets exhibited no systematic bias with respect to those defined at the hadron level.
The data sample was selected by requiring the following:
- the longitudinal position of the reconstructed vertex was in the range $|Z_{\rm vtx}|\le 40$ cm;
- $y_{\rm JB}\ge 0.2$, to reduce the contamination from beam-gas events, where $y_{\rm JB}$ is the Jacquet–Blondel estimator [@proc:epfacility:1979:391] of the inelasticity, $y$, the fraction of the incoming electron momentum carried by the photon;
- no scattered electron was observed in the CAL with $y_e\le0.85$, to remove background due to DIS, where $y_e$ is the electron-method estimator of $y$;
- $y_{\rm JB}\le 0.85$, which further removed contamination from DIS;
- $P_T^{\rm miss}/\sqrt{E_T}\le 2$ GeV$^{1/2}$, where $P_T^{\rm miss}$ is the missing transverse momentum and $E_T$ is the total transverse energy. This removed any background from charged current, cosmic and halo-muon events;
- at least three (or four, depending on the specific sample) jets were found with $|\eta^{\rm jet}|\le2.4$ and $E_T^{\rm jet}\ge 6~\rm{GeV}$.
Furthermore, the three- and four-jet analyses were conducted in low- and high-mass regions given by $25\le M_{nj}\le50$ GeV and $M_{nj}\ge50$ GeV, respectively, where $M_{nj}$ is the invariant mass of the $n$-jet system.
After all of the selection criteria had been applied, the low- and high-mass, three-jet data samples had 291646 and 38098 events, respectively, while the equivalent four-jet data samples had 31533 and 12525 events.
Photoproduction events are characterised by the low virtuality, $Q^2$, of the quasi-real exchanged photon, where $Q^2$ is the negative of the photon four-momentum squared. The kinematics are specified by $y$ and $x_\gamma$, the fraction of the photon momentum taking part in the interaction. The variable $x_\gamma$ can be approximated using the observable $x_\gamma^{\rm obs}$, defined as $$x_\gamma^{\rm obs}=\frac{\sum_{i=1}^{\rm n}E_{T,i}^{\rm jet}\exp(-\eta_i^{\rm jet})}{2yE_e},
\label{equ:xgammaobs}$$ where the sum runs over the $n$-jets considered in the event and $E_e=27.5$ GeV denotes the energy of the incoming electron.
An angular variable used to analyse three-jet events is $\psi_3$ [@Art:multijetkine], which, as shown in Fig. \[fig:angles\], is the angle in the three-jet centre-of-mass frame between the plane containing the highest energy jet and the beam, and the plane containing the three jets. The beam is described by the three-vector ${\bf p}_{\rm beam}={\bf p}_p-{\bf p}_e$, where ${\bf p}_p$ and ${\bf p}_e$ are the momenta of the proton and electron beams, respectively. The variable may be written as $$\cos(\psi_{\rm 3})=\frac{({\bf p}_{\rm beam}\times{\bf p}_{\rm 3})\cdot({\bf p}_{\rm 4}\times {\bf p}_{\rm 5})}{|{\bf p}_{\rm beam}\times{\bf p}_{\rm 3}||{\bf p}_{\rm 4}\times{\bf p}_{\rm 5}|},
\label{equ:cos_psi3}$$ where it is conventional to number the three jets, 3, 4 and 5, in order of decreasing energy. The $\psi_3$ angle reflects the orientation of the lowest-energy jet. In the case where this jet arises from initial-state gluon radiation, the coherence property of QCD will tend to orient the third jet close to the incoming proton or photon direction. The two planes shown in Fig. \[fig:angles\] will therefore tend to coincide leading to a $\psi_3$ distribution that peaks toward 0 and $\pi$.
Monte Carlo models
==================
Two MC generators were used to simulate photoproduction events, [Herwig]{} 6.505 [@Art:herwig; @*Art:herwig2; @*Art:herwig3] and [Pythia]{} 6.206 [@Art:pythia; @*Art:pythia2]. Both models include the LO ($2\rightarrow2$) matrix elements, approximate higher-order processes using initial-state and final-state parton showers and simulate hadronisation. Direct and resolved photoproduction samples were generated separately. The implementation of the parton-showers and hadronisation models in the generators differs [@Art:pythia; @*Art:pythia2; @Art:hrwPS1; @Art:hrwPS2; @Art:hrwPS3]. Hadronisation in [Herwig]{} is simulated using the cluster model [@Art:hrwHad1] while [Pythia]{} uses the Lund string model [@Art:pytHad1; @Art:pytHad2].
Underlying-event models
-----------------------
In the [Herwig]{} samples, MPIs were simulated using a separate program called [Jimmy]{} 4.0 [@Art:butterworthMPI3; @Art:butterworthMPI1; @Art:butterworthMPI2; @unp:JIMMY4]. The latter is based on a simple eikonal model that approximates the interacting hadrons as disks that are extended in the transverse plane. An impact parameter quantifies the degree to which the two disks overlap during the collision.
The [Pythia]{} MPI model that was used in this paper is known as the “simple model” [@Art:pythia; @*Art:pythia2]. It estimates the average number of MPIs per event, $\bar n$, as $\bar n = \sigma_H(\hat{p}_T^{\rm min})/\sigma_{\rm ND}(s)$, where $\sigma_H$ and $\sigma_{\rm ND}$ are the inclusive-hard and non-diffractive, inelastic cross sections, respectively, and $\hat{p}_T^{\rm min}$ is a minimal constraint applied to the $p_T$ of the partonic collision. The number of MPIs is assumed to obey a Poisson distribution. The probability that the secondary collisions occur with transverse momentum $p_T$ is given by $P(p_T)=(1/\sigma_{\rm ND}(s))({\rm d}\sigma_H/{\rm d}p_T)$. The simple model proceeds by ordering the secondary interactions in terms of $p_T$ and then calculates the probability of each successive scatter.
Monte Carlo samples
-------------------
The MC samples were used for two purposes: to correct the data for detector effects and to compare to the measured hadron-level cross sections. The MC samples used to correct the data included a full [Geant]{} 3.13 [@unp:geant] simulation of the ZEUS detector and three-level trigger. The resolved and direct MC samples were combined in the ratio that gave the best fit to the $x_\gamma^{\rm obs}$ distribution in the data, significantly improving the overall description of the data by the MC. The resolved and direct samples used to compare to the measured cross sections, however, were combined in the ratio predicted by the MC models.
The [Herwig]{} samples that were compared to the data were generated with $\hat{p}_T^{\rm min}$ set to 4 GeV and with the CTEQ5L [@Art:cteq5l] and GRV-G LO [@Art:grvgloTN] parameterisations for the proton and photon parton density functions (PDFs), respectively. All other [Herwig]{} parameters were set to default. To simulate the [Herwig]{} MPIs, the [Jimmy]{} model was run with the minimum $p_T$ of the secondary scatters, $\hat{p}_T^{\rm mpi}$, set to 2.2 GeV, the probability that the photon would resolve via a large hadron-like fluctuation set to $1/340$ and the effective transverse radius of the photon squared was increased by a factor of 4 from its default value. These adjusted MPI settings were the result of simultaneous tuning to the three- and four-jet $x_\gamma^{\rm obs}$ cross sections measured in this analysis, in the low- and high-mass regions.
The [Pythia]{} samples that were compared to the data were generated with $\hat{p}_T^{\rm min}=4.5$ GeV, $\hat{p}_T^{\rm mpi}=1.9$ GeV, which is the default value, and with the CTEQ5L and GRV-G LO parameterisations for the proton and photon PDFs, respectively. All other [Pythia]{} and simple-model parameters were set to default.
Monte Carlo scaling factors {#sec:MCmodels}
---------------------------
Both LO MC simulations underestimated the magnitude of the measured cross sections. To compare with the shape of the differential cross sections, the MC predictions were either area normalised to the equivalent data set or scaled by the factor required to area normalise the high-mass MC cross sections to the high-mass data, where MPIs are expected to be less influential. The choice of normalisation procedure will be clearly stated in Section \[sec:results\] and in the figure captions.
The scaling factors that were applied to the [Herwig]{} and [Pythia]{} cross sections are shown in Table \[tab:mc\_scl\]. The scaling factors for the [Pythia]{} model were significantly larger than for [Herwig]{}. Within each model, the factors applied to the three-jet predictions were similar whether MPIs were included or not, indicating that both models expect relatively little effect from MPIs in the high-mass three-jet sample. In the four-jet case, however, the scaling factors for the high-mass MPI samples were significantly smaller than those for the nominal samples, showing that both models predict a sizable contribution from MPIs in the high-mass four-jet sample.
The $\mathcal{O}(\alpha\alpha_s^2)$ pQCD calculation
====================================================
The three-jet differential cross sections were calculated at $\mathcal{O}(\alpha\alpha_s^2)$ using the program by Klasen, Kleinwort and Kramer [@Art:klasen1; @Art:klasen2]. The calculation is LO for this process. The renormalisation, $\mu_r$, and factorisation, $\mu_f$, scales were set to $\mu_r=\mu_f=\mu=E_T^{\rm max}$, the $E_T$ of the hardest parton. The theoretical scale uncertainty was evaluated by varying $\mu$, setting it to $2^{\rm \pm1}E_T^{\rm max}$. The value of $\alpha_s$ was calculated with one-loop precision and assuming five active flavours. A value of $\Lambda^{(5)}_{\rm \overline{MS}}=181$ MeV was taken. The CTEQ6L [@Art:cteq6l] proton and GRV-G LO [@Art:grvgnlo; @*Art:grvglo] photon PDFs were used for the calculation.
The theoretical calculations were corrected for both hadronisation and MPI effects. The correction factors were obtained using the [Herwig]{} and [Pythia]{} models. The hadronisation corrections, $\mathcal{C}_{\rm had}$, were calculated by taking the bin-by-bin ratio of the MC cross sections at the hadron ($\sigma_{\rm HL}$) and parton levels ($\sigma_{\rm PL}$), $\mathcal{C}_{\rm had}=\sigma_{\rm HL}/\sigma_{\rm PL}$. To obtain $\sigma_{\rm PL}$, the $k_T$ cluster algorithm was run over all partons produced by the parton shower, prior to hadronisation. The MPI corrections, $\mathcal{C}_{\rm MPI}$, were calculated by taking the bin-by-bin ratio of the hadron-level cross sections, with ($\sigma_{\rm HL}^{\rm MPI}$) and without MPIs ($\sigma_{\rm HL}^{\rm noMPI}$), $\mathcal{C}_{\rm MPI}=\sigma_{\rm HL}^{\rm MPI}/\sigma_{\rm HL}^{\rm noMPI}$. The hadronisation and MPI corrections that were applied to the calculation represented the average corrections taken from the two MC models. A symmetric uncertainty was associated with the correction factors equal to half the difference between the predictions of the two MC models. The hadronisation corrections were found to typically reduce the cross section by $\mathcal{O}(20\%)$, whereas the MPI corrections increased the cross section by $\mathcal{O}(50\%)$ in the low-mass region and by $\mathcal{O}(10\%)$ in the high-mass sample. The uncertainty associated with the MPI corrections was larger than that estimated for the hadronisation factors.
Acceptance corrections {#sec:unfold}
======================
Detector effects and trigger inefficiencies were accounted for by applying bin-by-bin correction factors, $\mathcal{C}_{\rm det}=\sigma_{\rm HL}/\sigma_{\rm DL}$, where $\sigma_{\rm DL}$ is the predicted detector-level cross section. The correction factors were extracted separately from the [Herwig]{} and [Pythia]{} MC samples with MPIs. Before calculating the corrections, the resolved components of the MC samples were reweighted in $y_{\rm JB}$ to improve the overall description of the data [@thesis:tim]. The typical bin-by-bin corrections applied to the low- and high-mass three-(four-)jet data were approximately $0.75$ ($0.45$) and $1.3$ ($0.95$), respectively. The largest correction factors were found in the lowest $E_T^{\rm jet}$ and the extreme $\eta^{\rm jet}$ bins.
Once corrected to the hadron level, the measured cross sections, $\sigma_{\rm \sqrt{s}}$, at $\sqrt{s}=300$ GeV and $\sqrt{s}=318$ GeV were combined using the following formula:
$$\sigma = \frac{\sigma_{\rm 300}\cdot(\sigma_{\rm 318}^{\rm mc}/\sigma_{\rm 300}^{\rm mc})\cdot\mathcal{L}_{\rm 300} + \sigma_{\rm 318}\cdot\mathcal{L}_{\rm 318}}{\mathcal{L}_{\rm 300}+\mathcal{L}_{\rm 318}},
\label{equ:epcom}$$
where $\mathcal{L}_{\rm \sqrt{s}}$ is the integrated luminosity and $\sigma_{\rm \sqrt{s}}^{\rm mc}$, the predicted cross section at $\sqrt{s}$. As such, the measured cross sections presented here correspond to $\sqrt{s}=318$ GeV. The $\sigma_{\rm 318}^{\rm mc}/\sigma_{\rm 300}^{\rm mc}$ ratios were $\sim 1.1$ for all but the low-mass four-jet sample, where it was around 10% larger.
The results presented here represent the average hadron-level cross sections obtained when the data is treated with [Herwig]{} or [Pythia]{}, with half the spread interpreted as symmetric systematic uncertainty. This systematic uncertainty was added in quadrature to those described below.
Systematic uncertainties {#sec:syst}
========================
A detailed study of the sources of systematic uncertainty associated with the measurement was performed using the [Herwig]{} MC sample [@thesis:tim]. The sources contributing to the quoted systematic uncertainties are,
- the CAL energy scale uncertainty is $\pm3\%$. The energy scale was changed in the MC simulation accordingly;
- the uncertainty due to the dependence of the measured cross sections on the detector-level selection criteria, which were varied up and down by the detector resolution for the data and the MC samples together. More specifically,
- the $E_T^{\rm jet}$ cut on each of the jets was changed by $\pm1$ GeV;
- the $|\eta^{\rm jet}|$ cut on each of the jets was changed by $\pm0.04$;
- the $M_{nj}$ selection criteria were changed by $\pm10\%$;
- the lower and upper cut on $y_{\rm JB}$ was changed by $\pm0.03$ and $\pm0.05$, respectively;
- the cut on $y_e$, used to differentiate between real and fake scattered-electron candidates, was changed by $\pm0.05$.
In addition, it was verified that the bin-by-bin acceptance corrections were not sensitive to variations in the relative amount of direct and resolved events or the exact shape of the $y_{\rm JB}$ distribution in the combined MC sample.
All the systematic uncertainties were added in quadrature except that associated with the CAL energy scale. The CAL energy scale uncertainty is highly correlated between bins and is displayed separately in the plots presented here.
The largest systematic uncertainty in the cross sections, except for the high-mass three-jet case, was that associated with the CAL energy scale, which led to an uncertainty of approximately $\pm10\%$ in both the low- and high-mass three-jet cross sections, and $\pm20\%$ and $\pm15\%$ in the low- and high-mass four-jet cross sections, respectively. The largest uncertainty in the high-mass three-jet cross section came from the choice of MC model used to calculate the acceptance corrections and was approximately $\pm10\%$. In the low-mass three-jet sample, the uncertainty due to the MC model was $\pm5\%$ and in the four-jet low- and high-mass samples the values were $\pm3\%$ and $\pm4\%$, respectively. Another significant source of systematic uncertainty in the low-mass samples was that associated with the $E_T^{\rm jet}$ selection criteria, which generated a $\pm4\%$ and $\pm8\%$ effect in the three- and four-jet cases, respectively. In the high-mass samples, varying the $M_{nj}$ selection criteria generated a $\pm8\%$ and $\pm6\%$ effect in the three- and four-jet cross sections, respectively. The high-mass four-jet cross section was also sensitive to varying the $E_T^{\rm jet}$ cut, which changed its value by about $\pm4\%$.
Results and discussion {#sec:results}
======================
The three- and four-jet photoproduction cross sections are presented here at the hadron level for jets with $E_T^{\rm jet}\ge6$ GeV and $|\eta^{\rm jet}|\le 2.4$, in the kinematic region given by $Q^2<1~{\rm GeV^2}$ and $0.2\le y\le0.85$, and in low- ($25\le M_{nj}<50$ GeV) and high-mass ($M_{nj}\ge50$ GeV) regions.
The ${\rm d}\sigma/{\rm d}M_{nj}$ cross section
-----------------------------------------------
The three- and four-jet cross sections are given as a function of $M_{nj}$ in Fig. \[fig:xsec:Mnj\] and Tables \[tab:Mnj\_3j\_\] and \[tab:Mnj\_4j\_\]. In general, both cross sections decrease exponentially with increasing $M_{nj}$. Also shown in Fig. \[fig:xsec:Mnj\] are the [Herwig]{} and [Pythia]{} predictions without MPIs, normalised to the high-mass region ($M_{nj}\ge50$ GeV). Both models incorrectly describe the $M_{nj}$ dependence of the cross section and significantly underestimate the low mass data. The discrepancy is larger in the four-jet case. With the inclusion of MPIs, both scaled MC predictions give a reasonably good description of the data over the full $M_{nj}$ ranges.
The ${\rm d}\sigma/{\rm d}x_\gamma^{\rm obs}$ cross section
-----------------------------------------------------------
The three- and four-jet cross sections are given as a function of $x_\gamma^{\rm obs}$ for the low- and high-mass samples in Fig. \[fig:xsec:Xgam\] and Tables \[tab:X\_gamma\_3j\_\] and \[tab:X\_gamma\_4j\_\]. The distributions exhibit a peak at $x_\gamma^{\rm obs}\approx0.9$ in all but the low-mass four-jet sample. The low-mass cross sections show enhancement at low $x_\gamma^{\rm obs}$ compared to those at high-mass, which is a consequence of the tighter kinematic constraints at high mass. The degree to which the low $x_\gamma^{\rm obs}$ region is enhanced is larger in the four-jet case. Shown also in Fig. \[fig:xsec:Xgam\] are the scaled MC predictions compared to the data.
The MC predictions show that, in the absence of MPIs, the three- and four-jet cross sections are expected to decrease with decreasing $x_\gamma^{\rm obs}$ in both mass regions. Moreover, the low-$x_\gamma^{\rm obs}$ suppression is expected to be marginally stronger in the four-jet case. The low-mass data, however, are in contradiction to both predictions. These data therefore suggest that some mechanism in addition to the processes modelled in the MC without MPIs, is contributing to the low-mass cross sections. One possible mechanism is MPIs.
A comparison of the data with the MC predictions, normalised to the high-mass data, shows that both those with and without MPIs are in reasonable agreement with the high-mass cross sections (Figs. \[fig:xsec:Xgam\]b and \[fig:xsec:Xgam\]d). In the low-mass low-$x_\gamma^{\rm obs}$ regions that are poorly described by the nominal MC models, the introduction of MPIs to the simulations aids the description of the data. More specifically, the [Herwig]{} prediction with tuned MPIs describes the data reasonably well in all of the samples. The default [Pythia]{} model tends to overestimate the data in the mid-$x_\gamma^{\rm obs}$ region ($0.4\lesssim x_\gamma^{\rm obs}\lesssim 0.8$) in the low-mass samples but describes the cross section reasonably well elsewhere.
Shown also in Fig. \[fig:xsec:Xgam\] is the contribution to the [Herwig]{} predictions attributed to direct processes, as defined in that LO model with parton showers. In all four samples, direct events are predicted to be found solely at high $x_\gamma^{\rm obs}$, although resolved processes are expected to make a significant, and in the four-jet case dominant, contribution even at high $x_\gamma^{\rm obs}$. This is in contrast to the behaviour of dijet photoproduction, where the high-$x_\gamma^{\rm obs}$ region is dominated by direct events [@Art:zeusxg].
The ${\rm d}\sigma/{\rm d}y$ cross section
------------------------------------------
The three- and four-jet cross sections are given as a function of $y$ in the low- and high-mass samples in Fig. \[fig:xsec:yJB\] and Tables \[tab:Yjb\_3j\_\] and \[tab:Yjb\_4j\_\]. The cross sections are observed to increase steeply with $y$ in all but the low-mass three-jet sample where it is approximately constant. The shapes of the distributions are governed by the available kinematic phase space and the energy distribution of the photon flux: while the phase space increases with $y$, the photon flux decreases with increasing $y$.
Also shown in Fig. \[fig:xsec:yJB\] are the predictions from [Herwig]{} and [Pythia]{}, with and without MPIs, which have been area normalised to the data. At high mass, the predictions with and without MPIs are similar and describe the data well. In the low-mass samples, MPIs are predicted to cause a more steeply increasing cross section. However, the data are not precise enough to differentiate between the various models and all of the MC predictions roughly describe the data.
The ${\rm d}\sigma/{\rm d}E_T^{\rm jet}$ and ${\rm d}\sigma/{\rm d}\eta^{\rm jet}$ cross sections
-------------------------------------------------------------------------------------------------
The cross sections are given as a function of the $E_T^{\rm jet}$ of each of the jets in both three-jet samples in Fig. \[fig:xsec:Et\_3j\] and Tables \[tab:Et3\_3j\_\] to \[tab:Et5\_3j\_\] and in both four-jet samples in Fig. \[fig:xsec:Et\_4j\] and Tables \[tab:Et3\_4j\_\] to \[tab:Et6\_4j\_\]. Similarly, the ${\rm d}\sigma/{\rm d}\eta^{\rm jet}$ cross sections are given in Figs. \[fig:xsec:eta\_3j\] and \[fig:xsec:eta\_4j\] and Tables \[tab:Eta3\_3j\_\] to \[tab:Eta6\_4j\_\]. In each case the jets have been ordered in descending $E_T^{\rm jet}$.
The $E_T^{\rm jet}$ distributions of all of the jets in all of the samples are observed to fall off approximately exponentially with increasing $E_T^{\rm jet}$. All of the samples are similarly distributed when binned in terms of $\eta^{\rm jet}$, regardless of the position of the jet in the $E_T^{\rm jet}$ ordering. The generic trend is an increasing cross section from $\eta^{\rm jet}\approx-1.4$ followed by a plateau and, in the high-mass samples, a continued rise above $\eta^{\rm jet}\approx1.4$. In the low-mass samples, the plateaux begin at $\eta^{\rm jet}\approx0.6$ and extend to the upper edge of the measured $\eta$ region. In the high-mass samples, the lower bounds of the $\eta^{\rm jet}$ ranges covered by the plateaux decrease with decreasing $E_T^{\rm jet}$, while the upper bounds all lie at $\eta^{\rm jet}\approx1.4$.
Shown also in Figs. \[fig:xsec:Et\_3j\] and \[fig:xsec:Et\_4j\] are ${\rm d}\sigma/{\rm d}E_T^{\rm jet}$ predictions from the [Pythia]{} and [Herwig]{} models, with and without MPIs, normalised to the data. The description of the data by each of the MC models is generally good for all of the jets in all of the samples, although the MC predictions do vary somewhat. In all four samples, the [Pythia]{} cross sections without MPIs give the poorest description, in particular of ${\rm d}\sigma/{\rm d}E_T^{\rm jet2}$. Overall, the description of the data is improved by the inclusion of MPIs.
Figures \[fig:xsec:eta\_3j\] and \[fig:xsec:eta\_4j\] show the comparison between ${\rm d}\sigma/{\rm d}\eta^{\rm jet}$ in the data and as predicted by the MC models. Each MC model largely describes both sets of high-mass data. The best description of the low-mass data is given by the [Herwig]{} model with tuned MPIs. The [Pythia]{} prediction with default MPIs also generally describes the data. The poorest description is of the three-jet $\eta^{\rm jet1}$ distribution. The introduction of MPIs into the simulations certainly aids the description. In general, the MC models without MPIs predict a cross section that falls off at high $\eta^{\rm jet}$ but MPIs reduce this effect in line with what is observed.
The ${\rm d}\sigma/{\rm d}\cos(\psi_3)$ cross section
-----------------------------------------------------
The three-jet cross section is given as a function of $\cos(\psi_3)$ for the low- and high-mass samples in Fig. \[fig:xsec:cp3\] and Table \[tab:cos\_psi3\_3j\_\]. The $\cos(\psi_3)$ cross section has a similar shape in both the low- and high-mass samples; relatively flat in the central $\cos(\psi_3)$ region and increasing rapidly as $|\cos(\psi_3)|\rightarrow 1$. Also shown in Fig. \[fig:xsec:cp3\] is a comparison of the MC models with the data. All of the area normalised MC predictions describe the data well.
The ${\rm d}\sigma/{\rm d}M_{3j}$ cross section compared with $\mathcal{O}(\alpha\alpha_s^2)$ pQCD
--------------------------------------------------------------------------------------------------
Figure \[fig:tree:Mnj\] shows an $\mathcal{O}(\alpha\alpha_s^2)$ prediction, corrected for hadronisation effects and MPIs, compared to the measured ${\rm d}\sigma/{\rm d}M_{3j}$ cross section. The hadronisation and MPI corrections, including their estimated uncertainties, are given in Fig. \[fig:tree:Mnj\]b and in Table \[tab:Mnj\_3j\_\]. The hadronisation corrections are constant in $M_{3j}$, while the MPI corrections increase significantly towards low $M_{3j}$. The theoretical uncertainties on both the MPI corrections and the pQCD predictions are large. The magnitude and shape of the calculation is consistent with the data within the large theoretical uncertainties. This is best seen in the data over theory ratio shown in Fig. \[fig:tree:Mnj\]c. The level of consistency between data and theory would be far worse at low $M_{3j}$ if it were not for the large MPI corrections.
The ${\rm d}\sigma/{\rm d}E_T^{\rm jet}$ cross section compared with $\mathcal{O}(\alpha\alpha_s^2)$ pQCD
---------------------------------------------------------------------------------------------------------
Figure \[fig:tree:Et\_3j\] shows the comparison between the measured and predicted ${\rm d}\sigma/{\rm d}E_T^{\rm jet}$ cross sections at $\mathcal{O}(\alpha\alpha_s^2)$ for each jet in the low- and high-mass three-jet samples. The pQCD predictions have been corrected for hadronisation effects and MPIs, which are detailed in Tables \[tab:Et3\_3j\_\] to \[tab:Et5\_3j\_\] (not shown in Fig. \[fig:tree:Et\_3j\]). Hadronisation is predicted to decrease the parton-level cross sections at low $E_T^{\rm jet}$, whereas the MPIs are expected to do the reverse in the low-mass region: at high mass they are small.
The calculation is largely consistent, within the large uncertainties, with both the low- and high-mass data. The exception to this is the rapid predicted falloff of the $E_T^{\rm jet1}$ cross section at high $E_T^{\rm jet1}$ in the low-mass sample, which is not observed in the data. This effect is due to the three-parton kinematics within the theory, which precludes the highest momentum parton from carrying more than half of the total centre-of-mass energy. Thus, for the low-mass sample, the $M_{nj}<50$ GeV criterium translates into an $E_T<25$ GeV constraint for the partons in the theory, as indeed observed.
The ${\rm d}\sigma/{\rm d}\eta^{\rm jet}$ cross section compared with $\mathcal{O}(\alpha\alpha_s^2)$ pQCD
----------------------------------------------------------------------------------------------------------
Figure \[fig:tree:eta\_3j\] shows the comparison between the measured and predicted ${\rm d}\sigma/{\rm d}\eta^{\rm jet}$ cross sections at $\mathcal{O}(\alpha\alpha_s^2)$ for each jet in the low- and high-mass three-jet samples. The hadronisation and MPI correction factors are given in Tables \[tab:Eta3\_3j\_\] to \[tab:Eta5\_3j\_\] (not shown in Fig. \[fig:tree:eta\_3j\]). Hadronisation is predicted to cause an overall decrease in the parton-level cross sections, most significantly at low $\eta^{\rm jet}$. The MPIs in the low-mass sample are predicted to cause a small increase in the cross section at low $\eta^{\rm jet}$, becoming larger as $\eta^{\rm jet}$ increases. The calculation is consistent with the $\eta^{\rm jet}$ distributions. There is some indication of a difference in shape at lower masses, however, the uncertainties are again large. The description of the low-mass data would be worse if it were not for the MPI corrections.
The ${\rm d}\sigma/{\rm d}\cos(\psi_3)$ cross section compared with $\mathcal{O}(\alpha\alpha_s^2)$ pQCD
--------------------------------------------------------------------------------------------------------
The comparison between the measured and the predicted ${\rm d}\sigma/{\rm d}\cos(\psi_3)$ cross sections at $\mathcal{O}(\alpha\alpha_s^2)$ is shown for the low-mass sample in Fig. \[fig:tree:cp3\_i\] and for the high-mass region in Fig. \[fig:tree:cp3\_h\]. The ratio of the low-mass data divided by the calculation is shown along with the hadronisation and MPI corrections, which are also given in Table \[tab:cos\_psi3\_3j\_\]. In both low- and high-mass regions, the QCD calculation is consistent with the data. There is some indication of a difference in shape at low masses; however, the theory has large uncertainties.
Summary
=======
Three- and four-jet states have been measured in hard $\gamma p$ collisions at HERA, using an integrated luminosity of $121~{\rm pb^{\rm -1}}$. The three- and four-jet cross sections have been measured for jets with $E_T^{\rm jet}>6$ GeV and $|\eta^{\rm jet}|<2.4$, in the kinematic region given by $Q^2<1~\rm{GeV^2}$ and $0.2\le y\le0.85$ and in two mass regions defined as $25\le M_{nj}<50$ GeV and $M_{nj}\ge50$ GeV. The three-jet events have been measured with over seven times the luminosity of the previous ZEUS publication and in a wider $M_{3j}$ region. The four-jet process described here has been measured for the first time at HERA.
In the high-mass regions, the shape of the three- and four-jet cross sections are reasonably well described by both the [Pythia]{} and [Herwig]{} models without MPIs. In the low-mass region, the MC models without MPIs underestimate the data when normalised to the measured high-mass cross section. When MPIs are added to the MC simulations, the agreement between the models and data is generally improved. Although the data have large uncertainties, the measured cross sections are potentially useful in the testing and tuning of MPI and underlying-event models.
The $\mathcal{O}(\alpha\alpha_s^2)$ pQCD calculation was compared to the three-jet data and is consistent within the large uncertainties. The MPI corrections typically improved the description of the data by the pQCD calculation. This data will provide a testing ground for higher-order calculations in photoproduction.
Acknowledgments
===============
The support and encouragement of the DESY Directorate has been invaluable and we are indebted to the HERA machine group for their diligent efforts. The design, construction and installation of the ZEUS detector involved many people both in and outside of DESY that are not listed as authors here but their contributions are acknowledged with great appreciation nonetheless.
‘=11 bibitem\#1[@bstsupport @iftail[\#1]{} [;]{} [@first.@bibitem[\#1]{}]{} @firstfalse @iftail[\#1]{} [@bibitem[\#1]{}]{}]{}‘=12
[10]{}
C.H. Llewellyn Smith, Phys. Lett. [**B79**]{}, 83 (1978) I. Kang and C.H. Llewellyn Smith, Nucl. Phys. [**B166**]{}, 413 (1980) J.F. Owens, Phys. Rev. [**D21**]{}, 54 (1980) M. Fontannaz, A. Mantrach and D. Schiff, Z. Phys. [**C6**]{}, 241 (1980) W.J. Stirling and Z. Kunszt, , R.D. Peccei (ed.), Vol. 2, p. 331. DESY, Hamburg, Germany (1987) M. Drees and F. Halzen, Phys. Rev. Lett. [**61**]{}, 275 (1988) M. Drees and R.M. Godbole, Phys. Rev. Lett. [**61**]{}, 682 (1988) M. Drees and R.M. Godbole, Phys. Rev. [**D39**]{}, 169 (1989) H. Baer, J. Ohnemus and J.F. Owens, Z. Phys. [**C42**]{}, 657 (1989) H. Baer, J. Ohnemus and J.F. Owens, Phys. Rev. [**D40**]{}, 2844 (1989) CDF Collab., F. Abe et al., Phys. Rev. [**D47**]{}, 4857 (1993) CDF Collab., F. Abe et al., Phys. Rev. Lett. [**79**]{}, 584 (1997) CDF Collab., F. Abe et al., Phys. Rev. [**D56**]{}, 3811 (1997) D0 Collab., V. M. Abazov et al., Phys. Rev. [**D67**]{}, 052001 (2003) ZEUS Coll., M. Derrick et al., Phys. Lett. [**B348**]{}, 665 (1995) H1 Collab., S. Aid et al., Z. Phys. [**C70**]{}, 17 (1996) A. Del Fabbro and D. Treleani, hep-ph/0301178 (unpublished) D. Acosta et al., CERN-CMS-NOTE-2006-067 (unpublished) ZEUS Coll., J. Breitweg et al., Phys. Lett. [**B443**]{}, 394 (1998) H1 Coll., I. Abt et al., Z. Phys. [**C61**]{}, 59 (1994) H1 Coll., C. Adloff et al., Phys. Lett. [**B515**]{}, 17 (2001) ZEUS Coll., S. Chekanov et al., Eur. Phys. J [**C44**]{}, 183 (2005) ZEUS Coll., S. Chekanov et al., Phys. Lett. [**B516**]{}, 273 (2002) ZEUS [Coll.]{}, M. Derrick [et al.]{}, Phys. Lett. [**B 297**]{}, 404 (1992) ZEUS [Coll.]{}, U. Holm (ed.), . Status Report (unpublished), DESY (1993), available on `http://www-zeus.desy.de/bluebook/bluebook.html` M. Derrick [et al.]{}, Nucl. Inst. Meth. [**A 309**]{}, 77 (1991) A. Andresen [et al.]{}, Nucl. Inst. Meth. [**A 309**]{}, 101 (1991) A. Bernstein [et al.]{}, Nucl. Inst. Meth. [**A 336**]{}, 23 (1993) A. Caldwell [et al.]{}, Nucl. Inst. Meth. [**A 321**]{}, 356 (1992) N. Harnew [et al.]{}, Nucl. Inst. Meth. [**A 279**]{}, 290 (1989) B. Foster [et al.]{}, Nucl. Phys. Proc. Suppl. [**B 32**]{}, 181 (1993) B. Foster [et al.]{}, Nucl. Inst. Meth. [**A 338**]{}, 254 (1994) J. Andruszków et al., Preprint DESY-92-066, DESY, 1992 (unpublished) ZEUS [Coll.]{}, M. Derrick [et al.]{}, Z. Phys. [**C 63**]{}, 391 (1994) ZEUS [Coll.]{}, J. Breitweg [et al.]{}, Eur. Phys. J. [**C 1**]{}, 109 (1998) ZEUS Coll., J. Brietweg et al., Eur. Phys. J. [**C6**]{}, 43 (1999) G.M. Briskin. PhD. Thesis, Tel Aviv University, Israel, 1998 (unpublished) ZEUS Coll., J. Brietweg et al., Eur. Phys. J. [**C11**]{}, 35 (1999) ZEUS Coll. S. Chekanov et al., Eur. Phys. J. [**C23**]{}, 615 (2002) S. Catani et al., Nucl. Phys. [**B406**]{}, 187 (1993) S.D. Ellis and D.E. Soper, Phys. Rev. [**D48**]{}, 3160 (1993) F. Jacquet and A. Blondel, , U. Amaldi (ed.), p. 391. Hamburg, Germany (1979). Also in preprint S. Geer and T. Asakawa, Phys. Rev. [**D53**]{}, 4793 (1996) G. Corcella et al., , 2000, available on `http://hepwww.rl.ac.uk/theory/seymour/herwig/` G. Corcella et al., , 2001, available on `http://hepwww.rl.ac.uk/theory/seymour/herwig/` G. Marchesini et al., Comput. Phys. Commun. [**67**]{}, 465 (1992) T. Sjöstrand et al., , 2002, available on `http://www.thep.lu.se/\simtorbjorn/Pythia.html` T. Sjöstrand et al., Comput. Phys. Commun. [**135**]{}, 238 (2001) G. Marchesini and B.R. Webber, Nucl. Phys. [**B310**]{}, 461 (1988) I.G. Knowles, Nucl. Phys. [**B310**]{}, 571 (1988) I.G. Knowles, Comput. Phys. Commun. [**58**]{}, 271 (1990) B.R. Webber, Nucl. Phys [**B238**]{}, 492 (1984) B. Andersson et al., Phys. Rep. [**97**]{}, 31 (1983) T. Sjöstrand, Phys. Lett. [**B142**]{}, 420 (1984) J.M. Butterworth, J.R. Forshaw and M.H. Seymour, Z. Phys. [**C72**]{}, 637 (1996) J.M. Butterworth and J.R. Forshaw, J. Phys. [**G19**]{}, 1657 (1993) J.M. Butterworth et al., J. Phys. [**G22**]{}, 883 (1996) J.M. Butterworth and M.H. Seymour, (unpublished), 2005, available on `http://hepforge.cedar.ac.uk/jimmy/` R. Brun et al., , CERN-DD/EE/84-1, 1987 (unpublished) CTEQ Coll., H.L. Lai et al., Eur. Phys. J. [**C12**]{}, 375 (2000) M. Gluck, E. Reya and A. Vogt, Z. Phys. [**C53**]{}, 127 (1992) M. Klasen, T. Kleinwort and G. Kramer, Eur. Phys. J. direct [**C1**]{}, 1 (1998) M. Klasen, Eur. Phys. J. [**C7**]{}, 225 (1999) J. Pumplin et al., JHEP [**0602**]{}, 032 (2006) M. Gluck, E. Reya and A. Vogt, Phys. Rev. [**D45**]{}, 3986 (1992) T. Namsoo. PhD. Thesis, University of Bristol, UK, 2006 (unpublished)
[|c|c|c|c|]{}\
3-jet no MPIs & 3-jet with MPIs & 4-jet no MPIs & 4-jet with MPIs\
1.7 & 1.7 & 3.1 & 2.1\
\
3-jet no MPIs & 3-jet with MPIs & 4-jet no MPIs & 4-jet with MPIs\
3.8 & 3.1 & 9.2 & 5.3\
----- ------------------- ------------ ------------ -------------- ---------------------- ---------------------- ------- ------------ ------ -----------
$d\sigma/dM_{3j}$ $\pm$stat. $\pm$syst. $\pm$E-scale
25 - 32 94 $\pm$ 0 $_{-9}^{+10}$ $_{-12}^{+12}$ 0.793 $\pm$0.004 1.88 $\pm$0.37
32 - 41 80.4 $\pm$ 0.3 $_{-8.2}^{+8.5}$ $_{-9.5}^{+9.5}$ 0.805 $\pm$0.010 1.44 $\pm$0.23
41 - 50 48.4 $\pm$ 0.3 $_{-5.4}^{+5.5}$ $_{-5.6}^{+5.6}$ 0.815 $\pm$0.015 1.25 $\pm$0.16
50 - 59 25.2 $\pm$ 0.2 $_{-3.1}^{+3.1}$ $_{-2.5}^{+2.5}$ 0.815 $\pm$0.017 1.15 $\pm$0.13
59 - 70 10.7 $\pm$ 0.1 $_{-1.4}^{+1.9}$ $_{-1.2}^{+1.2}$ 0.817 $\pm$0.008 1.09 $\pm$0.10
70 - 82 4.00 $\pm$ 0.08 $_{-0.52}^{+0.56}$ $_{-0.50}^{+0.50}$ 0.810 $\pm$0.009 1.07 $\pm$0.08
82 - 95 1.44 $\pm$ 0.04 $_{-0.27}^{+0.24}$ $_{-0.20}^{+0.20}$ 0.808 $\pm$0.007 1.04 $\pm$0.05
95 - 109 0.61 $\pm$ 0.02 $_{-0.21}^{+0.28}$ $_{-0.07}^{+0.07}$ 0.811 $\pm$0.014 1.03 $\pm$0.05
109 - 124 0.179 $\pm$ 0.012 $_{-0.066}^{+0.084}$ $_{-0.023}^{+0.023}$ 0.840 $\pm$0.019 1.03 $\pm$0.05
124 - 140 0.066 $\pm$ 0.008 $_{-0.019}^{+0.022}$ $_{-0.014}^{+0.014}$ 0.798 $\pm$0.008 1.00 $\pm$0.04
140 - 160 0.023 $\pm$ 0.005 $_{-0.010}^{+0.019}$ $_{-0.012}^{+0.012}$ 0.809 $\pm$0.012 1.00 $\pm$0.06
----- ------------------- ------------ ------------ -------------- ---------------------- ---------------------- ------- ------------ ------ -----------
: The measured three-jet differential cross-section $d\sigma/dM_{3j}$. The statistical, systematic and calorimeter energy scale (E-scale) uncertainties are shown separately. Also shown are the hadronisation and MPI corrections that were applied to the $\mathcal{O}(\alpha\alpha_s^2)$ prediction. \[tab:Mnj\_3j\_\]
----- ------------------- ------------ ------------ -------------- ------------------------ ------------------------
$d\sigma/dM_{4j}$ $\pm$stat. $\pm$syst. $\pm$E-scale
25 - 32 1.16 $\pm$ 0.04 $_{-0.20}^{+0.29}$ $_{-0.28}^{+0.28}$
32 - 41 5.8 $\pm$ 0.1 $_{-0.7}^{+0.6}$ $_{-1.1}^{+1.1}$
41 - 50 6.3 $\pm$ 0.1 $_{-0.7}^{+0.7}$ $_{-1.1}^{+1.1}$
50 - 59 4.49 $\pm$ 0.09 $_{-0.48}^{+0.65}$ $_{-0.73}^{+0.73}$
59 - 70 2.85 $\pm$ 0.07 $_{-0.59}^{+0.43}$ $_{-0.43}^{+0.43}$
70 - 82 1.25 $\pm$ 0.05 $_{-0.20}^{+0.27}$ $_{-0.22}^{+0.22}$
82 - 95 0.488 $\pm$ 0.027 $_{-0.063}^{+0.064}$ $_{-0.087}^{+0.087}$
95 - 109 0.237 $\pm$ 0.018 $_{-0.068}^{+0.052}$ $_{-0.042}^{+0.042}$
109 - 124 0.060 $\pm$ 0.009 $_{-0.013}^{+0.019}$ $_{-0.010}^{+0.010}$
124 - 140 0.027 $\pm$ 0.006 $_{-0.015}^{+0.019}$ $_{-0.006}^{+0.006}$
140 - 160 0.0024 $\pm$ 0.0016 $_{-0.0019}^{+0.0053}$ $_{-0.0012}^{+0.0012}$
----- ------------------- ------------ ------------ -------------- ------------------------ ------------------------
: The measured four-jet differential cross-section $d\sigma/dM_{4j}$. Other details as described in the caption to Table \[tab:Mnj\_3j\_\]. \[tab:Mnj\_4j\_\]
[|l c l||l l l l|]{} & $d\sigma/dx_{\gamma}^{{\rm obs}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
0.00 & - & 0.11 & 109 & $\pm$ 7 & $_{-53}^{+66}$ & $_{-31}^{+31}$\
0.11 & - & 0.22 & 1449 & $\pm$ 15 & $_{-324}^{+334}$ & $_{-289}^{+289}$\
0.22 & - & 0.33 & 2102 & $\pm$ 15 & $_{-230}^{+240}$ & $_{-361}^{+361}$\
0.33 & - & 0.47 & 1989 & $\pm$ 13 & $_{-119}^{+128}$ & $_{-292}^{+292}$\
0.47 & - & 0.62 & 1774 & $\pm$ 12 & $_{-97}^{+117}$ & $_{-234}^{+234}$\
0.62 & - & 0.75 & 1752 & $\pm$ 13 & $_{-99}^{+115}$ & $_{-199}^{+199}$\
0.75 & - & 0.85 & 2579 & $\pm$ 19 & $_{-184}^{+204}$ & $_{-253}^{+253}$\
0.850 & - & 0.935 & 3803 & $\pm$ 26 & $_{-600}^{+672}$ & $_{-227}^{+227}$\
0.935 & - & 1.000 & 730 & $\pm$ 12 & $_{-194}^{+203}$ & $_{-22}^{+22}$\
\
0.11 & - & 0.22 & 6.9 & $\pm$ 1.0 & $_{-5.0}^{+8.0}$ & $_{-2.0}^{+2.0}$\
0.22 & - & 0.33 & 51 & $\pm$ 3 & $_{-13}^{+10}$ & $_{-10}^{+10}$\
0.33 & - & 0.47 & 153 & $\pm$ 4 & $_{-42}^{+42}$ & $_{-21}^{+21}$\
0.47 & - & 0.62 & 310 & $\pm$ 6 & $_{-75}^{+87}$ & $_{-40}^{+40}$\
0.62 & - & 0.75 & 504 & $\pm$ 8 & $_{-86}^{+93}$ & $_{-60}^{+60}$\
0.75 & - & 0.85 & 871 & $\pm$ 12 & $_{-113}^{+101}$ & $_{-100}^{+100}$\
0.850 & - & 0.935 & 1695 & $\pm$ 18 & $_{-211}^{+202}$ & $_{-161}^{+161}$\
0.935 & - & 1.000 & 666 & $\pm$ 14 & $_{-164}^{+169}$ & $_{-52}^{+52}$\
[|l c l||l l l l|]{} & $d\sigma/dx_{\gamma}^{{\rm obs}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
0.00 & - & 0.11 & 4.9 & $\pm$ 1.7 & $_{-3.0}^{+5.1}$ & $_{-1.4}^{+1.4}$\
0.11 & - & 0.22 & 100 & $\pm$ 4 & $_{-37}^{+44}$ & $_{-21}^{+21}$\
0.22 & - & 0.33 & 210 & $\pm$ 5 & $_{-45}^{+46}$ & $_{-53}^{+53}$\
0.33 & - & 0.47 & 177 & $\pm$ 4 & $_{-18}^{+31}$ & $_{-38}^{+38}$\
0.47 & - & 0.62 & 136 & $\pm$ 3 & $_{-13}^{+21}$ & $_{-29}^{+29}$\
0.62 & - & 0.75 & 94 & $\pm$ 3 & $_{-17}^{+27}$ & $_{-14}^{+14}$\
0.75 & - & 0.85 & 113 & $\pm$ 3 & $_{-35}^{+42}$ & $_{-15}^{+15}$\
0.850 & - & 0.935 & 125 & $\pm$ 4 & $_{-33}^{+28}$ & $_{-9}^{+9}$\
0.935 & - & 1.000 & 22 & $\pm$ 2 & $_{-13}^{+6}$ & $_{-3}^{+3}$\
\
0.11 & - & 0.22 & 0.4 & $\pm$ 0.2 & $_{-0.2}^{+2.3}$ & $_{-0.2}^{+0.2}$\
0.22 & - & 0.33 & 25 & $\pm$ 2 & $_{-12}^{+14}$ & $_{-9}^{+9}$\
0.33 & - & 0.47 & 64 & $\pm$ 3 & $_{-23}^{+26}$ & $_{-11}^{+11}$\
0.47 & - & 0.62 & 103 & $\pm$ 3 & $_{-30}^{+26}$ & $_{-19}^{+19}$\
0.62 & - & 0.75 & 126 & $\pm$ 4 & $_{-24}^{+18}$ & $_{-21}^{+21}$\
0.75 & - & 0.85 & 170 & $\pm$ 5 & $_{-27}^{+31}$ & $_{-24}^{+24}$\
0.850 & - & 0.935 & 307 & $\pm$ 8 & $_{-40}^{+39}$ & $_{-35}^{+35}$\
0.935 & - & 1.000 & 122 & $\pm$ 5 & $_{-37}^{+69}$ & $_{-7}^{+7}$\
[|l c l||l l l l|]{} & $d\sigma/dy$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
0.200 & - & 0.272 & 2264 & $\pm$ 22 & $_{-392}^{+400}$ & $_{-152}^{+152}$\
0.272 & - & 0.344 & 2794 & $\pm$ 22 & $_{-382}^{+491}$ & $_{-286}^{+286}$\
0.344 & - & 0.417 & 2854 & $\pm$ 22 & $_{-282}^{+390}$ & $_{-321}^{+321}$\
0.417 & - & 0.489 & 2900 & $\pm$ 22 & $_{-238}^{+288}$ & $_{-340}^{+340}$\
0.489 & - & 0.561 & 2973 & $\pm$ 22 & $_{-261}^{+270}$ & $_{-361}^{+361}$\
0.561 & - & 0.633 & 2918 & $\pm$ 22 & $_{-242}^{+268}$ & $_{-363}^{+363}$\
0.633 & - & 0.706 & 2833 & $\pm$ 22 & $_{-218}^{+232}$ & $_{-366}^{+366}$\
0.706 & - & 0.778 & 2739 & $\pm$ 22 & $_{-216}^{+250}$ & $_{-380}^{+380}$\
0.778 & - & 0.850 & 2797 & $\pm$ 23 & $_{-326}^{+340}$ & $_{-415}^{+415}$\
\
0.200 & - & 0.272 & 156 & $\pm$ 7 & $_{-38}^{+24}$ & $_{-16}^{+16}$\
0.272 & - & 0.344 & 271 & $\pm$ 7 & $_{-13}^{+28}$ & $_{-24}^{+24}$\
0.344 & - & 0.417 & 455 & $\pm$ 10 & $_{-65}^{+78}$ & $_{-43}^{+43}$\
0.417 & - & 0.489 & 565 & $\pm$ 11 & $_{-91}^{+59}$ & $_{-52}^{+52}$\
0.489 & - & 0.561 & 684 & $\pm$ 12 & $_{-108}^{+134}$ & $_{-67}^{+67}$\
0.561 & - & 0.633 & 768 & $\pm$ 13 & $_{-131}^{+111}$ & $_{-81}^{+81}$\
0.633 & - & 0.706 & 905 & $\pm$ 14 & $_{-199}^{+171}$ & $_{-100}^{+100}$\
0.706 & - & 0.778 & 961 & $\pm$ 15 & $_{-206}^{+201}$ & $_{-113}^{+113}$\
0.778 & - & 0.850 & 1058 & $\pm$ 16 & $_{-224}^{+259}$ & $_{-130}^{+130}$\
[|l c l||l l l l|]{} & $d\sigma/dy$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
0.200 & - & 0.272 & 90 & $\pm$ 4 & $_{-25}^{+30}$ & $_{-8}^{+8}$\
0.272 & - & 0.344 & 113 & $\pm$ 4 & $_{-14}^{+23}$ & $_{-18}^{+18}$\
0.344 & - & 0.417 & 154 & $\pm$ 5 & $_{-24}^{+28}$ & $_{-29}^{+29}$\
0.417 & - & 0.489 & 171 & $\pm$ 5 & $_{-17}^{+30}$ & $_{-33}^{+33}$\
0.489 & - & 0.561 & 186 & $\pm$ 5 & $_{-23}^{+15}$ & $_{-37}^{+37}$\
0.561 & - & 0.633 & 210 & $\pm$ 5 & $_{-27}^{+30}$ & $_{-39}^{+39}$\
0.633 & - & 0.706 & 215 & $\pm$ 5 & $_{-28}^{+31}$ & $_{-42}^{+42}$\
0.706 & - & 0.778 & 218 & $\pm$ 6 & $_{-24}^{+23}$ & $_{-44}^{+44}$\
0.778 & - & 0.850 & 246 & $\pm$ 7 & $_{-43}^{+37}$ & $_{-52}^{+52}$\
\
0.200 & - & 0.272 & 16.8 & $\pm$ 2.4 & $_{-6.8}^{+5.3}$ & $_{-3.1}^{+3.1}$\
0.272 & - & 0.344 & 40 & $\pm$ 3 & $_{-13}^{+13}$ & $_{-6}^{+6}$\
0.344 & - & 0.417 & 81 & $\pm$ 4 & $_{-7}^{+13}$ & $_{-9}^{+9}$\
0.417 & - & 0.489 & 108 & $\pm$ 5 & $_{-9}^{+8}$ & $_{-15}^{+15}$\
0.489 & - & 0.561 & 136 & $\pm$ 5 & $_{-16}^{+13}$ & $_{-17}^{+17}$\
0.561 & - & 0.633 & 171 & $\pm$ 6 & $_{-17}^{+28}$ & $_{-24}^{+24}$\
0.633 & - & 0.706 & 228 & $\pm$ 7 & $_{-47}^{+33}$ & $_{-33}^{+33}$\
0.706 & - & 0.778 & 266 & $\pm$ 8 & $_{-66}^{+54}$ & $_{-40}^{+40}$\
0.778 & - & 0.850 & 266 & $\pm$ 8 & $_{-51}^{+60}$ & $_{-47}^{+47}$\
[|l c l||l l l l||l l l l|]{} & $d\sigma/dE_{T}^{{\rm jet1}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale & &\
& & & & &\
\
6.0 & - & 8.5 & 107 & $\pm$ 1 & $_{-17}^{+19}$ & $_{-11}^{+11}$ &0.744& $\pm$0.026&2.09& $\pm$0.70\
8.5 & - & 11.2 & 248 & $\pm$ 1 & $_{-20}^{+20}$ & $_{-29}^{+29}$ &0.791& $\pm$0.012&1.67& $\pm$0.34\
11.2 & - & 14.2 & 169 & $\pm$ 1 & $_{-10}^{+11}$ & $_{-21}^{+21}$ &0.819& $\pm$0.009&1.38& $\pm$0.17\
14.2 & - & 17.6 & 75.2 & $\pm$ 0.6 & $_{-7.2}^{+6.9}$ & $_{-9.7}^{+9.7}$ &0.842& $\pm$0.003&1.24& $\pm$0.12\
17.6 & - & 21.5 & 19.9 & $\pm$ 0.3 & $_{-2.3}^{+2.6}$ & $_{-2.6}^{+2.6}$ &0.834& $\pm$0.017&1.17& $\pm$0.09\
21.5 & - & 26.0 & 2.54 & $\pm$ 0.09 & $_{-0.32}^{+0.38}$ & $_{-0.37}^{+0.37}$ &0.774& $\pm$0.064&1.16& $\pm$0.11\
26.0 & - & 31.2 & 0.200 & $\pm$ 0.022 & $_{-0.077}^{+0.071}$ & $_{-0.040}^{+0.040}$ &0.725& $\pm$0.080&1.21& $\pm$0.20\
\
6.0 & - & 8.5 & 2.7 & $\pm$ 0.2 & $_{-1.0}^{+1.3}$ & $_{-0.6}^{+0.6}$ &0.515& $\pm$0.052&1.42& $\pm$0.29\
8.5 & - & 11.2 & 14.4 & $\pm$ 0.4 & $_{-4.7}^{+5.8}$ & $_{-1.5}^{+1.5}$ &0.646& $\pm$0.044&1.28& $\pm$0.20\
11.2 & - & 14.2 & 24.5 & $\pm$ 0.4 & $_{-5.6}^{+5.5}$ & $_{-2.5}^{+2.5}$ &0.759& $\pm$0.024&1.17& $\pm$0.16\
14.2 & - & 17.6 & 27.8 & $\pm$ 0.4 & $_{-4.0}^{+4.7}$ & $_{-2.7}^{+2.7}$ &0.839& $\pm$0.018&1.13& $\pm$0.13\
17.6 & - & 21.5 & 23.1 & $\pm$ 0.3 & $_{-2.1}^{+2.2}$ & $_{-2.2}^{+2.2}$ &0.878& $\pm$0.007&1.09& $\pm$0.09\
21.5 & - & 26.0 & 14.1 & $\pm$ 0.2 & $_{-1.3}^{+1.5}$ & $_{-1.5}^{+1.5}$ &0.864& $\pm$0.016&1.07& $\pm$0.08\
26.0 & - & 31.2 & 6.07 & $\pm$ 0.12 & $_{-0.58}^{+0.59}$ & $_{-0.72}^{+0.72}$ &0.859& $\pm$0.025&1.06& $\pm$0.07\
31.2 & - & 37.0 & 2.06 & $\pm$ 0.07 & $_{-0.18}^{+0.19}$ & $_{-0.32}^{+0.32}$ &0.848& $\pm$0.034&1.04& $\pm$0.05\
37.0 & - & 43.9 & 0.79 & $\pm$ 0.04 & $_{-0.16}^{+0.63}$ & $_{-0.22}^{+0.22}$ &0.838& $\pm$0.032&1.02& $\pm$0.04\
43.9 & - & 51.8 & 0.246 & $\pm$ 0.020 & $_{-0.046}^{+0.044}$ & $_{-0.066}^{+0.066}$ &0.847& $\pm$0.045&1.04& $\pm$0.07\
51.8 & - & 60.8 & 0.067 & $\pm$ 0.007 & $_{-0.031}^{+0.021}$ & $_{-0.013}^{+0.013}$ &0.881& $\pm$0.039&1.00& $\pm$0.04\
60.8 & - & 71.0 & 0.035 & $\pm$ 0.004 & $_{-0.019}^{+0.019}$ & $_{-0.001}^{+0.001}$ &0.870& $\pm$0.001&1.02& $\pm$0.07\
[|l c l||l l l l||l l l l|]{} & $d\sigma/dE_{T}^{{\rm jet2}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale & &\
& & & & &\
\
6.0 & - & 8.5 & 383 & $\pm$ 1 & $_{-47}^{+48}$ & $_{-43}^{+43}$ &0.772& $\pm$0.023&1.68& $\pm$0.39\
8.5 & - & 11.2 & 229 & $\pm$ 1 & $_{-9}^{+10}$ & $_{-28}^{+28}$ &0.829& $\pm$0.010&1.41& $\pm$0.18\
11.2 & - & 14.2 & 59.0 & $\pm$ 0.5 & $_{-2.7}^{+3.7}$ & $_{-7.4}^{+7.4}$ &0.868& $\pm$0.018&1.25& $\pm$0.10\
14.2 & - & 17.6 & 10.4 & $\pm$ 0.2 & $_{-1.0}^{+1.3}$ & $_{-1.3}^{+1.3}$ &0.869& $\pm$0.005&1.19& $\pm$0.06\
17.6 & - & 21.5 & 0.84 & $\pm$ 0.05 & $_{-0.23}^{+0.42}$ & $_{-0.11}^{+0.11}$ &0.861& $\pm$0.042&1.16& $\pm$0.09\
\
6.0 & - & 8.5 & 22.2 & $\pm$ 0.5 & $_{-8.0}^{+9.0}$ & $_{-2.6}^{+2.6}$ &0.626& $\pm$0.027&1.26& $\pm$0.22\
8.5 & - & 11.2 & 40 & $\pm$ 1 & $_{-9}^{+11}$ & $_{-4}^{+4}$ &0.759& $\pm$0.019&1.16& $\pm$0.15\
11.2 & - & 14.2 & 34.3 & $\pm$ 0.4 & $_{-4.0}^{+4.1}$ & $_{-3.6}^{+3.6}$ &0.861& $\pm$0.005&1.11& $\pm$0.12\
14.2 & - & 17.6 & 21.6 & $\pm$ 0.3 & $_{-1.4}^{+1.5}$ & $_{-2.3}^{+2.3}$ &0.891& $\pm$0.008&1.08& $\pm$0.09\
17.6 & - & 21.5 & 11.5 & $\pm$ 0.2 & $_{-0.6}^{+0.7}$ & $_{-1.3}^{+1.3}$ &0.898& $\pm$0.011&1.07& $\pm$0.07\
21.5 & - & 26.0 & 4.48 & $\pm$ 0.12 & $_{-0.54}^{+0.49}$ & $_{-0.47}^{+0.47}$ &0.888& $\pm$0.015&1.06& $\pm$0.07\
26.0 & - & 31.2 & 1.57 & $\pm$ 0.07 & $_{-0.12}^{+0.18}$ & $_{-0.16}^{+0.16}$ &0.881& $\pm$0.010&1.06& $\pm$0.07\
31.2 & - & 37.0 & 0.558 & $\pm$ 0.035 & $_{-0.091}^{+0.065}$ & $_{-0.077}^{+0.077}$ &0.892& $\pm$0.028&1.03& $\pm$0.05\
37.0 & - & 43.9 & 0.214 & $\pm$ 0.021 & $_{-0.053}^{+0.046}$ & $_{-0.025}^{+0.025}$ &0.898& $\pm$0.003&1.01& $\pm$0.03\
43.9 & - & 51.8 & 0.071 & $\pm$ 0.013 & $_{-0.020}^{+0.026}$ & $_{-0.008}^{+0.008}$ &0.898& $\pm$0.028&1.00& $\pm$0.05\
51.8 & - & 60.8 & 0.042 & $\pm$ 0.009 & $_{-0.025}^{+0.022}$ & $_{-0.004}^{+0.004}$ &0.765& $\pm$0.092&1.02& $\pm$0.10\
[|l c l||l l l l||l l l l|]{} & $d\sigma/dE_{T}^{{\rm jet3}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale & &\
& & & & &\
\
6.0 & - & 8.5 & 644 & $\pm$ 2 & $_{-57}^{+68}$ & $_{-76}^{+76}$ &0.799& $\pm$0.011&1.53& $\pm$0.28\
8.5 & - & 11.2 & 63.2 & $\pm$ 0.6 & $_{-2.8}^{+3.5}$ & $_{-8.5}^{+8.5}$ &0.859& $\pm$0.011&1.30& $\pm$0.08\
11.2 & - & 14.2 & 3.82 & $\pm$ 0.14 & $_{-0.70}^{+0.82}$ & $_{-0.68}^{+0.68}$ &0.887& $\pm$0.017&1.24& $\pm$0.02\
14.2 & - & 17.6 & 0.096 & $\pm$ 0.034 & $_{-0.090}^{+0.053}$ & $_{-0.050}^{+0.050}$ &0.834& $\pm$0.058&1.09& $\pm$0.21\
\
6.0 & - & 8.5 & 94 & $\pm$ 1 & $_{-20}^{+20}$ & $_{-11}^{+11}$ &0.767& $\pm$0.010&1.15& $\pm$0.14\
8.5 & - & 11.2 & 43.1 & $\pm$ 0.5 & $_{-4.4}^{+5.1}$ & $_{-4.4}^{+4.4}$ &0.864& $\pm$0.010&1.08& $\pm$0.09\
11.2 & - & 14.2 & 15.3 & $\pm$ 0.3 & $_{-1.3}^{+1.5}$ & $_{-1.7}^{+1.7}$ &0.914& $\pm$0.015&1.06& $\pm$0.07\
14.2 & - & 17.6 & 5.0 & $\pm$ 0.1 & $_{-1.0}^{+1.0}$ & $_{-0.5}^{+0.5}$ &0.915& $\pm$0.020&1.04& $\pm$0.05\
17.6 & - & 21.5 & 1.08 & $\pm$ 0.06 & $_{-0.15}^{+0.15}$ & $_{-0.16}^{+0.16}$ &0.919& $\pm$0.016&1.04& $\pm$0.06\
21.5 & - & 26.0 & 0.289 & $\pm$ 0.026 & $_{-0.068}^{+0.086}$ & $_{-0.034}^{+0.034}$ &0.898& $\pm$0.032&1.00& $\pm$0.04\
26.0 & - & 31.2 & 0.065 & $\pm$ 0.012 & $_{-0.034}^{+0.030}$ & $_{-0.006}^{+0.006}$ &0.887& $\pm$0.145&1.00& $\pm$0.08\
31.2 & - & 40.0 & 0.0185 & $\pm$ 0.0053 & $_{-0.0078}^{+0.0085}$ & $_{-0.0037}^{+0.0037}$ &0.921& $\pm$0.054&1.05& $\pm$0.15\
[|l c l||l l l l|]{} & $d\sigma/dE_{T}^{{\rm jet1}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
6.0 & - & 8.5 & 9.6 & $\pm$ 0.2 & $_{-1.6}^{+1.4}$ & $_{-1.7}^{+1.7}$\
8.5 & - & 11.2 & 19.6 & $\pm$ 0.3 & $_{-2.0}^{+1.9}$ & $_{-3.7}^{+3.7}$\
11.2 & - & 14.2 & 9.3 & $\pm$ 0.2 & $_{-1.1}^{+1.4}$ & $_{-1.8}^{+1.8}$\
14.2 & - & 17.6 & 2.66 & $\pm$ 0.10 & $_{-0.52}^{+0.40}$ & $_{-0.50}^{+0.50}$\
17.6 & - & 21.5 & 0.29 & $\pm$ 0.03 & $_{-0.06}^{+0.12}$ & $_{-0.06}^{+0.06}$\
\
6.0 & - & 8.5 & 1.00 & $\pm$ 0.16 & $_{-0.55}^{+0.48}$ & $_{-0.13}^{+0.13}$\
8.5 & - & 11.2 & 6.1 & $\pm$ 0.2 & $_{-1.7}^{+1.8}$ & $_{-0.9}^{+0.9}$\
11.2 & - & 14.2 & 8.3 & $\pm$ 0.2 & $_{-1.4}^{+1.4}$ & $_{-1.1}^{+1.1}$\
14.2 & - & 17.6 & 6.46 & $\pm$ 0.16 & $_{-0.56}^{+0.71}$ & $_{-0.95}^{+0.95}$\
17.6 & - & 21.5 & 3.93 & $\pm$ 0.11 & $_{-0.29}^{+0.43}$ & $_{-0.62}^{+0.62}$\
21.5 & - & 26.0 & 1.93 & $\pm$ 0.07 & $_{-0.21}^{+0.29}$ & $_{-0.33}^{+0.33}$\
26.0 & - & 31.2 & 0.74 & $\pm$ 0.04 & $_{-0.09}^{+0.09}$ & $_{-0.14}^{+0.14}$\
31.2 & - & 37.0 & 0.289 & $\pm$ 0.022 & $_{-0.052}^{+0.044}$ & $_{-0.051}^{+0.051}$\
37.0 & - & 43.9 & 0.100 & $\pm$ 0.014 & $_{-0.023}^{+0.028}$ & $_{-0.020}^{+0.020}$\
43.9 & - & 51.8 & 0.016 & $\pm$ 0.004 & $_{-0.011}^{+0.017}$ & $_{-0.004}^{+0.004}$\
51.8 & - & 60.8 & 0.0022 & $\pm$ 0.0014 & $_{-0.0022}^{+0.0040}$ & $_{-0.0010}^{+0.0010}$\
[|l c l||l l l l|]{} & $d\sigma/dE_{T}^{{\rm jet2}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
6.0 & - & 8.5 & 27.0 & $\pm$ 0.3 & $_{-3.6}^{+3.1}$ & $_{-4.9}^{+4.9}$\
8.5 & - & 11.2 & 15.3 & $\pm$ 0.3 & $_{-1.8}^{+2.2}$ & $_{-2.9}^{+2.9}$\
11.2 & - & 14.2 & 2.04 & $\pm$ 0.10 & $_{-0.30}^{+0.50}$ & $_{-0.33}^{+0.33}$\
14.2 & - & 17.6 & 0.066 & $\pm$ 0.017 & $_{-0.061}^{+0.050}$ & $_{-0.004}^{+0.004}$\
\
6.0 & - & 8.5 & 5.7 & $\pm$ 0.2 & $_{-1.5}^{+1.9}$ & $_{-0.7}^{+0.7}$\
8.5 & - & 11.2 & 12.7 & $\pm$ 0.3 & $_{-1.4}^{+1.8}$ & $_{-1.9}^{+1.9}$\
11.2 & - & 14.2 & 8.4 & $\pm$ 0.2 & $_{-0.7}^{+0.5}$ & $_{-1.4}^{+1.4}$\
14.2 & - & 17.6 & 3.40 & $\pm$ 0.13 & $_{-0.25}^{+0.48}$ & $_{-0.48}^{+0.48}$\
17.6 & - & 21.5 & 1.55 & $\pm$ 0.08 & $_{-0.51}^{+0.26}$ & $_{-0.24}^{+0.24}$\
21.5 & - & 26.0 & 0.55 & $\pm$ 0.05 & $_{-0.09}^{+0.17}$ & $_{-0.08}^{+0.08}$\
26.0 & - & 31.2 & 0.152 & $\pm$ 0.021 & $_{-0.032}^{+0.043}$ & $_{-0.036}^{+0.036}$\
31.2 & - & 37.0 & 0.058 & $\pm$ 0.016 & $_{-0.029}^{+0.036}$ & $_{-0.013}^{+0.013}$\
37.0 & - & 43.9 & 0.017 & $\pm$ 0.005 & $_{-0.009}^{+0.014}$ & $_{-0.005}^{+0.005}$\
[|l c l||l l l l|]{} & $d\sigma/dE_{T}^{{\rm jet3}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
6.0 & - & 8.5 & 40.8 & $\pm$ 0.4 & $_{-4.0}^{+4.7}$ & $_{-7.7}^{+7.7}$\
8.5 & - & 11.2 & 4.72 & $\pm$ 0.14 & $_{-0.72}^{+0.99}$ & $_{-0.85}^{+0.85}$\
11.2 & - & 14.2 & 0.054 & $\pm$ 0.019 & $_{-0.047}^{+0.060}$ & $_{-0.005}^{+0.005}$\
\
6.0 & - & 8.5 & 17.5 & $\pm$ 0.4 & $_{-2.6}^{+3.7}$ & $_{-2.6}^{+2.6}$\
8.5 & - & 11.2 & 13.1 & $\pm$ 0.3 & $_{-1.0}^{+1.4}$ & $_{-1.7}^{+1.7}$\
11.2 & - & 14.2 & 4.3 & $\pm$ 0.1 & $_{-1.5}^{+0.7}$ & $_{-0.6}^{+0.6}$\
14.2 & - & 17.6 & 0.99 & $\pm$ 0.07 & $_{-0.22}^{+0.21}$ & $_{-0.10}^{+0.10}$\
17.6 & - & 21.5 & 0.191 & $\pm$ 0.031 & $_{-0.049}^{+0.075}$ & $_{-0.051}^{+0.051}$\
21.5 & - & 26.0 & 0.035 & $\pm$ 0.009 & $_{-0.015}^{+0.017}$ & $_{-0.007}^{+0.007}$\
[|l c l||l l l l|]{} & $d\sigma/dE_{T}^{{\rm jet4}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
6.0 & - & 8.5 & 45.4 & $\pm$ 0.4 & $_{-2.4}^{+5.3}$ & $_{-8.6}^{+8.6}$\
8.5 & - & 11.2 & 0.53 & $\pm$ 0.05 & $_{-0.11}^{+0.15}$ & $_{-0.11}^{+0.11}$\
\
6.0 & - & 8.5 & 32.0 & $\pm$ 0.5 & $_{-3.8}^{+4.5}$ & $_{-4.9}^{+4.9}$\
8.5 & - & 11.2 & 5.02 & $\pm$ 0.17 & $_{-0.45}^{+0.37}$ & $_{-0.71}^{+0.71}$\
11.2 & - & 14.2 & 0.81 & $\pm$ 0.07 & $_{-0.24}^{+0.59}$ & $_{-0.09}^{+0.09}$\
14.2 & - & 20.0 & 0.086 & $\pm$ 0.014 & $_{-0.034}^{+0.044}$ & $_{-0.004}^{+0.004}$\
[|l c l||l l l l||l l l l|]{} & $d\sigma/d\eta^{{\rm jet1}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale & &\
& & & & &\
\
-1.686 & - & -1.371 & 3.3 & $\pm$ 0.4 & $_{-1.7}^{+1.4}$ & $_{-0.5}^{+0.5}$ &0.324& $\pm$0.012&1.15& $\pm$0.21\
-1.371 & - & -1.057 & 36.0 & $\pm$ 1.4 & $_{-9.9}^{+7.8}$ & $_{-2.9}^{+2.9}$ &0.538& $\pm$0.041&1.34& $\pm$0.23\
-1.057 & - & -0.743 & 123 & $\pm$ 2 & $_{-20}^{+18}$ & $_{-9}^{+9}$ &0.686& $\pm$0.044&1.33& $\pm$0.18\
-0.743 & - & -0.429 & 256 & $\pm$ 3 & $_{-37}^{+40}$ & $_{-19}^{+19}$ &0.763& $\pm$0.025&1.35& $\pm$0.19\
-0.429 & - & -0.114 & 394 & $\pm$ 4 & $_{-51}^{+55}$ & $_{-33}^{+33}$ &0.793& $\pm$0.025&1.35& $\pm$0.19\
-0.114 & - & 0.200 & 506 & $\pm$ 4 & $_{-63}^{+67}$ & $_{-51}^{+51}$ &0.803& $\pm$0.014&1.37& $\pm$0.21\
0.200 & - & 0.514 & 600 & $\pm$ 5 & $_{-65}^{+68}$ & $_{-65}^{+65}$ &0.820& $\pm$0.009&1.39& $\pm$0.21\
0.514 & - & 0.829 & 653 & $\pm$ 5 & $_{-57}^{+65}$ & $_{-74}^{+74}$ &0.816& $\pm$0.008&1.43& $\pm$0.22\
0.829 & - & 1.143 & 631 & $\pm$ 5 & $_{-49}^{+54}$ & $_{-76}^{+76}$ &0.814& $\pm$0.004&1.49& $\pm$0.25\
1.143 & - & 1.457 & 556 & $\pm$ 4 & $_{-58}^{+63}$ & $_{-72}^{+72}$ &0.814& $\pm$0.007&1.55& $\pm$0.28\
1.457 & - & 1.771 & 573 & $\pm$ 4 & $_{-50}^{+58}$ & $_{-85}^{+85}$ &0.821& $\pm$0.003&1.64& $\pm$0.31\
1.771 & - & 2.086 & 697 & $\pm$ 6 & $_{-84}^{+97}$ & $_{-103}^{+103}$ &0.829& $\pm$0.002&1.75& $\pm$0.36\
2.086 & - & 2.400 & 790 & $\pm$ 9 & $_{-186}^{+186}$ & $_{-107}^{+107}$ &0.843& $\pm$0.005&1.90& $\pm$0.42\
\
-1.686 & - & -1.371 & 2.4 & $\pm$ 0.5 & $_{-1.6}^{+1.7}$ & $_{-0.6}^{+0.6}$ &0.300& $\pm$0.040&1.19& $\pm$0.25\
-1.371 & - & -1.057 & 14.7 & $\pm$ 1.1 & $_{-4.0}^{+5.5}$ & $_{-2.0}^{+2.0}$ &0.498& $\pm$0.022&1.20& $\pm$0.15\
-1.057 & - & -0.743 & 42 & $\pm$ 2 & $_{-11}^{+12}$ & $_{-4}^{+4}$ &0.648& $\pm$0.030&1.13& $\pm$0.10\
-0.743 & - & -0.429 & 67 & $\pm$ 2 & $_{-15}^{+16}$ & $_{-6}^{+6}$ &0.757& $\pm$0.019&1.10& $\pm$0.10\
-0.429 & - & -0.114 & 84 & $\pm$ 2 & $_{-16}^{+18}$ & $_{-7}^{+7}$ &0.830& $\pm$0.019&1.08& $\pm$0.09\
-0.114 & - & 0.200 & 90 & $\pm$ 2 & $_{-10}^{+11}$ & $_{-8}^{+8}$ &0.863& $\pm$0.010&1.06& $\pm$0.06\
0.200 & - & 0.514 & 99 & $\pm$ 2 & $_{-16}^{+14}$ & $_{-9}^{+9}$ &0.867& $\pm$0.008&1.06& $\pm$0.07\
0.514 & - & 0.829 & 92 & $\pm$ 2 & $_{-9}^{+10}$ & $_{-10}^{+10}$ &0.840& $\pm$0.006&1.07& $\pm$0.08\
0.829 & - & 1.143 & 99 & $\pm$ 2 & $_{-13}^{+13}$ & $_{-11}^{+11}$ &0.830& $\pm$0.018&1.08& $\pm$0.09\
1.143 & - & 1.457 & 115 & $\pm$ 2 & $_{-19}^{+14}$ & $_{-14}^{+14}$ &0.812& $\pm$0.026&1.11& $\pm$0.12\
1.457 & - & 1.771 & 145 & $\pm$ 2 & $_{-21}^{+17}$ & $_{-18}^{+18}$ &0.808& $\pm$0.019&1.12& $\pm$0.12\
1.771 & - & 2.086 & 200 & $\pm$ 3 & $_{-34}^{+31}$ & $_{-23}^{+23}$ &0.818& $\pm$0.012&1.16& $\pm$0.16\
2.086 & - & 2.400 & 283 & $\pm$ 4 & $_{-54}^{+66}$ & $_{-32}^{+32}$ &0.829& $\pm$0.012&1.22& $\pm$0.17\
[|l c l||l l l l||l l l l|]{} & $d\sigma/d\eta^{{\rm jet2}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale & &\
& & & & &\
\
-1.686 & - & -1.371 & 9.1 & $\pm$ 0.8 & $_{-3.4}^{+7.4}$ & $_{-1.8}^{+1.8}$ &0.382& $\pm$0.005&1.17& $\pm$0.20\
-1.371 & - & -1.057 & 57 & $\pm$ 2 & $_{-10}^{+11}$ & $_{-9}^{+9}$ &0.592& $\pm$0.019&1.35& $\pm$0.27\
-1.057 & - & -0.743 & 168 & $\pm$ 3 & $_{-25}^{+21}$ & $_{-15}^{+15}$ &0.720& $\pm$0.024&1.40& $\pm$0.25\
-0.743 & - & -0.429 & 282 & $\pm$ 4 & $_{-30}^{+39}$ & $_{-26}^{+26}$ &0.766& $\pm$0.015&1.42& $\pm$0.26\
-0.429 & - & -0.114 & 408 & $\pm$ 4 & $_{-44}^{+41}$ & $_{-39}^{+39}$ &0.780& $\pm$0.012&1.43& $\pm$0.25\
-0.114 & - & 0.200 & 492 & $\pm$ 4 & $_{-46}^{+54}$ & $_{-50}^{+50}$ &0.808& $\pm$0.007&1.42& $\pm$0.22\
0.200 & - & 0.514 & 574 & $\pm$ 5 & $_{-57}^{+66}$ & $_{-67}^{+67}$ &0.817& $\pm$0.005&1.44& $\pm$0.22\
0.514 & - & 0.829 & 628 & $\pm$ 5 & $_{-61}^{+67}$ & $_{-73}^{+73}$ &0.817& $\pm$0.003&1.45& $\pm$0.21\
0.829 & - & 1.143 & 615 & $\pm$ 5 & $_{-57}^{+62}$ & $_{-72}^{+72}$ &0.816& $\pm$0.003&1.48& $\pm$0.22\
1.143 & - & 1.457 & 550 & $\pm$ 4 & $_{-34}^{+40}$ & $_{-66}^{+66}$ &0.818& $\pm$0.013&1.50& $\pm$0.25\
1.457 & - & 1.771 & 564 & $\pm$ 4 & $_{-32}^{+36}$ & $_{-77}^{+77}$ &0.819& $\pm$0.012&1.56& $\pm$0.29\
1.771 & - & 2.086 & 686 & $\pm$ 6 & $_{-82}^{+79}$ & $_{-96}^{+96}$ &0.833& $\pm$0.015&1.63& $\pm$0.33\
2.086 & - & 2.400 & 767 & $\pm$ 8 & $_{-181}^{+183}$ & $_{-96}^{+96}$ &0.844& $\pm$0.015&1.75& $\pm$0.38\
\
-2.000 & - & -1.686 & 0.48 & $\pm$ 0.16 & $_{-0.27}^{+0.44}$ & $_{-0.18}^{+0.18}$ &0.085& $\pm$0.007&1.00& $\pm$0.18\
-1.686 & - & -1.371 & 8.2 & $\pm$ 0.7 & $_{-2.5}^{+2.5}$ & $_{-0.9}^{+0.9}$ &0.347& $\pm$0.021&1.20& $\pm$0.22\
-1.371 & - & -1.057 & 43 & $\pm$ 2 & $_{-19}^{+16}$ & $_{-5}^{+5}$ &0.571& $\pm$0.025&1.20& $\pm$0.17\
-1.057 & - & -0.743 & 70 & $\pm$ 2 & $_{-18}^{+20}$ & $_{-7}^{+7}$ &0.710& $\pm$0.038&1.18& $\pm$0.14\
-0.743 & - & -0.429 & 92 & $\pm$ 2 & $_{-24}^{+22}$ & $_{-11}^{+11}$ &0.783& $\pm$0.014&1.14& $\pm$0.12\
-0.429 & - & -0.114 & 97 & $\pm$ 2 & $_{-15}^{+20}$ & $_{-10}^{+10}$ &0.822& $\pm$0.013&1.12& $\pm$0.09\
-0.114 & - & 0.200 & 96 & $\pm$ 2 & $_{-12}^{+12}$ & $_{-10}^{+10}$ &0.845& $\pm$0.006&1.10& $\pm$0.10\
0.200 & - & 0.514 & 104 & $\pm$ 2 & $_{-13}^{+12}$ & $_{-11}^{+11}$ &0.852& $\pm$0.007&1.09& $\pm$0.08\
0.514 & - & 0.829 & 106 & $\pm$ 2 & $_{-14}^{+16}$ & $_{-12}^{+12}$ &0.858& $\pm$0.001&1.09& $\pm$0.09\
0.829 & - & 1.143 & 108 & $\pm$ 2 & $_{-17}^{+14}$ & $_{-11}^{+11}$ &0.851& $\pm$0.009&1.09& $\pm$0.10\
1.143 & - & 1.457 & 103 & $\pm$ 2 & $_{-16}^{+16}$ & $_{-11}^{+11}$ &0.846& $\pm$0.011&1.09& $\pm$0.10\
1.457 & - & 1.771 & 113 & $\pm$ 2 & $_{-14}^{+19}$ & $_{-13}^{+13}$ &0.827& $\pm$0.018&1.10& $\pm$0.11\
1.771 & - & 2.086 & 157 & $\pm$ 3 & $_{-20}^{+26}$ & $_{-17}^{+17}$ &0.829& $\pm$0.013&1.12& $\pm$0.13\
2.086 & - & 2.400 & 232 & $\pm$ 4 & $_{-49}^{+51}$ & $_{-23}^{+23}$ &0.839& $\pm$0.010&1.14& $\pm$0.15\
[|l c l||l l l l||l l l l|]{} & $d\sigma/d\eta^{{\rm jet3}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale & &\
& & & & &\
\
-2.000 & - & -1.686 & 0.33 & $\pm$ 0.23 & $_{-0.33}^{+0.49}$ & $_{-0.08}^{+0.08}$ &0.090& $\pm$0.045&1.42& $\pm$0.82\
-1.686 & - & -1.371 & 15.2 & $\pm$ 1.0 & $_{-3.5}^{+3.4}$ & $_{-1.3}^{+1.3}$ &0.388& $\pm$0.019&1.30& $\pm$0.33\
-1.371 & - & -1.057 & 80 & $\pm$ 2 & $_{-13}^{+16}$ & $_{-7}^{+7}$ &0.631& $\pm$0.004&1.33& $\pm$0.33\
-1.057 & - & -0.743 & 187 & $\pm$ 3 & $_{-24}^{+25}$ & $_{-17}^{+17}$ &0.738& $\pm$0.007&1.38& $\pm$0.34\
-0.743 & - & -0.429 & 308 & $\pm$ 4 & $_{-39}^{+37}$ & $_{-29}^{+29}$ &0.791& $\pm$0.002&1.41& $\pm$0.31\
-0.429 & - & -0.114 & 396 & $\pm$ 4 & $_{-35}^{+46}$ & $_{-42}^{+42}$ &0.800& $\pm$0.006&1.44& $\pm$0.30\
-0.114 & - & 0.200 & 474 & $\pm$ 4 & $_{-44}^{+49}$ & $_{-53}^{+53}$ &0.826& $\pm$0.009&1.46& $\pm$0.27\
0.200 & - & 0.514 & 560 & $\pm$ 5 & $_{-60}^{+67}$ & $_{-67}^{+67}$ &0.829& $\pm$0.015&1.49& $\pm$0.26\
0.514 & - & 0.829 & 616 & $\pm$ 5 & $_{-74}^{+73}$ & $_{-72}^{+72}$ &0.818& $\pm$0.000&1.51& $\pm$0.22\
0.829 & - & 1.143 & 605 & $\pm$ 5 & $_{-58}^{+67}$ & $_{-71}^{+71}$ &0.807& $\pm$0.010&1.52& $\pm$0.21\
1.143 & - & 1.457 & 538 & $\pm$ 4 & $_{-29}^{+38}$ & $_{-64}^{+64}$ &0.806& $\pm$0.017&1.50& $\pm$0.19\
1.457 & - & 1.771 & 559 & $\pm$ 4 & $_{-30}^{+34}$ & $_{-72}^{+72}$ &0.808& $\pm$0.022&1.52& $\pm$0.21\
1.771 & - & 2.086 & 685 & $\pm$ 5 & $_{-75}^{+79}$ & $_{-91}^{+91}$ &0.822& $\pm$0.032&1.57& $\pm$0.24\
2.086 & - & 2.400 & 793 & $\pm$ 8 & $_{-177}^{+183}$ & $_{-103}^{+103}$ &0.835& $\pm$0.030&1.66& $\pm$0.29\
\
-2.000 & - & -1.686 & 1.9 & $\pm$ 0.3 & $_{-1.2}^{+1.2}$ & $_{-1.0}^{+1.0}$ &0.145& $\pm$0.014&1.01& $\pm$0.11\
-1.686 & - & -1.371 & 26.1 & $\pm$ 1.3 & $_{-6.9}^{+9.5}$ & $_{-3.1}^{+3.1}$ &0.456& $\pm$0.004&1.19& $\pm$0.22\
-1.371 & - & -1.057 & 70 & $\pm$ 2 & $_{-16}^{+14}$ & $_{-9}^{+9}$ &0.673& $\pm$0.003&1.20& $\pm$0.22\
-1.057 & - & -0.743 & 104 & $\pm$ 2 & $_{-22}^{+26}$ & $_{-11}^{+11}$ &0.788& $\pm$0.001&1.18& $\pm$0.19\
-0.743 & - & -0.429 & 100 & $\pm$ 2 & $_{-16}^{+20}$ & $_{-10}^{+10}$ &0.840& $\pm$0.005&1.14& $\pm$0.15\
-0.429 & - & -0.114 & 97 & $\pm$ 2 & $_{-16}^{+12}$ & $_{-10}^{+10}$ &0.876& $\pm$0.014&1.13& $\pm$0.14\
-0.114 & - & 0.200 & 92 & $\pm$ 2 & $_{-13}^{+11}$ & $_{-10}^{+10}$ &0.901& $\pm$0.033&1.11& $\pm$0.12\
0.200 & - & 0.514 & 89 & $\pm$ 2 & $_{-15}^{+19}$ & $_{-9}^{+9}$ &0.900& $\pm$0.036&1.10& $\pm$0.11\
0.514 & - & 0.829 & 98 & $\pm$ 2 & $_{-16}^{+17}$ & $_{-11}^{+11}$ &0.856& $\pm$0.022&1.12& $\pm$0.10\
0.829 & - & 1.143 & 95 & $\pm$ 2 & $_{-14}^{+16}$ & $_{-10}^{+10}$ &0.840& $\pm$0.002&1.11& $\pm$0.09\
1.143 & - & 1.457 & 89 & $\pm$ 2 & $_{-12}^{+11}$ & $_{-10}^{+10}$ &0.819& $\pm$0.012&1.12& $\pm$0.08\
1.457 & - & 1.771 & 103 & $\pm$ 2 & $_{-15}^{+16}$ & $_{-12}^{+12}$ &0.815& $\pm$0.023&1.11& $\pm$0.06\
1.771 & - & 2.086 & 138 & $\pm$ 3 & $_{-21}^{+22}$ & $_{-16}^{+16}$ &0.817& $\pm$0.030&1.11& $\pm$0.08\
2.086 & - & 2.400 & 222 & $\pm$ 4 & $_{-49}^{+44}$ & $_{-19}^{+19}$ &0.820& $\pm$0.042&1.12& $\pm$0.07\
[|l c l||l l l l|]{} & $d\sigma/d\eta^{{\rm jet1}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
-1.371 & - & -1.057 & 0.61 & $\pm$ 0.09 & $_{-0.36}^{+0.91}$ & $_{-0.08}^{+0.08}$\
-1.057 & - & -0.743 & 1.76 & $\pm$ 0.23 & $_{-0.67}^{+0.85}$ & $_{-0.30}^{+0.30}$\
-0.743 & - & -0.429 & 6.1 & $\pm$ 0.4 & $_{-1.5}^{+1.0}$ & $_{-1.0}^{+1.0}$\
-0.429 & - & -0.114 & 13.2 & $\pm$ 0.6 & $_{-3.0}^{+2.2}$ & $_{-2.1}^{+2.1}$\
-0.114 & - & 0.200 & 21.9 & $\pm$ 0.8 & $_{-3.6}^{+4.4}$ & $_{-4.0}^{+4.0}$\
0.200 & - & 0.514 & 36.7 & $\pm$ 1.1 & $_{-6.7}^{+5.7}$ & $_{-6.4}^{+6.4}$\
0.514 & - & 0.829 & 43.3 & $\pm$ 1.3 & $_{-4.6}^{+6.0}$ & $_{-7.5}^{+7.5}$\
0.829 & - & 1.143 & 55 & $\pm$ 1 & $_{-12}^{+12}$ & $_{-9}^{+9}$\
1.143 & - & 1.457 & 43 & $\pm$ 1 & $_{-8}^{+10}$ & $_{-8}^{+8}$\
1.457 & - & 1.771 & 44.5 & $\pm$ 1.1 & $_{-4.1}^{+8.1}$ & $_{-9.5}^{+9.5}$\
1.771 & - & 2.086 & 54 & $\pm$ 2 & $_{-9}^{+11}$ & $_{-12}^{+12}$\
2.086 & - & 2.400 & 68 & $\pm$ 3 & $_{-24}^{+25}$ & $_{-13}^{+13}$\
\
-1.371 & - & -1.057 & 1.10 & $\pm$ 0.29 & $_{-0.70}^{+0.63}$ & $_{-0.14}^{+0.14}$\
-1.057 & - & -0.743 & 4.9 & $\pm$ 0.6 & $_{-1.4}^{+2.3}$ & $_{-0.6}^{+0.6}$\
-0.743 & - & -0.429 & 10.9 & $\pm$ 0.8 & $_{-2.3}^{+3.7}$ & $_{-1.2}^{+1.2}$\
-0.429 & - & -0.114 & 20.6 & $\pm$ 0.9 & $_{-7.1}^{+5.1}$ & $_{-2.5}^{+2.5}$\
-0.114 & - & 0.200 & 20.9 & $\pm$ 1.0 & $_{-2.8}^{+3.7}$ & $_{-2.7}^{+2.7}$\
0.200 & - & 0.514 & 24.0 & $\pm$ 1.1 & $_{-2.5}^{+4.0}$ & $_{-3.3}^{+3.3}$\
0.514 & - & 0.829 & 24.4 & $\pm$ 1.1 & $_{-2.5}^{+4.8}$ & $_{-3.4}^{+3.4}$\
0.829 & - & 1.143 & 27.8 & $\pm$ 1.2 & $_{-3.4}^{+3.5}$ & $_{-3.6}^{+3.6}$\
1.143 & - & 1.457 & 28.0 & $\pm$ 1.0 & $_{-4.7}^{+4.2}$ & $_{-4.3}^{+4.3}$\
1.457 & - & 1.771 & 37.0 & $\pm$ 1.3 & $_{-3.7}^{+5.9}$ & $_{-6.4}^{+6.4}$\
1.771 & - & 2.086 & 47.0 & $\pm$ 1.7 & $_{-8.9}^{+7.6}$ & $_{-8.6}^{+8.6}$\
2.086 & - & 2.400 & 63 & $\pm$ 2 & $_{-17}^{+16}$ & $_{-11}^{+11}$\
[|l c l||l l l l|]{} & $d\sigma/d\eta^{{\rm jet2}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
-1.371 & - & -1.057 & 0.67 & $\pm$ 0.18 & $_{-0.50}^{+0.43}$ & $_{-0.11}^{+0.11}$\
-1.057 & - & -0.743 & 3.15 & $\pm$ 0.37 & $_{-0.73}^{+0.87}$ & $_{-0.35}^{+0.35}$\
-0.743 & - & -0.429 & 11.3 & $\pm$ 0.5 & $_{-6.9}^{+6.5}$ & $_{-1.0}^{+1.0}$\
-0.429 & - & -0.114 & 17.0 & $\pm$ 0.8 & $_{-2.8}^{+2.4}$ & $_{-1.9}^{+1.9}$\
-0.114 & - & 0.200 & 23.5 & $\pm$ 0.9 & $_{-3.6}^{+4.1}$ & $_{-3.5}^{+3.5}$\
0.200 & - & 0.514 & 37.9 & $\pm$ 1.2 & $_{-5.8}^{+4.9}$ & $_{-6.3}^{+6.3}$\
0.514 & - & 0.829 & 42.4 & $\pm$ 1.3 & $_{-5.7}^{+4.9}$ & $_{-7.5}^{+7.5}$\
0.829 & - & 1.143 & 46.2 & $\pm$ 1.3 & $_{-6.4}^{+4.2}$ & $_{-9.0}^{+9.0}$\
1.143 & - & 1.457 & 40.3 & $\pm$ 1.1 & $_{-6.9}^{+8.7}$ & $_{-8.3}^{+8.3}$\
1.457 & - & 1.771 & 43.3 & $\pm$ 1.1 & $_{-5.5}^{+8.7}$ & $_{-9.0}^{+9.0}$\
1.771 & - & 2.086 & 52 & $\pm$ 1 & $_{-7}^{+10}$ & $_{-11}^{+11}$\
2.086 & - & 2.400 & 57 & $\pm$ 2 & $_{-15}^{+19}$ & $_{-11}^{+11}$\
\
-1.686 & - & -1.371 & 0.08 & $\pm$ 0.04 & $_{-0.08}^{+0.21}$ & $_{-0.03}^{+0.03}$\
-1.371 & - & -1.057 & 3.1 & $\pm$ 0.3 & $_{-1.3}^{+5.2}$ & $_{-0.6}^{+0.6}$\
-1.057 & - & -0.743 & 7.6 & $\pm$ 0.7 & $_{-1.4}^{+1.9}$ & $_{-1.0}^{+1.0}$\
-0.743 & - & -0.429 & 15.2 & $\pm$ 0.9 & $_{-1.7}^{+2.3}$ & $_{-2.5}^{+2.5}$\
-0.429 & - & -0.114 & 22.9 & $\pm$ 1.2 & $_{-6.0}^{+4.2}$ & $_{-3.5}^{+3.5}$\
-0.114 & - & 0.200 & 20.9 & $\pm$ 1.1 & $_{-2.5}^{+3.8}$ & $_{-2.5}^{+2.5}$\
0.200 & - & 0.514 & 26.4 & $\pm$ 1.2 & $_{-5.2}^{+2.7}$ & $_{-3.2}^{+3.2}$\
0.514 & - & 0.829 & 33.8 & $\pm$ 1.3 & $_{-9.4}^{+5.6}$ & $_{-3.9}^{+3.9}$\
0.829 & - & 1.143 & 28.1 & $\pm$ 1.3 & $_{-3.0}^{+5.0}$ & $_{-3.7}^{+3.7}$\
1.143 & - & 1.457 & 25.4 & $\pm$ 1.1 & $_{-4.7}^{+5.5}$ & $_{-3.6}^{+3.6}$\
1.457 & - & 1.771 & 29.4 & $\pm$ 1.1 & $_{-3.6}^{+4.1}$ & $_{-4.4}^{+4.4}$\
1.771 & - & 2.086 & 43.3 & $\pm$ 1.6 & $_{-7.0}^{+8.4}$ & $_{-6.7}^{+6.7}$\
2.086 & - & 2.400 & 51 & $\pm$ 2 & $_{-12}^{+15}$ & $_{-10}^{+10}$\
[|l c l||l l l l|]{} & $d\sigma/d\eta^{{\rm jet3}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
-1.371 & - & -1.057 & 0.5 & $\pm$ 0.1 & $_{-0.4}^{+1.1}$ & $_{-0.3}^{+0.3}$\
-1.057 & - & -0.743 & 3.7 & $\pm$ 0.4 & $_{-0.9}^{+1.1}$ & $_{-0.9}^{+0.9}$\
-0.743 & - & -0.429 & 9.4 & $\pm$ 0.6 & $_{-1.5}^{+1.9}$ & $_{-1.2}^{+1.2}$\
-0.429 & - & -0.114 & 20.0 & $\pm$ 0.8 & $_{-7.8}^{+7.2}$ & $_{-3.2}^{+3.2}$\
-0.114 & - & 0.200 & 26.2 & $\pm$ 0.9 & $_{-3.3}^{+2.6}$ & $_{-4.6}^{+4.6}$\
0.200 & - & 0.514 & 37.2 & $\pm$ 1.2 & $_{-4.4}^{+6.4}$ & $_{-6.5}^{+6.5}$\
0.514 & - & 0.829 & 45.7 & $\pm$ 1.4 & $_{-5.8}^{+4.3}$ & $_{-7.8}^{+7.8}$\
0.829 & - & 1.143 & 47.8 & $\pm$ 1.4 & $_{-5.6}^{+5.0}$ & $_{-8.5}^{+8.5}$\
1.143 & - & 1.457 & 37.8 & $\pm$ 1.0 & $_{-4.1}^{+8.2}$ & $_{-7.3}^{+7.3}$\
1.457 & - & 1.771 & 40.0 & $\pm$ 1.0 & $_{-5.5}^{+7.2}$ & $_{-8.6}^{+8.6}$\
1.771 & - & 2.086 & 51 & $\pm$ 1 & $_{-6}^{+8}$ & $_{-11}^{+11}$\
2.086 & - & 2.400 & 61 & $\pm$ 2 & $_{-19}^{+16}$ & $_{-13}^{+13}$\
\
-1.686 & - & -1.371 & 0.59 & $\pm$ 0.20 & $_{-0.59}^{+0.97}$ & $_{-0.28}^{+0.28}$\
-1.371 & - & -1.057 & 6.1 & $\pm$ 0.6 & $_{-2.0}^{+2.0}$ & $_{-1.4}^{+1.4}$\
-1.057 & - & -0.743 & 19 & $\pm$ 1 & $_{-10}^{+5}$ & $_{-3}^{+3}$\
-0.743 & - & -0.429 & 20.4 & $\pm$ 1.1 & $_{-4.3}^{+4.6}$ & $_{-4.6}^{+4.6}$\
-0.429 & - & -0.114 & 21.0 & $\pm$ 1.1 & $_{-3.4}^{+4.0}$ & $_{-3.9}^{+3.9}$\
-0.114 & - & 0.200 & 22.7 & $\pm$ 1.2 & $_{-4.3}^{+4.8}$ & $_{-3.6}^{+3.6}$\
0.200 & - & 0.514 & 24.8 & $\pm$ 1.2 & $_{-4.1}^{+6.0}$ & $_{-3.7}^{+3.7}$\
0.514 & - & 0.829 & 27.6 & $\pm$ 1.3 & $_{-3.4}^{+3.8}$ & $_{-4.1}^{+4.1}$\
0.829 & - & 1.143 & 27.3 & $\pm$ 1.3 & $_{-4.1}^{+6.3}$ & $_{-4.1}^{+4.1}$\
1.143 & - & 1.457 & 24.0 & $\pm$ 1.1 & $_{-2.7}^{+4.1}$ & $_{-3.9}^{+3.9}$\
1.457 & - & 1.771 & 26.3 & $\pm$ 1.0 & $_{-2.7}^{+4.3}$ & $_{-3.9}^{+3.9}$\
1.771 & - & 2.086 & 33.5 & $\pm$ 1.4 & $_{-2.9}^{+4.3}$ & $_{-4.8}^{+4.8}$\
2.086 & - & 2.400 & 55 & $\pm$ 2 & $_{-13}^{+11}$ & $_{-5}^{+5}$\
[|l c l||l l l l|]{} & $d\sigma/d\eta^{{\rm jet4}}$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale\
&\
\
-1.371 & - & -1.057 & 1.3 & $\pm$ 0.2 & $_{-1.1}^{+1.0}$ & $_{-0.3}^{+0.3}$\
-1.057 & - & -0.743 & 4.4 & $\pm$ 0.4 & $_{-1.6}^{+1.1}$ & $_{-0.4}^{+0.4}$\
-0.743 & - & -0.429 & 18 & $\pm$ 1 & $_{-13}^{+10}$ & $_{-3}^{+3}$\
-0.429 & - & -0.114 & 18.8 & $\pm$ 0.8 & $_{-1.9}^{+2.1}$ & $_{-3.6}^{+3.6}$\
-0.114 & - & 0.200 & 26.3 & $\pm$ 0.9 & $_{-2.4}^{+4.0}$ & $_{-4.5}^{+4.5}$\
0.200 & - & 0.514 & 36.6 & $\pm$ 1.2 & $_{-4.3}^{+4.5}$ & $_{-6.4}^{+6.4}$\
0.514 & - & 0.829 & 44.1 & $\pm$ 1.3 & $_{-6.3}^{+7.1}$ & $_{-7.7}^{+7.7}$\
0.829 & - & 1.143 & 45.0 & $\pm$ 1.4 & $_{-5.0}^{+7.3}$ & $_{-7.9}^{+7.9}$\
1.143 & - & 1.457 & 37.8 & $\pm$ 1.0 & $_{-4.1}^{+7.9}$ & $_{-7.0}^{+7.0}$\
1.457 & - & 1.771 & 39.2 & $\pm$ 1.0 & $_{-3.5}^{+7.5}$ & $_{-8.2}^{+8.2}$\
1.771 & - & 2.086 & 48.3 & $\pm$ 1.3 & $_{-6.8}^{+7.6}$ & $_{-9.8}^{+9.8}$\
2.086 & - & 2.400 & 60 & $\pm$ 2 & $_{-15}^{+14}$ & $_{-12}^{+12}$\
\
-1.686 & - & -1.371 & 2.8 & $\pm$ 0.4 & $_{-1.8}^{+2.0}$ & $_{-2.2}^{+5.6}$\
-1.371 & - & -1.057 & 7.3 & $\pm$ 0.7 & $_{-2.2}^{+2.8}$ & $_{-1.0}^{+1.0}$\
-1.057 & - & -0.743 & 15.2 & $\pm$ 0.9 & $_{-2.2}^{+3.8}$ & $_{-2.0}^{+2.0}$\
-0.743 & - & -0.429 & 21.6 & $\pm$ 1.2 & $_{-4.0}^{+4.1}$ & $_{-3.5}^{+3.5}$\
-0.429 & - & -0.114 & 22.7 & $\pm$ 1.2 & $_{-5.7}^{+5.5}$ & $_{-3.5}^{+3.5}$\
-0.114 & - & 0.200 & 24.3 & $\pm$ 1.1 & $_{-4.1}^{+4.1}$ & $_{-3.9}^{+3.9}$\
0.200 & - & 0.514 & 23.3 & $\pm$ 1.2 & $_{-3.1}^{+4.9}$ & $_{-3.7}^{+3.7}$\
0.514 & - & 0.829 & 27.3 & $\pm$ 1.2 & $_{-4.0}^{+4.1}$ & $_{-4.0}^{+4.0}$\
0.829 & - & 1.143 & 34.6 & $\pm$ 1.3 & $_{-9.3}^{+6.3}$ & $_{-4.7}^{+4.7}$\
1.143 & - & 1.457 & 25.8 & $\pm$ 1.0 & $_{-3.5}^{+2.8}$ & $_{-3.7}^{+3.7}$\
1.457 & - & 1.771 & 25.6 & $\pm$ 1.0 & $_{-4.2}^{+5.3}$ & $_{-4.2}^{+4.2}$\
1.771 & - & 2.086 & 34.0 & $\pm$ 1.4 & $_{-5.8}^{+7.4}$ & $_{-5.4}^{+5.4}$\
2.086 & - & 2.400 & 44.3 & $\pm$ 2.0 & $_{-9.7}^{+9.9}$ & $_{-6.1}^{+6.1}$\
[|l c l||l l l l||l l l l|]{} & $d\sigma/dcos(\psi_{3})$ & $\pm$stat. & $\pm$syst. & $\pm$E-scale & &\
& & & & &\
\
-1.00 & - & -0.88 & 1494 & $\pm$ 13 & $_{-175}^{+217}$ & $_{-188}^{+188}$ &0.864& $\pm$0.012&1.76& $\pm$0.43\
-0.88 & - & -0.72 & 856 & $\pm$ 9 & $_{-105}^{+117}$ & $_{-107}^{+107}$ &0.821& $\pm$0.017&1.51& $\pm$0.28\
-0.72 & - & -0.56 & 804 & $\pm$ 8 & $_{-77}^{+77}$ & $_{-99}^{+99}$ &0.793& $\pm$0.019&1.43& $\pm$0.20\
-0.560 & - & -0.336 & 766 & $\pm$ 7 & $_{-76}^{+75}$ & $_{-94}^{+94}$ &0.776& $\pm$0.008&1.43& $\pm$0.19\
-0.336 & - & -0.112 & 759 & $\pm$ 7 & $_{-64}^{+70}$ & $_{-94}^{+94}$ &0.763& $\pm$0.005&1.42& $\pm$0.20\
-0.112 & - & 0.112 & 746 & $\pm$ 6 & $_{-57}^{+70}$ & $_{-92}^{+92}$ &0.761& $\pm$0.006&1.41& $\pm$0.19\
0.112 & - & 0.336 & 812 & $\pm$ 7 & $_{-80}^{+80}$ & $_{-98}^{+98}$ &0.767& $\pm$0.012&1.42& $\pm$0.20\
0.336 & - & 0.560 & 851 & $\pm$ 7 & $_{-98}^{+101}$ & $_{-99}^{+99}$ &0.780& $\pm$0.015&1.48& $\pm$0.22\
0.56 & - & 0.72 & 886 & $\pm$ 9 & $_{-103}^{+99}$ & $_{-100}^{+100}$ &0.809& $\pm$0.017&1.48& $\pm$0.22\
0.72 & - & 0.88 & 898 & $\pm$ 9 & $_{-107}^{+114}$ & $_{-102}^{+102}$ &0.839& $\pm$0.013&1.51& $\pm$0.27\
0.88 & - & 1.00 & 1475 & $\pm$ 13 & $_{-196}^{+212}$ & $_{-173}^{+173}$ &0.894& $\pm$0.006&1.64& $\pm$0.38\
\
-1.00 & - & -0.88 & 542 & $\pm$ 9 & $_{-94}^{+99}$ & $_{-55}^{+55}$ &0.824& $\pm$0.018&1.15& $\pm$0.15\
-0.88 & - & -0.72 & 263 & $\pm$ 5 & $_{-45}^{+44}$ & $_{-29}^{+29}$ &0.812& $\pm$0.031&1.11& $\pm$0.11\
-0.72 & - & -0.56 & 215 & $\pm$ 5 & $_{-40}^{+38}$ & $_{-25}^{+25}$ &0.807& $\pm$0.021&1.10& $\pm$0.10\
-0.560 & - & -0.336 & 164 & $\pm$ 3 & $_{-24}^{+26}$ & $_{-19}^{+19}$ &0.790& $\pm$0.026&1.10& $\pm$0.09\
-0.336 & - & -0.112 & 156 & $\pm$ 3 & $_{-30}^{+31}$ & $_{-18}^{+18}$ &0.793& $\pm$0.012&1.10& $\pm$0.10\
-0.112 & - & 0.112 & 149 & $\pm$ 3 & $_{-27}^{+33}$ & $_{-17}^{+17}$ &0.785& $\pm$0.013&1.11& $\pm$0.11\
0.112 & - & 0.336 & 143 & $\pm$ 3 & $_{-22}^{+25}$ & $_{-15}^{+15}$ &0.784& $\pm$0.009&1.12& $\pm$0.11\
0.336 & - & 0.560 & 151 & $\pm$ 3 & $_{-28}^{+25}$ & $_{-16}^{+16}$ &0.802& $\pm$0.017&1.12& $\pm$0.11\
0.56 & - & 0.72 & 169 & $\pm$ 4 & $_{-34}^{+32}$ & $_{-17}^{+17}$ &0.823& $\pm$0.015&1.10& $\pm$0.11\
0.72 & - & 0.88 & 217 & $\pm$ 5 & $_{-45}^{+39}$ & $_{-22}^{+22}$ &0.847& $\pm$0.006&1.10& $\pm$0.11\
0.88 & - & 1.00 & 370 & $\pm$ 7 & $_{-61}^{+67}$ & $_{-38}^{+38}$ &0.865& $\pm$0.009&1.12& $\pm$0.12\
![ A schematic representation of an event exhibiting a secondary scatter (MPI). \[fig:MPI\]](DESY-07-102_1.eps)
![ A schematic representation of the three-jet centre-of-mass system and the $\psi_3$ angle. \[fig:angles\]](DESY-07-102_2.eps)
![ Measured cross section as a function of (a) $M_{3j}$ and (b) $M_{4j}$ for the three- and four-jet samples (solid circles). The inner error bars represent the statistical uncertainty. The outer error bars represent the statistical plus systematic uncertainties added in quadrature. The shaded band represents the calorimeter energy scale uncertainty. The predictions from the [Herwig]{} and [Pythia]{} models are also shown both with and without MPIs, as is the direct component of the [Herwig]{} cross sections. Each Monte Carlo cross section has been area normalised to the high-mass ($M_{nj}\ge50$ GeV) measured cross section by scaling the predictions by the factors indicated in the legend. \[fig:xsec:Mnj\]](DESY-07-102_0.eps)
![ Measured cross section as a function of (a) $M_{3j}$ and (b) $M_{4j}$ for the three- and four-jet samples (solid circles). The inner error bars represent the statistical uncertainty. The outer error bars represent the statistical plus systematic uncertainties added in quadrature. The shaded band represents the calorimeter energy scale uncertainty. The predictions from the [Herwig]{} and [Pythia]{} models are also shown both with and without MPIs, as is the direct component of the [Herwig]{} cross sections. Each Monte Carlo cross section has been area normalised to the high-mass ($M_{nj}\ge50$ GeV) measured cross section by scaling the predictions by the factors indicated in the legend. \[fig:xsec:Mnj\]](DESY-07-102_3.eps)
(a)
![ Measured cross section as a function of (a) $M_{3j}$ and (b) $M_{4j}$ for the three- and four-jet samples (solid circles). The inner error bars represent the statistical uncertainty. The outer error bars represent the statistical plus systematic uncertainties added in quadrature. The shaded band represents the calorimeter energy scale uncertainty. The predictions from the [Herwig]{} and [Pythia]{} models are also shown both with and without MPIs, as is the direct component of the [Herwig]{} cross sections. Each Monte Carlo cross section has been area normalised to the high-mass ($M_{nj}\ge50$ GeV) measured cross section by scaling the predictions by the factors indicated in the legend. \[fig:xsec:Mnj\]](DESY-07-102_4.eps)
(b)
![ Measured cross section as a function of $x_\gamma^{obs}$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:Xgam\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $x_\gamma^{obs}$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:Xgam\]](DESY-07-102_5.eps "fig:") ![ Measured cross section as a function of $x_\gamma^{obs}$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:Xgam\]](DESY-07-102_6.eps "fig:")
(a)(b)
![ Measured cross section as a function of $x_\gamma^{obs}$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:Xgam\]](DESY-07-102_7.eps "fig:") ![ Measured cross section as a function of $x_\gamma^{obs}$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:Xgam\]](DESY-07-102_8.eps "fig:")
(c)(d)
![ Measured cross section as a function of $y$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. The predictions from the [Herwig]{} and [Pythia]{} models are also shown both with and without MPIs, as is the direct component of the [Herwig]{} cross sections. Each Monte Carlo cross section has been area normalised to the data. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:yJB\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $y$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. The predictions from the [Herwig]{} and [Pythia]{} models are also shown both with and without MPIs, as is the direct component of the [Herwig]{} cross sections. Each Monte Carlo cross section has been area normalised to the data. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:yJB\]](DESY-07-102_9.eps "fig:") ![ Measured cross section as a function of $y$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. The predictions from the [Herwig]{} and [Pythia]{} models are also shown both with and without MPIs, as is the direct component of the [Herwig]{} cross sections. Each Monte Carlo cross section has been area normalised to the data. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:yJB\]](DESY-07-102_10.eps "fig:")
(a)(b)
![ Measured cross section as a function of $y$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. The predictions from the [Herwig]{} and [Pythia]{} models are also shown both with and without MPIs, as is the direct component of the [Herwig]{} cross sections. Each Monte Carlo cross section has been area normalised to the data. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:yJB\]](DESY-07-102_11.eps "fig:") ![ Measured cross section as a function of $y$ in the three-jet (a) low- and (b) high-mass samples and the four-jet, (c) low- and (d) high-mass samples. The predictions from the [Herwig]{} and [Pythia]{} models are also shown both with and without MPIs, as is the direct component of the [Herwig]{} cross sections. Each Monte Carlo cross section has been area normalised to the data. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:xsec:yJB\]](DESY-07-102_12.eps "fig:")
(c)(d)
![ Measured cross section as a function of $E_T^{jet}$ in the three-jet (a) low- and (b) high-mass samples. The jets are ordered such that $E_T^{jet1}>E_T^{jet2}>E_T^{jet3}$. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:yJB\]. \[fig:xsec:Et\_3j\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $E_T^{jet}$ in the three-jet (a) low- and (b) high-mass samples. The jets are ordered such that $E_T^{jet1}>E_T^{jet2}>E_T^{jet3}$. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:yJB\]. \[fig:xsec:Et\_3j\]](DESY-07-102_13.eps)
(a)
![ Measured cross section as a function of $E_T^{jet}$ in the three-jet (a) low- and (b) high-mass samples. The jets are ordered such that $E_T^{jet1}>E_T^{jet2}>E_T^{jet3}$. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:yJB\]. \[fig:xsec:Et\_3j\]](DESY-07-102_14.eps)
(b)
![ Measured cross section as a function of $E_T^{jet}$ in the four-jet (a) low- and (b) high-mass samples. The jets are ordered such that $E_T^{jet1}>E_T^{jet2}>E_T^{jet3}>E_T^{jet4}$. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:Et\_3j\]. \[fig:xsec:Et\_4j\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $E_T^{jet}$ in the four-jet (a) low- and (b) high-mass samples. The jets are ordered such that $E_T^{jet1}>E_T^{jet2}>E_T^{jet3}>E_T^{jet4}$. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:Et\_3j\]. \[fig:xsec:Et\_4j\]](DESY-07-102_15.eps)
(a)
![ Measured cross section as a function of $E_T^{jet}$ in the four-jet (a) low- and (b) high-mass samples. The jets are ordered such that $E_T^{jet1}>E_T^{jet2}>E_T^{jet3}>E_T^{jet4}$. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:Et\_3j\]. \[fig:xsec:Et\_4j\]](DESY-07-102_16.eps)
(b)
![ Measured cross section as a function of $\eta^{jet}$ in the three-jet (a) low- and (b) high-mass samples. Each cross section has been shifted upwards by the amount given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:Et\_3j\]. \[fig:xsec:eta\_3j\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $\eta^{jet}$ in the three-jet (a) low- and (b) high-mass samples. Each cross section has been shifted upwards by the amount given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:Et\_3j\]. \[fig:xsec:eta\_3j\]](DESY-07-102_17.eps)
(a)
![ Measured cross section as a function of $\eta^{jet}$ in the three-jet (a) low- and (b) high-mass samples. Each cross section has been shifted upwards by the amount given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:xsec:Et\_3j\]. \[fig:xsec:eta\_3j\]](DESY-07-102_18.eps)
(b)
![ Measured cross section as a function of $\eta^{jet}$ in the four-jet (a) low- and (b) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:eta\_3j\]. \[fig:xsec:eta\_4j\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $\eta^{jet}$ in the four-jet (a) low- and (b) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:eta\_3j\]. \[fig:xsec:eta\_4j\]](DESY-07-102_19.eps)
(a)
![ Measured cross section as a function of $\eta^{jet}$ in the four-jet (a) low- and (b) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:eta\_3j\]. \[fig:xsec:eta\_4j\]](DESY-07-102_20.eps)
(b)
![ Measured cross section as a function of $\cos(\psi_{\rm 3})$ in the three-jet (a) low- and (b) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:yJB\]. \[fig:xsec:cp3\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $\cos(\psi_{\rm 3})$ in the three-jet (a) low- and (b) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:yJB\]. \[fig:xsec:cp3\]](DESY-07-102_21.eps)
(a)
![ Measured cross section as a function of $\cos(\psi_{\rm 3})$ in the three-jet (a) low- and (b) high-mass samples. Other details as in the caption to Fig. \[fig:xsec:yJB\]. \[fig:xsec:cp3\]](DESY-07-102_22.eps)
(b)
![ (a) Measured cross section as a function of $M_{3j}$ in the three-jet sample compared with an $\mathcal{O}(\alpha\alpha_s^2)$ prediction, corrected for hadronisation and MPI effects. (b) The hadronisation and MPI corrections as a function of $M_{3j}$. (c) The ratio of the $M_{3j}$ cross section divided by the corrected $\mathcal{O}(\alpha\alpha_s^2)$ prediction. The theoretical uncertainty is represented by the dashed bands. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:tree:Mnj\]](DESY-07-102_0.eps)
![ (a) Measured cross section as a function of $M_{3j}$ in the three-jet sample compared with an $\mathcal{O}(\alpha\alpha_s^2)$ prediction, corrected for hadronisation and MPI effects. (b) The hadronisation and MPI corrections as a function of $M_{3j}$. (c) The ratio of the $M_{3j}$ cross section divided by the corrected $\mathcal{O}(\alpha\alpha_s^2)$ prediction. The theoretical uncertainty is represented by the dashed bands. Other details as in the caption to Fig. \[fig:xsec:Mnj\]. \[fig:tree:Mnj\]](DESY-07-102_23.eps)
(a)
(b)
(c)
![ Measured cross section as a function of $E_T^{jet}$ for each jet in the three-jet (a) low- and (b) high-mass samples compared with $\mathcal{O}(\alpha\alpha_s^2)$ predictions, corrected for both hadronisation and MPI effects. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:Et\_3j\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $E_T^{jet}$ for each jet in the three-jet (a) low- and (b) high-mass samples compared with $\mathcal{O}(\alpha\alpha_s^2)$ predictions, corrected for both hadronisation and MPI effects. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:Et\_3j\]](DESY-07-102_24.eps)
(a)
![ Measured cross section as a function of $E_T^{jet}$ for each jet in the three-jet (a) low- and (b) high-mass samples compared with $\mathcal{O}(\alpha\alpha_s^2)$ predictions, corrected for both hadronisation and MPI effects. Each cross section has been multiplied by $10^n$, where $n$ is given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:Et\_3j\]](DESY-07-102_25.eps)
(b)
![ Measured cross section as a function of $\eta^{jet}$ for each jet in the three-jet (a) low- and (b) high-mass samples compared with $\mathcal{O}(\alpha\alpha_s^2)$ predictions, corrected for both hadronisation and MPI effects. Each cross section has been shifted upwards by the amount given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:eta\_3j\]](DESY-07-102_0.eps)
![ Measured cross section as a function of $\eta^{jet}$ for each jet in the three-jet (a) low- and (b) high-mass samples compared with $\mathcal{O}(\alpha\alpha_s^2)$ predictions, corrected for both hadronisation and MPI effects. Each cross section has been shifted upwards by the amount given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:eta\_3j\]](DESY-07-102_26.eps)
(a)
![ Measured cross section as a function of $\eta^{jet}$ for each jet in the three-jet (a) low- and (b) high-mass samples compared with $\mathcal{O}(\alpha\alpha_s^2)$ predictions, corrected for both hadronisation and MPI effects. Each cross section has been shifted upwards by the amount given on the plot, to aid visibility. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:eta\_3j\]](DESY-07-102_27.eps)
(b)
![ (a) Measured cross section as a function of $\cos(\psi_3)$ in the low-mass three-jet sample. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:cp3\_i\]](DESY-07-102_0.eps)
![ (a) Measured cross section as a function of $\cos(\psi_3)$ in the low-mass three-jet sample. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:cp3\_i\]](DESY-07-102_28.eps)
(a)
(b)
(c)
![ (a) Measured cross section as a function of $\cos(\psi_3)$ in the high-mass three-jet sample. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:cp3\_h\]](DESY-07-102_0.eps)
![ (a) Measured cross section as a function of $\cos(\psi_3)$ in the high-mass three-jet sample. Other details as in the caption to Fig. \[fig:tree:Mnj\]. \[fig:tree:cp3\_h\]](DESY-07-102_29.eps)
(a)
(b)
(c)
[^1]: From now on the word “electron” is used as a generic term for electrons and positrons, unless stated otherwise.
[^2]: The ZEUS coordinate system is a right-handed Cartesian system, with the $Z$ axis pointing in the proton beam direction, referred to as the “forward direction”, and the $X$ axis pointing left towards the centre of HERA. The coordinate origin is at the nominal interaction point.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Author: Peter Wilson'
- 'Maintainer: Will Robertson'
date: '2009/09/02'
title: 'The package[^1]'
---
Usage
=====
In LaTeX, a new counter called, say ‘[*ctr*]{}’, is created by the command [[`\gobble`]{}]{}[`{`[*ctr*]{}`}`]{}[`[`[*within*]{}`]`]{}. If the optional [*within*]{} argument is given the the counter ‘[*ctr*]{}’ is reset to zero each time the counter ‘[*within*]{}’ changes. The command [[`\gobble`]{}]{} typesets the value of the counter [*ctr*]{}. This is automatically defined by [[`\gobble`]{}]{} and is initialised to typeset arabic numerals.
It is sometimes desireable to change a counter that has been defined by [[`\gobble`]{}]{}[`{`[*ctr*]{}`}`]{} to act as though it had been defined as [[`\gobble`]{}]{}[`{`[*ctr*]{}`}`]{}[`[`[*within*]{}`]`]{}. The package provides the command [[`\gobble`]{}]{}[`{`[*ctr*]{}`}`]{}[`{`[*within*]{}`}`]{} that accomplishes this. By default, it also redefines the [[`\gobble`]{}]{} command so that it typesets values in the style `\gobble.\gobble{ctr}`. The starred version of the command suppresses the redefinition of [[`\gobble`]{}]{} (e.g., [[`\gobblewithin*`]{}]{}`{ctr}{within}`).
Likewise, the command [[`\gobble`]{}]{}[`{`[*ctr*]{}`}`]{}[`{`[*within*]{}`}`]{} changes a counter that has been created by [[`\gobble`]{}]{}[`{`[*ctr*]{}`}`]{}[`[`[*within*]{}`]`]{} to act as though it had been created by [[`\gobble`]{}]{}[`{`[*ctr*]{}`}`]{}. By default it also redefines the [[`\gobble`]{}]{} command so that it just typesets an arabic numeral. The starred version of the command suppresses the redefinition of [[`\gobble`]{}]{}.
Any number of [[`\gobble`]{}]{}[`{`[*ctr*]{}`}`]{}[`{`[*…*]{}`}`]{} and [[`\gobbleout`]{}]{}[`{`[*ctr*]{}`}`]{}[`{`[*…*]{}`}`]{} commands can be issued for a given counter, [*ctr*]{}, if you wish to toggle between the two styles. The current value of [*ctr*]{} is unaffected by [[`\gobble`]{}]{} and [[`\gobble`]{}]{}. If you want to change the value after one of these commands, use [[`\gobblecounter`]{}]{}[`{`[*ctr*]{}`}`]{}[`{`[*…*]{}`}`]{}, and to change the typeseting style use [[`\gobblecommand`]{}]{}`{`[[`\gobble`]{}]{}`}`[`{`[*…*]{}`}`]{}.
Change History:
===============
------- ---------------------------------
v1.0a New maintainer (Will Robertson)
------- ---------------------------------
[^1]: This manual corresponds to v1.0a, dated 2009/09/02.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present here a new conjecture for the nature of the Mersenne prime numbers by connecting it with the Collatz-Kakutani problem. By introducing a natural *path length* on the basis of the Collatz-Kakutani tree, we conjecture that this path length of a Mersenne prime from the root of the Collatz-Kakutani tree is approximately proportional to the index of the Mersenne prime. We also discuss difference of behaviors between Mersenne numbers and Mersenne primes.'
author:
- Toru Ohira and Hiroshi Watanabe
title: 'A Conjecture on the Collatz-Kakutani Path Length for the Mersenne Primes'
---
Introduction
============
Prime numbers have attracted mathematically oriented minds. They are fountains of the interesting problems which are left unresolved to the present. As well as prime numbers, natural numbers themselves sometimes show unexpected behaviors in spite of their simple appeal. In this note, we would like to present a simple property arising from the combination of two unsolved problems in number theory; the Mersenne prime numbers and the Collatz-Kakutani conjecture. Namely, we show that there exists an approximate linear relation between an index of the Mersenne prime and its Collatz-Kakutani path length, both of which are defined in the next section.
Mersenne primes and Collatz-Kakutani Conjecuture
================================================
Let us start with a brief descriptions of the Mersenne primes and the Collatz-Kakutani conjecture[@mathlib]. A Mersenne number is a positive integer given by $$M_n = 2^n -1,$$ where $n$ is a positive integer, called “index”. A Mersenne prime is a Mersenne number that is prime. It has been shown that if $M_n$ is a Mersenne prime, $n$ must be prime. However, the converse is not true. In fact, only forty-seven Mersenne primes have been found up to date with the largest one given when $n=43,112,609$. This is also the largest known prime number. There are many fundamental questions left unresolved such as whether there are infinitely many Mersenne primes. The behaviors of Mersenne primes are irregular and unpredictable as seen in other types of prime numbers. Therefore, the recent discoveries of the Mersenne primes were achieved with the help of massively distributed computers, and the effort for finding new Mersenne primes has been made continuously [@gimps]. The Mersenne primes are known for their relationships with the perfect numbers, and they are also applied to create pseudorandom number generators [@mt].
The Collatz-Kakutani conjecture is also one of the unsolved problems in number theory. This conjecture is a halting problem that the following operations will stop for an arbitrary positive number. Consider a positive integer $X$.
If $X$ is odd, $3X+1$ is the next integer.
If $X$ is even, $X/2$ is the next integer.
If we repeat this process, it will eventually reach the halting state $X=1$ for all positive integers. For example, if we start from $X=7$, we have a sequence as
$$7, 22, 11, 34, 17, 52, 26, 13, 40, 20, 10, 5, 16, 8, 4, 2, 1.$$
This process can be visualized as the Collatz-Kakutani tree as shown in Fig. 1A. The conjecture states that this tree covers all the positive integers. Even though this conjecture is unsolved, it is generally believed as true from arguments through probability theories and through computational verifications.
Let us introduce the “Collatz-Kakutani path length” $D(X)$ for a number of steps for $X$ to reach $X=1$ through the above operations, i.e., a number of operational steps needed to reach $1$ (“the root”) on the Collatz-Kakutani tree. For example, $D(7) = 16$. The relationship between $X$ and $D(X)$ is highly irregular and no simple law for the path length is found[^1]. As a result, the plot of $D(X)$ versus $X$ produces a irregular graph as shown in Fig. 1B. While the general relationship between $X$ and $D(X)$ is unknown, there are some trivial relations such as $D(2^n) = n$. Note that, the number $2^n$ is just one larger than the Mersenne number $M_n =2^n - 1$.
![A. Schematic View of Collatz-Kakutani tree. B. Plot of the path length $D(X)$ on the Collatz-Kakutani tree to the root of tree $X=1$ for integers up to X=2000.[]{data-label="tree"}](collatztree.eps){width="0.95\columnwidth"}
Main Results
============
Our main finding to report is the fact that a path length of a Mersenne prime is approximately proportional to its index for large n, namely, $$D(M_n) \approx 13.45 n.$$ This is shown in Fig. 2, which we computed up to the largest known 47th Mersenne prime, ${M_p}(47) = M_{43112609}$. The behavior of the path length $D$ is non–monotonic for small indices, e.g., $D(M_{89}) > D(M_{107})$. We expect, however, beyond $n=107$, the path length increases monotonically with $n$ for the Mersenne primes. Also, there seems no regular path to reach the root of the tree, $X=1$. For example, we have examined the path for two nearby Mersenne primes, ${M_p}(16) = M_{2203}$ and ${M_p}(17) = M_{2281}$. While they both have the path length $D$ of approximately $30,000$, they share the last 16 steps, namely,
$22, 11, 34, 17, 52, 26, 13, 40, 20, 10, 5, 16, 8, 4, 2,$ and $1$.
The relations between indices of the Mersenne primes and their path lengths are summarized in Table \[tbl\_pathlength\].
---- ----------- ----------- ---------
1 2 7 3.5
2 3 16 5.33333
3 5 106 21.2
4 7 46 6.57143
5 13 158 12.1538
6 17 224 13.1765
7 19 177 9.31579
8 31 450 14.5161
9 61 860 14.0984
10 89 1454 16.3371
11 107 1441 13.4673
12 127 1660 13.0709
13 521 6769 12.9923
14 607 8494 13.9934
15 1279 17094 13.3651
16 2203 29821 13.5365
17 2281 30734 13.4739
18 3217 43478 13.5151
19 4253 55906 13.1451
20 4423 60716 13.7273
21 9689 129608 13.3768
22 9941 134345 13.5142
23 11213 153505 13.6899
24 19937 265860 13.335
25 21701 293161 13.5091
26 23209 312164 13.4501
27 44497 598067 13.4406
28 86243 1158876 13.4373
29 110503 1482529 13.4162
30 132049 1771117 13.4126
31 216091 2906179 13.4489
32 756839 10197081 13.4732
33 859433 11568589 13.4607
34 1257787 16927967 13.4585
35 1398269 18807193 13.4503
36 2976221 40055567 13.4585
37 3021377 40663017 13.4584
38 6972593 93778449 13.4496
39 13466917 181209792 13.4559
40 20996011 282515044 13.4557
41 24036583 323346876 13.4523
42 25964951 349304386 13.4529
43 30402457 409093991 13.456
44 32582657 438465334 13.457
45 371566673 499902411 13.4539
46 42643801 573966881 13.4596
47 43112609 580260946 13.4592
---- ----------- ----------- ---------
: Path lengths of the Mersenne primes. []{data-label="tbl_pathlength"}
This linear behavior can be understood by the following heuristic arguments. Suppose $N$ is a large random number. The standard heuristic arguments for the path length for $N$ gives $$D(N) \approx (3/(\ln 4/3)) \ln N.$$ A Mersenne number $2^n - 1$ becomes $3 \times {2^{n-1}} - 1$ after two steps in Collatz-Katutani tree. Therefore, A Mersenne Number $2^n - 1$ becomes $3^n-1$ after $2n$ steps. If we consider the number $N = 3^n-1 \sim 3^n$ to be a large random number for large $n$, then the path length of the number is given by $$D(3^n-1) \approx (3/(\ln 4/3)) n \ln 3.$$ Finally, we obtain the heuristic estimation of the path length of the Mersenne number to be $$D(2^n -1) \approx 2n + (3/(\ln 4/3)) n \ln 3 \approx 13.45652 n,$$ which is consistent with our numerical results [@referee1; @lagarias; @kontorovich]. This arguments implies that our finding is true not only for the Mersenne primes, but also for the Mersenne numbers. Our main conjecture, however, is that this linear relationship for the Mersenne primes are better than that for the general Mersenne numbers, or other general sequence of integers.
In order to show the difference between the Mersenne numbers in general and the Mersenne primes, the ratio of the path length to the index, $D(M_n)/n$, is plotted against $n$ in Fig. 2B. If the distance increases linearly to the index, then $D(M_n)/n$ becomes a constant. Figure 2B shows that the Mersenne primes show better linearity than the general Mersenne numbers.
![A. Plot of $D(M_n)$ for Mersenne primes with index $n$ up to around ${M_p(47)}$. B. Plot of $D(M_n)/n$ for the Mersenne primes (Blue) and the Mersenne numbers with randomly chosen index (Red) up to around ${M_p(31)}$.[]{data-label="linear"}](collatzlinear47.eps){width="1.0\columnwidth"}
To be more quantitative, we have performed the following statistical analyses. We have chosen thirteen Mersenne primes from the $26^{\text{th}}$ (${M_p}(26) = M_{23, 209}$) to the $38^{\text{th}}$ (${M_p}(38) = M_{6, 972, 593}$), and computed the mean and the variance of $D(M_n)/n$. Let us call the prime indices of the Mersenne primes as the Mersenne prime indices. We compare the result with that from the following set of thirteen points approximately in the same interval.
A\) Thirteen prime indices which is the next smallest to thirteen Mersenne prime indices. We note the Mersenne numbers associated with this set is not prime numbers themselves.
B\) Thirteen indices which are near the mid-point of two successive Mersenne prime indices.
C\) Thirteen indices which are twice the Mersenne prime indices. These are all even numbers. We have selected some which are close to the Mersenne prime indices.
D\) Thirteen indices which is on the least squre best-fit line based on the heuristics that, given the $k$-th Mersenne prime ${M_p}(k)$, the plot of $\log_{2} (\log_{2} ( {M_p}(k) )$ versus $k$ lies approximately on the straight line (Figure 3). We have selected some which are close to the Mersenne prime indices [@schroeder].
![Plot of $\log_{2} ( \log_{2} ( {M_p}(k) )$ versus $k$ for all the known Mersenne prime up to ${M_p(47)}$. The straight line is obtained by the least square error fit and $\sim 0.92757 + 0.55715 k $.[]{data-label="shroeder"}](shroeder.eps){width="0.5\columnwidth"}
The results of these comparison are summarized in Table \[tbl\_statistics\], showing that the path lengths of the Mersenne primes have the smaller variance than other sets. However, this difference is not enough to single out a Mersenne prime index.
In spite of the above, we look more detail into this property. We have examined the plot of $D(X=2^n-1)$ versus $n$ more closely and noted that they are composed of a collection of “flat” regions with jumps like a staircase. Correspondingly, the plot of ${D(X)/n}$ versus $n$ becomes collection of stripes, which interestingly appears near parallel. They are plotted in Figure 4. These plots are made for the sampled prime indices near three Mersenne prime indices, $23209, 110503, 216091$ which give $26^{\text{th}}$, $29^{\text{th}}$, and $31^{\text{th}}$ Mersenne primes respectively. It is also interesting that there are occasional “overlaps” of “stair steps”.
---------------- ---------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------- --------- -------------
Mersenne Prime 23209, 44497, 86243, 110503, 132049, 216091, 756839, 859433, 1257787, 1398269, 2976221, 3021377, 6972593 312164, 598067, 1158876, 1482529, 1771117, 906179, 10197081, 11568589, 16927967, 18807193, 40055567, 40663017, 93778449 13.4473 0.0002977
A 23227, 44501, 86249, 110527, 132059, 216103, 756853, 859447, 1257827, 1398281, 2976229, 3021407, 6972607 312182, 598071, 1158882, 1482553, 1771127, 2906191, 10197095, 11568603, 16928007, 18807205, 40055575, 40663047, 93778463 13.4460 0.0003194
B 22455, 33853, 65370, 98373, 121276, 174070, 486465, 808136, 1058610, 1328028, 2187245, 2998799, 4996985 299801, 457438, 875438, 1327329, 1633743, 2344640, 6524449, 10868120, 14246657, 17876449, 29428265, 40364153, 67195624 13.4485 0.0017853
C 22426, 43402, 46418, 88994, 172486, 221006, 264098, 432182, 1513678, 1718866, 2515574, 2796538, 5952442 299772, 584422, 627877, 1201650, 2320161, 2974984, 3556035, 5828307, 20384499, 23124964, 33827530, 37632788, 80085173 13.4618 0.00132591
D 20160, 43644, 64216, 94484, 139021, 204550, 442830, 651562, 958682, 1410567, 2075452, 3053739, 4493150 270868, 587280, 866299, 1265873, 1871202, 2748585, 5947053, 8769774, 12911249, 18991590, 27939124, 41095221, 60441877 13.4515 0.000502943
---------------- ---------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------- --------- -------------
: Statistical analyses on path lengths of Mersenne primes and Mersenne numbers. E\[$x$\] denotes the arithmetic average of $x$ and Var\[$x$\] denotes the variance of $x$, respectively. []{data-label="tbl_statistics"}
![Plot of $D(M_n)$ (top) and $D(M_n)/n$ (bottom) versus $n$. The Mersenne primes are marked by arrows, and horizontal line is at $D(M_n)/n = 13.45$. A. Prime indices near $n = 23209$ (${M_p(26)})$, B. Every fifth prime indices near $n = 110503$ (${M_p(29)}$), C. Every fifth prime indices near $n = 216091$ (${M_p(31)})$.[]{data-label="fig4"}](Fig4b.eps){width="1.0\columnwidth"}
The nature and mechanism of these properties appeared in Figure 4 need to be studied further. We, nevertheless, conjecture that these qualitative behavior will continue to larger Mersenne numbers, and the Mersenne prime appears with its ${D(X)/n} \approx 13.45$. Unfortunately, again, this is clearly not sufficient to select out the Mersenne primes. If this conjecture is true, however, we can heuristically rule out points that are further away from the horizontal line of ${D(X)/n} \approx 13.45$ from candidates of the Mersenne primes.
Discussion
==========
We have used the Collatz-Kakutani tree as a provider of “length” for natural numbers, and obtained a conjecture for the Mersenne primes. This may be yet another example where prime numbers can give rise to an emergence of unexpected order, such as Ulam Spiral [@ulam64; @ulam67].
Though the reason behind such simple behavior is left unclear, this approach of using the Collatz-Kakutani tree can be extended to other types of set of numbers known for irregularities, possibly leading to interesting insights[@sinyor].
#### Acknowledgments
Authors would like to thank Dr. Masaru Suzuki, Dr. Atsushi Kamimura, and Shigenori Matsumoto for useful discussions. T.O. would like to thank Emeritus Prof. Yuji Ito of Keio University, for his account on his advisor, late Prof. Shizuo Kakutani of Yale University.
[9]{} The Math Forum Internet Mathematics Library, available at <http://mathforum.org/library/>.
Great Internet Mersenne Prime Search:GIMPS, available at <http://www.mersenne.org/default.php/>.
M. Matsumoto, T. Nishimura, Mersenne twister: A 623-dimensionally equidistributed uniform pseudorandom number generator, *ACM Trans. on Modeling and Computer Simulations* **8** (1998) 3.
This heuristic arguments of this paragraph is provided by one of the anonymous reviewers.
J. C. Lagarias, The $3x+1$ Problem and Its Generalizations, *Amer. Math. Monthly* **92** (1985) 3
A. V. Kontorovich, J. C. Lagarias, Stochastic Models for the $3x+1$ and $5x+1$ Problems, arXive:09101944 \[math.NT\] (2009)
M. R. Schroeder, Where Is the Next Mersenne Prime Hiding? *Mathematical Intelligencer* **5** (1983) 31
M. L. Stein, S. M. Ulam, M. B. Wells, A Visual Display of Some Properties of the Distribution of Primes, *Amer. Math. Monthly* **71** (1964) 516 M. L. Stein, S. M. Ulam, An Observation on the Distribution of Primes, *Amer. Math. Monthly* **74** (1967) 43. After the initial submission to the arXive, we are pointed to the following work by the author. J. Sinyor, The $3x+1$ Problem as a String Rewriting System, *Int. J. of Math. and Math. Science* **2010** (2010) 458563
*Graduate School of Mathematics, Nagoya University, Nagoya, Japan.\
[email protected]*
*The Institute of Solid State Physics, The University of Tokyo, Kashiwa, Japan.*
[^1]: In fact, it may not be defined for some $X$ if the Collatz-Kakutani conjecture is false
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Nathan R. Finney$^{1*}$'
- 'Matthew Yankowitz$^{2*}$'
- 'Lithurshanaa Muraleetharan$^{1}$'
- 'K. Watanabe$^{3}$'
- 'T. Taniguchi$^{3}$'
- 'Cory R. Dean$^{2\dagger}$'
- 'James Hone$^{1\dagger}$'
bibliography:
- 'references.bib'
title: 'Tunable crystal symmetry in graphene–boron nitride heterostructures with coexisting moiré superlattices'
---
**In heterostructures consisting of atomically thin crystals layered on top of one another, lattice mismatch or rotation between the layers results in long-wavelength moiré superlattices. These moiré patterns can drive significant band structure reconstruction of the composite material, leading to a wide range of emergent phenomena including superconductivity [@Cao2018b; @Yankowitz2019; @Chen2019], magnetism [@Sharpe2019], fractional Chern insulating states [@Spanton2018], and moiré excitons [@Seyler2019; @Tran2019; @Jin2019; @Alexeev2019]. Here, we investigate monolayer graphene encapsulated between two crystals of boron nitride (BN), where the rotational alignment between all three components can be varied. We find that band gaps in the graphene arising from perfect rotational alignment with both BN layers can be modified substantially depending on whether the relative orientation of the two BN layers is 0 or 60 degrees, suggesting a tunable transition between the absence or presence of inversion symmetry in the heterostructure. Small deviations ($<1^{\circ}$) from perfect alignment of all three layers leads to coexisting long-wavelength moiré potentials, resulting in a highly reconstructed graphene band structure featuring multiple secondary Dirac points. Our results demonstrate that the interplay between multiple moiré patterns can be utilized to controllably modify the electronic properties of the composite heterostructure.**
The ability to combine diverse vdW materials into a heterostructure enables engineering of new properties not observed in the constituent materials alone. A unique degree of freedom within these vdW heterostructures is the twist angle between layers, and changing this angle can strongly modify the material properties owing to the formation of moiré patterns. In graphene–BN heterostructures, the moiré pattern introduces a spatially-periodic effective potential that modifies the graphene band structure, giving rise to emergent secondary Dirac points (SDPs) at finite energy [@Yankowitz2012; @Ponomarenko2013; @Dean2013; @Hunt2013] and band gaps at the charge neutrality point and valence band SDP [@Hunt2013; @Woods2014; @Chen2014; @Wang2015; @Wang2016; @Song2013; @Bokdam2014; @Moon2014; @Wallbank2015; @Jung2015; @SanJose2014; @Slotman2015; @Jung2017; @Yankowitz2018]. In twisted bilayer graphene (tBLG), correlated insulating states and superconductivity emerge at a twist angle of $\sim$1.1$^{\circ}$ where the lowest energy moiré bands become exceptionally flat [@Cao2018a; @Cao2018b; @Yankowitz2019; @Sharpe2019; @Lu2019]. However, typical vdW heterostructures comprising many flakes possess numerous crystal interfaces, and in principle multiple long-wavelength moiré patterns may coexist within a single heterostructure, likely with profound consequences on moiré-driven physics. For example, topological bands have been shown to potentially arise in tBLG aligned to BN [@Sharpe2019]. So far, little has been done to controllably tune the alignment of multiple pairs of crystals within a single device, and it is not well understood how multiple moiré patterns interact to influence the properties of the vdW heterostructure.
{width="3.5"}
In a heterostructure where graphene is encapsulated on both sides by BN, there are a number of qualitatively distinct stacking orders that can be realized by independently controlling the twist angle of the graphene relative to the bottom BN, $\theta_b$, and the top BN, $\theta_t$ (angles illustrated schematically in Fig. \[fig:1\]a). In this work, we fabricate devices in which the graphene is aligned to the bottom BN ($\theta_b=0^{\circ}\pm0.15^{\circ})$, but $\theta_t$ can be tuned to arbitrary angle. In this structure there are two unique positions in which the top BN layer may be “aligned” to the graphene — $\theta_t=0^{\circ}$ and $60^{\circ}$ — which have distinct symmetry due to the inequivalence of the boron and nitrogen triangular sublattices of the BN unit cell. Additionally, we study the case of small misalignment of the top layer (*i.e.* small but non-zero values of $\theta_t$) such that the two interfaces give rise to separate and incommensurate moiré patterns but with similar wavelengths.
{width="6.9"}
As illustrated in Fig. \[fig:1\]b, we utilize a dynamically rotatable heterostructure to vary $\theta_t$ while maintaining fixed $\theta_b$. Graphene is first aligned to the bottom BN layer and shaped into a Hall bar geometry. The rotational alignment of the graphene and BN is initially determined by Raman spectroscopy prior to making electrical contact, and subsequently confirmed precisely by measuring the charge carrier density of the SDPs in electrical transport measurements (see Methods). The top BN is separately patterned into a circular shape with rectangular “handles," and is subsequently transferred onto the graphene Hall bar such that it covers the entire active area of the channel (Fig. \[fig:1\]b). The top BN can be mechanically rotated to any angle ($0^{\circ}<\theta_{t}<360^{\circ}$) using an AFM tip, and electrical transport measurements can be performed simultaneously [@Chari2016; @Ribeiro2018]. We are able to measure changes in the twist angle of the top BN ($\Delta\theta_t$) to better than 0.1$^{\circ}$ from AFM topographs.
Fig. \[fig:1\]c shows spectra of the graphene 2D Raman peak at different values of $\theta_t$. In Fig. \[fig:1\]d we plot the full width at half maximum of the 2D peak (FWHM$_{2D}$) versus $\theta_t$ over a 60$^{\circ}$ range. Over most of this range the FWHM$_{2D}$ exhibits a constant value that is $\sim$20 cm$^{-1}$ larger than isolated graphene. This behavior is consistent with the presence of a single long-wavelength moiré pattern resulting from the fixed zero-angle alignment between the graphene and bottom BN [@Eckmann2013; @Ribeiro2018]. When $\theta_t$ approaches $0^{\circ}$ or $60^{\circ}$, the linewidth is broadened by an additional $\sim$20 cm$^{-1}$, with FWHM$_{2D}$ near 55 cm$^{-1}$. We interpret this additional broadening to result from rotational alignment of the top BN layer to the graphene/bottom BN, and therefore provides an absolute measure of the top BN layer orientation. In previous studies, the broadening of the 2D mode was understood to arise from moiré-scale relaxations of the graphene lattice [@Eckmann2013]. Our observation of an approximate doubling of this broadening indicates that the graphene couples similarly to the moiré patterns from both the top and bottom encapsulating BN layers.
While the 2D Raman peak is 60$^{\circ}$ periodic, the BN–graphene–BN trilayer lattice structure is not precisely equivalent upon 60$^{\circ}$ rotations of the top BN crystal. As illustrated schematically in Figs. \[fig:2\]a-b, for $\theta_t=0^{\circ}$ the moiré pattern is three-fold rotationally symmetric, whereas for $\theta_t=60^{\circ}$ the moiré pattern is six-fold rotationally symmetric. The lattice structures at the high-symmetry points of the moiré patterns (outer schematics in Figs. \[fig:2\]a-b) highlight the important difference between these two cases. For $\theta_t=0^{\circ} $ the top layer B (N) atoms sit atop the bottom layer B (N) atoms and the overall structure breaks inversion symmetry. In contrast, for $\theta_t=60^{\circ}$ the top layer B (N) atoms sit atop N (B) atoms and the structure hosts inversion symmetry. The nature of inversion symmetry in graphene–BN heterostructures has previously been tied to band structure reconstruction of the graphene [@Hunt2013; @Wang2015], suggesting that the electronic properties of our devices may vary substantially with $\theta_t$.
We first investigate room temperature electrical transport of our device, and in particular observe notable differences between $\theta_t=0^{\circ}$ and 60$^{\circ}$. Figs. \[fig:2\]c-f show the device resistivity, $\rho$, as a function of charge carrier density, $n$, for various values of $\theta_t$. When the top BN is far from alignment (Fig. 2e), we observe both a sharp resistance peak at the primary Dirac point (PDP) and broad resistance peaks at finite density corresponding to the moiré-induced SDPs. We note that the resistivity of the hole-band SDP is larger than the electron-band SDP but smaller than the PDP, consistent with previous room temperature studies of graphene aligned to a single BN layer [@Wang2015]. As $\theta_t$ approaches zero (Figs. \[fig:2\]c-d), both the PDP and SDP grow, but their relationship inverts with $\rho_{SDP}$ exceeding $\rho_{PDP}$ very near $\theta_{t}=0^{\circ}$. By contrast, near $\theta_{t}=60^{\circ}$ (Fig. \[fig:2\]f), the SDP again grows, but less so, and the PDP appears slightly suppressed. The angle dependence of the PDP and hole-band SDP peaks are shown in more detail in Figs. \[fig:2\]g-h, where the peak resistance values are plotted over a full $360^{\circ}$ rotation. In particular, we note that both the PDP and hole-band SDP are maximal at the angle we label $\theta_{t}=0^{\circ}$. While translation of the top BN with respect to the bottom would change the overall stacking configuration, we observe nearly equivalent transport every $\Delta\theta_t=120^{\circ}$. This suggests translation does not play a significant role, and that the stacking configuration with no translational offset is the structural ground state at the aligned positions.
In graphene aligned to a single BN layer, the staggered sublattice potential of the BN breaks inversion symmetry in the graphene layer for both 0$^{\circ}$ and 60$^{\circ}$ “aligned” orientations, resulting in a band gap at the PDP whose value is expected to scale with the magnitude of effective superlattice potential [@Hunt2013; @Woods2014; @Chen2014; @Wang2015; @Wang2016; @Song2013; @Bokdam2014; @Moon2014; @Wallbank2015; @Jung2015; @SanJose2014; @Slotman2015; @Jung2017; @Yankowitz2018]. We conjecture that the asymmetry between $\theta_{t}=0^\circ$ and 60$^\circ$ in our device correlates with the transition between the broken inversion symmetry structure at $\theta_{t}=\theta_{b}=0^\circ$ (Fig. \[fig:2\]a) — in which the PDP gap is likely to be largest — and the inversion symmetric structure at $\theta_{t}=60^\circ$ (Fig. \[fig:2\]b) — in which no PDP gap would be expected within a single-particle model.
{width="5.5"}
Figs. \[fig:3\]a-d compare the temperature dependence of the PDP and SDP at $\theta_t=0^{\circ}$ and 60$^{\circ}$. Strongly insulating behavior is observed in both at $\theta_t=0^{\circ}$, whereas only weakly insulating behavior is observed for $\theta_t=60^{\circ}$. A linear fit to the thermally activated regime (black dashed lines in Figs. \[fig:3\]e-f) gives a measure of the activation gap, $\Delta$. In Fig. \[fig:3\]g we plot the gaps for 5 different values of $\theta_t$ (square markers). Both the PDP and SDP gaps are notably enhanced at $\theta_t=0^{\circ}$, whereas at $\theta_t=60^{\circ}$ both are notably reduced. The gaps have little dependence on $\theta_t$ at all other angles. Our observation of $\Delta_{PDP}>60$ meV is so far the largest gap observed in a pristine graphene device, and may potentially be significantly further enhanced under pressure [@Yankowitz2018]. Although the gaps extracted from an Arrhenius fit are not identically zero at $\theta_t=60^{\circ}$, the device only exhibits activated behavior over well less than a decade change in conductance, hence we expect our reported gaps to be an upper bound at this angle.
The band structure modification anticipated from symmetry considerations is illustrated schematically in Fig. \[fig:3\]h, where we anticipate the largest gaps for $\theta_t=\theta_b=0^{\circ}$ owing to the doubled moiré potential, while at $\theta_t=60^{\circ}$ inversion symmetry protects the Dirac crossings. Following this simple expectation, the measured gaps corroborate our previous assignments of $\theta_t=0^{\circ}$ and $60^{\circ}$. We additionally measure the band gaps in two “stationary” devices (*i.e.* without the ability to dynamically rotate the top BN, see Supplementary Section 2), which also exhibit large broadening of the FWHM$_{2D}$ Raman peak. The gaps are similarly either enhanced or suppressed (diamond and triangle markers in Fig. \[fig:3\]c), suggesting this effect is generic for samples in which both BN layers are aligned to graphene.
{width="6.3"}
Finally, we investigate the small angle limit ($0^{\circ}<\theta_t<1^{\circ}$) where the top and bottom BN yield moiré patterns with only slightly different period. Fig. \[fig:4\]b shows the hole-doped transport for three values of $\theta_t\neq0^{\circ}$ in this regime. In addition to a peak in the resistivity at the PDP, we observe two sizable peaks at finite hole-doped densities denoted $n_{SDP,1}$ and $n_{SDP,2}$. $n_{SDP,1}$ is independent of $\theta_t$ and corresponds to the SDP arising from the moiré potential of the perfectly aligned graphene and bottom BN ($\theta_b\approx0^{\circ}$), while the resistance peaks at $n_{SDP,2}$ correspond to the moiré potential from the top BN. The position of $n_{SDP,2}$ as a function of $\theta_t$ is shown the inset of Fig. \[fig:4\]b, and is in good quantitative agreement with the anticipated dispersion (see Supplementary Section 4) [@Yankowitz2012]. We note that in the small angle limit, thermally activated behavior is observed at both $n_{SDP,1}$ and $n_{SDP,2}$, suggesting band gaps associated with each SDP with typical values of $\sim$20 meV and $\sim$5 meV, respectively (see Supplementary Fig. 5). This implies the reconstructed band structure shown schematically in Fig. \[fig:4\]c in which an isolated narrow band emerges between the two superlattice gaps $\Delta_{SDP,1}$ and $\Delta_{SDP,2}$. Further theoretical and experimental effort is necessary to fully explore the consequence of this band reconstruction, however the appearance of a flat band whose width varies with rotation angle provides the intriguing possibility of hosting tunable correlated states at low temperature [@Cao2018a; @Chen2019].
In a magnetic field, we observe sequences of quantum oscillations emerging from the Dirac points at $n=0$, $n_{SDP,1}$, and $n_{SDP,2}$ (Fig. \[fig:4\]d-e, for $\theta_t=59.4^{\circ}$). We further observe weak signatures of resistive states adjacent to each of the three Dirac points, marked by orange arrows. To identify these resistive states more clearly, we plot $d\rho/dT$ versus $T$ and $n$ in Fig. \[fig:4\]f. In addition to the insulating states previously discussed at the PDP and the two SDPs, we observe a negative temperature dependence at the same densities marked by orange arrows in Fig. \[fig:4\]d, suggesting the presence of new insulating-like states. The two coexisting moiré patterns may in principle interfere to produce a second-order moiré pattern with a very long period (Fig. \[fig:4\]a), inducing an additional resistive state at low density [@Wang2019]. However, the position of the low density resistive feature does not agree precisely with quantitative theoretical estimates for all measured $\theta_t$ (see Supplementary Section 6). The disagreement may arise due to unexpected structural reconstructions of the second order moiré pattern, however at this point we are unable to understand these new resistive states in detail.
In conclusion, we demonstrate the ability to induce and control multiple moiré patterns within a BN–graphene–BN heterostructure. In particular, we are able to dynamically tune the crystal symmetry of the composite material by realizing distinct stacking configurations of the three layers, and further induce coexisting moiré patterns which combine to strongly modify the graphene band structure. Our techniques for engineering multiple moiré patterns within a single vdW heterostructure are quite general and can easily be extended to a wide variety of other device structures, motivating a new class of experiments in which the twist angle of multiple crystal interfaces can be tuned to realize novel material properties.
Methods {#methods .unnumbered}
=======
All heterostructures are assembled using standard dry-transfer techniques with a poly-propylene carbonate (PPC) film on a polydimethyl siloxane (PDMS) stamp [@Wang2013], and rest atop a Si/SiO$_2$ substrate. The device fabrication of the rotatable devices largely follows the techniques developed in Refs. [@Chari2016; @Ribeiro2018]. Device R1 consists of a graphene Hall bar on a $\sim$44 nm thick BN resting on a $\sim$11 nm thick graphite gate. A $\sim$56 nm thick BN flake capped by $\sim$40 nm of hydrogen silsesquioxane (HSQ) is subsequently transferred onto the graphene Hall bar. The HSQ cap acts as the etch mask to shape the rotating BN, and also provides additional durability during AFM pushes. “Stationary” devices consist of a graphene Hall bar fully encapsulated by BN, all atop a graphite gate. Electrical contact to all devices is made by Cr/Au (2 nm/100 nm). Supplementary Sections 1-2 and Supplementary Figs. 1-2 provide full details of the device fabrication.
Raman spectroscopy measurements are performed at room temperature in air. All measurements are acquired using a 532 nm wavelength laser with a power $<$ 1 mW. Measurements to extract the graphene/BN twist angle in stationary devices are performed before the addition of the graphite back gate. In the rotatable device, the encapsulated graphene region sits atop the graphite gate. To isolate the Raman response from the graphene, we separately acquire a Raman spectrum from a nearby region of the graphite gate without the graphene and subtract this background response. We extract FWHM$_{2D}$ from a Lorentzian fit (see Supplementary Section 3 and Supplementary Fig. 3 for full details).
Transport measurements are conducted in a four-terminal geometry with ac current excitation of 10-100 nA using standard lock-in technique at 17.7 Hz. The graphene contact regions (which extend beyond the graphite bottom gate) are gated to a high carrier density by applying a bias to the silicon substrate (typically 5-50 V for SiO$_2$ thickness of $\sim$285 nm) to reduce the contact resistance. We extract $n(V_G)$ by fitting the dispersion of the quantum Hall states in high magnetic field as $n=\nu eB/h$, where $\nu$ is the filling factor, $h$ is Planck’s constant, and $e$ is the elementary charge. The moiré wavelength, $\lambda$, is calculated using the geometric relation $\lambda^2 = 8/(n_{SDP}\sqrt{3})$ where n$_{SDP}$ is the density at full filling of the moiré unit cell.
acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank Rebeca Ribeiro-Palau, Changjian Zhang, and Shaowen Chen for technical support, as well as Jeil Jung, Mikito Koshino and Chris Marianetti for helpful discussions. This work was primarily supported by the NSF MRSEC program through Columbia in the Center for Precision Assembly of Superstratic and Superatomic Solids (DMR-1420634). Sample device design and fabrication was partially supported by DoE Pro-QM EFRC (DE-SC0019443). N.F. acknowledges support from the Stewardship Science Graduate Fellowship program provided under cooperative agreement number DE-NA0002135. CRD acknowledges the support of the David and Lucile Packard Foundation. K.W. and T.T. acknowledge support from the Elemental Strategy Initiative conducted by the MEXT, Japan and the CREST (JPMJCR15F3), JST.
Author contributions {#author-contributions .unnumbered}
====================
N.R.F. and L.M. fabricated the devices. N.R.F. and M.Y. performed the measurements and analyzed the data. K.W. and T.T. grew the hBN crystals. C.R.D. and J.H. advised on the experiments. The manuscript was written with input from all authors.
Competing interests {#competing-interests .unnumbered}
===================
The authors declare no competing interests.
Supplementary Information {#supplementary-information .unnumbered}
=========================
{width="7.0"}
{width="7.0"}
Fabrication of rotatable devices
--------------------------------
To fabricate rotatable devices R1 and R2, we begin by sequentially picking up flakes of BN, graphite, BN, and graphene using a PPC/PDMS stamp. The PPC film is then mechanically removed from the PDMS stamp and placed onto a Si/SiO$_2$ substrate, such that the graphene layer is exposed at the top of the heterostructure. The underlying PPC is removed by vacuum annealing at $T=350^{\circ}$ C. In all processed devices the silicon is used as a global gate for the graphene contacts, and the graphite is used as a local gate for the graphene channel (Fig. \[fig:S1\]a-b). The use of a graphite gates has previously been demonstrated to significantly improve the charge homogeneity of graphene devices [@Zibrov2017].
Devices with good rotational alignment of the graphene and BN are next identified using Raman spectroscopy [@Eckmann2013; @Ribeiro2018]. Devices in which the graphene and BN are well aligned are then shaped into a Hall bar geometry using an oxygen plasma etch through a PMMA mask defined by standard electron-beam lithography (Fig. \[fig:S1\]c). In order to realize more robust electrical contacts, the BN surrounding the graphene contacts is further etched using a CHF$_3$/O$_2$ plasma through a PMMA mask (Fig. \[fig:S1\]d). All electrical contacts in device R1 show no significant change over 8 thermal cycles ($T$ = 300 K - 1.5 K), while many of the contacts in device R2 (in which no BN etch is performed) became open after thermal cycling.
We exfoliate BN onto a separate Si/SiO$_2$ substrate and etch an array of BN “rotators” using a protective HSQ mask (Figs. \[fig:S1\]f-g). The rotators are picked up using a separate PPC/PDMS stamp (Fig. \[fig:S1\]h) and transferred onto the graphene Hall bar (Fig. \[fig:S1\]i). A single rotator is aligned as desired onto the graphene Hall bar by positioning with an AFM tip in contact mode. The device is annealed a second time before deposition of Cr/Au (2 nm/100 nm) metal electrodes (Fig. \[fig:S1\]j). The graphene channel is further cleaned by sequentially pushing a BN rotator across the surface to collect and remove interfacial contamination.
Fabrication of stationary devices
---------------------------------
To fabricate stationary devices S1-S4, we begin by identifying flakes of BN which have fractured into two pieces during the mechanical exfoliation procedure (Fig. \[fig:S2\]a). These fractured BN pieces share common edge profiles which can be easily optically aligned. One piece of the fractured BN is first picked up by a PPC/PDMS stamp (Fig. \[fig:S2\]b-c), leaving the remaining piece on the Si/SiO$_2$ wafer. A flake of graphene is then added to the heterostructure, and is chosen and aligned to fit completely within the BN area (Fig. \[fig:S2\]d). Next, the two BN flakes are rotationally aligned optically (with precision better than 0.5$^{\circ}$) and the remaining BN is added to the heterostructure (Fig. \[fig:S2\]e). The heterostructure is transferred to a Si/SiO$_2$ substrate and annealed at $T=350^{\circ}$ C. No effort is made to align crystal edges of graphene to the BN, however, because the graphene rests entirely within both BN flakes it tends to rotate to alignment with the BN during the transfer or annealing [@Wang2015]. Devices with good rotational alignment of the graphene and BN are identified with Raman spectroscopy. A new PPC/PDMS stamp is then used to pick up the heterostructure, followed by a graphite flake (Fig. \[fig:S2\]f), and the heterostructure is transferred to a separate Si/SiO$_2$ substrate. The device is then shaped into a Hall bar geometry and electrically contacted following the procedures described for devices R1 and R2 (Figs. \[fig:S2\]g-h).
{width="6.5"}
Analysis of Raman spectroscopy
------------------------------
Fig. \[fig:S3\]a plots Raman spectra near the 2D graphene peak in device R1. The measured signal (purple curve) features contributions from both the graphene channel and the graphite back gate. To isolate the graphene response (black curve), we separately acquire a Raman spectrum on the graphite gate alone (green curve) and subtract that contribution. The insets show the position of the laser (green dot) during acquisitions of the purple and green curves. We fit the graphene 2D peak with a Lorentzian (red dashed curve) to extract FWHM$_{2D}$. The reported values of FWHM$_{2D}$ are averaged across a few (3-6) measured background spots. Fig. \[fig:S3\]b shows the 2D Raman peak for stationary BN encapsulated graphene devices in three distinct stacking configurations. These spectra are acquired prior to the addition of graphite back gates.
Fig. \[fig:S3\]c plots the graphene FWHM$_{2D}$ at several alignment configurations for two different rotatable devices. The open triangle markers are taken from Ref. [@Ribeiro2018], in which the graphene is misaligned from the bottom BN. The closed circle markers reproduce the data from device R1 shown in Fig. \[fig:1\]d of the main text. The dashed lines show FWHM$_{2D}$ of stationary devices S1 (graphene aligned to both BN layers), S3 (graphene aligned to a single BN layer), and S4 (graphene not aligned to either BN layer).
{width="7.0"}
[|C[75pt]{}|C[75pt]{}|C[80pt]{}|C[80pt]{}|C[80pt]{}|C[45pt]{}|C[45pt]{} N|]{} $\theta_t$ & n$_{SDP}$ (10$^{12}$ cm$^{-2}$) (doubly-aligned) & n$_{SDP,1}$ (10$^{12}$ cm$^{-2}$) & n$_{SDP,2}$ (10$^{12}$ cm$^{-2}$) & $\lambda$ (nm) (doubly-aligned)& $\lambda_1$ (nm)& $\lambda_2$ (nm)&\
\[10pt\] unencapsulated & n/a & 2.05 & n/a & n/a & 15.0 & n/a &\
-0.7$^{\circ}$ & n/a & 2.05 & 3.22 & n/a & 15.0 & 12.0 &\
0$^{\circ}$ & 2.00 & n/a & n/a & 15.2 & n/a & n/a &\
9.9$^{\circ}$ & n/a & 2.05 & n/a & n/a & 15.0 & n/a &\
59.1$^{\circ}$ & n/a & 2.05 & 3.00 & n/a & 15.0 & 12.4 &\
59.4$^{\circ}$ & n/a & 2.05 & 3.72 & n/a & 15.0 & 11.1 &\
60$^{\circ}$ & 2.00 & n/a & n/a & 15.2 & n/a & n/a &\
Determination of moiré wavelength and zero top BN angle assignment
------------------------------------------------------------------
We extract the moiré wavelength in our devices via the geometric relation $\lambda^2 = 8/(n_{SDP}\sqrt{3})$. We accurately determine $n_{SDP}$ by projecting the quantum oscillations from the SDP Landau fan to $B=0$. Table \[table:table1\] lists $n_{SDP}$ and $\lambda$ for several values of $\theta_t$ in device R1. We additionally measure $n_{SDP}$ in a region of the graphene Hall bar which is not encapsulated by the top BN rotator, such that we are sensitive only to $\theta_b$, and extract a corresponding $\lambda$ = 15.0 nm. This implies nearly perfect alignment of the graphene and bottom BN. The maximum uncertainty in moiré wavelength calculated at each configuration is estimated to be $\pm0.1$ nm.
We observe identical $n_{SDP}$ for all $\theta_t$ except at the angles labeled 0$^{\circ}$ and 60$^{\circ}$, in which $n_{SDP}$ becomes marginally smaller. We take the smallest measured value of $n_{SDP}$ to correspond to perfect alignment of the graphene and top BN. Using $\theta_t=0^{\circ}$ and the calculated $\lambda$ = 15.2 nm, we use the geometric relation derived in Ref. [@Yankowitz2012] to extract the lattice mismatch between the graphene and BN to be $\delta\approx$ 1.65%. We then estimate the misalignment in $\theta_b$ to be less than $\pm$0.15$^{\circ}$, although we are not sensitive to the sign of the misalignment. We use these values to generate the black dashed curve in the inset of Fig. \[fig:4\]b in the main text.
We note that if we instead assume $\theta_b\equiv0^{\circ}$, then the corresponding $n_{SDP,1}=2.05\times10^{12}$ cm$^{-2}$ sets a lattice mismatch of $\delta\approx$ 1.67%. The larger 15.2 nm wavelength computed for the $\theta_t = 0^{\circ}$ and 60$^{\circ}$ configurations (where $n_{SDP}=2.00\times10^{12}$ cm$^{-2}$) could instead be understood to arise owing to a decrease in the lattice mismatch ($\sim$1.3%), indicating the onset of a commensurate transition, which has not yet been observed in exfoliated graphene–BN heterostructures. However, we are not able to reliably distinguish between these two possibilities within our experimental resolution.
The values of $\theta_b$ and $\theta_t$ are more challenging to precisely determine in the stationary devices owing to the lack of dynamic rotational control. Our method of using fractured BN flakes for both the top and bottom encapsulating layers ensures that the relative twist angle of the two BN layers can be determined optically to better than 1$^{\circ}$. We use the broadened Raman 2D peak to first identify that the graphene is well aligned to both BN layers. However, because stacking sequence of bulk BN is AA’ (in which boron and nitrogen atoms alternate between stacked layers) [@Constantinescu2013], there is ambiguity in whether $\theta_t=0^{\circ}$ or 60$^{\circ}$ in our stationary devices depending on whether the number of layers in the BN flake is even or odd. We therefore rely on the measured band gap — in particular whether it is enhanced or suppressed from the typical value of a graphene device aligned with a single BN — to distinguish between $\theta_t=0^{\circ}$ and 60$^{\circ}$.
Figs. \[fig:S4\]a-b plot Landau fan diagrams for stationary devices S1 and S2. Graphene is aligned to both BN layers in these devices, with $\theta_t=0^{\circ}$ in device S1 and 60$^{\circ}$ in device S2. Despite the difference in stacking configuration, we observe qualitatively similar features in both fan diagrams, with sequences of quantum oscillations emerging from the PDP and from the SDPs. Although there may be quantitative differences in the Landau level gaps at the SDPs compared to devices in which graphene is aligned to only a single BN layer due to the doubled moiré potential strength, we have not investigated these effects in detail in this study.
{width="3.9"}
Band gaps for small top BN twist angles
---------------------------------------
Fig. \[fig:S5\] shows the band gaps measured by thermal activation at the PDP and both valence band SDPs in the case of small, non-zero $\theta_t$. In the case of perfect alignment, only a single SDP gap is observed, as discussed in the main text. We find that the $\Delta_{SDP,2}$ are much smaller than the corresponding $\Delta_{SDP,1}$, however we have not measured these gaps at enough values of $\theta_t$ to conclusively determine a trend with twist angle.
{width="7.0"}
{width="3.25"}
Second order moiré patterns
---------------------------
Coexisting moiré patterns with different wavelengths may in principle interfere to produce a second-order moiré pattern. We determine the wavelength, $\Lambda$, using the dispersion in Ref. [@Yankowitz2012] with lattice constants taken as the moiré wavelength of the graphene and bottom BN, $\lambda_b$, and the graphene and top BN, $\lambda_t$. The twist angle between the constituent moiré patterns, $\Theta$, is a function of the twist angle of the top and bottom moiré patterns relative to the graphene ($\phi_t$ and $\phi_b$, respectively), determined by $\theta_t$, $\theta_b$, and $\delta$ [@Yankowitz2012]. Values of $|\phi_t -\phi_b|$, must correctly map to values of $\Theta$ on a 60$^{\circ}$ interval centered at 0$^{\circ}$ for the dispersion to predict $\Lambda$ [@Wang2019]. Accounting for this, we have $$\Lambda = \frac{(1+D)a^2}{\sqrt{2(1+D)(1-cos\Theta) + D^2}}$$ where $$D = \frac{\lambda_b-\lambda_t}{\lambda_t},\quad \lambda_t \leq \lambda_b$$ $$\tan{\phi_{t,b}} = \frac{-\sin{\theta_{t,b}}}{(1+D)-\cos{\theta_{t,b}}}$$ $$\Theta = \left(|\phi_t-\phi_b|-30^{\circ}\right) \textbf{mod}\:(60^{\circ}) - 30^{\circ}$$
The modular division used here to find $\Theta$ reproduces the output of the piecewise conditional reported in [@Wang2019], with a negligible error produced at the endpoints of the interval.
Fig. \[fig:S6\]a shows second order moiré patterns for two combinations of top and bottom BN twist angles, with values of $\Lambda$ calculated using the scheme described above. We note that there is an inequivalence in second order moiré wavelength for $\theta_b=+$0.15$^{\circ}$ and $\theta_b=-$0.15$^{\circ}$, owing to the difference in moiré twist angle $\phi_b$ for each case. Therefore for the top BN twist angle configuration shown ($\theta_t =$ 0.6$^{\circ}$), there are two possible second order moiré wavelengths possible if the sign of the bottom BN twist angle is unknown.
Fig. \[fig:S6\]b plots the density at full filling of the second order moiré unit cell, $N$, determined by $N$ = 8/($\Lambda^2\sqrt{3}$) as a function of $\theta_t$ and $\theta_b$, with an assumed graphene/BN lattice mismatch of $\delta\approx$ 1.65%. Our transport measurements exhibit dominant resistive peaks corresponding to densities $n_{SDP,1}$ and $n_{SDP,2}$ (Figs. \[fig:4\]b-f and \[fig:S6\]c-h) with at least weakly activated behavior and corresponding sequences of quantum oscillations. Furthermore, the measured top BN twist angles corresponding to $n_{SDP,2}$ are consistent with top BN twist angles predicted by the dispersion in Ref. [@Yankowitz2012], and we observe that $n_{SDP,1}$ is fixed by $\theta_b$ for all values of $\theta_t$. We therefore conduct our analysis of possible second order moiré effects on the basis that the densities labeled $n_{SDP,1}$ and $n_{SDP,2}$ correspond to the density at full filling of the bottom and top moiré unit cells, respectively.
To compute the expected values of $N$ for both $\theta_b>0^{\circ}$ and $\theta_b<0^{\circ}$, we first take the uncertainty in $\delta$, $n_{SDP,1}$, and $n_{SDP,2}$ into account. We estimate the uncertainty in $n_{SDP,1}$ and $n_{SDP,2}$ to be $\sim\pm0.02 \times 10^{12}$cm$^{-2}$, corresponding to the uncertainty in fitting the trajectories of the quantum oscillations in the Landau fan diagram. Using a standard differential error propagation we estimate the range of values of $\theta_b$ and $\theta_t$ over a range of possible $\delta$ (ranging from $\sim$1.637% to $\sim$1.654%). This sets a range of possible values of $N$ at each value of $\theta_t$.
Figs. \[fig:S6\]c-e show $d\rho/dT$ as a function of $T$ and $n$ for $\theta_t =$ 59.1$^{\circ}$, -0.7$^{\circ}$ and 59.4$^{\circ}$. We observe regions of negative $d\rho/dT$ (orange arrows, indicating insulating-like behavior) in the vicinity of the expected values of $N$. Figs. \[fig:S6\]f-h show $\rho$ as a function of $n$ at $T$ = 1.7 K for the same values of $\theta_t$. The bounds of the negative $d\rho/dT$ regions observed in Figs. \[fig:S6\]c-e are marked by orange dashed lines. The computed range of possible values of $N$ at each value of $\theta_t$ are shown for $\theta_b<0^{\circ}$ (red shaded region) and $\theta_b>0^{\circ}$ (blue shaded region).
We find that there exists some overlap between the computed values of $N$ and the observed regions of negative $d\rho/dT$, suggesting a possible correspondence with second-order moiré features. However, for fixed $\delta$ and $n_{SDP,1}$, there appears to be no combination of choices for $n_{SDP,2}$ that generate values of $N$ matching negative $d\rho/dT$ features for all three configurations ($\theta_t=59.1^{\circ}$, -0.7$^{\circ}$, and 59.4$^{\circ}$) simultaneously. Furthermore, there are additional regions of negative $d\rho/dT$ marked by orange arrows in Fig. 4f of the main text at higher densities which remain unexplained. While these may be related to replica features of the low density state around SDPs, their density spacing from SDP features is not identical to $N$.
We additionally plot the transport over a full range of accessible gate voltage in Fig. \[fig:S7\]. We observe three features for electron-type doping roughly corresponding to the two SDPs and a lower density feature (black arrows). However, unexpectedly the exact densities of these features do not exactly correspond with their hole-doped counterparts. Therefore, a more detailed study with small $\theta_t$ is necessary to understand all of these extra resistive states in more detail, and in particular how they relate to second order moiré patterns.
{width="7.0"}
Room temperature transport in device R2
---------------------------------------
We measure the room temperature transport of a second rotatable device, R2. Fig. \[fig:S8\]a-b shows $\rho_{PDP}$ and $\rho_{SDP}$ as a function of $\theta_t$. Similar to device R1, we observe extrema in both $\rho_{PDP}$ and $\rho_{SDP}$ with 60$^{\circ}$ periodicity. Following the same procedure, we define $\theta_t=0^{\circ}$ as the angle corresponding to the maximum values observed in both $\rho_{PDP}$ and $\rho_{SDP}$ over a given 120$^{\circ}$ interval. We observe minima in $\rho_{PDP}$ as well as suppressed enhancement of $\rho_{SDP}$ near $\theta_t=60^{\circ}$ in both devices.
Fig. \[fig:S8\]c plots gate sweeps at various $\theta_t$ in device R2. Fig. \[fig:S8\]d shows a zoomed-in view of the PDP peaks from Fig. \[fig:S8\]c. We observe a slight suppression in $\rho_{PDP}$ very near $\theta_t=180^{\circ}$ compared with its value at large misalignment (151.4$^{\circ}$). Comparable behavior is observed in device R1, however at present we do not have a complete model to understand the room-temperature transport response, which is complicated by scattering from acoustic phonons in the graphene and from polar optical phonons in the BN substrate.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Cybercriminals have leveraged the popularity of a large user base available on Online Social Networks (OSNs) to spread spam campaigns by propagating phishing URLs, attaching malicious contents, etc. However, another kind of spam attacks using phone numbers has recently become prevalent on OSNs, where spammers advertise phone numbers to attract users’ attention and convince them to make a call to these phone numbers. The dynamics of phone number based spam is different from URL-based spam due to an inherent trust associated with a phone number. While previous work has proposed strategies to mitigate URL-based spam attacks, phone number based spam attacks have received less attention.
In this paper, we aim to detect spammers that use phone numbers to promote campaigns on Twitter. To this end, we collected information (tweets, user meta-data, etc.) about $3,370$ campaigns spread by $670,251$ users. We model the Twitter dataset as a [*heterogeneous network*]{} by leveraging various interconnections between different types of nodes present in the dataset. In particular, we make the following contributions – (i) We propose a simple yet effective metric, called [*Hierarchical Meta-Path Score*]{} ([*HMPS*]{}) to measure the proximity of an unknown user to the other known pool of spammers. (ii) We design a [*feedback-based active learning strategy*]{} and show that it significantly outperforms three state-of-the-art baselines for the task of spam detection. Our method achieves 6.9% and 67.3% higher F1-score and AUC, respectively compared to the best baseline method. (iii) To overcome the problem of less training instances for supervised learning, we show that our proposed [*feedback strategy*]{} achieves 25.6% and 46% higher F1-score and AUC respectively than other oversampling strategies. Finally, we perform a case study to show how our method is capable of detecting those users as spammers who have not been suspended by Twitter (and other baselines) yet.
author:
- Srishti Gupta
- Abhinav Khattar
- Arpit Gogia
- Ponnurangam Kumaraguru
- Tanmoy Chakraborty
bibliography:
- 'ref.bib'
title: 'Collective Classification of Spam Campaigners on Twitter: A Hierarchical Meta-Path Based Approach'
---
<ccs2012> <concept> <concept\_id>10002951</concept\_id> <concept\_desc>Information systems</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002978</concept\_id> <concept\_desc>Security and privacy</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10010405</concept\_id> <concept\_desc>Applied computing</concept\_desc> <concept\_significance>300</concept\_significance> </concept> </ccs2012>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe a robust technique based on the ULySS IDL code for measuring velocity dispersions of galaxies observed with the MMT’s fiber-fed spectrograph, Hectospec. This procedure is applicable to all Hectospec spectra having a signal-to-noise $\ga$5 and weak emission lines. We estimate the internal error in the Hectospec velocity dispersion measurements by comparing duplicate measurements of 171 galaxies. For a sample of 984 galaxies with a median z=0.10, we compare velocity dispersions measured by Hectospec through a 1.5$^{\prime\prime}$ diameter optical fiber with those measured by the Sloan Digital Sky Survey (SDSS) and Baryon Oscillation Spectral Survey (BOSS) through 3$^{\prime\prime}$ and 2$^{\prime\prime}$ diameter optical fibers, respectively. The systematic differences between the Hectospec and the SDSS/BOSS measurements are $<$7% for velocity dispersions between 100 and 300 km s$^{-1}$, the differences are no larger than the differences among the three BOSS velocity dispersion reductions. We analyze the scatter about the fundamental plane and find no significant redshift dependent systematics in our velocity dispersion measurements to z$\sim$0.6. This analysis also confirms our estimation of the measurement errors. In one hour in good conditions, we demonstrate that we achieve 30 km s$^{-1}$ velocity dispersion errors for galaxies with an SDSS r fiber magnitude of 21.'
author:
- 'Daniel Fabricant, Igor Chilingarian, Ho Seong Hwang, Michael J. Kurtz and Margaret J. Geller'
- 'Ian P. Del’Antonio'
- 'Kenneth J. Rines'
title: Measuring Galaxy Velocity Dispersions with Hectospec
---
INTRODUCTION {#sect:intro}
============
Measurements of stellar velocity dispersions in galaxies have broad application including estimation of galaxy masses, galaxy classification, and distance measurements. The use of velocity dispersion measurements to determine galaxy properties and their evolution requires a clear understanding of the statistical and systematic errors in these measurements over a wide redshift range. Comparison of large samples of measurements obtained with different instruments and different techniques constrains these errors.
These broad applications of stellar velocity dispersion in galaxies motivate our investigation of velocity dispersion measurements from moderate to low signal-to-noise (SN) spectra originally obtained for redshift measurements. Hectospec, the MMT’s fiber-fed spectrograph [@fab05], has obtained spectra of $\sim$700,000 unique objects; an appreciable fraction of these spectra can yield reliable velocity dispersions. Hectospec’s fibers subtend 1.5$^{\prime\prime}$ on the sky. Most spectra, obtained with a 270 line mm$^{-1}$ grating, have 5 Å FWHM resolution. We discuss results obtained with direct spectral fitting, an approach first described by @rix92. Direct spectral fitting is conceptually straightforward and allows simple masking of portions of the spectrum contaminated with strong skylines, bad pixels, or emission lines. We use a IDL-based software package, ULySS, developed by @kol09 to perform direct fitting of Hectospec spectra. ULySS fits observed spectra to model spectra of synthesized galaxy stellar populations broadened by the instrumental line spread function and a velocity dispersion.
An extensive literature reporting velocity dispersion techniques and measurements follows the pioneering work of @min54 who reported measurements of M31’s velocity dispersion. The earliest measurements using optical fiber front-ends were performed at the AAT [@col87; @luc88]. @jor95 made extensive comparisons of velocity dispersions from the fiber-fed OCTOPUS spectrograph with slit spectrograph measurements, and obtained consistent results. Beginning with the first data release [@aba03], the Sloan Digital Sky Survey (SDSS) has made available extensive catalogs of well-calibrated velocity dispersions obtained with fiber-fed spectrographs. The only serious disadvantage of measuring velocity dispersions through fibers rather than directly with a slit spectrograph is the sacrifice of spatial resolution for multiplex advantage.
In Section 2 we describe application of ULySS to Hectospec spectra with weak nebular line emission allowing uncontaminated measurement of stellar absorption features. We calculate the internal errors in our velocity dispersion measurements in Section 3. In Section 4 we compare our velocity dispersion measurements for 984 galaxies with SDSS/Baryon Oscillation Spectral Survey (BoSS) Date Release 9 (DR9) pipeline measurements. We plot the scatter about the fundamental plane for a sample of 1857 galaxies to a redshift of 0.6 in Section 5 to demonstrate that our velocity dispersion errors are accurate at higher redshifts. We describe how to plan velocity dispersion measurements with Hectospec in Section 6, and give our conclusions in Section 7. We adopt cosmological parameters H$_0$=70, $\Omega_M$=0.3, and $\Omega_\Lambda$=0.7.
Hectospec Velocity Dispersion Measurements
==========================================
Introduction {#introduction}
------------
The ULySS algorithms [@kol09] are based on an earlier IDL package, NBURSTS, developed by @chi07. NBURSTS in turn is based on the pPXF package developed by @cap04. ULySS simultaneously fits a spectrum with the internal galaxy dynamics and parameters describing a star formation history. To extract accurate velocity dispersions, the model spectrum must be convolved to the same wavelength dependent spectral resolution as the data. ULySS then applies a multiplicative polynomial to account for errors in flux calibration of the data and for inaccuracies in the model prediction.
The galaxy spectral models we use are single age stellar populations (SSP) parameterized by age and metallicity. We used a precomputed grid of SSP models [@pru11] calculated with the PegaseHR code [@leb04] from the MILES stellar library [@san06], assuming a Salpeter initial mass function and solar neighborhood abundances. This grid of models is available on the ULyss web site (ulyss.univ-lyon1.fr). We adopt a cosmology with $H_0$=70, $\Omega_M$=0.3, and $\Omega_\Lambda$=0.7.
ULySS Parameters
----------------
The first step in our analysis is relative flux calibration of the Hectospec spectra following the techniques described in @fab08. We write out the deredshifted fluxed spectra in a FITS format compatible with ULySS. As described in @fab08, the flux calibration is quite stable over time, and the ULySS multiplicative polynomial accounts for any small errors in flux calibration.
We experimented extensively with the available ULySS parameters. We obtain the lowest velocity dispersion errors and stablest results by restricting the spectral range to 4100-5500 Å, by restricting the model metallicities (log(${model}\over{solar}$)) between -0.5 and 0.5, and by using a third order multiplicative polynomial. Accessing lower metallicities in model fits allows an unphysical degeneracy between age and metallicity for low SN spectra [@wor95]. Higher order multiplicative polynomials do not meaningfully reduce chi squared and sometimes attempt to null real spectral features on low SN spectra.
Hectospec’s Line Spread Function
--------------------------------
We determine Hectospec’s line spread function (LSF) using the ULySS routine uly\_lsf. This routine derives the line spread function by fitting twilight flat spectra to a high resolution model (R=10000) of the solar spectrum provided in the ULySS distribution. As supplied, uly\_lsf fits only a velocity shift and a Gaussian LSF, but is easily modified to fit h3 and h4 Hermite polynomials to describe a non-Gaussian LSF. Although Hectospec’s optical fibers do provide a somewhat flat topped LSF, measurements of h3 and h4 with uly\_lsf scatter closely about 0. We therefore fix the h3 and h4 terms at zero.
Hectospec’s fiber feed guarantees a consistent input aperture, and the line spread function should be quite stable; the only time dependent changes should arise from focus variations and changes in the fiber focal ratio degradation. The Hectospec focus is regularly checked and is maintained within a tight range, and Hectospec was carefully designed to minimize focal ratio degradation [@fab05]. In addition to time dependent changes, variation in the image quality of Hectospec’s optics and the flatness and alignment of the CCD detectors can introduce fiber to fiber variations in the line spread function. These spatially and time dependent variations in the line spread function can be recovered from the twilight flats for each night of data and each fiber. A new pipeline under development will correct for these issues. We show that very acceptable results can be obtained from the current pipeline by using a line spread function averaged over fiber and time.
We have measured the Hectospec line spread function for each of the 300 fibers on three randomly chosen nights: 13 October 2007, 20 November 2008, and 15 October 2009. For these 900 measurements we calculate the LSF in 17 wavelength bins, each 200 Å wide, centered between 3800 to 8900 Å. For each bin, we calculate the mean Gaussian (1 $\sigma$) LSF in km s$^{-1}$ and the standard deviation in the LSF. The LSF ranges between 172 km s$^{-1}$ at 3800 Å, 105 km s$^{-1}$ at 6000 Å, and 78 km s$^{-1}$ at 9000 Å, or 5.1, 4.9, 5.5 Å FWHM, respectively. The standard deviation in these measurements is typically 3 km s$^{-1}$, or 2% of the LSF at the blue end and 5% of the LSF at the red end of Hectospec’s spectral range. These standard deviations are small enough to support use of the average LSF. Spectral regions affected by strong atmospheric absorption (particularly between 6800 and 7600 Å) yield incorrect LSFs from the twilight flats and must be rejected. We fit a third order polynomial to the valid data points; the coefficients of the fit are in Table \[lsf\].
To analyze the fluxed Hectospec spectra shifted to rest frame wavelengths, we use the LSF appropriate to the original observed wavelengths. In addition, we correct for the intrinsic resolution of the MILES stellar library, 2.51 Å FWHM [@fal11], or $\sigma$=1.066 Å (see also [@pru11; @bei11]). We subtract this resolution in quadrature from the redshift-shifted Hectospec LSF to produce a final LSF for each spectrum. Figure \[ulyss\] shows sample Hectospec spectra with the ULySS fits overplotted.
Internal Errors
===============
We examine the internal errors in our velocity dispersions using a sample of 171 pairs of measurements from the SHELS survey [@gel10] where the error in each measurement is $<$30 km s$^{-1}$. Figure \[interr\] is a histogram of the dispersion differences in km s$^{-1}$. The expected RMS dispersion difference calculated from the ULySS errors is 21.7 km s$^{-1}$, we measure an RMS dispersion difference of 20.5 km s$^{-1}$ for the 171 pairs. We show Gaussian of 18 km s$^{-1}$ $\sigma$ fit to the binned data for reference in Figure \[interr\]. Our repeated measurements confirm the ULySS error estimates, and we adopt these errors for subsequent analysis.
External Errors - Comparison with SDSS/BOSS DR9
===============================================
We assess our external errors using Hectospec velocity dispersion measurements for a sample of 984 galaxies with high SN velocity dispersion measurements (estimated dispersion errors $<$20 km s$^{-1}$ in both cases) in the 9th SDSS/Baryon Oscillation Spectroscopic Survey (BOSS) data release (DR9, @ahn12); 843 of these spectra were obtained with the SDSS fiber-fed spectrographs and 141 with the updated BOSS fiber-fed spectrographs [@sme13]. The median redshift of the combined sample is 0.147. Galaxies with \[OII\]$\lambda$3727 equivalent widths $>$5Å were excluded from this sample to avoid contamination from emission lines.
The Hectospec spectra are drawn from three surveys: SHELS [@gel10; @gel12; @hwa13], the Hectospec Cluster Survey(HeCS) [@rin13], and HectoMap [@gel11]. These surveys include $\sim$15,000, $\sim$22,000, and $\sim$52,000 Hectospec spectra, respectively, all with the same Hectospec configuration. SHELS is a magnitude limited survey complete to R=20.6; here we use the entire sample including some galaxies fainter than the R=20.6 limit. HeCS surveys red sequence galaxies with r$<$21 in 58 galaxy clusters. HectoMAP is a survey of red selected galaxies in a 50 deg$^2$ strip to r$<$21.3. By design, only a small fraction of these galaxies overlap with one of the SDSS surveys.
The SDSS-I/II fibers subtend 3$^{\prime\prime}$, the SDSS-III/BOSS fibers subtend 2$^{\prime\prime}$. To compare the Hectospec measurements directly with the DR9 velocity dispersions, we apply an aperture correction to scale the Hectospec velocity dispersions downwards to match the larger SDSS or BOSS apertures. We use the aperture correction from @cap06:
$${{\sigma_1}\over{\sigma_2}}=\left({{r_1}\over{r_2}}\right)^{-0.066}$$
The aperture corrections to transform to the SDSS and BOSS apertures are thus 0.955 and 0.981, respectively. Figure \[exterr\] shows the 984 pairs of Hectospec and DR9 [@dr9pipe] pipeline velocity dispersions. We remove two galaxies from analysis with dispersions that disagree by more than 5 $\sigma$. Although the two measurements of independent spectra analyzed with rather different software show remarkable agreement, it is clear that the Hectospec/ULySS velocity dispersions are systematically larger at large dispersions. The solid line in Figure \[exterr\] shows the error-weighted best fit line relating the two measurements; the line (valid only between 100 and 300 km s$^{-1}$) has an intercept of -20.5 km s$^{-1}$ and a slope of 1.139.
We can also study the error distribution. The expected RMS dispersion difference calculated from the ULySS and DR9 pipeline errors is 17.1 km s$^{-1}$, but we measure a larger RMS dispersion difference of 24.7 km s$^{-1}$ for the 984 pairs. Here, we have removed the linear relation described above. The measured RMS dispersion difference suggests that the errors are underestimated by $\sqrt{2}$. However, various systematic errors like aperture centration and aperture corrections probably contribute. Because our repeated measurements agree within the ULySS errors, systematic errors may dominate the apparent excess error.
We next consider the systematic differences in velocity dispersions. We can explore the effect of different velocity dispersion analysis techniques by comparing the three BOSS DR9 reductions; the pipeline reduction [@dr9pipe], the Portsmouth reduction [@dr9port], and the Wisconsin reduction [@dr9wisc]. Figure \[port\] shows the velocity dispersions of the subset of 12759 BOSS galaxies with velocity dispersion errors of $<$10 km s$^{-1}$ in both analyses. The solid line in Figure \[exterr\] shows the error-weighted best fit line relating the two measurements; the line (valid only between 100 and 300 km s$^{-1}$) has an intercept of -10.4 km s$^{-1}$ and a slope of 1.093. The systematic difference between the two analyses is negligible at 100 km s$^{-1}$, rising to $\sim$6% at 300 km s$^{-1}$. Figure \[wisc\] shows the velocity dispersions of the subset of 6638 BOSS galaxies with velocity dispersion errors of $<$20 km s$^{-1}$ in both analyses. The solid line in Figure \[exterr\] shows the error-weighted best fit line relating the two measurements; the line (valid only between 100 and 300 km s$^{-1}$) has an intercept of 15.6 km s$^{-1}$ and a slope of 0.947. The systematic difference between the two analyses is $\sim$10% at 100 km s$^{-1}$, and negligible at 300 km s$^{-1}$. @dr9port presents plots comparing the various DR9 BOSS reductions with a less restrictive cut on the velocity dispersion errors.
Table \[dispdiff\] summarizes the ratios of the velocity dispersions from the Hectospec analysis, the Portsmouth, and the Wisconsin reductions relative to the DR9 pipeline. The DR9 reductions reflect only differing analysis techniques; the underlying spectra are identical. It is interesting to note that the systematic deviations of the Hectospec/ULySS velocity dispersions relative to the DR9 pipeline are almost identical to those of the Portsmouth reductions relative to the DR9 pipeline. The Portsmouth analysis uses the pPXF [@cap04] code that is functionally identical to ULySS and very similar stellar population models. In contrast, both the Wisconsin and pipeline analyses use some form of principal component analysis. We conclude that the systematic differences between Hectospec and DR9 pipeline dispersions are no larger than the differences among the three different DR9 reductions.
External Validation Using the Fundamental Plane
===============================================
Night sky features contribute differently to velocity dispersion errors as a function of the redshift of the measured galaxy. The MgI b triplet at low redshift is in a spectral region where the night sky is relatively smooth and easily subtracted; at a redshift of 0.6 this feature is well inside the “forest” of OH night-sky emission lines. The direct external validation by comparison with the SDSS in section 4 does not contain a sufficient number of high redshift objects to show the presence or absence of systematic tendencies at redshifts above z $\sim$0.2. We perform an indirect validation by observing the scatter and offset about the fundamental plane as a function of redshift.
It has long been known [@fab76; @djo87] that early type galaxies define a narrow relation in velocity dispersion ($\sigma$), surface brightness ($\mu$) and effective radius ($r_e$), the fundamental plane (FP). $$log(r_e) = a log(\sigma) + b \mu + c$$ We use this relation to test for redshift dependant systematics in the dispersions and their errors. Previous studies have shown: (1) that the FP parameterization does not vary with redshift, at least for z$\la$0.6 [@kel97; @jor96; @van96] and (2) that the intrinsic scatter about the FP relation is stable [@hyd09].
We use spectral measurements from the SHELS survey [@gel10] and photometric measurements from the SDSS DR9 [@ahn12] to make that test. We adopt the r band C-model FP calibration of [@sau13] (a=1.041, b= 0.30 c=-7.76) which uses cModelMag\_r and deVRad\_r, and we follow their prescription almost exactly for correcting the input measures for aperture differences, the k correction, and evolution. The two differences from their analysis are: (1) that we do not correct the measured radius of each galaxy for its ellipticity, and (2) for redshifts from 0.5 to 0.6 we use an extrapolation of the modified k correction of @chi10 used by [@sau13]. Our extrapolation assumes an elliptical galaxy spectrum.
We choose “red” galaxies by requiring that deVfrac\_r $>$0.6 and OII 3727 emission $<$5 Å and $D_{\lambda}4000$ $>$1.65. In addition we require a fractional error in the velocity dispersion $<$20%. Figure \[FP\] shows the residuals in the FP as a function of redshift. We show all 1857 objects which pass the selection. The mean offset is 0.0011 with a standard deviation of 0.1181. We compute the number of objects, the mean offset and the standard deviation in three redshift bins: (0$<$z$<$0.25; n=719; offset=0.0077; s.d.=0.1176), (0.25$<$z$<$0.25); n=600; offset= -0.0020, s.d.=0.1195), (0.35$<$z$<$0.60; n = 483; offset=-0.0075, s.d.=0.1211). The offsets of the whole sample and the redshift selected subsamples are consistent with zero. Thus there is no significant redshift dependent error in the measured velocity dispersions.
We externally validate the errors in the velocity dispersion measurements by computing the intrinsic scatter in the FP with two different cuts in fractional velocity dispersion error. Because we compute the intrinsic width of the FP by subtracting (in quadrature) the error expected by propagating the individual measurement errors from the measured scatter, under(over) estimates of the measurement errors will produce a computed intrinsic width that differs for samples with different error cuts.
To eliminate any redshift dependent effects we narrow our analysis to the central redshift bin: 0.25$<$z$<$0.35. We compute the FP intrinsic width from two samples: (1) galaxies with a maximum fractional dispersion error of 0.20, and (2) galaxies with a fractional dispersion error $<$0.10. For the 20% error sample we compute an intrinsic width of 0.0999 from 600 galaxies; for the 10%error sample we compute an intrinsic width of 0.1008 from 182 galaxies. The difference in the computed widths is small. If the difference resulted entirely from a misestimation of the velocity dispersion errors, the errors would have to be overestimated by 5%, in basic agreement with the internal error analysis in section 3.
Planning Hectospec Dispersion Measurements
==========================================
Here we compute the velocity dispersion errors expected as a function of magnitude for an exposure time of 3600 s. We have insufficient experience with significantly longer exposures to determine the errors as a function of exposure time.
We examine a sample of 1262 SHELS galaxies observed for 3600 s with velocity dispersion errors $<$50 km s$^{-1}$. We expect the errors to correlate best with 1.5$^{\prime\prime}$ aperture magnitudes corresponding to the Hectospec fiber diameter. Figure \[errmag\] plots the Hectospec velocity dispersion errors as a function of R magnitude in a 1.5$^{\prime\prime}$ aperture [@wit02; @wit06]. The correlation reflects a range of observing conditions including seeing, cloud cover, and moon illumination. The lower envelope of the distribution corresponds to observations during dark conditions with good seeing and clear skies. Under these conditions, with a 3600 s observation we expect a 20 km s$^{-1}$ velocity dispersion error for a galaxy with an R aperture magnitude of 20.5, and a 30 km s$^{-1}$ velocity dispersion error for a galaxy with an R aperture magnitude of 21.
For most observers, the best easily available proxy for Hectospec fiber magnitudes is SDSS fiber magnitudes measured in a 3$^{\prime\prime}$ aperture. These magnitudes do not correlate as well with the Hectospec velocity dispersion errors as the 1.5$^{\prime\prime}$ aperture magnitudes (Figure \[errmag\]), but they are still useful (Figure \[errmagsdss\]). Under the best conditions, we expect a 20 km s$^{-1}$ velocity dispersion error for a galaxy with an r SDSS fiber magnitude of 20.4, and a 30 km s$^{-1}$ velocity dispersion error for a galaxy with an r fiber magnitude of 21. The differences in filter bandpass and aperture roughly cancel.
Conclusions
===========
The main goal of our investigation is to enable studies of fundamental galaxy properties and their evolution using Hectospec data. Careful comparisons of velocity dispersion measurements made with independent instruments are also of more general interest to establish the accuracy of our large data sets in an era of “precision cosmology".
We describe the use of publicly available software, ULySS [@kol09], to obtain velocity dispersions for Hectospec galaxy spectra. We compare our measurements to those from the SDSS DR9 pipeline for 984 galaxies in common with velocity dispersion errors of $<$20 km s$^{-1}$. The systematic differences in the two measurements are $<$7% for galaxies with dispersions between 100 and 300 km s$^{-1}$. These differences are comparable to the systematic differences among the three velocity dispersion reductions for the DR9 BOSS data.
By analyzing the scatter about the fundamental plane we show that there are no significant systematics in our velocity dispersion measures as a function of redshift, for z$\la$0.6. Additionally we confirm that our estimation of the measurement errors is correct, within narrow tolerances.
In one hour in good conditions, we can expect 20 km s$^{-1}$ velocity dispersion errors for a galaxy with an r SDSS fiber magnitude of 20.4, and 30 km s$^{-1}$ velocity dispersion errors for a galaxy with an r fiber magnitude of 21.
[*Facility:*]{}
ACKNOWLEDGMENTS
===============
We thank Marijn Franx, Nelson Caldwell and Daniel Eisenstein for helpful comments.
Observations reported here were obtained at the MMT Observatory, a joint facility of the Smithsonian Institution and the University of Arizona. We thank the Hectospec engineering team including Robert Fata, Tom Gauron, Marc Lacasse, Mark Mueller, and Joe Zajac, and the instrument specialists Perry Berlind and Michael Calkins. We are grateful for the contributions of the members of the CfA’s Telescope Data Center including Warren Brown, Anne Matthews, John Roll, Susan Tokarz and Sean Moran. The entire staff of the MMT Observatory under the direction of G. Grant Williams has provided outstanding support for Hectospec operations. Nelson Caldwell has ably scheduled Hectospec’s queue operations.
Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the U.S. Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, the Max Planck Society, and the Higher Education Funding Council for England. The SDSS Web Site is http://www.sdss.org/. The SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions. Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/.
Abazajian, K. et al. 2003, , 126, 2081 Ahn, C. P., 2012, , 203, 21 Beifiori, A. et al. 2011, , 531, 109 Bolton, A. et al. 2012, , 144, 144 Chen, Y.-M. et al. 2012, , 421, 314 Cappellari, M. & Emsellem, E. 2004, , 116, 138 Cappellari, M. et al. 2006,, 366, 1126 Chilingarian, I. et al. 2007, in IAU Symp. 241, 175 Chilingarian,I. et al. 2010, , 405, 1409 Colless, M. & Hewett, P. 1987, , 224, 453 Djorgovski, S., & Davis, M. 1987, , 313, 59 Fabricant, D., et al. 2005, , 117, 1411 Faber, S. M., & Jackson, R. E. 1976, , 204, 668 Fabricant et al. 2008, , 120, 1222 Falcón-Barroso, J. et al. 2011, , 532, 95 Geller, M. J. et al. 2010, , 709, 832 Geller, M. J., Diaferio, A., & Kurtz, M. J. 2011, , 142, 133 Geller, M. J. et al. 2012, , 143, 102 Hyde, J. B., & Bernardi, M. 2009, , 396, 1171 Hwang, H. S. & Geller, M. J. 2013, , 769, 116 J[ø]{}rgensen, I., Franx, M., & Kj[æ]{}rgaard, P. 1995, , 276, 1341 J[ø]{}rgensen, I., Franx, M., & Kj[æ]{}rgaard, P. 1996, , 280, 167 Kelson, D. D. et al. 1997, , 478, L13 Le Borgne, D. et al. 2004, , 425, 881 Koleva, M. et al. 2009, , 501, 1269 Lucey, J.R. & Carter, D. 1988, , 235, 1177 Minkowski, 1954, Carnegie Yearbook, 26 Prugniel, P., Vauglin, I., & Koleva, M. 2011, , 521, 165 Rines, K. et al. 2013, , 767, 15 Rix,H.-W. & White, S. 1992, , 254, 389 Sanchez-Blazquez, P. 2006, , 371, 703 Saulder, C., et al. 2013, arXiv:1306.0285 Smee, S. et al. 2013, , in press Thomas, D. et al. 2013, , 431, 1383 van Dokkum, P. G., & Franx, M. 1996, , 281, 985 Wittman, D. M., et al. 2002, , 4836, 73 Wittman, D., et al. 2006, , 643, 128 Worthey, G. 1995, , 95, 107
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[ulyss\] Sample ULySS fits (bold line) overplotted on Hectospec spectra (fine line). The spectrum in the bottom panel yields a velocity dispersion of 101$\pm$6 km s$^{-1}$ and the spectrum in the top panel yields a velocity dispersion of 281$\pm$8 km s$^{-1}$.](fig1.eps "fig:"){height="6in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[interr\] Hectospec internal velocity dispersion differences for repeated measurements of 171 galaxies where the error in each measurement is $<$30 km s$^{-1}$. The expected RMS dispersion difference calculated from the errors is 22 km s$^{-1}$ and the measured RMS difference is 21 km s$^{-1}$. A Gaussian of 18 km s$^{-1}$ $\sigma$ fit to the binned data is shown for reference.](fig2.eps "fig:"){height="6in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[exterr\] A comparison of independent Hectospec and SDSS measurements of 984 galaxies with velocity dispersion errors $<$20 km s$^{-1}$. An aperture correction has been applied to the Hectospec data (see text). The short solid line segment is a error weighted fit of a line to the plotted data points. This line has an intercept of -20.5 and a slope of 1.139. Measurements with the original SDSS spectrograph and 3$^{\prime\prime}$ fibers are plotted with filled symbols (843 galaxies) and measurements with the updated BOSS spectrograph with 2$^{\prime\prime}$ fibers are plotted with open symbols (141 galaxies).](fig3.eps "fig:"){height="7in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[port\] A comparison of DR9 Portsmouth and DR9 pipeline measurements of 12759 galaxies with velocity dispersion errors $<$10 km s$^{-1}$. The short solid line segment is a error weighted fit of a line to the plotted data points. This line has an intercept of -10.4 and a slope of 1.093.](fig4.eps "fig:"){height="7in"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[wisc\] A comparison of DR9 Wisconsin and DR9 pipeline measurements of 6638 galaxies with velocity dispersion errors $<$20 km s$^{-1}$. The short solid line segment is a error weighted fit of a line to the plotted data points. This line has an intercept of 15.6 and a slope of 0.947.](fig5.eps "fig:"){height="7in"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[FP\] The scatter of 1857 galaxies from the SHELS survey [@gel10] about the C-model fundamental plane relationship of @sau13 as a function of redshift. Here $r_e$ is deVRad\_r and $\mu$ is computed as $\mu = 2.5 log(2\pi) + 5 log(r_e) + cModelMag\_r$. The sample selection is described in the text.](fig6.eps "fig:"){height="6in"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[errmag\] Hectospec velocity dispersion errors for a 3600 s observation as a function of R magnitude within a 1.5$^{\prime\prime}$ Hectospec fiber aperture. The points reflect observations of galaxies with velocity dispersions between 100 and 300 km s$^{-1}$ during a wide range of conditions including seeing and transparency. Although we expect the errors to depend on the galaxy dispersion, this dependence is obscured by the seeing and transparency variations that we cannot accurately remove. The lower envelope reflects observations during the best conditions.](fig7.eps "fig:"){height="7in"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[errmagsdss\] Hectospec velocity dispersion errors for a 3600 s observation as a function of r magnitude within a 3$^{\prime\prime}$ aperture (SDSS fibermag). The points reflect observations during a wide range of conditions including seeing and transparency. The lower envelope reflects observations during the best conditions.](fig8.eps "fig:"){height="7in"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[rc]{} constant & 4.64400 $\times$ 10$^{2}$\
linear & -1.15515 $\times$ 10$^{-1}$\
quadratic& 1.16604 $\times$ 10$^{-5}$\
cubic & -3.99359 $\times$ 10$^{-10}$\
[cccc]{} 100& 0.93 & 0.99 & 1.10\
150& 1.00 & 1.02 & 1.05\
200& 1.04 & 1.04 & 1.02\
250& 1.06 & 1.05 & 1.01\
300& 1.07 & 1.06 & 1.00\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The quantum XY, Heisenberg, and transverse field Ising models on hyperbolic lattices are studied by means of the Tensor Product Variational Formulation algorithm. The lattices are constructed by tessellation of congruent polygons with coordination number equal to four. The calculated ground-state energies of the XY and Heisenberg models and the phase transition magnetic field of the Ising model on the series of lattices are used to estimate the corresponding quantities of the respective models on the Bethe lattice. The hyperbolic lattice geometry induces mean-field-like behavior of the models. The ambition to obtain results on the non-Euclidean lattice geometries has been motivated by theoretical studies of the anti-de Sitter/conformal field theory correspondence.'
address: 'Institute of Physics, Slovak Academy of Sciences, SK-845 11, Bratislava, Slovakia'
author:
- Michal Daniška and Andrej Gendiar
title: Analysis of quantum spin models on hyperbolic lattices and Bethe lattice
---
[*Keywords*]{}: tensor product state, quantum spin systems, non-Euclidean geometry, phase transition
Introduction
============
Many analytical and computational techniques have been developed to study quantum spin models on two-dimensional (Euclidean) lattices. Among such techniques, let us mention the corner transfer matrix approach [@Baxter], the coordinate Bethe Ansatz [@Bethe], the algebraic Bethe Ansatz [@Fad], the vertex operator approach [@AA], including numerical algorithms based on tensor product states and tensor networks [@Orus; @TPS1; @TPS2; @TPS3; @HOTRG], all of which have been successfully applied to the description of the energy spectrum and matrix elements of local operators in either integrable lattice models and quantum spin chains or non-integrable quantum spin systems. However, the task of finding an appropriate approach to analyze the quantum models on hyperbolic lattices, which belongs to challenging problems related to the correspondence between the anti-de Sitter space and the conformal field theory [@QG], still remains an open question of the quantum gravity. A remarkable demand for an appropriate numerical tool persists. Implementation of the Monte Carlo simulations fails due to exponential increase of the number of the lattice sites for models on hyperbolic lattices with respect to the expanding lattice size from the lattice center [@MC1; @MC2]. Our desire is to propose a novel and sufficiently accurate numerical algorithm, which originates from the solid state physics and inherits the typical features coming from widely accepted renormalization group approaches, especially based on the Density Matrix Renormalization Group [@White; @Uli1; @Uli2].
Recently, we modified the Tensor Product Variational Formulation (TPVF) [@DaniskaGendiar], which is an algorithm combining an ansatz for the ground-state in the form of the Tensor product state (TPS) [@Orus] with the Corner transfer matrix renormalization group (CTMRG) scheme [@Nishino]. This algorithm can be used to study quantum spin systems in the thermodynamic limit on regular hyperbolic lattices of constant negative Gaussian curvature [@Sadoc]. The hyperbolic lattices are constructed by tessellation of congruent $p$-sided polygons (with the coordination number fixed to four). We applied the modified TPVF algorithm to the Euclidean square ($p = 4$) and hyperbolic pentagonal ($p = 5$) lattices in order to analyze the critical phenomena of the XY, Heisenberg and transverse field Ising model (TFIM). On the square lattice numerical inaccuracy varied from $1.2\%$ in the XY model to $3.7\%$ in TFIM at the phase transition. This observation originates in the mean-field-like behavior induced by the TPS ansatz, which, as a consequence, cannot accurately approximate the correct ground state of the TFIM on the two-dimensional Euclidean lattice, which belongs to the Ising universality class. On the contrary, since the Hausdorff dimension of the hyperbolic lattices is infinite, spin models on these lattices belong to the mean-field universality class due to short range correlations, even though the mean-field approximation of the Hamiltonian is not applied [@Baxter]. We conjectured that TPVF was more suitable for models on the pentagonal hyperbolic lattice due to off-critical and weakly correlated characteristics [@DaniskaGendiar].
In this work we expand the set of hyperbolic lattices investigated by the TPVF to a series of lattices constructed from congruent $p$-sided polygons, where $p \in \{5,6, \dots, 11\}$. In analogy to our previous studies of classical spin models on these hyperbolic lattices [@hctmrg-Ising-5-4; @hctmrg-Ising-p-4; @hctmrg-Ising-3-q; @hctmrg-Ising-3-qn], we expect fast convergence of the phase transition magnetic field of the quantum TFIM as well as the ground-state energies of the quantum XY and Heisenberg models toward the asymptotic case $p \rightarrow \infty$, which represents the Bethe lattice [@hctmrg-Ising-p-4]. Numerical results presented in the following sections are in complete agreement with the expectations. The key feature of this work is the consequent indirect analysis of the quantum TFIM, XY, and Heisenberg models on the Bethe lattice with coordination number four, which has not been considered yet.
The article is organized as follows. In Sec. II we define the three Hamiltonians on the respective hyperbolic lattices and give a brief description of the principles of the TPVF algorithm, which have been discussed in [@DaniskaGendiar]. An accurate analysis of the numerical results is presented in Sec. III and we summarize them in Sec. IV.
The Model
=========
\
\
We study the ground-state properties and the phase transition of the quantum TFIM, XY, and Heisenberg models in the thermodynamic limit on a series of hyperbolic lattices. Each hyperbolic lattice is made from equivalent congruent $p$-sided polygons. The polygon vertices coincide with the lattice spin sites, where a single spin is positioned. Each spin site has four nearest-neighboring spin sites, which is commonly referred to as the coordination number equal to four. We investigate the three models on a set of regular hyperbolic lattices of infinite size with the lattice parameter $p \in \{5,6,\dots, 11\}$. Apart from the set, we include two additional cases: $p=4$ being the Euclidean square lattice and the asymptotic case $p\to\infty$, which is associated to the Bethe lattice. Figure \[Fig1\] depicts the typical structure of the lattices. The square lattice serves as a reference lattice. The three spin models and the TPVF algorithm have been described in detail in [@DaniskaGendiar], and we focus only on the substantial aspects of the models on the hyperbolic lattices in the following.
In general, the Hamiltonian ${\cal H}$ of any of the three models can be written as a sum of local Hamiltonians $G_{k}^{(p)}$ of the $p$-sided polygonal shape, in particular, $${\cal H} = \sum\limits_{{\langle k \rangle}_p}^{~} G_{k}^{(p)}\, ,
\label{Hm1}$$ where $k$ labels the polygons and the sum runs over the set of all indices of the lattice polygons ${\langle k \rangle}_p$. The local Hamiltonian takes the form $$\begin{aligned}
\nonumber
G_{k}^{(p)} & = & -\frac{1}{2}\sum\limits_{i=1}^{p} \bigg[
J_{xy} \left( S_{k_i}^{x} S_{k_{i+1}}^{x} + S_{k_i}^{y} S_{k_{i+1}}^{y} \right) \\
& + & J_{z} S_{k_i}^{z} S_{k_{i+1}}^{z}
+ \frac{h}{4} \left( S_{k_i}^{x} + S_{k_{i+1}}^{x}\right) \bigg] \, ,
\label{Hm2}\end{aligned}$$ where $k_1,k_2,\dots,k_p$ label the spin positions on the $k^{\rm th}$ $p$-sided polygon (noticing that $k_{p+1} \equiv k_1$), and $S_{k_i}^x$, $S_{k_i}^y$, $S_{k_i}^z$ denote the corresponding Pauli spin-$\frac{1}{2}$ operators. We consider constant nearest-neighbor couplings $J_{xy}, J_{z}$ and a uniform external magnetic field $h$. By setting $J_{xy} = 0$ and $J_{z} = 1$ we obtain the TFIM at the transverse magnetic field $h$, whereas the choice $J_{xy} = 1,J_{z} = h = 0$ gives the XY model and $J_{xy} = - J_{z} = 1, h = 0$ the Heisenberg model [@DaniskaGendiar].
Our task is to calculate an approximate ground-state of the system in the thermodynamic limit in the product form $$| \Psi_p\rangle = \lim\limits_{N\to\infty}
\sum\limits_{\sigma_1^{~}\sigma_2^{~}\cdots\sigma_{N}^{~}}^{~}
\prod\limits_{{\langle k \rangle}_p}^{~} W_p(\{\sigma_k\})
|\sigma_1^{~}\sigma_2^{~}\cdots\sigma_N^{~}\rangle \, ,
\label{Psi}$$ where $N$ stands for the total number of the lattice spins. The basis $\sigma_j$ for $j=1, \dots, N$ denotes a binary state, for which we use the arrow notation ${\downarrow}$ or ${\uparrow}$ in the following. The summation runs over the $2^N$ base spin states $|\sigma_1^{~}\sigma_2^{~}\cdots\sigma_N^{~}\rangle$, and $W_p(\{ \sigma_k \})$ are the elements of the $p$-rank tensor $W_p$ depending on $p$ spins $\sigma_{k_1}, \dots, \sigma_{k_p}$ on the $k^{th}$ lattice polygon. The symbol $\{\sigma_k\}$ stands for one of the $2^p$ base configurations of a multi-spin variable representing the group of spins $\sigma_{k_1}, \dots, \sigma_{k_p}$. All the tensors $W_p$ are considered to be identical, therefore, the set of $2^p$ tensor elements $W_p(\{\sigma\})$, where the subscript $k$ has been omitted due to the uniformity of the tensors $W_p$, uniquely describes the state $ |\Psi_p\rangle$, i.e. $|\Psi_p\rangle = |\Psi_p[W_p(\{\sigma\})]\rangle$.
We regard $|\Psi_p^{*}\rangle$ as the best approximation of the ground-state within the class of TPS $|\Psi_p\rangle$, if the minimum of the energy normalized per bond, $${ E}_0^{(p)} \equiv \min\limits_{\Psi_p}\lim\limits_{N_b\to\infty}\frac{1}{N_b}
\frac{\langle \Psi_p | {\cal H} | \Psi_p \rangle}
{\langle \Psi_p | \Psi_p \rangle}
\, ,
\label{Eg}$$ is obtained for $|\Psi_p^{*}\rangle$. Here, $N_b$ denotes the total number of bonds in the system. The energy ${E}_0^{(p)}$, due to its variational origin, serves as an upper bound of the true ground-state energy per bond ${\cal E}_{0}^{(p)}$.
Since the structure of every local Hamiltonian $G_{k}^{(p)}$ does not depend on $k$ (we investigate the system in the thermodynamic limit), the variational problem in is equivalent to minimization of the local energy per bond of an arbitrary polygon in the lattice center (in order to avoid boundary effects) $${E}_0^{(p)} = \min\limits_{\Psi_p}
\frac{2}{p}\frac{\langle \Psi_p | {G_\ell^{(p)}} | \Psi_p \rangle}
{\langle \Psi_p | \Psi_p \rangle} \, ,
\label{Egbond}$$ where $\ell$ is the index of the selected polygon and the normalization factor $2/p$ reflects the fact that the $p$ bonds of each polygon are shared with its neighbors. Moreover, if we utilize the tensor product structure of the state $|\Psi_p\rangle$, we can express the denominator $\langle \Psi_p | \Psi_p \rangle \equiv {\cal D}(W_p(\{\sigma\}))$ and the numerator $\langle \Psi_p | {G_\ell^{(p)}} | \Psi_p \rangle \equiv {\cal N}(W_p(\{\sigma\}))$ as functions of the tensor elements $W_p(\{\sigma\})$ only. Consequently, our variational problem transforms onto a multi-dimensional minimization over $2^p$ tensor elements $W_p(\{\sigma\})$ $${E}_0^{(p)} = \min\limits_{W_p(\{\sigma\})}
\frac{2}{p}\frac{{\cal N}(W_p(\{\sigma\}))}{{\cal D}(W_p(\{\sigma\}))} \, .
\label{e0}$$
Furthermore, symmetries of the local Hamiltonian $G_{\ell}^{(p)}$ may significantly reduce the dimension of the problem. Rotational and spin-ordering symmetries are present in all the three spin models. As a typical example, let us consider a hexagonal lattice ($p=6$) and its particular base configuration of spins on the lattice polygon $\{\sigma^{*}\} = \{{\uparrow}{\downarrow}{\uparrow}{\uparrow}{\downarrow}{\downarrow}\}$. Rotational symmetry requires that the tensor elements corresponding to the set of configurations $\{{\downarrow}{\uparrow}{\downarrow}{\uparrow}{\uparrow}{\downarrow}\}$, $\{{\downarrow}{\downarrow}{\uparrow}{\downarrow}{\uparrow}{\uparrow}\}$, $\{{\uparrow}{\downarrow}{\downarrow}{\uparrow}{\downarrow}{\uparrow}\}$, $\{{\uparrow}{\uparrow}{\downarrow}{\downarrow}{\uparrow}{\downarrow}\}$, $\{{\downarrow}{\uparrow}{\uparrow}{\downarrow}{\downarrow}{\uparrow}\}$ are identical to $W_{p=6}(\{\sigma^{*}\})$. Next, let us consider a spin-ordering operation, which reverses the order of the polygon spins. In particular, if the spins are labelled clockwise, the operation reorders them in the anti-clockwise direction. It means that the configuration $\{{\uparrow}{\downarrow}{\uparrow}{\uparrow}{\downarrow}{\downarrow}\}$ is equivalent to $\{{\downarrow}{\downarrow}{\uparrow}{\uparrow}{\downarrow}{\uparrow}\}$ by the spin-ordering symmetry and to all the rotations of the latter configuration ($\{{\uparrow}{\downarrow}{\downarrow}{\uparrow}{\uparrow}{\downarrow}\}$, $\{{\downarrow}{\uparrow}{\downarrow}{\downarrow}{\uparrow}{\uparrow}\}$, $\{{\uparrow}{\downarrow}{\uparrow}{\downarrow}{\downarrow}{\uparrow}\}$, $\{{\uparrow}{\uparrow}{\downarrow}{\uparrow}{\downarrow}{\downarrow}\}$, $\{{\downarrow}{\uparrow}{\uparrow}{\downarrow}{\uparrow}{\downarrow}\}$) by a composition of the spin-ordering and the rotational symmetry. As a result, the 12 tensor elements $W_6(\{\sigma\})$ corresponding to the configuration $\{\sigma^{*}\}$ and its 11 equivalent configurations are represented by a single variational parameter, as they share the same value.
By performing a similar analysis on the set of all $2^p$ configurations $\{\sigma\}$ we can factorize it into $N_{\rm Ising}^{(p)}$ classes of equivalence with representatives $\theta_j$, where $j \in \{1, \dots, N_{\rm Ising}^{(p)}\}$ [@DaniskaGendiar]. Thus, in case of a system with the rotational and the spin-ordering symmetry (as in the TFIM), there are only $N_{\rm Ising}^{(p)}$ free variational parameters $W_p(\theta_j)$ within the set of $2^p$ tensor elements $W_p(\{\sigma\})$. If there is no preferred spin alignment in the system (such as in the XY model, the Heisenberg model, as well as in the TFIM at and above the phase transition magnetic field), the spin-inversion symmetry appears. For instance, if $p=4$, the configuration $\{{\uparrow}{\uparrow}{\uparrow}{\downarrow}\}$ is equivalent to $\{{\downarrow}{\downarrow}{\downarrow}{\uparrow}\}$, which is obtained by flipping each spin. Such an additional symmetry results in consequent reduction of the set of the free variational parameters, the size of which drops to $N_{\rm Heis}^{(p)} < N_{\rm Ising}^{(p)}$. The numbers of the free variational parameters $N_{\rm Ising}^{(p)}$ and $N_{\rm Heis}^{(p)}$ with respect to the lattice parameter $p$ are summarized in table \[Tab1\]. In addition, one more variational parameter can be eliminated from each set of the free variational parameters by setting it to $1$, being the normalization condition in $W_p(\{\sigma\})$ and $| \Psi_p \rangle$, consequently.
$p$ $4$ $5$ $6$ $7$ $8$ $9$ $10$ $11$
---------------------------------- ----- ----- ------ ------ ------ ------ ------ -------
$N_{\rm Heis}^{(p)}$ $4$ $4$ $8$ $9$ $18$ $23$ $44$ $63$
\[0.04cm\] $N_{\rm Ising}^{(p)}$ $6$ $8$ $13$ $18$ $30$ $46$ $78$ $126$
\[0.04cm\]
: The numbers of the free variational parameters $N_{\rm Heis}^{(p)}$ (for the XY and the Heisenberg models) and $N_{\rm Ising}^{(p)}$ (for the TFIM) including the normalization parameter.
\[Tab1\]
The free variational parameters $W_p(\theta_j)$ are optimized numerically by TPVF [@DaniskaGendiar]. It is based on the fact that the product structure of the state $|\Psi_p\rangle$ allows to calculate the numerator ${\cal N}(W_p(\{\sigma\}))$ and the denominator ${\cal D}(W_p(\{\sigma\}))$ in for the given set of the tensor elements $W_p(\{\sigma\})$ by an appropriate modification of the CTMRG algorithm. Having applied the CTMRG as the effective and accurate numerical tool for calculation of the ratio in , a multi-dimensional minimizer is used for optimizing the variational parameters $W_p(\theta_j)$ [@gsl1; @gsl2; @NM].
Numerical results
=================
XY and Heisenberg models
------------------------
We study the XY and the Heisenberg models at zero magnetic field, where these models are known to be critical in the Euclidean space. Therefore, there is no preferred direction (the spin alignment) in the system on the Euclidean lattice at $h \geq 0$, and the spin-inversion symmetry is present. We expect that the models on hyperbolic lattices also exhibit the spin-inversion symmetry. It enables to reduce the number of the free variational parameters $W_p(\theta_j)$ within the TPVF minimization part down to $N_{\rm Heis}^{(p)}$ as listed in table \[Tab1\]. Despite the significant reduction, the number of the free parameters $N_{\rm Heis}^{(p)}$ still grows fast with respect to the increasing lattice parameter $p$. The computational time of the minimization algorithm is significantly prolonged due to (at least) linear dependence on the increasing number of the free variational parameters. Also, the algorithm may possibly be trapped in a local energy minimum and thus a series of initial conditions has to be tested in order to obtain the global energy minimum (or, at least, a sufficiently good approximation of it). For all these reasons, the calculations were stopped at $p=11$ with respect to the constraints of our computational resources and time.
---------- --------------- --------------
XY Heisenberg
$4$ $-1.08456618$ $-1.3089136$
$5$ $-1.08151200$ $-1.2912704$
$6$ $-1.08097046$ $-1.2925639$
$7$ $-1.08086301$ $-1.2918936$
$8$ $-1.08084068$ $-1.2919769$
$9$ $-1.08083585$ $-1.2919403$
$10$ $-1.08083478$ $-1.2919460$
$11$ $-1.08083453$ $-1.2919437$
$\infty$ $-1.08083446$ $-1.291944$
---------- --------------- --------------
: The ground-state energies per bond $E_0^{(p)}$ listed with respect to $p$ for the Heisenberg and XY models. The number of block spin states [@White; @Nishino] kept was $m=20$ for $4\leq p \leq 10$ and $m=10$ for $p=11$. The asymptotic estimate of $E_0^{(\infty)}$ corresponds to the model on the Bethe lattice.
\[Tab2\]
The ground-state energies $E_0^{(p)}$ obtained by the TPVF algorithm for both the XY and the Heisenberg models are summarized in table \[Tab2\]. The energies $E_0^{(p)}$ remained identical even if the larger set of $N_{\rm Ising}^{(p)}$ free variational parameters $W_p(\theta_j)$ in TPVF was used, whereby the optimal values of the parameters $W^{*}_p(\theta_j)$ coupled by spin-inversion symmetry were equal. These results witness the spin-inversion symmetry of the models on hyperbolic lattices. Recall that $E_0^{(p)}$ represents only an upper estimate of the true ground-state energy ${\cal E}_0^{(p)}$. We have shown that the energies $E_0^{(4)}$ of the referencing Euclidean square lattice calculated by TPVF were higher if compared to the Monte Carlo simulations (the relative errors for the XY and the Heisenberg models, respectively, are $1.2\%$ and $2.2\%$) [@DaniskaGendiar]. This observation can be explained by suppression of the quantum long-range correlations induced by the TPS approximation of the low-dimensional uniform tensors $W_4$, which cannot correctly reproduce the divergence of the correlation length in the models on the square lattice. On the other hand, any quantum spin model on hyperbolic lattice belongs to the mean-field universality class, because the hyperbolic lattices exhibit the infinite Hausdorff dimension, which significantly exceeds the critical lattice dimension $D_c=3$ [@Baxter]. Because of the mean-field-like character of the TPS approximation, the TPVF algorithm is expected to be more accurate whenever a hyperbolic lattice geometry is considered [@DaniskaGendiar; @hctmrg-Ising-p-4].
![The ground-state energy $E_0^{(p)}$ of the XY model with respect to the lattice parameter $p \in \{4,5,\dots, 11\}$. The inset shows the zoomed-in energy including the details of the fitting function.[]{data-label="Fig2"}](Fig2.pdf){width="70.00000%"}
![The ground-state energy $E_0^{(p)}$ of the Heisenberg model with respect to the lattice parameter $p \in \{4,5,\dots, 11\}$. The fitting function parameters are shown in the inset.[]{data-label="Fig3"}](Fig3.pdf){width="70.00000%"}
Figure \[Fig2\] illustrates a monotonically increasing and rapidly saturating curve of the energy $E_0^{(p)}$ for the XY model with respect to the lattice parameter $p$. The inset depicts the tail of the curve in detail together with an exponential fit applied to the five energies $E_0^{(7)}, \dots, E_0^{(11)}$. The case $p=4$, where the TPVF algorithm is not sufficiently accurate for the reasons mentioned above, was excluded from the extrapolation analysis. The fitting function is proposed in the form $$E_0^{(p)} = E_0^{(\infty)} +a_1 \exp(a_2 p ) \, ,
\label{fitXY}$$ where $E_0^{(\infty)}$, $a_1$, and $a_2$ are the fitting parameters, which were determined in the following way. First we defined a function $f(E)$, which returns the residual sum of squares ($RSS$) of the linear regression $\ln\vert E-E_0^{(p)}\vert = \ln\vert a_1\vert +a_2 p $. Then, $E_0^{(\infty)}$ was chosen as the argument, which minimizes the function $f(E)$. The corresponding linear regression $\ln\vert E_0^{(\infty)}-E_0^{(p)}\vert = \ln\vert a_1\vert + a_2 p $ specifies the parameters $a_1$ and $a_2$. If considering another way, $E_0^{(\infty)}$ is such a value that the curve $\ln\vert E_0^{(\infty)}-E_0^{(p)}\vert$ is as close as possible to a line, where the closeness is measured by the $RSS$. Thus obtained parameters $E_0^{(\infty)}$, $a_1$, and $a_2$ are listed in the inset of figure \[Fig2\], where the dot-dashed line represents the estimate of the ground-state energy per bond of the quantum XY model on the Bethe lattice $E_0^{(\infty)}=\lim\limits_{p \rightarrow \infty} E_0^{(p)}=-1.08083446$.
Analogously, the ground-state energies $E_0^{(p)}$ of the Heisenberg model are plotted in figure \[Fig3\]. Again, rapid convergence of the energy to the asymptotic values is obvious from the data. Although we have not clarified the physical origin of the non-monotonic convergence (saw-like pattern) of $E_0^{(p)}$ yet, a detailed analysis indicates that the exponential fitting function in can successfully describe the data, if applied separately onto two sets: those with even $p \in \{6, 8, 10\}$ (the lower branch shown in the inset) and the odd $p \in \{5, 7, 9, 11\}$ (the upper branch). The fitting parameters of the two regressions are listed in the inset of figure \[Fig3\]. The lower and the upper branches yield the energies $E_0^{(\infty)}$ $-1.2919443$ and $E_0^{(\infty)}-1.2919440$, respectively. With respect to an independent application of additional analogous fits, we found $E_0^{(\infty)}=-1.291944$ (all the digits are valid) to be considered as the correct estimate of the ground-state energy per bond of the Heisenberg model on the Bethe lattice.
We have not found any theoretical reasoning for the exponential convergence of the ground-state energies $E_0^{(p)}$ yet. However, if a power-law fitting function was applied instead, we obtained a less accurate fitting and greater $RSS$.
Transverse field Ising model
----------------------------
The TFIM undergoes a quantum phase transition at a nonzero magnetic field $h_t^{(p)}>0$, where we explicitly emphasize its dependence on the lattice geometry. The nonzero spontaneous magnetization in the ordered phase at $h<h_t^{(p)}$ breaks the spin-inversion symmetry, which results in approximately twice larger set of the free variational parameters $N_{\rm Ising}^{(p)}$ in the TPVF algorithm if compared to $N_{\rm Heis}^{(p)}$ in the XY and Heisenberg models, cf. table \[Tab1\]. The computational time for a particular fixed field $h$ is, therefore, significantly prolonged. Moreover, in order to screen the vicinity of the phase transition field $h_t^{(p)}$, multiple calculations for a sequence of magnetic fields $h$ had to be performed. As a consequence, in order to restrict the total computational time, we have analyzed the TFIM on the hyperbolic lattices up to $p=10$ only. (Notice that the number of block spins states kept was $m=20$ for $p \in \{4,5, \dots, 8\}$, and only $m=4$ for $p \in \{9,10\}$, which was sufficient due to exponentially weak correlations caused by the hyperbolic lattice geometry [@hctmrg-Ising-3-q]; any further increase of the states kept $m$ has not improved the numerical calculations significantly).
We have analyzed the phase transition of the TFIM by the expectation value of the spontaneous magnetization $\langle S_p^{z} \rangle$ as well as by the magnetic susceptibility $\chi_p$. Solving the minimization problem in , we received the optimal tensor elements $W^*_p(\{\sigma\})$, which uniquely define the approximative ground state $|\Psi_p^*\rangle$ via . Once $|\Psi_p^*\rangle$ has been constructed, we evaluated the spontaneous magnetization $$\langle S_p^{z} \rangle = \frac{\langle \Psi^*_p | S_{c}^{z} |\Psi^*_p \rangle}
{\langle \Psi^*_p | \Psi^*_p \rangle} \, ,
\label{sz}$$ where ${c}$ labels an arbitrary spin in the central polygon of the lattice in order to suppress boundary effects. The resulting dependence of the magnetization $\langle S_p^{z}\rangle$ with respect to the magnetic field $h$ near the phase transition field $h_t^{(p)}$ is plotted in the upper graph of figure \[Fig4\]. The quantum phase transition of the TFIM is characterized by a non-analytic behavior of the magnetization curve, when $\langle S_p^{z}\rangle \to 0$ if approaching the phase transition field $h \to h_t^{(p)}$ from the ordered phase ($h < h_t^{(p)}$). The phase transition exponent $\beta_p$, which depends on the lattice geometry, describes the singularity through the scaling relation in the ordered phase $$\langle S_p^z (h) \rangle \propto {\left(h_t^{(p)} - h \right)}^{\beta_p} \, .
\label{powerlaw}$$ Figure \[Fig4\] (the lower graph) shows the squared transversal magnetization $ {\langle S_p^{z}\rangle}^{2}$, where we point out the linearity of the squared magnetization if approaching the phase transition field $h_t^{(p)}$. Such a dependence confirms the mean-field exponent $\beta_p=\frac{1}{2}$ regardless of the lattice parameter $p$, which results in the mean-field-like behavior of the TFIM if approaching the phase transition.
![The spontaneous magnetization $\langle S_p^{z}\rangle$ (the upper graph) and its square ${\langle S_p^{z}\rangle}^{2}$ (the lower graph) in the vicinity of the phase transitions with respect to the magnetic field $h$ for $p \in \{4,5, \dots, 10\}$. The inset shows the detailed zoomed-in behavior for higher values of $p$.[]{data-label="Fig4"}](Fig4.pdf){width="70.00000%"}
![The phase transition field $h_t^{(p)}$ of the TFIM with respect to the lattice parameter $p$. The horizontal dot-dashed line represents the estimated asymptotic value $h_t^{(\infty)}=3.29332$.[]{data-label="Fig5"}](Fig5.pdf){width="70.00000%"}
The phase transition fields $h_t^{(p)}$, calculated according to the method described in [@DaniskaGendiar], are summarized in table \[Tab3\] together with their errors $\Delta^{(p)}$. Notice that $\Delta^{(p)}$ represents only the error of the method providing that the calculated magnetization $\langle S_p^{z}\rangle$ is considered accurate. The data are graphically plotted in figure \[Fig5\], whereas the error bars are too small to be displayed. Using an analogous exponential fitting function applied on the critical magnetic fields $h_t^{(p)}$ for $p \in \{6, \dots, 10\}$, (cf. ), we calculated the asymptotic phase transition field of the TFIM on the Bethe lattice $h_t^{(\infty)}=3.29332$ as listed in Tab. \[Tab3\].
$p$ $4$ $5$ $6$ $7$
---------------- ------------------ ------------------ ------------------ ------------------
$h_t^{(p)}$ $3.158034$ $3.263825$ $3.285405$ $3.291055$
$\Delta^{(p)}$ $1\times10^{-6}$ $1\times10^{-6}$ $1\times10^{-6}$ $1\times10^{-6}$
$p$ $8$ $9$ $10$ $\infty$
$h_t^{(p)}$ $3.292647$ $3.293113$ $3.293263$ $3.29332$
$\Delta^{(p)}$ $2\times10^{-6}$ $2\times10^{-6}$ $5\times10^{-6}$ $1\times10^{-5}$
: The phase transition fields $h_t^{(p)}$ of the TFIM including the estimated errors $\Delta^{(p)}$ with respect to the lattice parameter $p$.
\[Tab3\]
Another independent way of obtaining (and confirming) the phase transition fields $h_t^{(p)}$ can be carried out by analyzing the magnetic susceptibility $$\chi_p=-\frac{\partial^2 E_0^{(p)}}{\partial h^2}\, .
\label{chi}$$ The functional dependence of the susceptibility on the magnetic field $h$ is shown in figure \[Fig6\]. A non-diverging discontinuity of $\chi_p$ occurs at the identical phase transition fields $h_t^{(p)}$, which we have determined above by the spontaneous magnetization analysis and are depicted by the vertical dot-dashed lines. The inaccuracy comes from performing the second derivative in numerically, and the additional improvement rests in decreasing the spacing interval $\delta h$, i.e, in shrinking the distance between the magnetic fields, at which the ground-state energy is evaluated by TPVF. In the limit $\delta h \to 0$, the magnetic susceptibility undergoes a discontinuous jump at $h_t^{(p)}$ [@DaniskaGendiar]. It is obvious that there is no significant difference between the phase transition magnetic fields $h_t^{(p)}$ obtained by the analysis of the transverse magnetization $\langle S_p^{z}\rangle$ and the magnetic susceptibility $\chi_p$.
Except for the analysis of the phase transition by the spontaneous magnetization $\langle S_p^{z}\rangle$ and the magnetic susceptibility $\chi_p$, the field dependence of the set of the optimal free variational parameters $W^*_p(\theta_j)$ also provides helpful information about the phase transition $h_t^{(p)}$. The pairs of the optimal variational parameters $W^*_p(\theta_j)$ coupled by spin-inversion symmetry smoothly collapse onto a single curve exactly at the phase transition for all considered lattice geometries. This process follows the identical behavior as we have presented in [@DaniskaGendiar]. However, due to the large number of the variational parameters $N_{\rm Ising}^{(p)}$, we do not plot the $h$-dependence of $W^*_p(\theta_j)$ since the behavior remains qualitatively unchanged.
![The magnetic susceptibility $\chi_p$ of the TFIM as a function of the magnetic field $h$ for the hyperbolic lattices with $p \in \{5, \dots,10\}$. The vertical dot-dashed lines serve as guides for the eye and correspond to the phase transitions $h_t^{(p)}$.[]{data-label="Fig6"}](Fig6.pdf){width="70.00000%"}
Conclusions
===========
We have investigated three quantum spin-$\frac{1}{2}$ models (Heisenberg, XY, and TFIM) on a series of hyperbolic lattices by means of the numerical algorithm Tensor Product Variational Formulation [@DaniskaGendiar]. The series of lattices is constructed by tessellation of regular $p$-sided polygons with the fixed coordination number equal to four, where $p \in \{5, \dots,11\}$. The Euclidean square lattice ($p=4$) has been also considered as a reference lattice, although we have discussed in [@DaniskaGendiar] that TPVF applied to the models on the square lattice is less accurate than on the hyperbolic lattices ($p>4$).
The ground-state energies $E_0^{(p)}$ of the XY and the Heisenberg models have been studied in the absence of magnetic field on the series of the regular lattices with $4\leq p \leq 11$. Since no spontaneous symmetry breaking occurs in the two models at $h=0$, the spin-inversion symmetry helps to accelerate the TPVF algorithm due to significant reduction of the number of the free variational parameters. The resulting dependence of the ground-state energy per bond $E_0^{(p)}$ on the lattice geometry $p$ differs considerably for the two models. While the energies $E_0^{(p)}$ of the XY model form a monotonically increasing and exponentially saturated sequence with increasing $p$, the Heisenberg model induces a saw-like dependence containing the separated upper (odd $p$) and the lower (even $p$) branches, both of them converging exponentially fast to the common asymptotic value $E_0^{(\infty)}$ which corresponds to the ground-state energy on the Bethe lattice with the coordination number four.
We have analyzed the phase transition magnetic fields $h_t^{(p)}$ of the TFIM by the expectation value of the spontaneous magnetization $\langle S_p^z \rangle$, the associated magnetic exponent $\beta_p$, and the magnetic susceptibility $\chi_p$. We have calculated a sequence of the phase transition magnetic fields $h_t^{(p)}$, which is a strictly monotonous and increasing function, which converges exponentially to the asymptotic value $h_t^{(\infty)}$. This feature is completely analogous to a fast exponential saturation of the critical temperatures $T_c^{(p)}$ we had observed for the classical Ising model on the identical series of hyperbolic lattices in our earlier studies [@hctmrg-Ising-5-4; @hctmrg-Ising-p-4]. However, we have not found physical interpretation of this phenomenon yet. The quantum spin systems (as well as the classical ones) investigated on the hyperbolic lattices belong to the mean-field universality class, since infinite Hausdorff dimension of the hyperbolic lattice geometry exceeds the critical lattice dimensions $D_c = 3$ (for quantum models) and $D_c=4$ (for the classical ones). The linearity of the squared magnetization in the vicinity of the phase transition confirms the mean-field-like behavior, in which the associated magnetic exponents $\beta_p = \frac{1}{2}$.
Although the set of the calculated phase transition magnetic fields $h_t^{(p)}$ and the ground-state energies $E_0^{(p)}$ are restricted to $4 \leq p \leq 11$, which is far away from the asymptotics $p\to\infty$, the fast convergence and the exponential character of $h_t^{(p)}$ and $E_0^{(p)}$ with increasing $p$ enables to estimate the respective quantities of the quantum spin models on the Bethe lattice ($p \to \infty$). In particular, we conjecture that the phase transition field of the TFIM on the Bethe lattice is positioned at $h_t^{(\infty)} = 3.29332$ and the ground-state energies per bond of the XY and the Heisenberg models, respectively, occur at $E_0^{(\infty)} = -1.08083446 $ and $-1.291944$. The three quantum spin models have not yet been considered on the Bethe lattice with the coordination number four.
The support received from the Grants QIMABOS APVV-0808-12 and VEGA-2/0130/15 is acknowledged.
References {#references .unnumbered}
==========
[99]{} R. J. Baxter, Exactly Solved Models in Statistical Mechanics (Academic Press, London, 1982). H. Bethe, Zeitschrift für Physik, [**71**]{}, 205 (1931); R. Orbach, Phys. Rev. [**112**]{}, 309 (1958); L. R. Walker, Phys. Rev. [**116**]{}, 1089 (1959); C. N, Yang and C. P. Yang, Phys. Rev. [**150**]{}, 321 (1966). E. K. Sklyanin, L. A. Takhtajan, and L. D. Faddeev, Theor. Math. Phys. [**40**]{}, 688 (1979); L. A. Takhtajan and L. D. Faddeev, Russ. Math. Surveys. [**34**]{}, 11 (1979). M. Jimbo and T. Miwa, Algebraic analysis of solvable lattice models (AMS, 1995). R. Orus, Annals of Physics [**349**]{}, 117 (2014). F. Verstraete, V. Murg, I. Cirac, Adv. in Phys. [**57**]{}, 143 (2008). P. Corboz, G. Evenbly, F. Verstraete, G. Vidal, Phys. Rev. A [**81**]{}, 010303(R) (2010); C. V. Kraus, N. Schuch, F. Verstraete, J. I. Cirac, Phys. Rev. A [**81**]{}, 052338 (2010) M. Levin and C. P. Nave, Phys. Rev. Lett. [**99**]{}, 120601 (2007). Z.Y. Xie, J. Chen, M.P. Qin, J.W. Zhu, L.P. Yang, T. Xiang, Phys. Rev. B [**86**]{}, 045139 (2012). J. Maldacena, Int. J. Theor. Phys. [**38**]{}, 1113 (1999). S. K. Baek, P. Minnhagen, H. Shima, and B. J. Kim, Phys. Rev. E [**80**]{}, 011133 (2009). S. K. Baek, H. Shima, and B. J. Kim, Phys. Rev. E [**79**]{}, 060106(R) (2009). S. R. White, Phys. Rev. Lett. [**69**]{}, 2863 (1992); Phys. Rev. B [**48**]{}, 10345 (1993). U. Schollwöck, Rev. Mod. Phys. [**77**]{}, 259 (2005). U. Schollwöck, Annals of Physics [**326**]{}, 96 (2011). M. Daniška and A. Gendiar, J. Phys. A: Math. Theor. [**48**]{}, 435002 (2015). T. Nishino, J. Phys. Soc. Jpn. [**65**]{}, 891 (1996). R. Mosseri and J. F. Sadoc, J. Physique Lett. [**43**]{}, 249 (1982). K. Ueda, R. Krčmár, A. Gendiar, and T. Nishino, J. Phys. Soc. Japan [**76**]{}, 084004 (2007). R. Krčmár, A. Gendiar, K. Ueda, and T. Nishino, J. Phys. A: Math. Gen. [**41**]{}, 125001 (2008). A. Gendiar, R. Krčmár, S. Andergassen, M. Daniška, and T. Nishino, Phys. Rev. E [**86**]{}, 021105 (2012). A. Gendiar, M. Daniška, R. Krčmár, and T. Nishino, Phys. Rev. E [**90**]{}, 012122 (2014). http://www.gnu.org/software/gsl/ M. Galassi [*et al.*]{}, GNU Scientific Library Reference Manual (3rd edn), Network Theory, ISBN 0954612078 (2009). J. A. Nelder and R. Mead, Computer Journal [**7**]{}, 308 (1965).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Quickly moving to a new area of research is painful for researchers due to the vast amount of scientific literature in each field of study. One possible way to overcome this problem is to summarize a scientific topic. In this paper, we propose a model of summarizing a single article, which can be further used to summarize an entire topic. Our model is based on analyzing others’ viewpoint of the target article’s contributions and the study of its citation summary network using a clustering approach.'
author:
- |
Vahed Qazvinian\
School of Information\
University of Michigan\
Ann Arbor, MI\
[[email protected]]{} Dragomir R. Radev\
Department of EECS and\
School of Information\
University of Michigan\
Ann Arbor, MI\
[[email protected]]{}
bibliography:
- 'ref.bib'
title: Scientific Paper Summarization Using Citation Summary Networks
---
Introduction
============
It is quite common for researchers to have to quickly move into a new area of research. For instance, someone trained in text generation may want to learn about parsing and someone who knows summarization well, may need to learn about question answering. In our work, we try to make this transition as painless as possible by automatically generating summaries of an entire research topic. This enables a researcher to find the chronological order and the progress in that particular field of study. An ideal such system will receive a topic of research, as the user query, and will return a summary of related work on that topic. In this paper, we take the first step toward building such a system.
Studies have shown that different citations to the same article often focus on different aspects of that article, while none alone may cover a full description of its entire contributions. Hence, the set of citation summaries, can be a good resource to understand the main contributions of a paper and how that paper affects others. The *citation summary* of an article ($A$), as defined in , is a the set of citing sentences pointing to that article. Thus, this source contains information about $A$ from others’ point of view. Part of a sample citation summary is as follows:
[*In the context of DPs, this method was proposed by [**(Eisner, 1996)**]{}.*]{}
[*[**Eisner (1996)**]{} gave a with a algorithm based on an edge factorization of trees.*]{}
[*Eisner [**(Eisner, 1996)**]{} proposed an .*]{}
[*If the parse has to be projective, Eisner’s can be used for the search.*]{}
The problem of summarizing a whole scientific topic, in its simpler form, may reduce to summarizing one particular article. A citation summary can be a good resource to make a summary of a target paper. Then using each paper’s summary and some knowledge of the citation network, we’ll be able to generate a summary of an entire topic. Analyzing citation networks is an important component of this goal, and has been widely studied before [@Newman01a].
Our main contribution in this paper is to use citation summaries and network analysis techniques to produce a summary of a single scientific article as a framework for future research on topic summarization. Given that the citation summary of any article usually has more than a few sentences, the main challenge of this task is to find a *subset* of these sentences that will lead to a better and shorter summary.
Related Work {#sect:relatedwork}
------------
Although there has been work on analyzing citation and collaboration networks [@teufel2006; @Newman01a] and scientific article summarization [@simone2002], to the knowledge of the author there is no previous work that study the text of the citation summaries to produce a summary. [@Bradshaw03; @Bradshaw02] get benefit from citations to determine the content of articles and introduce “Reference Directed Indexing” to improve the results of a search engine.
In other work, [@NanbaKO00; @NanbaEtal04] analyze citation sentences and automatically categorize citations into three groups using 160 pre-defined phrase-based rules. This categorization is then used to build a tool for survey generation. also discuss the same citation categorization to support a system for writing a survey. report that co-citation implies similarity by showing that the textual similarity of co-cited papers is proportional to the proximity of their citations in the citing article.
Previous work has shown the importance of the citation summaries in understanding what a paper says. The *citation summary* of an article $A$ is the set of sentences in other articles which cite $A$. performed a large-scale study on citation summaries and their importance. They conducted several experiments on a set of $2,497$ articles from the free PubMed Central (PMC) repository[^1]. Results from this experiment confirmed that the “Self Cohesion” of a citation summary of an article is consistently higher than the that of its abstract.
Cluster Nodes Edges
--------- ------- -- -- -------
DP 167 323
PBMT 186 516
Summ 839 1425
QA 238 202
TE 56 44
: Clusters and their citation network size[]{data-label="tbl:cluster"}
also conclude that citation summaries are more focused than abstracts, and that they contain additional information that does not appear in abstracts. [@Kupiec95] use the abstracts of scientific articles as a target summary, where they use 188 Engineering Information summaries that are mostly indicative in nature. Abstracts tend to summarize the document’s topics well, however, they don’t include much use of metadata. [@kan2002] use annotated bibliographies to cover certain aspects of summarization and suggest guidelines that summaries should also include metadata and critical document features as well as the prominent content-based features.
Siddharthan and Teufel describe a new task to decide the scientific attribution of an article [@simone2007] and show high human agreement as well as an improvement in the performance of Argumentative Zoning [@teufel2005]. Argumentative Zoning is a rhetorical classification task, in which sentences are labeled as one of [Own, Other, Background, Textual, Aim, Basis, Contrast]{} according to their role in the author’s argument. These all show the importance of citation summaries and the vast area for new work to analyze them to produce a summary for a given topic.
ACL-ID Title Year CS Size
-- ---------- --------------------------------------------------------------------------------------------- ------ ---------
C96-1058 Three New Probabilistic Models For Dependency Parsing: An Exploration 1996 66
P97-1003 Three Generative, Lexicalized Models For Statistical Parsing 1997 55
P99-1065 A Statistical Parser For Czech 1999 54
P05-1013 Pseudo-Projective Dependency Parsing 2005 40
P05-1012 On-line Large-Margin Training Of Dependency Parsers 2005 71
N03-1017 Statistical Phrase-Based Translation 2003 180
W03-0301 An Evaluation Exercise For Word Alignment 2003 14
J04-4002 The Alignment Template Approach To Statistical Machine Translation 2004 50
N04-1033 Improvements In Phrase-Based Statistical Machine Translation 2004 24
P05-1033 A Hierarchical Phrase-Based Model For Statistical Machine Translation 2005 65
A00-1043 Sentence Reduction For Automatic Text Summarization 2000 19
A00-2024 Cut And Paste Based Text Summarization 2000 20
C00-1072 The Automated Acquisition Of Topic Signatures For Text Summarization 2000 19
W00-0403 Centroid-Based Summarization Of Multiple Documents: Sentence Extraction, ... 2000 31
W03-0510 The Potential And Limitations Of Automatic Sentence Extraction For Summarization 2003 15
A00-1023 A Question Answering System Supported By Information Extraction 2000 13
W00-0603 A Rule-Based Question Answering System For Reading Comprehension Tests 2002 19
P02-1006 Learning Surface Text Patterns For A Question Answering System 2002 74
D03-1017 Towards Answering Opinion Questions: Separating Facts From Opinions ... 2003 42
P03-1001 Offline Strategies For Online Question Answering: Answering Questions Before They Are Asked 2003 27
D04-9907 Scaling Web-Based Acquisition Of Entailment Relations 2004 12
H05-1047 A Semantic Approach To Recognizing Textual Entailment 2005 8
H05-1079 Recognising Textual Entailment With Logical Inference 2005 9
W05-1203 Measuring The Semantic Similarity Of Texts 2005 17
P05-1014 The Distributional Inclusion Hypotheses And Lexical Entailment 2005 10
Data {#sect:Data}
====
The ACL Anthology is a collection of papers from the Computational Linguistics journal, and proceedings from ACL conferences and workshops and includes almost $11,000$ papers. To produce the [**A**]{}CL [**A**]{}nthology [**N**]{}etwork (AAN), manually performed some preprocessing tasks including parsing references and building the network metadata, the citation, and the author collaboration networks. The full AAN includes all citation and collaboration data within the ACL papers, with the citation network consisting of $8,898$ nodes and $38,765$ directed edges.
Clusters
--------
We built our corpus by extracting small clusters from the AAN data. Each cluster includes papers with a specific phrase in the title or content. We used a very simple approach to collect papers of a cluster, which are likely to be talking about the same topic. Each cluster consists of a set of articles, in which the topic phrase is matched within the title or the content of papers in AAN. In particular, the five clusters that we collected this way, are: [**D**]{}ependency [**P**]{}arsing (DP), [**P**]{}hrased [**B**]{}ased [**M**]{}achine [**T**]{}ranslation (PBMT), Text [**Summ**]{}arization (Summ), [**Q**]{}uestion [**A**]{}nswering (QA), and [**T**]{}extual [**E**]{}ntailment (TE). Table \[tbl:cluster\] shows the number of articles and citations in each cluster. For the evaluation purpose we chose five articles from each cluster. Table \[tbl:paperchoice\] shows the title, year, and citation summary size for the 5 papers chosen from each cluster. The citation summary size of a paper is the size of the set of citation sentences that cite that paper.
Analysis
========
Fact Distribution {#sect:factdist}
-----------------
We started with an annotation task on 25 papers, listed in Table \[tbl:paperchoice\], and asked a number of annotators to read the citation summary of each paper and extract a list of the main contributions of that paper. Each item on the list is a *non-overlapping contribution* ([***fact***]{}) perceived by reading the citation summary. We asked the annotators to focus on the citation summary to do the task and not on their background on this topic.
As our next step we manually created the union of the shared and similar facts by different annotators to make a list of facts for each paper. This fact list made it possible to review all sentences in the citation summary to see which facts each sentence contained. There were also some unshared facts, facts that only appear in one annotator’s result, which we ignored for this paper.
Table \[tbl:collins99\] shows the shared and unshared facts extracted by four annotators for P99-1065.
Fact Occurrences
-- ---------------------------------- -------------
$f_4$: “Czech DP” 10
$f_1$: “lexical rules” 6
$f_3$: “POS/ tag classification” 6
$f_2$: “constituency parsing” 5
$f_5$: “Punctuation” 2
$f_6$: “Reordering Technique” 2
$f_7$: “Flat Rules” 2
“Dependency conversion”
“80% UAS”
“97.0% F-measure”
“Generative model”
‘Relabel coordinated phrases’’
‘Projective trees’’
“Markovization”
: Facts of P99-1065
\[tbl:collins99\]
The manual annotation of P99-1065 shows that the fact “Czech DP” appears in 10 sentences out of all 54 sentences in the citation summary. This shows the importance of this fact, and that “Dependency Parsing of Czech” is one of the main contributions of this paper. Table \[tbl:collins99\] also shows the number of times each shared fact appears in P99-1065’s citation summary sorted by occurrence.
After scanning through all sentences in the citation summary, we can come up with a [***fact distribution matrix***]{} for an article. The rows of this matrix represent sentences in the citation summary and the columns show facts. A $1$ value in this matrix means that the sentence covers the fact. The matrix $D$ shows the fact distribution of P99-1065. IDs in each row show the citing article’s ACL ID, and the sentence number in the citation summary. These matrices, created using annotations, are particularly useful in the evaluation process.
$$D= \left(
\begin{array}{c|ccccccc}
& f_1 & f_2 & f_3& f_4 & f_5& f_6 & f_7\\ \hline
\tiny W06\textrm{-}2935\textrm{:}1 &1 &0 &0 &0 &0 &0 &0\\
W06\textrm{-}2935\textrm{:}2 &0 &0 &0 &0 &0 &0 &0\\
W06\textrm{-}2935\textrm{:}3 &0 &0 &1 &1 &0 &0 &0\\
W06\textrm{-}2935\textrm{:}4 &0 &0 &0 &0 &0 &0 &1\\
W06\textrm{-}2935\textrm{:}5 &0 &0 &0 &0 &0 &0 &0\\
W06\textrm{-}2935\textrm{:}6 &0 &0 &0 &0 &1 &0 &0\\
W05\textrm{-}1505\textrm{:}7 &0 &1 &0 &1 &0 &0 &0\\
W05\textrm{-}1505\textrm{:}8 &0 &0 &0 &0 &0 &1 &0\\
\vdots & \vdots & & &\vdots & & &\vdots \\
W05\textrm{-}1518\textrm{:}54 &0 &0 &0 &0 &0 &0 &0\\
\end{array}
\right)$$
Similarity Measures {#sect:similarity}
-------------------
We want to build a network with citing sentences as nodes and similarities of two sentences as edge weights. We’d like this network to have a nice community structure, whereby each cluster corresponds to a fact. So, a similarity measure in which we are interested is the one which results in high values for pairs of sentences that cover the same facts. On the other hand, it should return a low value for pairs that do not share a common contribution of the target article.
The following shows two sample sentences from P99-1065 that cover the same fact and we want the chosen similarity measure to return a high value for them:
[*So, Collins et al (1999) proposed a for parsing the treebank.*]{}
[*The of Collins et al (1999) was run on a different data set... .* ]{}
Conversely, we’d like the similarity of the two following sentences that cover no shared facts, to be quite low:
[*Collins (1999) explicitly added features to his parser to improve accuracy.*]{}
[*The trees are then transformed into Penn Treebank style constituencies- using the technique described in (Collins et al, 1999).* ]{}
We used P99-1065 as the training sample, on which similarity metrics were trained, and left the others for the test. To evaluate a similarity measure for our purpose we use a simple approach. For each measure, we sorted the similarity values of all pairs of sentences in P99-1065’s citation summary in a descending order. Then we simply counted the number of pairs that cover the same fact (out of $172$ such fact sharing pairs) in the top $100$, $200$ and $300$ highly similar ones out of total $2,862$ pairs. Table \[tbl:similarities\] shows the number of fact sharing pairs among the top highest similar pairs.
Measure Top $100$ Top $200$ Top $300$
------------------ ----------- ----------- -----------
tf-idf (General) $34$ $66$ $74$
tf-idf (AAN) $34$ $56$ $74$
LCSS $26$ $37$ $54$
tf $24$ $34$ $46$
tf2gen $13$ $26$ $35$
tf-idf (DP) $16$ $26$ $28$
Levenshtein $2$ $9$ $16$
: Different similarity measures and their performances in favoring fact-sharing sentences. Each column shows the number of fact-sharing pairs among top highly similar pairs.[]{data-label="tbl:similarities"}
Table \[tbl:similarities\] shows how cosine similarity that uses a tf-idf measure outperforms the others. We tried three different policies for computing IDF values to compute cosine similarity: a general IDF, an AAN-specific IDF where IDF values are calculated only using the documents of AAN, and finally DP-specific IDF in which we only used all-DP data set. Table \[tbl:similarities\] also shows the results for an asymmetric similarity measure, generation probability [@erkan:2006:HLT-NAACL06-Main] as well as two string edit distances: the longest common substring and the Levenshtein distance [@levenshtein66].
Methodology
===========
In this section we discuss our graph clustering method for article summarization, as well as other baseline methods used for comparisons.
Network-Based Clustering
------------------------
*The Citation Summary Network* of an article $A$ is a network in which each sentence in the citation summary of $A$ is a node. This network is a *complete undirected weighted graph* where the weight of an edge between two nodes shows the *similarity* of the two corresponding sentences of those nodes. The similarity that we use, as described in section \[sect:similarity\], is such that sentences with the same facts have high similarity values. In other words, strong edges in the citation summary network are likely to indicate a shared fact between two sentences.\
A graph clustering method tries to cluster the nodes of a graph in a way that the average intra-cluster similarity is maximum and the average inter-cluster similarity is minimum. To find the communities in the citation summary network we use [@Clauset04], a hierarchical agglomeration algorithm which works by greedily optimizing the modularity in a linear running time for sparse graphs.\
To evaluate how well the clustering method works, we calculated the *purity* for the clusters found of each paper. Purity [@manning07] is a method in which each cluster is assigned to the class with the majority vote in the cluster, and then the accuracy of this assignment is measured by dividing the number of correctly assigned documents by $N$. More formally: $$\mbox{purity}( \Omega,\mathbb{C} ) = \frac{1}{N} \sum_k \max_j \vert\omega_k \cap c_j\vert$$ where $\Omega = \{ \omega_1, \omega_2, \ldots, \omega_K \}$ is the set of clusters and $\mathbb{C} = \{ c_1,c_2,\ldots,c_J \}$ is the set of classes. $\omega_k$ is interpreted as the set of documents in $\omega_k$ and $c_j$ as the set of documents in $c_j$. For each evaluated article, Table \[tbl:purity\] shows the number of real facts ($|\mathbb{C}|= J$), the number of clusters ($|\Omega| = K$) and $purity( \Omega,\mathbb{C} )$ for each evaluated article. Figure \[fig:final1\] shows the clustering result for J04-4002, in which each color (number) shows a real fact, while the boundaries and capital labels show the clustering result.
ACL-ID \#Facts $|\mathbb{C}|$ \#Clusters $|\Omega|$ $Purity( \Omega,\mathbb{C} )$
-- ---------- ------------------------ ----------------------- -------------------------------
C96-1058 $4$ $4$ $0.636$
P97-1003 $5$ $5$ $0.750$
P99-1065 $7$ $7$ $0.724$
P05-1013 $5$ $3$ $0.689$
P05-1012 $7$ $5$ $0.500$
N03-1017 $8$ $4$ $0.464$
W03-0301 $3$ $3$ $0.777$
J04-4002 $5$ $5$ $0.807$
N04-1033 $5$ $4$ $0.615$
P05-1033 $6$ $5$ $0.650$
A00-1043 $5$ $4$ $0.812$
A00-2024 $5$ $2$ $0.333$
C00-1072 $3$ $4$ $0.857$
W00-0403 $6$ $4$ $0.682$
W03-0510 $4$ $3$ $0.727$
A00-1023 $3$ $2$ $0.833$
W00-0603 $7$ $4$ $0.692$
P02-1006 $7$ $5$ $0.590$
D03-1017 $7$ $4$ $0.500$
P03-1001 $6$ $4$ $0.500$
D04-9907 $7$ $3$ $0.545$
H05-1047 $4$ $3$ $0.833$
H05-1079 $5$ $3$ $0.625$
W05-1203 $3$ $3$ $0.583$
P05-1014 $4$ $2$ $0.667$
: Number of real facts, clusters and purity for each evaluated article[]{data-label="tbl:purity"}
[|@l@ | l|]{} ID & Sentence\
\
W05-1505:9 & 3 Constituency Parsing for Dependency Trees A pragmatic justification for using constituency- based parser in order\
& to predict dependency struc- tures is that currently the best Czech dependency- tree parser is a constituency-based parser (Collins et al, 1999; Zeman, 2004).\
W04-2407:27 & However, since most previous studies instead use the mean attachment score per word (Eisner, 1996; Collins et al, 1999), we will give this measure as well.\
J03-4003:33 & 3 We find lexical heads in Penn Treebank data using the rules described in Appendix A of Collins (1999).\
H05-1066:51 & Furthermore, we can also see that the MST parsers perform favorably compared to the more powerful\
& lexicalized phrase-structure parsers, such as those presented by Collins et al (1999) and Zeman (2004) that use expensive O(n5) parsing al- gorithms.\
E06-1011:21 & 5.2 Czech Results For the Czech data, we used the predefined train- ing, development and testing split\
& of the Prague Dependency Treebank (Hajiˇc et al, 2001), and the automatically generated POS tags supplied with the data,\
&which we reduce to the POS tag set from Collins et al (1999).\
\
P05-1012:16 & The Czech parser of Collins et al (1999) was run on a different data set and most other dependency parsers are evaluated using English.\
W04-2407:26 & More precisely, parsing accuracy is measured by the attachment score, which is\
& a standard measure used in studies of dependency parsing (Eisner, 1996; Collins et al, 1999).\
W05-1505:14 & In an attempt to extend a constituency-based pars- ing model to train on dependency trees,\
& Collins transforms the PDT dependency trees into con- stituency trees (Collins et al, 1999).\
P06-1033:31 & More specifi- cally for PDT, Collins et al (1999) relabel coordi- nated phrases after converting dependency struc- tures to phrase\
& structures, and Zeman (2004) uses a kind of pattern matching, based on frequencies of the parts-of-speech of conjuncts and conjunc- tions.\
P05-1012:17 & In par- ticular, we used the method of Collins et al (1999) to simplify part-of-speech tags since\
& the rich tags used by Czech would have led to a large but rarely seen set of POS features.\
### Sentence Extraction {#sect:extraction}
Once the graph is clustered and communities are formed, to build a summary we extract sentences from the clusters. We tried these two different simple methods:
- [**[*Cluster Round-Robin (C-RR)*]{}**]{}\
We start with the largest cluster, and extract sentences in the order they appear in each cluster. So we extract first, the first sentences from each cluster, then the second ones, and so on, until we reach the summary length limit $|S|$. Previously, we mentioned that facts with higher weights appear in greater number of sentences, and clustering aims to cluster such fact-sharing sentences in the same communities. Thus, starting with the largest community is important to ensure that the system summary first covers the facts that have higher frequencies and therefore higher weights.
- [**[*Cluster Lexrank (C-lexrank)*]{}**]{}\
The second method we used was Lexrank inside each cluster. In other words, for each cluster $\Omega_i$ we made a lexical network of *the sentences in that cluster* ($N_i$) . Using Lexrank we can find the most central sentences in $N_i$ as salient sentences of $\Omega_i$ to include in the main summary. We simply choose, for each cluster $\Omega_i$, the most salient sentence of $\Omega_i$, and if we have not reached the summary length limit, we do that for the second most salient sentences of each cluster, and so on. The way of ordering clusters is again by decreasing size.
Table \[tbl:systemSumm\] shows the two system summaries generated with C-RR and C-lexrank methods for P99-1065. The sentences in the table appear as they were extracted automatically from the text files of papers, containing sentence fragments and malformations occurring while doing the automatic segmentation.
![Each node is a sentence in the citation summary for paper J04-4002. Colors (numbers) represent facts and boundaries show the clustering result[]{data-label="fig:final1"}](final1.pdf)
Baseline Methods
----------------
We also conducted experiments with two baseline approaches. To produce the citation summary we used *Mead*’s Random Summary and *Lexrank* on the entire citation summary network as baseline techniques. Lexrank is proved to work well in multi-document summarization . It first builds a lexical network, in which nodes are sentences and a weighted edge between two nodes shows the lexical similarity. Once this network is built, Lexrank performs a random walk to find the most central nodes in the graph and reports them as summary sentences.
Experimental Setup
==================
Evaluation Method
-----------------
Fact-based evaluation systems have been used in several past projects [@jimmy06; @Marton-naacl-2006], especially in the TREC question answering track. [@jimmy06] use stemmed unigram similarity of responses with nugget descriptions to produce the evaluation results, whereas [@Marton-naacl-2006] uses both human judgments and human descriptions to evaluate a response.
An ideal summary in our model is one that covers more facts and more important facts. Our definition for the properties of a “good” summary of a paper is one that is relatively short and consists of the main contributions of that paper. From this viewpoint, there are two criteria for our evaluation metric. First, summaries that contain more important facts are preferred over summaries that cover fewer relevant facts. Second, facts should not be equally weighted in this model, as some of them may show more important contributions of a paper, while others may not.
To evaluate our system, we use the pyramid evaluation method [@nenkova2004ecs] at sentence level. Each fact in the citation summary of a paper is a *summarization content unit (SCU)* [@nenkova2004ecs], and the fact distribution matrix, created by annotation, provides the information about the importance of each fact in the citation summary.
The score given by the pyramid method for a summary is a ratio of the sum of the weights of its facts to the sum of the weights of an optimal summary. This score ranges from 0 to 1, and high scores show the summary content contain more heavily weighted facts. We believe that if a fact appears in more sentences of the citation summary than another fact, it is more important, and thus should be assigned a higher weight. To weight the facts we build a pyramid, and each fact falls in a tier. Each tier shows the number of sentences a fact appears in. Thus, the number of tiers in the pyramid is equal to the citation summary size. If a fact appears in more sentences, it falls in a higher tier. So, if the fact $f_i$ appears $|f_i|$ times in the citation summary it is assigned to the tier $T_{|f_i|}$.
The pyramid score formula that we use is computed as follows. Suppose the pyramid has $n$ tiers, $T_i$, where tier $T_n$ on top and $T_1$ on the bottom. The weight of the facts in tier $T_i$ will be $i$ (i.e. they appeared in $i$ sentences). If $|T_i|$ denotes the number of facts in tier $T_i$, and $D_i$ is the number of facts in the *summary* that appear in $T_i$, then the total fact weight for the summary is $D = \sum_{i=1}^{n} i\times D_i$. Additionally, the optimal pyramid score for a summary with $X$ facts, is
$ Max = \sum_{i=j+1}^n i\times |T_i| + j\times (X-\sum_{i=j+1}^n |T_i|)$ where $j= \max_i (\sum_{t=i}^n|T_t|\geq X)$. Subsequently, the pyramid score for a summary is calculated as $ P = \frac{D}{Max} $.
Results and Discussion
----------------------
Based on the described evaluation method we conducted a number of experiments to evaluate different summaries of a given length. In particular, we use a gold standard and a random summary to determine how good a system summary is. The gold standard is a summary of a given length that covers as many highly weighted facts as possible. To make a gold summary we start picking sentences that cover new and highly weighted facts, until the summary length limit is reached. On the other hand, in the random summary sentences are extracted from the citation summary in a random manner. We expect a good system summary to be closer to the gold than it is to the random one.
Table \[tbl:results\] shows the value of pyramid score $P$, for the experiments on the set of 25 papers. A $P$ score of less than $1$ for a gold shows that there are more facts than can be covered with a set of $|S|$ sentences.
This table suggests that C-lexrank has a higher average score, $P$, for the set of evaluated articles comparing C-RR and Lexrank.
As mentioned earlier in section \[sect:extraction\], once the citation summary network is clustered in the C-RR method, the sentences from each cluster are chosen in a round robin fashion, which will not guarantee that a fact-bearing sentence is chosen.
This is because all sentences, whether they cover any facts or not, are assigned to some cluster anyway and such sentences might appear as the first sentence in a cluster. This will sometimes result in a low $P$ score, for which P05-1012 is a good example.
-- -------------- -------------- -------------- -------------- -------------- --------------- --
C96-1058 1.00 0.27 0.73 0.73 0.73
P97-1003 1.00 0.08 0.40 0.60 0.40
P99-1065 0.94 0.30 0.54 0.82 0.67
P05-1013 1.00 0.15 0.69 0.97 0.67
P05-1012 0.95 0.14 0.57 0.26 0.62
N03-1017 0.96 0.26 0.36 0.61 0.64
W03-0301 1.00 0.60 1.00 1.00 1.00
J04-4002 1.00 0.33 0.70 0.48 0.48
N04-1033 1.00 0.38 0.38 0.31 0.85
P05-1033 1.00 0.37 0.77 0.77 0.85
A00-1043 1.00 0.66 0.95 0.71 0.95
A00-2024 1.00 0.26 0.86 0.73 0.60
C00-1072 1.00 0.85 0.85 0.93 0.93
W00-0403 1.00 0.55 0.81 0.41 0.70
W03-0510 1.00 0.58 1.00 0.83 0.83
A00-1023 1.00 0.57 0.86 0.86 0.86
W00-0603 1.00 0.33 0.53 0.53 0.60
P02-1006 1.00 0.49 0.92 0.49 0.87
D03-1017 1.00 0.00 0.53 0.26 0.85
P03-1001 1.00 0.12 0.29 0.59 0.59
D04-9907 1.00 0.53 0.88 0.65 0.94
H05-1047 1.00 0.83 0.66 0.83 1.00
H05-1079 1.00 0.67 0.78 0.89 0.56
W05-1203 1.00 0.50 0.71 1.00 0.71
P05-1014 1.00 0.44 1.00 0.89 0.78
[**Mean**]{} [**0.99**]{} [**0.41**]{} [**0.71**]{} [**0.69**]{} [**0.75** ]{}
-- -------------- -------------- -------------- -------------- -------------- --------------- --
: Evaluation Results ($|S|=5$)[]{data-label="tbl:results"}
Conclusion and Future Work
==========================
In this work we use the citation summaries to understand the main contributions of articles. The citation summary size, in our experiments, ranges from a few sentences to a few hundred, of which we pick merely a few ($5$ in our experiments) most important ones.
As a method of summarizing a scientific paper, we propose a clustering approach where communities in the citation summary’s lexical network are formed and sentences are extracted from separate clusters. Our experiments show how our clustering method outperforms one of the current state-of-art multi-document summarizing algorithms, Lexrank, on this particular problem.
A future improvement will be to use a reordering approach like Maximal Marginal Relevance (MMR) to re-rank clustered documents within each cluster in order to reduce the redundancy in a final summary. Another possible approach is to assume the set of sentences in the citation summary as sentences talking about the same event, yet generated in different sources. Then one can apply the method inspired by to identify common phrases across sentences and use language generation to form a more coherent summary. The ultimate goal, however, is to produce a topic summarizer system in which the query is a scientific topic and the output is a summary of all previous works in that topic, preferably sorted to preserve chronology and topicality.
Acknowledgments
===============
The authors would like to thank Bonnie Dorr, Jimmy Lin, Saif Mohammad, Judith L. Klavans, Ben Shneiderman, and Aleks Aris from UMD, Bryan Gibson, Joshua Gerrish, Pradeep Muthukrishnan, Arzucan Özgür, Ahmed Hassan, and Thuy Vu from University of Michigan for annotations.
This paper is based upon work supported by the National Science Foundation grant “iOPENER: A Flexible Framework to Support Rapid Learning in Unfamiliar Research Domains”, jointly awarded to U. of Michigan and U. of Maryland as IIS 0705832. Any opinions, findings, and conclusions or recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the National Science Foundation.
[^1]: http://www.pubmedcentral.gov
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we describe SemEval-2013 Task 4: the definition, the data, the evaluation and the results. The task is to capture some of the meaning of English noun compounds via paraphrasing. Given a two-word noun compound, the participating system is asked to produce an explicitly ranked list of its free-form paraphrases. The list is automatically compared and evaluated against a similarly ranked list of paraphrases proposed by human annotators, recruited and managed through Amazon’s Mechanical Turk. The comparison of raw paraphrases is sensitive to syntactic and morphological variation. The “gold” ranking is based on the relative popularity of paraphrases among annotators. To make the ranking more reliable, highly similar paraphrases are grouped, so as to downplay superficial differences in syntax and morphology. Three systems participated in the task. They all beat a simple baseline on one of the two evaluation measures, but not on both measures. This shows that the task is difficult.'
author:
- |
[**Iris Hendrickx**]{}\
Radboud University Nijmegen &\
Universidade de Lisboa\
\
[**Preslav Nakov**]{}\
QCRI, Qatar Foundation\
\
[**Stan Szpakowicz**]{}\
University of Ottawa &\
Polish Academy of Sciences\
\
\
University of Southern California\
\
\
[**Diarmuid Ó Séaghdha**]{}\
University of Cambridge\
\
[**Tony Veale**]{}\
University College Dublin\
title: 'SemEval-2013 Task 4: Free Paraphrases of Noun Compounds'
---
Introduction
============
A noun compound (NC) is a sequence of nouns which act as a single noun [@Downing:1977], as in these examples: *colon cancer*, *suppressor protein*, *tumor suppressor protein*, *colon cancer tumor suppressor protein*, etc. This type of compounding is highly productive in English. NCs comprise 3.9% and 2.6% of all tokens in the Reuters corpus and the British National Corpus (BNC), respectively [@baldwin-tanaka:04].
The frequency spectrum of compound types follows a Zipfian distribution [@OSeaghdha:PhD], so many NC tokens belong to a “long tail" of low-frequency types. More than half of the two-noun types in the BNC occur exactly once [@Kim:Baldwin:06]. Their high frequency and high productivity make robust NC interpretation an important goal for broad-coverage semantic processing of English texts. Systems which ignore NCs may give up on salient information about the semantic relationships implicit in a text. Compositional interpretation is also the only way to achieve broad NC coverage, because it is not feasible to list in a lexicon all compounds which one is likely to encounter. Even for relatively frequent NCs occurring 10 times or more in the BNC, static English dictionaries provide only 27% coverage [@Tanaka:Baldwin:2003].
In many natural language processing applications it is important to understand the syntax and semantics of NCs. NCs often are structurally similar, but have very different meaning. Consider *caffeine headache* and *ice-cream headache*: a lack of caffeine causes the former, an excess of ice-cream – the latter. Different interpretations can lead to different inferences, query expansion, paraphrases, translations, and so on. A question answering system may have to determine whether *protein acting as a tumor suppressor* is an accurate paraphrase for *tumor suppressor protein*. An information extraction system might need to decide whether *neck vein thrombosis* and *neck thrombosis* can co-refer in the same document. A machine translation system might paraphrase the unknown compound *WTO Geneva headquarters* as *WTO headquarters located in Geneva*. Research on the automatic interpretation of NCs has focused mainly on common two-word NCs. The usual task is to classify the semantic relation underlying a compound with either one of a small number of predefined relation labels or a paraphrase from an open vocabulary. Examples of the former take on classification include [@moldovan-EtAl:2004:HLTNAACL; @Girju:07; @OSeaghdha:Copestake:08; @Tratz:Hovy:10]. Examples of the latter include [@Nakov:08a; @Nakov:08MT; @Nakov:Hearst:08; @Butnariu:Veale:08] and a previous NC paraphrasing task at SemEval-2010 [@butnariu-EtAl:2010:SemEval], upon which the task described here builds.
The assumption of a small inventory of predefined relations has some advantages – parsimony and generalization – but at the same time there are limitations on expressivity and coverage. For example, the NCs *headache pills* and *fertility pills* would be assigned the same semantic relation (*PURPOSE*) in most inventories, but their relational semantics are quite different [@Downing:1977]. Furthermore, the definitions given by human subjects can involve rich and specific meanings. For example, reports that a subject defined the NC *oil bowl* as “the bowl into which the oil in the engine is drained during an oil change”, compared to which a minimal interpretation *bowl for oil* seems very reductive. In view of such arguments, linguists such as , and have argued for a fine-grained, essentially open-ended space of interpretations.
The idea of working with fine-grained paraphrases for NC semantics has recently grown in popularity among NLP researchers [@Butnariu:Veale:08; @Nakov:Hearst:08; @Nakov:08MT]. Task 9 at SemEval-2010 [@butnariu-EtAl:2010:SemEval] was devoted to this methodology. In that previous work, the paraphrases provided by human subjects were required to fit a restrictive template admitting only verbs and prepositions occurring between the NC’s constituent nouns. Annotators recruited through Amazon Mechanical Turk were asked to provide paraphrases for the dataset of NCs. The gold standard for each NC was the ranked list of paraphrases given by the annotators; this reflects the idea that a compound’s meaning can be described in different ways, at different levels of granularity and capturing different interpretations in the case of ambiguity.
For example, a *plastic saw* could be a *saw made of plastic* or a *saw for cutting plastic*. Systems participating in the task were given the set of attested paraphrases for each NC, and evaluated according to how well they could reproduce the humans’ ranking.
The design of this task, SemEval-2013 Task 4, is informed by previous work on compound annotation and interpretation. It is also influenced by similar initiatives, such as the English Lexical Substitution task at SemEval-2007 [@mccarthy:2007:SemEval-2007], and by various evaluation exercises in the fields of paraphrasing and machine translation. We build on SemEval-2010 Task 9, extending the task’s flexibility in a number of ways. The restrictions on the form of annotators’ paraphrases was relaxed, giving us a rich dataset of close-to-freeform paraphrases (Section \[sec:data\]). Rather than ranking a set of attested paraphrases, systems must now both generate and rank their paraphrases; the task they perform is essentially the same as what the annotators were asked to do. This new setup required us to innovate in terms of evaluation measures (Section \[sec:eval\]).
We anticipate that the dataset and task will be of broad interest among those who study lexical semantics. We believe that the overall progress in the field will significantly benefit from a public-domain set of free-style NC paraphrases. That is why our primary objective is the challenging endeavour of preparing and releasing such a dataset to the research community. The common evaluation task which we establish will also enable researchers to compare their algorithms and their empirical results.
Task description {#sec:task}
================
This is an English NC interpretation task, which explores the idea of interpreting the semantics of NCs via free paraphrases. Given a noun-noun compound such as *air filter*, the participating systems are asked to produce an explicitly ranked list of free paraphrases, as in the following example:
1 filter for air\
2 filter of air\
3 filter that cleans the air\
4 filter which makes air healthier\
5 a filter that removes impurities from the air\
…
Such a list is then automatically compared and evaluated against a similarly ranked list of paraphrases proposed by human annotators, recruited and managed via Amazon’s Mechanical Turk. The comparison of raw paraphrases is sensitive to syntactic and morphological variation. The ranking of paraphrases is based on their relative popularity among different annotators. To make the ranking more reliable, highly similar paraphrases are grouped so as to downplay superficial differences in syntax and morphology.
Data collection {#sec:data}
===============
We used Amazon’s *Mechanical Turk* service to collect diverse paraphrases for a range of “gold-standard” NCs.[^1] We paid the workers a small fee (\$0.10) per compound, for which they were asked to provide five paraphrases. Each paraphrase should contain the two nouns of the compound (in singular or plural inflectional forms, but not in another derivational form), an intermediate non-empty linking phrase and optional preceding or following terms. The paraphrasing terms could have any part of speech, so long as the resulting paraphrase was a well-formed noun phrase headed by the NC’s head.
We gave the workers feedback during data collection if they appeared to have misunderstood the nature of the task. Once raw paraphrases had been collected from all workers, we collated them into a spreadsheet, and we merged identical paraphrases in order to calculate their overall frequencies. Ill-formed paraphrases – those violating the syntactic restrictions described above – were manually removed following a consensus decision-making procedure; every paraphrase was checked by at least two task organizers. We did not require that the paraphrases be semantically felicitous, but we performed minor edits on the remaining paraphrases if they contained obvious typos.
The remaining well-formed paraphrases were sorted by frequency separately for each NC. The most frequent paraphrases for a compound are assigned the highest rank 0, those with the next-highest frequency are given a rank of 1, and so on.
Paraphrases with a frequency of 1 – proposed for a given NC by only one annotator – always occupy the lowest rank on the list for that compound.
We used 174+181 noun-noun compounds from the NC dataset of . The trial dataset, which we initially released to the participants, consisted of 4,255 human paraphrases for 174 noun-noun pairs; this dataset was also the training dataset. The test dataset comprised paraphrases for 181 noun-noun pairs. The “gold standard” contained 9,706 paraphrases of which 8,216 were unique for those 181 NCs. Further statistics on the datasets are presented in Table \[stat-table\].
-------------------- --------------- -------------------------
[**Total**]{} [**Min / Max / Avg**]{}
\[1ex\]
paraphrases 6,069 1 / 287 / 34.9
unique paraphrases 4,255 1 / 105 / 24.5
\[1ex\]
paraphrases 9,706 24 / 99 / 53.6
unique paraphrases 8,216 21 / 80 / 45.4
-------------------- --------------- -------------------------
: Statistics of the trial and test datasets: the total number of paraphrases with and without duplicates, and the minimum / maximum / average per noun compound.[]{data-label="stat-table"}
Compared with the data collected for the SemEval-2010 Task 9 on the interpretation of noun compounds, the data collected for this new task have a far greater range of variety and richness. For example, the following (selected) paraphrases for *work area* vary from parsimonious to expansive:
area for work
area of work
area where work is done
area where work is performed
…
an area cordoned off for persons responsible for work
an area where construction work is carried out
an area where work is accomplished and done
area where work is conducted
office area assigned as a work space
…
Scoring {#sec:eval}
=======
Noun compounding is a generative aspect of language, but so too is the process of NC interpretation: human speakers typically generate a range of possible interpretations for a given compound, each emphasizing a different aspect of the relationship between the nouns. Our evaluation framework reflects the belief that there is rarely a single right answer for a given noun-noun pairing. Participating systems are thus expected to demonstrate some generativity of their own, and are scored not just on the accuracy of individual interpretations, but on the overall breadth of their output.
For evaluation, we provided a scorer implemented, for good portability, as a Java class. For each noun compound to be evaluated, the scorer compares a list of system-suggested paraphrases against a “gold-standard” reference list, compiled and rank-ordered from the paraphrases suggested by our human annotators. The score assigned to each system is the mean of the system’s performance across all test compounds. Note that the scorer removes all determiners from both the reference and the test paraphrases, so a system is neither punished for not reproducing a determiner or rewarded for producing the same determiners.
The scorer can match words identically or non-identically. A match of two identical words $W_{gold}$ and $W_{test}$ earns a score of 1.0. There is a partial score of $(2~|P|~/~(|PW_{gold}|$ + $|PW_{test}|))^2$ for a match of two words $PW_{gold}$ and $PW_{test}$ that are not identical but share a common prefix $P$, $|P|>2$, e.g., $wmatch(\emph{cutting}, \emph{cuts})$ = $(6/11)^2$ = 0.297.
Two $n$-grams $N_{gold}$ = \[$GW_1$, …, $GW_n$\] and $N_{test}$ = \[$TW_1$, …, $TW_n$\] can be matched if $wmatch(GW_i,~TW_i) > 0$ for all $i$ in $1 .. n$. The score assigned to the match of these two $n$-grams is then $\sum_i wmatch(GW_i,~TW_i)$. For every $n$-gram $N_{test}$ = \[$TW_1$, …, $TW_n$\] in a system-generated paraphrase, the scorer finds a matching $n$-gram $N_{gold}$ = \[$GW_1$, …, $GW_n$\] in the reference paraphrase $Para_{gold}$ which maximizes this sum.
The overall $n$-gram overlap score for a reference paraphrase $Para_{gold}$ and a system-generated paraphrase $Para_{test}$ is the sum of the score calculated for all $n$-grams in $Para_{test}$, where $n$ ranges from 1 to the size of $Para_{test}$.
This overall score is then normalized by dividing by the maximum value among the $n$-gram overlap score for $Para_{gold}$ compared with itself and the $n$-gram overlap score for $Para_{test}$ compared with itself. This normalization step produces a paraphrase match score in the range \[0.0 1.0\]. It punishes a paraphrase $Para_{test}$ for both over-generating (containing more words than are found in $Para_{gold}$) and under-generating (containing fewer words than are found in $Para_{gold}$). In other words, $Para_{test}$ should ideally reproduce everything in $Para_{gold}$, and nothing more or less.
The reference paraphrases in the “gold standard” are ordered by rank; the highest rank is assigned to the paraphrases which human judges suggested most often. The rank of a reference paraphrase matters because a good participating system will aim to reproduce the top-ranked “gold-standard” paraphrases as produced by human judges. The scorer assigns a multiplier of $R/(R + n)$ to reference paraphrases at rank $n$; this multiplier asymptotically approaches 0 for the higher values of $n$ of ever lower-ranked paraphrases. We choose a default setting of $R=8$, so that a reference paraphrase at rank 0 (the highest rank) has a multiplier of 1, while a reference paraphrase at rank 5 has a multiplier of $8/13$ = 0.615.
When a system-generated paraphrase $Para_{test}$ is matched with a reference paraphrase $Para_{gold}$, their normalized $n$-gram overlap score is scaled by the rank multiplier attaching to the rank of $Para_{gold}$ relative to the other reference paraphrases provided by human judges. The scorer automatically chooses the reference paraphrase $Para_{gold}$ for a test paraphrase $Para_{test}$ so as to maximize this product of normalized $n$-gram overlap score and rank multiplier.
The overall score assigned to each system for a specific compound is calculated in two different ways: using *isomorphic matching* of suggested paraphrases to the “gold-standard’s” reference paraphrases (on a *one-to-one* basis); and using *non-isomorphic matching* of system’s paraphrases to the “gold-standard’s” reference paraphrases (in a potentially *many-to-one* mapping).
*Isomorphic matching* rewards both precision and recall. It rewards a system for accurately reproducing the paraphrases suggested by human judges, and for reproducing as many of these as it can, and in much the same order.
In isomorphic mode, system’s paraphrases are matched 1-to-1 with reference paraphrases on a first-come first-matched basis, so ordering can be crucial.
*Non-isomorphic* matching rewards only precision. It rewards a system for accurately reproducing the top-ranked human paraphrases in the “gold standard”. A system will achieve a higher score in a non-isomorphic match if it reproduces the top-ranked human paraphrases as opposed to lower-ranked human paraphrases. The ordering of system’s paraphrases is thus not important in non-isomorphic matching.
Each system is evaluated using the scorer in both modes, *isomorphic* and *non-isomorphic*. Systems which aim only for precision should score highly on non-isomorphic match mode, but poorly in isomorphic match mode. Systems which aim for precision *and* recall will face a more substantial challenge, likely reflected in their scores.
#### A naïve baseline
\
We decided to allow preposition-only paraphrases, which are abundant in the paraphrases suggested by human judges in the crowdsourcing Mechanical Turk collection process. This abundance means that the top-ranked paraphrase for a given compound is often a preposition-only phrase, or one of a small number of very popular paraphrases such as *used for* or *used in*. It is thus straightforward to build a naïve baseline generator which we can expect to score reasonably on this task, at least in *non-isomorphic matching* mode. For each test compound *M H*, the baseline system generates the following paraphrases, in this precise order: *H of M*, *H in M*, *H for M*, *H with M*, *H on M*, *H about M*, *H has M*, *H to M*, *H used for M*, *H used in M*.
This naïve baseline is truly unsophisticated. No attempt is made to order paraphrases by their corpus frequencies or by their frequencies in the training data. The same sequence of paraphrases is generated for each and every test compound.
Results
=======
Three teams participated in the challenge, and all their systems were supervised. The MELODI system relied on semantic vector space model built from the UKWAC corpus (window-based, 5 words). It used only the features of the right-hand head noun to train a maximum entropy classifier.
The IIITH system used the probabilities of the preposition co-occurring with a relation to identify the class of the noun compound. To collect statistics, it used Google $n$-grams, BNC and ANC.
The SFS system extracted templates and fillers from the training data, which it then combined with a four-gram language model and a MaxEnt reranker. To find similar compounds, they used Lin’s WordNet similarity. They further used statistics from the English Gigaword and the Google $n$-grams.
Table \[tbl:results\] shows the performance of the participating systems, SFS, IIITH and MELODI, and the naïve baseline. The baseline shows that it is relatively easy to achieve a moderately good score in non-isomorphic match mode by generating a fixed set of paraphrases which are both common and generic: two of the three participating systems, SFS and IIITH, under-perform the naïve baseline in non-isomorphic match mode, but outperform it in isomorphic mode. The only system to surpass this baseline in non-isomorphic match mode is the MELODI system; yet, it under-performs against the same baseline in isomorphic match mode. No participating team submitted a system which would outperform the naïve baseline in both modes.
--------------------------------------- -- --
**Team & isomorphic & non-isomorphic\
SFS & 23.1 & 17.9\
IIITH & 23.1 & 25.8\
MELODI-Primary & 13.0 & 54.8\
MELODI-Contrast &13.6 & 53.6\
*Naive Baseline* & *13.8* & *40.6*\
**
--------------------------------------- -- --
: \[tbl:results\] Results for the participating systems; the baseline outputs the same paraphrases for all compounds.
Conclusions
===========
The conclusions we draw from the experience of organizing the task are mixed. Participation was reasonable but not large, suggesting that NC paraphrasing remains a niche interest – though we believe it deserves more attention among the broader lexical semantics community and hope that the availability of our freeform paraphrase dataset will attract a wider audience in the future.
We also observed a varied response from our annotators in terms of embracing their freedom to generate complex and rich paraphrases; there are many possible reasons for this including laziness, time pressure and the fact that short paraphrases are often very appropriate paraphrases. The results obtained by our participants were also modest, demonstrating that compound paraphrasing is both a difficult task and a novel one that has not yet been “solved”.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work has partially supported by a small but effective grant from Amazon; the credit allowed us to hire sufficiently many Turkers – thanks! And a thank-you to our additional annotators Dave Carter, Chris Fournier and Colette Joubarne for their complete sets of paraphrases of the noun compounds in the test data.
Timothy Baldwin and Takaaki Tanaka. 2004. Translation by machine of complex nominals: Getting it right. , Barcelona, Spain, 24-31.
Cristina Butnariu and Tony Veale. 2008. A concept-centered approach to noun-compound interpretation. , Manchester, UK, 81-88.
Cristina Butnariu, Su Nam Kim, Preslav Nakov, Diarmuid Ó Séaghdha, Stan Szpakowicz, and Tony Veale. 2010. SemEval-2010 Task 9: The interpretation of noun compounds using paraphrasing verbs and prepositions. , Uppsala, Sweden, 39-44.
Seana Coulson. 2001. . Cambridge University Press, Cambridge, UK.
Pamela Downing. 1977. On the creation and use of English compound nouns. , [**53**]{}(4): 810-842.
Roxana Girju. 2007. Improving the interpretation of noun phrases with cross-linguistic information. , Prague, Czech Republic, 568-575.
Su Nam Kim and Timothy Baldwin. 2006. Interpreting semantic relations in noun compounds via verb semantics. , Sydney, Australia, 491-498.
Diana McCarthy and Roberto Navigli. 2007. Semeval-2007 task 10: English lexical substitution task. , Prague, Czech Republic, 48-53.
Dan Moldovan, Adriana Badulescu, Marta Tatu, Daniel Antohe, and Roxana Girju. 2004. Models for the semantic classification of noun phrases. Dan Moldovan and Roxana Girju, eds., [*HLT-NAACL 2004: Workshop on Computational Lexical Semantics*]{}, Boston, MA, USA, 60-67.
Preslav Nakov and Marti Hearst. 2008. Solving relational similarity problems using the Web as a corpus. , Columbus, OH, USA, 452-460.
Preslav Nakov. 2008a. Improved statistical machine translation using monolingual paraphrases. , Patras, Greece, 338-342.
Preslav Nakov. 2008b. Noun compound interpretation using paraphrasing verbs: Feasibility study. , Varna, Bulgaria, [*Lecture Notes in Computer Science*]{} [**5253**]{}, Springer, 103-117.
Diarmuid [Ó Séaghdha]{}. 2007. Designing and Evaluating a Semantic Annotation Scheme for Compound Nouns. In [*Proceedings of the 4th Corpus Linguistics Conference*]{}, Birmingham, UK.
Diarmuid [Ó]{} S[é]{}aghdha. 2008. . thesis, Computer Laboratory, University of Cambridge. Published as University of Cambridge Computer Laboratory Technical Report 735.
Diarmuid [Ó]{} S[é]{}aghdha and Ann Copestake. 2009. Using lexical and relational similarity to classify semantic relations. , Athens, Greece, 621-629.
Diarmuid [Ó Séaghdha]{} and Ann Copestake. 2008. Semantic classification with distributional kernels. In [*Proc. 22nd International Conference on Computational Linguistics (COLING-08)*]{}, Manchester, UK.
Mary Ellen Ryder. 1994. . University of California Press, Berkeley, CA, USA.
Takaaki Tanaka and Tim Baldwin. 2003. Noun-noun compound machine translation: A feasibility study on shallow processing. , Sapporo, Japan, 17-24.
Stephen Tratz and Eduard Hovy. 2010. A taxonomy, dataset, and classifier for automatic noun compound interpretation. , Uppsala, Sweden, 678-687.
[^1]: Since the annotation on Mechanical Turk was going slowly, we also recruited four other annotators to do the same work, following exactly the same instructions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'High resolution G-band images of the interior of a supergranulation cell show ubiquitous Bright Points (some 0.3 BPs per Mm$^2$). They are located in intergranular lanes and often form chains of elongated blobs whose smallest dimension is at the resolution limit (135 km on the Sun). Most of them live for a few minutes, having peak intensities from 0.8 to 1.8 times the mean photospheric intensity. These BPs are probably tracing intense magnetic concentrations, whose existence has been inferred in spectro-polarimetric measurements. Our finding provides a new convenient tool for the study of the inter-network magnetism, so far restricted to the interpretation weak polarimetric signals.'
author:
- 'J. Sánchez Almeida, I. Márquez, J. A. Bonet, I. Domínguez Cerdeña, [and]{} R. Muller'
title: ' Bright points in the inter-network quiet Sun '
---
Introduction {#intro}
============
The presence of Bright Points (BPs) in the solar photosphere was first reported by @dun73 and @meh74. They were identified with small magnetic concentrations, which are expected to be bright [@spr77][^1]. BPs appear in all magnetic structures: around sunspots, in plage regions, and in the network [e.g. @mul94]. This Letter shows how they are also very common inside supergranulation network cells.
The new result refers to the magnetism of the so-called Inter-Network (IN), and may be consequential. The seemingly inactive IN provides a large fraction of the (unsigned) magnetic flux and energy existing on the solar surface at any given time [e.g., @ste82; @san02b; @san04; @tru04]. Therefore, it may be an important element to understand the global magnetic properties of the Sun (e.g., the structure of the quiet corona, @sch03b; the chromospheric heating, @goo04; the sources of the solar wind, @woo97b [@hu03]; or the emergence of magnetic flux, @dep02). We still have a very primitive knowledge of the properties of the IN fields, which mostly comes from the model-dependent analysis of weak spectro-polarimetric signals [for a recent review, see @san04]. The finding of G-band BPs provides a new simple and direct mean to study the IN magnetism using unpolarized light. In addition, the presence of BPs strongly suggests that part of the IN fields have kG field strengths. Spectro-polarimetric observations indicate IN field strengths going all the way from zero to kG [see, e.g., the introduction of @san03c]. The presence of kG field strengths has been particularly controversial, and this work provides independent support.
Observations and image restoration {#observations}
==================================
A quiet Sun IN region was observed at the solar disk center with the SST (Swedish Solar Telescope), a new 1-m instrument equipped with adaptive optics [@sch03c; @sch03d]. The Field-Of-View (FOV) was selected using Ca [ii]{} H images to avoid network magnetic concentrations. We image through a 10.8 Å wide filter centered in the G-band at 4305.6 Å. Although magnetic concentrations are bright at all wavelengths, they become particularly conspicuous in this CH band [@mul84; @ber95]. The optical setup included a Phase Diversity (PD) device for post-facto restoration of the optical aberrations induced by the terrestrial atmosphere and the telescope [@gon79; @pax92]. The series of PD images was taken on September 30, 2003, from 11:33 UT to 12:23 UT. We use a 10-bit Kodak Megaplus 1.6 camera (pixel size 0041, exposure time 8 msec). Real-time frame selection provided us with the 4 sharpest images every 20 sec, which are needed in our PD inversion code (@bon03, based on @lof94, and @pax96). First, each image of the series is divided in a mosaic of 96 overlapping isoplanatic patches, which we restore independently. Then the mosaic is assembled, and the 4 images of each selection interval are combined to form a restored snapshot. The process of splitting, restoring and merging renders a FOV of 23x 35; see Fig. \[chickenpox\]. Excellent and rather bad moments of seeing alternated during the observing run. Therefore several sets of 4 images could not be successfully restored, leaving gaps in our series. It finally has 100 restored snapshots during its 50 min duration. The PD code includes a self-regulated high-spatial-frequency noise removal filter. Given the conditions of the data, the largest cutoffs set by this filter are 014, i.e., only slightly worst than the diffraction limit (009). Note that the cutoff differs for each PD restoration, and so, it varies within a single restored snapshot. The cutoff approximately yields the Full Width Half Maximum (FWHM) of the point spread function, a rule of thumb used in § \[twogauss\].
Ca images allowed us to center the G-band FOV within a supergranulation cell. Figure \[ines\]a shows one of these Ca images, which includes a box with the G-band FOV outline. Note how the brightest Ca patches remain outside the G-band contour. Only a small patch in the lower right corner lies within the bounds. In order to stress the IN character of our target, we pinpoint the observed FOV in a full disk MDI magnetogram[^2] taken 18 minutes before the beginning of the series. It is shown in Fig. \[ines\]b. The figure includes an inset with the SST Ca image of the region. Figure \[ines\]b was set up to emphasize the continuity between the MDI magnetogram and the Ca images, and so, it displays the unsigned polarimetric signals. Figure \[ines\]c shows the true MDI magnetogram saturated at $\pm~50$ G to reveal the weakest network features. A small blob appears within the G-band FOV. The rest remains field-free as far as the MDI magnetogram is concerned. The MDI signals within the FOV are $(-0.9\pm 7.1)$ G, with the standard deviation matching the MDI noise (e.g., 6.9 G for @hag01). The magnetic patch of the MDI magnetogram is very conspicuous in the G-band; see the conglomerate of BPs at the lower right part of Fig. \[chickenpox\], coordinates \[16,3\].
Analysis and Results {#all}
====================
Area coverage and number density {#byeye}
--------------------------------
The sharpest snapshot of the time series was selected for in-depth study (Fig. \[chickenpox\], left). Although this reference snapshot was chosen by visual inspection, it has one of the largest contrasts of the full series (14.3%). Then, playing the series back and forth, we single out all those BPs which (a) were in the reference snapshot, (b) persisted three or more snapshots, and (c) stay in an intergranular lane. This subjective selection of BPs was complemented with an automatic determination of their areas. We apply the segmentation algorithm by @str94 [ Chapter 8], and then those patches overlaying a visually selected BP were chosen as the area of the BP. The result is shown in Fig. \[chickenpox\], right. We identify 126 individual points, some of which are part of long chains. Given the area of the FOV, we detect 0.3 points per Mm${^{2}}$. The density of BPs increases towards the lower left part of the FOV (Fig. \[chickenpox\]), i.e., when approaching the network patch (Fig. \[ines\]c). The area associated with these BPs covers some 0.5% of the FOV. This area depends on the segmentation algorithm, and a more meaningful estimate is carried out in § \[twogauss\].
Size and Brightness {#twogauss}
-------------------
The BPs stay in intergranular lanes, which are narrow dark wells of the intensity distribution. @tit96 show how this fact biases any measurement of size unless the shape of the background is considered. They propose a 2-Gaussian fit to decontaminate from the influence of the intergranular background. We adopt this approach for estimating the widths of our BPs. Those pixels selected by the segmentation algorithm were fitted using $$\begin{aligned}
%\begin{equation}
&f=c+aG(x,y,x_0,y_0,\sigma_x,\sigma_y,\theta)
-bG(x,y,x_0,y_0,\sigma^\prime_x,\sigma^\prime_y,\theta),\cr
\label{eq1}
%\end{equation}
%with
&G(x,y,x_0,y_0,\sigma_x,\sigma_y,\theta)=
\exp{-(r_x^2+r_y^2)},\cr
&r_x=[(x-x_0)\cos\theta+(y-y_0)\sin\theta]/\sigma_x,\cr
&r_y=[-(x-x_0)\sin\theta+(y-y_0)\cos\theta]/\sigma_y.%\nonumber\end{aligned}$$ The 2-Gaussian function $f$ depends on the spatial coordinates $x$ and $y$, as well as on 10 free parameters: $a, b,$ and $c$ for the amplitudes of two Gaussians and a background, $x_0$ and $y_0$ for the core of the BP, $\theta$ for a global orientation, $\sigma_x$ and $\sigma_y$ for the widths of the BP and, finally, $\sigma^\prime_x$ and $\sigma^\prime_y$ for the widths of the intergranular lane. By means of a non-linear least squares algorithm, we tried to reproduce to all BPs selected in Sec.\[byeye\]. Some 60% of the trials converged. Figure \[example\] shows four representative cases, which indicate how the elliptical contours of the 2-Gaussian functions really reproduce the observed BP shapes.
Figure \[tripode\]a contains the histograms of the FWHM of the 2-dimensional Gaussians representing the BP (FWHM = $1.67\sigma_x$ or $1.67\sigma_y$). The minor axis FWHM peaks at some 135 km, and it seldom exceeds 200 km. We believe that this 135 km represents the mean spatial resolution of the snapshot. As we discuss in § \[observations\], the resolution varies along the FOV, with a lower limit of some 014 or 100 km. The width of the minor axis histogram is mostly set by uncertainties in the non-linear least squares fit (some 30 km, as provided by the fitting routine). The interpretation of the areas derived in § \[byeye\] is ambiguous. We prefer a straight definition based on the 2-Gaussian fits, namely, the area within the elliptical contour embraced by the FWHM (i.e., $2.18\sigma_x\sigma_y$). This definition yields areas 40% larger than the segmentation algorithm. Then the 0.5% coverage in § \[byeye\] turns out to be 0.7%.
Figure \[tripode\]b shows a scatter plot of the peak G-band intensity (i.e., $a-b+c$ in Eq. \[\[eq1\]\]) versus minor axis FWHM. There is no obvious correlation between brightness and size, which can be interpreted as the unresolved character of most of the features [e.g., @ber95]. Figure \[tripode\]b also shows many BPs with an intensity smaller than that of the mean photosphere. On top of this, a few BPs reach 1.8 . In a different snapshot of the series, the network clump peak intensity scores 2.3.
Lifetimes
---------
Lifetimes were estimated by visual inspection of the time series. The BPs in the reference snapshot were visually tracked to find out when they appear and disappear. This time interval defines the lifetime. A histogram of lifetimes is represented in Fig. \[tripode\]c, the solid line. Most lifetimes are shorter than 10 min. In addition, some BPs live as much as we can measure (longer than 30 min). Keep in mind that the lifetimes are strongly biased. First, BPs often appear or end within one of the gaps of the time series, which underestimate the true values. Second, the detection criteria bias the estimates towards long-lasting BPs (§ \[byeye\]). Although these caveats should be kept in mind, we believe that most BPs are really short-lived. Figure \[tripode\]c, the dotted line, shows a histogram based on BPs whose birth and death we witnessed, and they have short lifetimes.
Comments and conclusions {#conclusions}
========================
We have detected many Bright Points (BPs) in the interior of a supergranulation cell (i.e., in the Inter-Network or IN). According to the current paradigm, these bright points trace intense magnetic concentrations (§ \[intro\]). Our finding has two main consequences. First, it provides a new convenient tool for the study of the IN magnetism, so far restricted to the interpretation of low spatial resolution weak polarimetric signals. Second, it supports a result based on Zeeman magnetometry indicating that part of the IN magnetic fields are not weak but have kG magnetic field strengths [@san00; @soc02; @soc03]. The same polarimetric measurements also point out how these kG concentrations carry the body of the IN magnetic flux and energy [@san03c; @san04]. Consequently, unpolarized imaging may allow us to detect and study a significant fraction of the IN magnetism.
The detected BPs cover some 0.7% of the surface, which is not enough to account for the magnetic signals found by (@dom03a, [-@dom03b], @san03d). Several reasons may explain this difference. We miss many BPs since the detectability critically depends on the resolution [@tit96], and our BPs remain unresolved (§ \[twogauss\]). On the other hand, some strong fields may not be bright. The actual brightness depends on subtleties of the magnetic concentration and its non-magnetic environment, and sometimes modeling shows faint kG features [@san01]. Yet another possibility is the random fluctuation of properties of different cell interiors. The work was funded by the Spanish project AYA2001-1649 and the EC contract HPRN-CT-2002-00313. Thanks are due to F. Kneer for continuous support, and to A. Sainz Dalda for help with MDI data (courtesy of the SOHO/MDI consortium by ESA & NASA). The SST is operated by the Institute for Solar Physics, Stockholm, at the ORM of the IAC.
, T. E., [Schrijver]{}, C. J., [Shine]{}, R. A., [Tarbell]{}, T. D., [Title]{}, A. M., & [Scharmer]{}, G. 1995, , 454, 531
, J. A., & [Márquez]{}, I. 2003, in ASP Conf. Ser., Vol. 307, Solar Polarization 3, ed. J. [Trujillo-Bueno]{} & J. [Sánchez Almeida]{} (San Francisco: ASP), 137
, B. 2002, , 569, 474
, I., [Kneer]{}, F., & [Sánchez Almeida]{}, J. 2003a, , 582, L55
, I., [Sánchez Almeida]{}, J., & [Kneer]{}, F. 2003b, , 407, 741
, R. B., & [Zirker]{}, J. B. 1973, , 33, 281
, R. A., & [Childlaw]{}, R. 1979, in Proc. Soc. Photo-Opt. Instrum. Eng., Vol. 207, Applications of Digital Image Processing III, ed. A. G. [Tescher]{}, 32
, M. L. 2004, , in press
, H. J. 2001, , 555, 448
, Y. Q., [Habbal]{}, S. R., [Chen]{}, Y., & [Li]{}, X. 2003, , 108(A10), 1377
, D., [Rutten]{}, R. J., & [Plez]{}, B. 2001, in IAU Symp., Vol. 203, Recent Insights into the Physics of the Sun and Heliosphere, ed. P. [Brekke]{}, B. [Fleck]{}, & J. B. [Gurman]{} (San Francisco: ASP), 287
, M. G., & [Scharmer]{}, G. B. 1994, , 107, 243
, J. P. 1974, , 38, 43
, R. 1994, in Solar Surface Magnetism, ed. R. J. [Rutten]{} & C. J. [Schrijver]{}, NATO ASI Ser. 433 (Dordrecht: Kluwer), 55
, R., & [Roudier]{}, T. 1984, , 94, 33
, R. G., [Schulz]{}, T. J., & [Fienup]{}, J. R. 1992, JOSA, A9, 1072
, R. G., [Seldin]{}, J. H., [Löfdahl]{}, M. G., [Scharmer]{}, G. B., & [Keller]{}, C. U. 1996, , 466, 1087
, J. 2003a, , 411, 615
, J. 2003b, in AIP Conf. Proc., Vol. 679, Solar Wind 10, ed. M. [Velli]{}, R. [Bruno]{}, & F. [Malara]{} (New York: AIP), 293
, J. 2004, in The Solar-B Mission and the Forefront of Solar Physics, ed. T. [Sakurai]{} & T. [Sekii]{}, ASP Conf. Ser. (San Francisco: ASP), in press (astro-ph/0404053)
, J., [Asensio Ramos]{}, A., [Trujillo Bueno]{}, J., & [Cernicharo]{}, J. 2001, , 555, 978
, J., [Dom[í]{}nguez Cerde[\~ n]{}a]{}, I., & [Kneer]{}, F. 2003, , 597, L177
, J., & [Lites]{}, B. W. 2000, , 532, 1215
, M., [Shelyag]{}, S., [Berdyugina]{}, S., [V[" o]{}gler]{}, A., & [Solanki]{}, S. K. 2003, , 597, L173
, G. B., [Bjelksjö]{}, K., [Korhonen]{}, T. K., [Lindberg]{}, B., & [Petterson]{}, B. 2003a, Proc. SPIE, 4853, 341
, G. B., [Dettori]{}, P. M., [Löfdahl]{}, M. G., & [Shand]{}, M. 2003b, Proc. SPIE, 4853, 370
, C. J., & [Title]{}, A. M. 2003, , 597, L165
, H., & [Sánchez Almeida]{}, J. 2002, , 565, 1323
, H., & [Sánchez Almeida]{}, J. 2003, , 593, 581
, H. C. 1977, , 55, 3
, O., [Hauschildt]{}, P. H., & [Bruls]{}, J. 2001, , 372, L13
, J. O. 1982, , 80, 209
, L. 1994, Ph.D. thesis, Utrecht University, Utrecht
, A. M., & [Berger]{}, T. E. 1996, , 463, 797
, J., [Shchukina]{}, N. G., & Asensio Ramos, A. 2004, , submitted
, R., & [Habbal]{}, S. R. 1997, , 24, 1159
[^1]: The highly magnetized plasma of a magnetic concentration is more transparent than the mean photosphere, allowing photons to escape from deep (and usually) hot sub-photospheric layers.
[^2]: http://soi.stanford.edu
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a lot of systems, charge transport is governed by local features rather than being a global property as suggested by extracting a single resistance value. Consequently, techniques that resolve local structure in the electronic potential are crucial for a detailed understanding of electronic transport in realistic devices. Recently, we have introduced a new potentiometry method based on low-energy electron microscopy (LEEM) that utilizes characteristic features in the reflectivity spectra of layered materials [@Kautz-LEEP]. Performing potentiometry experiments in LEEM has the advantage of being fast, offering a large field of view and the option to zoom in and out easily, and of being non-invasive compared to scanning-probe methods. However, not all materials show clear features in their reflectivity spectra. Here we, therefore, focus on a different version of low-energy electron potentiometry (LEEP) that uses the mirror mode transition, i.e. the drop in electron reflectivity around zero electron landing energy when they start to interact with the sample rather than being reflected in front of it. This transition is universal and sensitive to the local electrostatic surface potential (either workfunction or applied potential). It can consequently be used to perform LEEP experiments on a broader range of material compared to the method described in Ref. [@Kautz-LEEP]. We provide a detailed description of the experimental setup and demonstrate LEEP on workfunction-related intrinsic potential variations on the Si(111) surface and for a metal-semiconductor-metal junction with external bias applied. In the latter, we visualize the Schottky effect at the metal-semiconductor interface. Finally, we compare how robust the two LEEP techniques discussed above are against image distortions due to sample inhomogeneities or contamination.'
address:
- 'Huygens-Kamerlingh Onnes Laboratorium, Leiden University, P.O. Box 9504, NL-2300 RA Leiden, Netherlands'
- 'Department of Physics, Columbia University, New York, New York 10027, USA'
- 'IBM T.J. Watson Research Center, 1101 Kitchawan Road, P.O. Box 218, Yorktown Heights, New York 10598, USA'
author:
- 'Johannes Jobst$^\star$'
- 'Jaap Kautz$^\star$'
- Maria Mytiliniou
- 'Rudolf M. Tromp'
- Sense Jan van der Molen
bibliography:
- 'ZZ\_own-papers-2017\_LEEP-Ultramicroscopy.bib'
title: 'Low-Energy Electron Potentiometry'
---
Low-energy electron microscopy ,LEEM ,Potentiometry ,Work function ,Transport properties
Introduction {#sec:introduction}
============
While conductance measurements are commonplace in material characterization, such measurements are typically limited to averages over entire samples whereas local variations of conductance are lost. This poses limitations in the analysis of local effects which often govern global transport. For example, in topological insulators the current is completely carried by edge channels. Moreover, random charge inhomogeneity causes lateral variations in the doping level of semiconducting materials and thin films while step edges and grain or domain boundaries can locally influence their conductivity. This is particularly important for two-dimensional electron systems that are easily perturbed due to their large surface/bulk ratio. With the emergence of a novel material class of atomically layered crystals such as graphene, this is more relevant than ever. A good understanding of transport properties therefore requires a technique which can locally probe electrical conductance.
Several tools can be used to study charge transport by locally probing the electrical potential at the surface. Photoelectron emission microscopy (PEEM) [@Ballarotto-PEEM-dopants; @Giesen-PEEM-pn] and synchrotron-based scanning PEEM [@Phaneuf-SPEM-pn], for example, were used to study doping levels and band bending in semiconductor p-n junctions. And while scanning probe techniques such as Kelvin probe force microscopy [@tzalenchuk-kelvin-probe], scanning squid microscopy [@nowack-imaging-SHE] and scanning tunneling microscopy [@Park-STM-pn; @ross-steps] have provided exciting new results, their acquisition times are long and their field of view is limited due to the scanning nature of the methods. In the faster scanning electron microscopy, potentiometry measurements can be performed by studying the effect of the local sample potential on the secondary-electron trajectories [@Oatley1957; @Menzel1983; @Girard1988]. However, the secondary-electron emission is strongly material dependent [@Nakamae1981; @Girard1988] and the high landing energy (0.5–10keV) of the electrons can severely influence the local conductivity[@Nakamae1981; @Girard1988], making the results difficult to interpret. This problem can be overcome in low-energy electron microscopy (LEEM) where the electron landing energy is typically in the range of 0–30eV. LEEM-based potentiometry (LEEP) thus provides a fast alternative with a large field of view. Anderson and Kellogg used LEEM to probe potential differences in doped silicon samples [@Kellogg-pn], and we showed how the sensitivity of LEEP to the local potential can be further enhanced [@Kautz-LEEP] by studying the full energy dependence of electron reflection.
In the present paper we describe a generalized version of LEEP, which uses the LEEM Mirror-Mode Transition (MMT) – a steep drop in intensity as we change the sample potential relative to the electron gun potential from a more negative value where all electrons are reflected above the sample, to a more positive value where they interact with the surface. We therefore, call this method mirror-mode LEEP (M-LEEP). The MMT is present for all materials while clear LEED (Low Energy Electron Diffraction) IV-curves as used in Ref. [@Kautz-LEEP] are not generally available. Consequently, M-LEEP can be applied to a broader range of materials, of which we show two applications in this paper. First, we demonstrate M-LEEP by measuring the intrinsic potential differences between different surface reconstructions of the Si(111) surface [@Hannon-phase-transition]. Second, we locally probe the potential distribution within a metal-semiconductor-metal junction over which an external bias is applied. This allows us to directly probe the Schottky effect without the use of metallic probe contacts, which could easily disturb such a measurement.
![ (a) For negative voltages [$V_0$]{}, all electrons are reflected and the reflection intensity is high for all positions. (b) For positive [$V_0$]{}, electrons scatter off the surface reducing the reflection intensity. (c) When a bias voltage [$V_{\text{bias}}$]{} is applied between source and drain, the landing energy and the reflection intensity become position dependent. At the MMT, i.e., where the landing energy is zero, the intensity drops sharply. (d) By varying [$V_0$]{}, the position of the MMT shifts. By determining the value of [$V_0$]{} of the MMT for each position $r$, the local potential [$V(\mathbf{r})$]{} can be determined. []{data-label="fig:mleep"}](Fig1_mleep){width="\columnwidth"}
Potentiometry in LEEM {#sec:potentiometryinleem}
=====================
In LEEM, the sample is illuminated with a parallel beam of low-energy electrons. The electrons are extracted from the gun and accelerated to 15keV. While they travel through most of the electron-optical system at this high energy, their low kinetic landing energy ${\ensuremath{E_{0}}}\approx 0$–30eV is achieved by decelerating them towards the sample. For that purpose, the sample is lifted to $-15\text{\,kV} + {\ensuremath{V}}$. By changing the voltage [$V$]{}, the landing energy [$E_{0}$]{} of the electrons can thus be accurately tuned. A bright-field LEEM image is formed by the cathode objective lens, projecting the specularly reflected electrons with energy [$E_{0}$]{} onto a channelplate-intensified detector.
The reflectivity, and therefore, the image intensity $I$, in LEEM experiments is a strong function of [$E_{0}$]{}. The intensity of the (0,0) LEED beam with energy (the so-called LEED-IV curve) is highly sensitive to surface atomic structure [@hannon-LEEM-IV; @Flege-IV], as well as the unoccupied band structure [@Jobst-ARRES; @Jobst-ARRES-GonBN] of the sample under investigation. One feature present in the IV-curves for all materials is the MMT when the landing energy is changed from negative to positive values. In mirror mode (MM), i.e. for negative [$E_{0}$]{}, all incoming electrons are repelled and reflected before they reach the sample (Fig. \[fig:mleep\]a). This leads to a high intensity on the detector. For positive [$E_{0}$]{}, the electrons are decelerated less strongly and reflect from the sample surface (Fig. \[fig:mleep\]b) causing a reduced intensity on the detector. The MMT is therefore characterized by a steep drop in specular reflection intensity that can easily be recognized in the IV-curve. This intensity drop can be described by an error function (erf), if you assume a Gaussian energy distribution of the incident electrons, the center of which defines the voltage [$V_{\text{MMT}}$]{} for which the average landing energy is zero.
During regular LEEM imaging of a homogeneous sample, the landing energy and therefore also [$V_{\text{MMT}}$]{} is the same for the whole illuminated area. In a typical LEEP experiment, in contrast, the surface potential [$V(\mathbf{r})$]{} is inhomogeneous over the sample and therefore, the local landing energy [$E_0(\mathbf{r})$]{} becomes a function of position $${\ensuremath{E_0(\mathbf{r})}}= {\ensuremath{V_0}}- e \left[{\ensuremath{\Delta \Phi (\mathbf{r})}}+ {\ensuremath{U(\mathbf{r})}}\right],$$ where [$\Delta \Phi (\mathbf{r})$]{} is the spatial variation of the workfunction and [$U(\mathbf{r})$]{} the spatial variation of the surface potential due to an applied bias voltage [$V_{\text{bias}}$]{} between $\mathbf{r}$ and a reference point [$\mathbf{r}_0$]{} (e.g. in Sec. \[sec:WF\], a point on the $7\times 7$-reconstruction). The decelerating voltage [$V_0$]{} is defined such that ${\ensuremath{V_0}}:= 0$ when ${\ensuremath{E_0(\mathbf{r}_0)}}= 0$. A position dependence of [$\Delta \Phi (\mathbf{r})$]{} can arise due to workfunction differences within the surface, e.g. by the coexistence of different surface reconstructions. While [$\Delta \Phi (\mathbf{r})$]{} is an intrinsic property of the studied surface, [$U(\mathbf{r})$]{} is defined by how [$V_{\text{bias}}$]{} drops over the sample (Fig. \[fig:mleep\]c) and therefore, contains information about local conductance. Both these properties can be investigated by measuring the landing energy locally resolved.
In M-LEEP we measure [$E_0(\mathbf{r})$]{} by acquiring LEEM images while scanning the voltage [$V_0$]{} and using the steep intensity drop at the MMT to determine [$V_{\text{MMT}}$]{} for each position. The difference in [$V_{\text{MMT}}$]{} is then exactly the difference in local landing energy and hence local surface potential [$V(\mathbf{r})$]{} between those two points. As an example, we sketch in Fig. \[fig:mleep\]c and d how the position where ${\ensuremath{E_0(\mathbf{r})}}= 0$ shifts over the sample for different values of [$V_0$]{}, when an external bias is applied.
In this article, we discuss the two fundamental cases. First we apply no external bias ${\ensuremath{U(\mathbf{r})}}= 0$ and study workfunction variations within the sample. Second, we apply an external bias over a sample with no lateral workfunction variations ${\ensuremath{\Delta \Phi (\mathbf{r})}}= 0$. For the first case, we study the silicon (111) surface that shows different reconstructions with different surface potentials. For the second case, we look at the voltage drop between metallic electrodes on unreconstructed silicon.
![ (a) Construction drawing of the LEEP puck with four pins making electrical contact to the sample corners. They are pulled down by spring-loaded screws (blue) to maintain good contact during heating experiments. One of them is connected to the cap with a screw to serve as local potential reference. (b) A standard SPECS molybdenum cap [@Tromp-stage] that shields sharp corners at the sample and the contacts is fixed on the puck by screws such that a 200[m]{} wide gap isolates it from the contact pins. A tungsten filament (red) below the sample serves as electron bombardment heater. []{data-label="fig:puck"}](Fig2_resub_puck){width="\columnwidth"}
LEEP sample holder and electronics {#sec:design}
==================================
Our experiments are performed at the ESCHER facility [@ESCHER], which is based on the commercially available, aberration-corrected FE-LEEM P90 instrument (SPECS GmbH, Berlin) designed by IBM [@tromp-AC1]. In order to perform a LEEP experiment, a number of additional requirements compared to standard LEEM have to be fulfilled and the experimental setup has to be adjusted accordingly. First, an in-plane bias voltage has to be applied across the sample while imaging. Second, the current through the sample needs to be measured to be able to compute the local conductivity from the derived surface potential. Third, 4-probe voltage measurements have to be performed *in situ* to compare the LEEP results to the global conductance. For that purpose, we developed special measurement electronics and a new puck that holds the sample and makes electrical contact to it. The puck is milled from aluminum oxide and thus, is non-magnetic and compatible with ultra-high-vacuum (UHV). It contains an electron bombardment heater for temperatures up to 1450K, fits onto the existing SPECS sample holder (described in Ref. [@Tromp-stage]) and is only $\approx\,1\,\text{cm}^3$ small (cf. Fig. \[fig:puck\]a). A bayonet fitting at the side of the cap allows for *in situ* mounting of the puck on the macor carrier piece [@Tromp-stage].
The sample is mounted onto the puck and is electrically contacted by molybdenum pins, which are pressed onto the corners of the sample by springs (Fig. \[fig:puck\]b). Metallic contact pads in the corners of the sample lead towards its center where the devices are structured and LEEP microscopy is performed. The contact pins are concealed below the standard SPECS sample cap to prevent field enhancement at their edges that could cause arcs. The cap is lifted to $\approx -15$kV where the precise voltage between cap and objective lens controls the landing energy of the electrons in LEEM/LEEP (see Sec. \[sec:potentiometryinleem\]). In order to make the cap potential the local reference potential (as in standard LEEM), the cap is fixed onto the puck such that it is electrically connected to one of the contact pins. It is insulated from the other contacts by a 200[m]{} wide gap (cf. Fig. \[fig:puck\]c). Apart from the four connections to the sample, two more contacts lead to the tungsten filament of the electron bombardment heater.
Under operating conditions of the microscope, the sample is kept at a potential of $\approx -15$kV and thus cannot be directly connected to a conventional power supply to apply [$V_{\text{bias}}$]{}. Consequently, the electronics are divided in a low-voltage and a high-voltage part that are electronically decoupled through fiber-optical connections. The applied voltage as well as the measured voltages and currents are optically transmitted by a frequency modulated signal via these fibers. All electronic components that are in direct contact with the sample are equipped with an overvoltage protection consisting of a spark gap placed parallel to two Zener diodes. It diverts the current to the local ground in the case of arcing due to the strong electric field between sample and objective lens.
![ The electronics are divided in a high- and a low-voltage part, which are optically decoupled. The high-voltage part applies [$V_{\text{bias}}$]{} over the sample and measures the current through the sample as well as the two-point and four-point voltage drop. []{data-label="fig:electronics"}](Fig3_electronics){width="\columnwidth"}
Results {#sec:results}
=======
Intrinsic Potential Differences of the Si(111) Surface\[sec:WF\]
----------------------------------------------------------------
We use the Si(111) surface that exhibits two surface structures with intrinsic surface potential differences [@Hannon2003] as a test case to verify the potential resolving capabilities of LEEM. Above 1135K, Si(111) shows the unreconstructed 1$\times$1 phase, while a 7$\times$7 reconstruction is formed when the temperature is lowered. Around the phase transition temperature, the surface forms a mixture of 7$\times$7 and 1$\times$1 reconstructed domains [@Hannon2003]. We prepare these mixed surfaces by flash annealing Si(111) samples several times to $\approx 1450$K for 10 seconds to remove the native oxide and then cool down to the desired temperature.
![ (a) Bright-field LEEM micrograph (${\ensuremath{E_{0}}}= 1.6$eV) showing the almost completely 7$\times$7 reconstructed surface (bright) at 870K. Small 1$\times$1 areas exist only at step edges (dark lines from bottom left to top right) and domain boundaries (dark lines perpendicular to step edges). (b) At 1040K a mixture of triangular 7$\times$7 domains (bright) and 1$\times$1 reconstructed areas (dark) are observed (${\ensuremath{E_{0}}}= 2.0$eV). (c) IV-curves taken from two different areas marked in (b) show a clear MMT (solid red and green for 7$\times$7 and 1$\times$1 reconstruction, respectively). An error function (dashed) is fitted to the data to determine [$V_{\text{MMT}}$]{}.[]{data-label="fig:mixedmm"}](Fig4_resub_mixed){width="\columnwidth"}
Figure \[fig:mixedmm\]a shows a real-space bright-field image of a Si(111) surface at 870K. The surface is widely 7$\times$7 reconstructed with 1$\times$1 domains only at the step edges. Domain boundaries are visible as dark lines perpendicular to the step edges. At 1040K, around the phase transition temperature, the surface exhibits triangular 7$\times$7 domains surrounded by 1$\times$1 reconstructed areas (Fig. \[fig:mixedmm\]b).
We measure the difference in the surface potential of these surface reconstructions, by varying the sample potential via [$V_0$]{} while imaging the sample as described in Sec. \[sec:potentiometryinleem\]. In Fig. \[fig:mixedmm\]c, the IV-curves for the positions marked in Fig. \[fig:mixedmm\]b clearly show the MMT as a strong drop in intensity. The fact that this transition is not infinitely sharp, is caused by the energy spread of the electrons leaving the gun. If we assume this spread to be Gaussian, we can deduce the standard deviation of the electron energy distribution as 0.15eV from the width of the fitted error function (dashed lines in Fig. \[fig:mixedmm\]c). The center of the fitted error function defines [$V_{\text{MMT}}$]{} for which the local landing energy ${\ensuremath{E_0(\mathbf{r})}}= 0$. Repeating this procedure for each pixel individually, we obtain a map of the surface potential [$V(\mathbf{r})$]{} of the sample as shown in Fig. \[fig:potmap\]a. Since no external bias is applied over the sample, the local variations are caused by workfunction differences between the different surface reconstructions. The domains of unreconstructed and 7$\times$7-reconstructed that are visible in Fig. \[fig:mixedmm\]b, clearly show different workfunction values in this LEEP image. In the vicinity of step edges and impurities, however, artifacts caused by lateral field inhomogeneity is observed as well.
![ (a) A potential map of the area in Fig. \[fig:mixedmm\] clearly resolves that 7$\times$7 reconstructed domains have a higher surface potential than the surrounding 1$\times$1 areas. (b) Histograms of the potential values in (a) shows that this difference is $0.12 \pm 0.02$V. This is more clear from the top histogram gathered from the area marked black in (a), whereas two more peaks (black lines) are visible in the bottom histogram that is taken over the full field of view. They arise from artifacts due to lateral fields at domain boundaries (visible as bright blue and green lines in (a)). (c) From a line scan of the area indicated in (a), the spatial resolution, defined as the lateral distance between 20% and 80% of the step height, is determined to be 40nm. The determined workfunction difference is 0.13eV. []{data-label="fig:potmap"}](Fig5_resub_potmap){width="\columnwidth"}
In order to quantify the workfunction difference between different domains, we plot histograms of the local potential values for this sample in Fig. \[fig:potmap\]b. The top histogram is taken from the area indicated in black in Fig. \[fig:potmap\]a that is free from field-induced artifacts. It shows one peak from the $7 \times 7$ reconstruction at lower surface potential and a second one from the $1 \times 1$ areas, clearly separated from one another. From this splitting, we determine the difference in surface potential to be $0.12 \pm 0.02$V, which is comparable with the $0.15 \pm 0.03$eV found by Hannon *et al.* [@Hannon-phase-transition]. The bottom histogram in Fig. \[fig:potmap\]b is generated from the full field of view. In addition to the 1$\times$1 and the 7$\times$7 peaks, two more peaks are needed to describe the histogram. Those values are caused by the abovementioned artifacts and can obscure the data evaluation easily. The spatial resolution of M-LEEP is determined by looking at a line scan over the 7$\times$7-1$\times$1 interface (see Fig. \[fig:potmap\]c). The height of the step is $0.13 \pm 0.02$eV in agreement with the evaluation from the histogram in Fig. \[fig:potmap\]b and literature. From the lateral spacing between 20% and 80% lines of the step in potential, a resolution of 40nm is extracted. While here, the LEEM images are slightly out of focus, we expect the spatial resolution of the potentiometry measurements to be closer to the optimal LEEM resolution ($< 2$nm) [@schramm-thesis] for optimal alignment conditions. With M-LEEP, however, this optimal microscopy resolution cannot be reached since the M-LEEP resolution will always be limited by the presence of a lateral electric field above the 7$\times$7-1$\times$1 interface, induced by the difference in surface potential [@Kennedy-MM-distortions; @Kennedy-MM-distortions2]. We estimate 10nm to be the resolution limit of M-LEEP. This experiment nevertheless clearly demonstrates how M-LEEP can be used to measure local changes in the workfunction of a surface.
Electron Transport Through a Metal-Semiconductor-Metal Junction \[sec:Schottky\]
--------------------------------------------------------------------------------
In order to study the potential distribution due to an external applied bias with M-LEEP, we prepare an undoped Si(111) sample ($\rho=10\,$k$\Omega$cm) with two metal contacts. Dust particles are removed from the sample with acetone and isopropanol. Then two contacts separated by 4m are fabricated by shadow evaporation of 5nm of chromium and 30nm of Au using a tungsten wire as a mask, thereby minimizing contamination of the sample. Subsequently, the sample is cleaned by nitric acid ($97\%$ Sigma Aldrich) for 10 seconds, rinsing in demineralized water and a 30 second treatment in buffered oxide etch \[7:1 ammonium fluoride (40% in water) and hydrofluoric acid (49% in water), J.T. Baker\]. The latter removes the native oxide and passivates the Si surface with hydrogen. The sample is then directly loaded into the LEEM to prevent oxidation and contamination and heated to $\approx$600K *in situ*.
A LEEM image of the resulting two-probe structure is shown in Fig. \[fig:slices\]a with the drain and source contacts on the left and right, respectively. The central Si(111) surface is clean and free of oxide as confirmed by the LEED pattern in Fig. \[fig:slices\]b, proving the effectiveness of the cleaning method used. Figure \[fig:slices\]a is recorded at ${\ensuremath{V_0}}{} = 0.3$V with an external bias of ${\ensuremath{V_{\text{bias}}}}{} = 2$V applied, resembling the situation in Fig. \[fig:mleep\]c. As a consequence, the left side of the sample appears much brighter than the right side because the former is in MM where all electrons are reflected (bright), while the latter is in imaging conditions where electrons scatter from the sample surface (darker). In between, the position of the MMT can be found, or in other words, a position where the local landing energy is zero. This MM-position shifts over the sample when [$V_0$]{} is changed as sketched in Fig. \[fig:mleep\]c,d. The evolution of the MM-position is visualized in Fig. \[fig:slices\]c where intensity profiles taken parallel to the current paths (along the red line in Fig. \[fig:slices\]a) are plotted for different [$V_0$]{}. It is noteworthy that this is similar to plotting IV-curves for every position on the line. Figure \[fig:slices\]d, for example, shows IV-curves for the three points indicated by circles in Fig. \[fig:slices\]a. For low values of [$V_0$]{}, the whole sample is in MM and the intensity is high everywhere in (cf. left side of Fig. \[fig:slices\]c), while for high values of [$V_0$]{} the intensity is lower throughout the device. The boundary between dark and bright is the MMT which is a direct visualization of the local potential ${\ensuremath{V(\mathbf{r})}}{}$. Two features are evident in the lateral change of the local potential shown in Fig. \[fig:slices\]c. First, a linear potential gradient over the Si is visible as expected for a homogeneous material. Second, the potential exhibits pronounced steps at the metal-Si interfaces.
![ (a) LEEM image of the metal-semiconductor-metal junction at [$V_0$]{}=0.3eV and an applied bias of ${\ensuremath{V_{\text{bias}}}}= 2$V. The center of the gap just goes through the MMT. (b) LEED pattern of the Si(111) surface within the junction. (c) Line scans at the line indicated in (a) for different values of [$V_0$]{} visualize the influence of the bias voltage. A linear voltage drop over the silicon and a large jump in voltage at the right metal-Si interface are apparent. (d) IV-curves taken at the positions indicated by circles in (a) corresponding to vertical cuts in (c). The voltage difference between the different points is visible as an energy shift of the MMT. []{data-label="fig:slices"}](Fig6_AuSi_300dpi){width="\columnwidth"}
For a more rigorous, quantitative investigation, the automated determination of the MMT via IV-curves that is described in Sec. \[sec:WF\] is used for every pixel in Fig. \[fig:slices\]a to calculate a map of the local electrical potential (Fig. \[fig:potmapgap\]a). Figure \[fig:potmapgap\]b shows a line scan parallel to the current flow (along the red line in Fig. \[fig:potmapgap\]a). It confirms the linear voltage drop over the silicon and the steep potential step at the metal-Si interface of the source contact that are already visible in Fig. \[fig:slices\]c. From this step we estimate the lateral resolution of M-LEEP to be $\sim$100nm using the same 80%-20%-criterion as in Fig. \[fig:potmap\]c. Here, the lateral resolution is mainly limited by the large height difference of $\approx 35$nm between Si surface and Au electrode and a not perfectly clean sample. In addition, the lateral electrical field due to the difference in surface potential deforms the image. These effects are particularly pronounced for M-LEEP because electron trajectories are affected by asperity and impurities of the surface strongest around the MM as then, the electron energy is minimal [@Kennedy-MM-distortions; @Kennedy-MM-distortions2].
![ (a) Potential map of metal-Si interface at the biased source electrode. (b) Line scan along the line indicated in (a). From the errors in the fits to the line scan, a voltage resolution of 40mV and 60mV is derived for the Si and Au areas, respectively. The spatial resolution is determined to be 100nm from the lateral spacing between 20% and 80% of the step height. []{data-label="fig:potmapgap"}](Fig7_potmap-electrodes){width="\columnwidth"}
We take the measured potential fluctuations around the expected trends as an estimate of the resolution in electrical potential. To obtain those, we fit linear trends to the data within the Au contact and the pure Si area, respectively. We take the standard deviation of the residuals of that fit as an upper bound to the resolution in potential. For the Au and Si surface this yields a resolution of 60mV and 40mV respectively. The resolution for the Au surface is limited by the surface roughness of the contact while the resolution for the Si surface is limited by the slight upwards bend of the potential. The latter is probably an artifact of the measurement (as the pronounced dip left of the contact is) and not caused by band bending, an effect which one would expect to occur within tens of nanometers from the interface. The observed bending might be caused by the lateral electric field over the Si area due to the [$V_{\text{bias}}$]{} applied between the electrodes. This field influences the trajectories of the imaging electrons by adding in-plane momentum and therefore, modifying the sample voltage at which the MMT occurs [@Jobst-ARRES; @Jobst-ARRES-GonBN].
![ Intensity as function of position and [$V_0$]{} of the line indicated in Fig. \[fig:slices\]a. For positive bias voltages, there is a voltage drop over the right metal-Si interface, while there is none over the left. For negative bias voltages, the situation is reversed. []{data-label="fig:vbsweep"}](Fig8_resub_slices){width="\columnwidth"}
In order to understand the potential step at the metal-Si interface of the source electrode in more detail, we perform M-LEEP measurements as a function of the applied bias [$V_{\text{bias}}$]{}. It is worth noting that such a repeated potentiometry measurement is only feasible because acquiring an M-LEEP data set takes only one minute in contrast to potential measurements with scanning probe techniques. Figure \[fig:vbsweep\] shows IV-curves for all pixels on a line across the junction, similar to Fig. \[fig:slices\]c for bias voltages from ${\ensuremath{V_{\text{bias}}}}= -2$V to ${\ensuremath{V_{\text{bias}}}}= 2$V. For all [$V_{\text{bias}}$]{} used, potential steps at both electrodes are visible that are composed of a constant part and a part that depends on [$V_{\text{bias}}$]{}. The constant part is the workfunction difference between the Au electrodes and the Si(111) surface (similar to Sec. \[sec:WF\]) and can be determined from the case of ${\ensuremath{V_{\text{bias}}}}= 0$ as $\Delta\Phi_{\text{Au-Si}} \approx 0.35$eV in agreement with literature [@WF-Si; @WF-Au]. The [$V_{\text{bias}}$]{}-dependent part has a more complicated, non-linear behavior: For small positive [$V_{\text{bias}}$]{}, most of the voltage drops over the right metal-Si interface, while the Si surface and the left electrode remain at an unchanged, constant potential. Only for ${\ensuremath{V_{\text{bias}}}}> 0.6$V, a clear voltage gradient over the silicon is observed while the steep potential drop over the right metal-Si interface persists. For negative [$V_{\text{bias}}$]{} the same behavior can be observed at the left metal-Si interface. This non-linear behavior can be explained by Schottky contacts formed between the metallic electrodes and the semiconducting Si. As a consequence, the interface is transparent for the current if it is forward biased causing only a negligible voltage drop at the interface while a sizable fraction of the applied voltage drops over the Schottky contact when it is reverse biased. In the full metal-Si and Si-metal device studied, the metal-Si and the Si-metal Schottky junctions at the left and right side, respectively, behave like two diodes with inverted polarities placed in series. Consequently, for positive [$V_{\text{bias}}$]{}, the right metal-Si interface is reverse biased (large potential step), while the left metal-Si interface is forward biased (small potential step). For negative bias voltages, the situation is reversed. In other words, the junctions are reverse biased when a positive voltage is applied to the metal contact. This suggests that our Si substrate is slightly p-doped [@Laughton2002] due to the cleaning and annealing procedures used.
These experiments clearly demonstrate how M-LEEP provides detailed information on local conductance properties. This is particularly important as these local contact effects are invisible in a conventional 2-point resistance measurement. The current versus [$V_{\text{bias}}$]{}-measurement shown in the last panel of Fig. \[fig:vbsweep\] is almost perfectly linear, averaging over both Schottky diodes and thus, obscuring the true physics that govern the device performance.
Robustness of M-LEEP
--------------------
In Sec. \[sec:WF\] and \[sec:Schottky\], we have shown that M-LEEP is a versatile tool to determine surface potentials, due to intrinsic workfunction differences or due to externally applied voltages. The advantage of using the steep MMT is that it is universally available for all materials. A notable complication arises, however, if the material studied has a bandgap right at the vacuum level. This bandgap causes low reflectivity over its whole energy range, which leads to an apparent shift of the MMT [@Flege-IV; @fujikawa-silver].
We have seen in Fig. \[fig:potmapgap\] that the electrons are easily affected by surface roughness, three-dimensionality of the sample or lateral electric fields due to their low kinetic energy around the MM. This limits the lateral resolution of M-LEEP to values well below the optimal resolution of LEEM [@Kennedy-MM-distortions; @Kennedy-MM-distortions2]. Various materials, on the other hand, show other pronounced features in their IV-curves at higher energies where the electron trajectories are less affected by these distortions. We will demonstrate in the following that those features, instead of the MMT, can be used to determine the local potential, and that this approach leads to better lateral resolution [@Kautz-LEEP].
![ (a) LEEM image (${\ensuremath{V_0}}= 4.1$V) of monolayer (ML), bilayer (BL) and trilayer (TL) graphene with ${\ensuremath{V_{\text{bias}}}}= -3$V applied from left to right (electrodes are not shown). (b) Same area recorded close to MMT (${\ensuremath{V_0}}= 2.4$V). The apparent size of imperfections (e.g. arrows) increases close to the MMT. (c) IV-curves taken at the points marked in red in a,b exhibit a clear difference in the exact shape of the drop of the MMT. (d) The normalized IV-curves from c coincide for higher energies, while the shift in MM is even enhanced. (e) The local potential, determined by using the MMT (black squares) or shape of the IV-curve (red circles) as criterion. The former shows strong variations with position although the measurement spots lie on what is to be expected an equipotential line, while the latter exhibits much less variation. The data underlying this figure is also published as Supporting Material of Ref. [@Kautz-LEEP]. []{data-label="fig:MMTvsIV"}](Fig9_MMTvsIV_300dpi){width="\columnwidth"}
Figure \[fig:MMTvsIV\]a shows a LEEM image of an area of monolayer, bilayer and trilayer graphene grown on insulating silicon carbide as described in Ref. [@emtsev-natmat] with ${\ensuremath{V_{\text{bias}}}}= -3$V applied from left to right. The structure of the sample is clearly visible at the used ${\ensuremath{V_0}}= 4.1$V, while in Fig. \[fig:MMTvsIV\]b, which is recorded close to the MMT, most of the detail is lost. For example, the apparent size of the imperfections marked with arrows is greatly increased in Fig. \[fig:MMTvsIV\]b compared to a. We record IV-curves for points around an imperfection that lie on an equipotential line (Fig. \[fig:MMTvsIV\]c). As expected, the IV-curves look very similar and no large shift is observed in their MMT. However, the exact shape of the MMT varies from spot to spot indicating that problems could occur if the M-LEEP method as described in Sec. \[sec:potentiometryinleem\] is used. In fact, if one zooms in on the normalized IV-curves in Fig. \[fig:MMTvsIV\]d, this problem becomes immediately apparent. As a consequence, the local potential as determined by M-LEEP exhibits a pronounced lateral variation (black squares in Fig. \[fig:MMTvsIV\]e). This is unexpected as the sampled points lie on an anticipated equipotential line. If the minima and maxima at higher energy are used as a criterion to calculate the shift of the IV-curves rather than the MMT, no such potential variation is observed (red circles in Fig. \[fig:MMTvsIV\]e) as it is expected from the placement of the measurement points. This comparison confirms that features of the IV-curve at higher energy, if available, are more robust as a basis for LEEP [@Kautz-LEEP]. This is particularly important as samples for such transport measurements are often processed by nanofabrication and thus contaminated with residues of lithography resists and other chemicals.
A way to perform M-LEEP on such processed devices from materials with featureless IV-curves (e.g. silicon as seen in Sec. \[sec:Schottky\]) is outlined in Fig. \[fig:MMTonContacts\]. A Hall bar with a graphene channel (Fig. \[fig:MMTonContacts\]a) is patterned by electron-beam lithography and subsequent oxygen plasma etching. Source and drain contacts as well a multiple voltage probes are defined from Cr/Au (5nm/30nm) using lithography and evaporation. The source and drain contacts are connected to the LEEP sample holder as described in Sec. \[sec:potentiometryinleem\] and a bias voltage [$V_{\text{bias}}$]{} is applied between them. The metallic probes to measure the local voltage drop, on the other hand, are too small to be connected to external measurement equipment but can easily be resolved in LEEM (cf. Fig. \[fig:MMTonContacts\]b). Performing M-LEEP on these 1m$^2$ big contact pads yields values for the local [$V_{\text{MMT}}$]{}. These values are plotted as a function of the applied [$V_{\text{bias}}$]{} for contacts 1 to 9 in Fig. \[fig:MMTonContacts\]c. The expected linear trend following Ohm’s law is clearly visible. The slope of these linear fits yields the percentage of [$V_{\text{bias}}$]{} that drops between the according contact and drain. In Fig. \[fig:MMTonContacts\]d, these values are plotted over the contact number that is also the distance (in micrometer) from drain. A linear trend is apparent, the slope of which matches the geometry of the Hall bar. Its slight offset is caused by the resistance of the cables, the connections on the chip and the contact resistance between the contacts and graphene, all in series with the channel. While this behavior indicates the reliability of M-LEEP, the outlier of contact 8, which is caused by dirt on the contact, also strengthens the point that cleanliness is of utmost importance for M-LEEP measurements.
![ (a) Scanning electron micrograph of a graphene Hall bar with Cr/Au electrodes. (b) LEEM image showing some contact pads. IV curves are taken in the marked areas. (c) The voltage [$V_{\text{MMT}}$]{} for which the MMT occurs, is a linearly dependent on the applied bias [$V_{\text{bias}}$]{}. Lines are linear fits. (d) The potential determined by M-LEEP as the slope of the fits in a on the center of contact pads shows the expected linear voltage drop. []{data-label="fig:MMTonContacts"}](Fig10_MMTonContacts){width="\columnwidth"}
Conclusions
===========
Even for simple devices, charge transport can be dominated by local features making it crucial to resolve the potential landscape laterally. Here we show that LEEM-based potentiometry offers a wide range of opportunities as a versatile tool for that purpose. We demonstrate that the MMT, available in the IV-curve of any material, allows us to determine the surface potential of virtually any sample (with the small exception of materials that have a bandgap directly at the vacuum level) that can be imaged in LEEM. Introducing M-LEEP, we study the intrinsic surface potential changes induced by the different workfunctions of the $1\times1$ and the $7\times7$ reconstruction of the Si(111) surface as well as the potential drop induced by applying an external bias over a metal-semiconductor-metal junction. We conclude from the latter that a Schottky contact is formed at the metal-Si interface. Remarkably, this local conductance property is invisible in conventional two-point resistance measurements, while we can directly visualize its behavior for various applied voltages using M-LEEP. We demonstrate a lateral resolution of 40nm and a potential resolution of 40mV using M-LEEP. The resolution here is limited by the fact that the electron energy around the MM is minimal and hence their trajectories are easily deformed by impurities or asperity of the surface. The robustness against such distortions of LEEP can consequently be improved by using features of the IV-curve at higher energy if such features exist. We demonstrate that for the example of few layer graphene that exhibits pronounced oscillations in the IV-curve between 0eV and 5eV. We conclude that while M-LEEP cannot reach the lateral resolution of such IV-LEEP, M-LEEP is more versatile in terms of the materials that can be studied since a clear MMT can be observed on nearly all surfaces that can be imaged with LEEM.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to Ruud van Egmond, Martijn Witlox, Raymond Koehler, Leendert Prevo, Marcel Hesselberth, and Daan Boltje for technical assistance, and to Alexander van der Torren, Daniël Geelen, Aniket Thete, Jan van Ruitenbeek, and Jan Aarts for useful discussions. This work was supported by the Netherlands Organization for Scientific Research (NWO) via an NWO-Groot grant (ESCHER) and a VIDI grant (680-47-502, SJvdM) and a VENI grant (680-47-447, J.J.) as well as by the FOM foundation via the Physics in 1D program.
References {#references .unnumbered}
==========
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Rafael Wisniewski[^1]'
bibliography:
- 'AbstractionsLit.bib'
title: Combinatorial Abstractions of Dynamical Systems
---
Extended Abstract
=================
Formal verification has been successfully developed in computer science for verifying combinatorial classes of models and specifications [@Clarke:1986:AVF:5397.5399]. In like manner, formal verification methods have been developed for dynamical systems [@tabuada2009verification]. However, the verification of system properties, such as safety, is based on reachability calculations, which are the sources of insurmountable complexity. This talk addresses indirect verification methods, which are based on abstracting the dynamical systems by models of reduced complexity and preserving central properties of the original systems.
Specifically, in this talk, I consider a dynamical system $\mathcal C = (M,\xi)$, where $M$ is the state space - a closed manifold, and $\xi$ is a smooth vector field on $M$.
We denote a flow line of $\xi$ by $\phi_x(t) \equiv \phi_x^{\xi} (t)$, that is $$\frac{d}{dt} \phi_x (t) = \xi\left({\phi_x(t)}\right) \hbox{ with } \phi_x(0) =
x.$$ The manifold $M$ is compact; thus, the vector field $\xi$ generates a 1-parameter group $\phi_t: M \rightarrow M,~t \in \Bbb R,$ of diffeomorphisms. The smooth flow map $\phi: \Bbb R \times M
\rightarrow M$ is related to $\phi_t$ in the following way $$\phi(t,x) \equiv \phi_t(x) \equiv \phi_x(t).$$
We will examine examples of candidates for the combinatorial system $\mathcal D$ that mirrors the behaviour of $C$. For now, the combinatorial system $\mathcal D$ is a pair $(Z,\Phi)$ consisting of a finite set $Z$, and a function $\Phi: \mathbb R \times Z \to 2^Z$, where $2^Z$ denotes the power set of $Z$. We think about $Z$ as a discrete state space and about $\Phi$ as a discrete flow map. Subsequently, we will discuss methods of converting the dynamical system $\mathcal C$ to a combinatorial object $\mathcal D$.
For $z \in Z$, the cell $[z] = \mathcal A^{-1}(z) \subset M$. If the cells are disjoint, the collection $K = \{[z]|~z \in Z\}$ is called a partition of the state space $M$; whereas, if a pair $[z] \cap [z'] \neq \emptyset$, the collection is called a cover.
An abstraction is an over-approximation if for any $(t,x) \in \mathbb R_{\geq 0} \times M$ $$\mathcal A \circ \phi(t,x) \subseteq \Phi(t, \mathcal A(x));$$ $\mathcal A$ is an under-approximation if $$\Phi(t, \mathcal A(x)) \subseteq \mathcal A \circ \phi(t,x).$$ If $\mathcal A$ is a both under- and an over-approximation, then it is called a complete abstraction. For the questions related to safety, one might choose an over-approximation; whereas, for the questions corresponding to reachability, one might work with an under-approximation. Conservativeness of the abstraction, say over-approximation, is measured by the volume, $$\sup_{t \in \mathbb R_{\geq 0}} \max_{z \in Z} \mathrm{vol} (\Phi(t, z) \setminus \mathcal A \circ \phi(t,[z])).$$
Below, we sketch a number of examples discussed during the talk.
Suppose $\{U_z | z \in Z\}$ is a finite family of subsets covering $M$. Let $\mathcal D$ be given by $Z$ and $\Phi(t,z) = \mathcal A \circ \phi(t,[z])$. Pick an order on Z. We define the abstraction $\mathcal A$ by $$\label{TrivialAbstraction} \mathcal A: x \mapsto \min\{z \in Z |~x \in U_z \}.$$ As a consequence of the definition of $\Phi$, the abstraction $\mathcal A$ is an over-approximation. In this example, the computation of $\Phi$ might be tedious if not impossible. Therefore, an approximation is in place.
To this end, we define $$\mathrm{pol}\{v_1,...v_l\} = \left\{\sum_{i=1}^l \alpha_i v_i(x) \left |~\alpha_i \geq 0 \hbox{ and } \sum_{i=1}^l \alpha_i^2 = 1\right.\right\}.$$ Let $L = \{L_i|~i = 1, \hdots, l\}$ be a family of linear vector fields, and define multivalued map $F(x) = \mathrm{pol} L(x)$. Suppose that $\xi \in F(x)$, and define $$\Phi(t,z) = \mathcal A \circ \mathrm{pol}\{ \phi^{L_1}(t,[z]), \hdots, \phi^{L_l}(t,[z]) \}.$$ The over-approximation might be relatively conservative, but the computation is simplified as the flow maps are linear in the second argument. The algorithm can be additionally simplified if the sets $U_{k}$ are polyhedral (in local patches).
Suppose that there exists a Finsler-Lyapunov (smooth) function [@ForliSepulchre] $V: TM \to \mathbb R$ (where $\pi: TM \to M$ is the tangent bundle) such that
1. $V(v) > 0$ for all $v \in TM \setminus 0_M$.
2. There is $p \in \mathbb N$ such that $V(\lambda v) = \lambda^p V(v)$ for all $v \in TM$ and $\lambda > 0$.
3. There is $p \in \mathbb N$ such that $V(v+w)^{\frac{1}{p}} < V(v)^{\frac{1}{p}} + V(w)^{\frac{1}{p}}$ for all $v, w \in TM$ with $\pi(v) = \pi(w)$.
The function $V$ defines metric $\rho$ on $M$ [@Tamassy2008483] $$\rho(x_1,x_2) = \inf_{\gamma \in \Gamma(x_1,x_2)} \int_{I} V(\dot \gamma)^{\frac{1}{p}} ds,$$ where $I = [0, 1]$, $\dot \gamma = \gamma_*(d/dt)$, $\Gamma(x_1,x_2)$ is the set of curves $I \to M$ with $\gamma(0) = x_1$ and $\gamma(1) = x_2$. Following Theorem 1 in \[Forni and Sepulchre\], if $dV: TM \to T^*(TM) $ satisfies the following inequality written in local coordinates $$DV(x,w)(\xi(x), D\xi(x)w) \leq - \alpha (V(x,w)),~ \hbox{ for all } (x,v) \in TM.$$ where $\alpha$ is a non-decreasing continuous function. Then $\rho(\phi(t,x_1), \phi(t,x_2)) \leq \alpha (\rho(x_1,x_2))$. Hence, the system incrementally stable [@Angeli02].
Since the state space $M$ is compact, it is possible to cover $M$ by the finite family $\{D(x_z,r_z) |~z \in Z\}$ of disks $D(x,r) = \{y \in M |~\rho(x,y) < r \}$ [@Frehse2008]. We define the abstraction $\mathcal A$ as in , and the combinatorial system $\mathcal D$ by $Z$ and $\Phi(t,z) = \mathcal A \phi(t,x_z)$. The abstraction $\mathcal A$ is an over-approximation. We note that computation of $\Phi$ amounts to simulating the dynamical system $\mathcal C$ for a finite number of initial conditions $x_z$.
Let $\xi$ be a Morse-Smale vector field on $M$ \[Palis and de Melo\]. Recall, a vector field $\xi \in \frak X^r(M)$ will be called Morse-Smale provided it satisfies the following five conditions:
1. $\xi$ has a finite number of singular points, say $\beta_1,...,\beta_k$, each hyperbolic,
2. $\xi$ has a finite number of closed orbits (periodic solutions), say $\beta_{k+1},...,\beta_N$, each hyperbolic;
3. For any $x \in M,~\alpha(x) = \beta_i$ and $\omega(x) = \beta_j$ for some $i$ and $j$;
4. $\Omega(\xi) = \{\beta_1,...,\beta_N \}$;
5. The stable and unstable manifolds associated with the $\beta_i$ have transversal intersection.
The sets $\beta_1,...,\beta_N$ will be called the singular elements of the vector field $\xi$. The set of the singular elements of $\xi$ will be denoted by ${\cal C}r(\xi)$. The stable (unstable) manifold of $\xi$ at a singular element $\beta$ is denoted by $ W^s(\beta_i)$ ($W^u(\beta_i)$).
We define a partial order relation on the singular elements of a Morse-Smale vector field: $\beta_i \succ \beta_j$ will mean that $W(\beta_i,\beta_j) \equiv W^u(\beta_i) \cap W^s(\beta_j) \neq \emptyset$. Consequently, each $W(\beta_i, \beta_j)$ is a cell, with the property that if $x \in W(\beta_i, \beta_j)$ then $\phi(t,x) \in W(\beta_i, \beta_j)$ for all $t \in \mathbb R$. Since the number of singular elements is finite, we can define $\mathcal D$ by $$Z = \{W(\beta_i,\beta_j)|~\beta_i \succ \beta_j \} \hbox{ and } \Phi(t,z) = z.$$
On the state space $M$, we define a family of functions $\{V_i: M \to \mathbb R|~i = 1, \hdots, l\}$ that satisfy
1. $d V_i(\xi)(x) \leq 0$.
2. Let $\mathrm{Reg}(V_i)$ be the set of regular values of $V_i$. For any singular element $\beta$ of $\xi$,
- if $V_i^{-1}(\mathrm{Reg}(V_i)) \cap W^s(\beta) \neq \emptyset$ then $W^u(\beta) \subset V_i^{-1}(V_i(\beta))$;
- if $V_i^{-1}(\mathrm{Reg}(V_i)) \cap W^u(\beta) \neq \emptyset$ then $W^s(\beta) \subset V_i^{-1}(V_i(\beta))$.
For each function $V_i$, we associate a family of regular values $A^i \equiv \{a^i_0, \hdots, a^i_k |~a^i_{k-1} < a^i_k\} \subset \mathbb R \cup \{-\infty, + \infty\}$. For $a^i_j \in A^i$, we define a shift operator $\sigma \equiv \sigma^i: a^i_j \mapsto a^i_{j-1}$ We use the notation $z = (z_1, \hdots, z_l)$ and define a cells $[z]$ with $z_i \in A^i$ by $$[z] = \bigcap V_i^{-1}([\sigma z_{i}, z_i])$$
Let $\mathbb R_{\infty} \equiv \mathbb R \cup \{-\infty,+\infty\}$. For each $z \in Z \equiv A^1 \times \hdots \times A^l$, we define a cube $\Box_{z} \equiv [\underline b_{z_1} \overline b_{z_1}] \times \hdots \times [\underline b_{z_l} \overline b_{z_l}] \subset \mathbb R_{\infty}^l$ with $\underline b_{z_i}$ ($\overline b_{z_i}$) being the minimal (maximal) time over the trajectories staring at $V_i^{-1}(\sigma z_i)$ and leaving $V_i^{-1}(z_i)$ (If $V_i^{-1}([\sigma z_i, z_i])$ is a positive invariant set, this time is set to $+\infty$). We denote the set of cubes in $\mathbb R^l$ by $\mathrm{Box}$. As a consequence, the combinatorial system is characterised by a map $\Box: Z \to \mathrm{Box}$ defined by $z \mapsto \Box_z$.
The following operator $L$ will be instrumental: $L = (L_1, \hdots, L_l) \to \mathbb R_{\infty}^l$, where $L_i = \partial \circ \pi_i$, $\pi_i$ is the projection on the $i$th component, and $\partial [\underline b, \overline b] = \overline b - \underline b$.
We define, a combinatorial system $\mathcal D$ by $Z$ and $\Phi$ as $$\begin{aligned}
\Phi(t,z) &=& \max \{z' \in A^1 \times \hdots \times A^l |~
\Box_z \equiv \Box_{z^0} < \Box_{z^1} \hdots < \Box_{z^m} \equiv \Box_{z'},~ \\
&~& L(\Box_{z^0} + \hdots + \Box_{z^m}) \leq (t, \hdots, t), \hbox{ and } z^{i-1} = \sigma z^i \hbox{ for } i = 1, \hdots ,m\}.\end{aligned}$$
By [@Sloth201380], this abstraction is complete.
[^1]: This work was supported by MT-LAB, a VKR Centre of Excellence for the Modeling of Information Technology.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An inhomogeneous Kaluza-Klein compactification to four dimensions, followed by a conformal transformation, results in a system with position dependent mass (PDM). This origin of a PDM is quite different from the condensed matter one. A substantial generalization of a previously studied nonlinear oscillator with variable mass is obtained, wherein the position dependence of the mass of a nonrelativistic particle is due to a dilatonic coupling function emerging from the extra dimension. Previously obtained solutions for such systems can be extended and reinterpreted as nonrelativistic particles interacting with dilaton fields, which, themselves, can have interesting structures. An application is presented for the nonlinear oscillator, where within the new scenario the particle is coupled to a dilatonic string.'
author:
- 'J. R. Morris'
title: New scenarios for classical and quantum mechanical systems with position dependent mass
---
Introduction
============
Quantum mechanical systems involving nonrelativistic particles with position dependent masses have been studied by numerous authors (see, e.g., [@Mathews]-[@epjp]). Such systems can arise in condensed matter settings, as pointed out, for example, in Ref.[@cari], but are of mathematical interests in their own right. An example of such a system is a previously studied prototype[@Mathews],[@cari] that derives from a classical lagrangian$$L=\frac{1}{2}\frac{m_{0}}{(\lambda x^{2}+1)}\left( \dot{x}^{2}-\alpha
^{2}x^{2}\right) \label{a1}$$
which can also be extended to higher dimensions[@refa],[@refb],[@refd],[@jmp],[@epjp]. The particle described by (\[a1\]) has an effective position dependent mass (PDM)$$m(x)=\frac{m_{0}}{(\lambda x^{2}+1)}\label{a2}$$
and is subject to an effective spring constant $k(x)=m(x)\alpha^{2}$. This problem has been studied at both the classical and quantum mechanical level, and in various space dimensions. (See, e.g., [@cari] and [@axel1]-[@epjp] for exact solutions of the quantum mechanical problem.)
One problem that arises for systems with a PDM is the question of how to proceed from a classical description to a quantum one, since an ordering ambiguity for momentum operators becomes apparent. The quantization can be accomplished with several approaches[@young],[@cari],[@refe]. One approach, that has been used in several studies involving quantum mechanical particles with position dependent masses, is one where the classical kinetic energy is factorized into two symmetric parts, each involving a canonical momentum. The classical momentum is then canonically replaced with its quantum counterpart, $\vec{p}\rightarrow-i\hbar\nabla$. A classical kinetic energy of the form $T=\frac{1}{2m_{0}}f^{-1}(\vec{x})\vec{p}\cdot\vec{p}=\frac{1}{2m_{0}}(f^{-1/2}\vec{p})\cdot(f^{-1/2}\vec{p})$ then takes its quantum form[@cari],[@axel1],[@epjp]$$T\rightarrow\hat{T}=-\frac{\hbar^{2}}{2m_{0}}[f^{-1/2}(\vec{x})\nabla
]\cdot\lbrack f^{-1/2}(\vec{x})\nabla]\label{0}$$
where $m_{0}$ is a constant mass related to an effective position dependent mass $m(\vec{x})$ by $m(\vec{x})=f(\vec{x})m_{0}$. Other approaches exist as well, and different quantization prescriptions can result in different Hamiltonians[@refe],[@young].
We can generalize the system in (\[a1\]) by writing a lagrangian in the form (in cartesian coordinates)$$L=\frac{1}{2}m(x)\left[ \delta_{ij}u^{i}u^{j}-U_{0}(\vec{x})\right]
\label{3}$$
where $m(x)U_{0}(\vec{x})$ represents a nonlinear potential energy, in general, and the summation convention is used. It is to be demonstrated here, that this form of lagrangian emerges naturally from an inhomogeneous Kaluza-Klein compactification of an extra spacetime dimension to yield an effective 4D theory. The extra space dimension, when compactified, gives rise to an effective mass $m(x)$ in the effective $4$-dimensional theory. An inhomogeneous compactification is manifest as a spacetime dependent scale factor $b(x^{\mu})$ for the extra dimension, so that there can be regions of spacetime where the size of the extra dimension becomes larger or smaller. This can result in some interesting physical effects and objects, for example, gravitational bags[@bags], dimension bubbles[@bubbles], and scattering from dimensional boundaries[@scatter]. However, interest here is focused upon the emergence and treatment of a low energy quantum mechanical system where a position dependent mass is involved. In the 4D theory the function $b(x^{\mu})$ is related to a dilaton coupling function which determines how strongly the dilaton field couples to matter particles.
We begin with an action describing a classical particle moving in a five dimensional spacetime with pure Einstein gravity and the inclusion of a possible cosmological constant $\Lambda$. We then dimensionally reduce this 5D theory to an effective 4D one, where the extra dimensional scale factor $b(x)$ is related to a scalar field $\phi(x)$ that is nonminimally coupled to the 4D Ricci scalar $\tilde{R}[\tilde{g}_{\mu\nu}]$. A conformal transformation from the 4D Jordan frame with metric $\tilde{g}_{\mu\nu}$ to an Einstein frame with metric $g_{\mu\nu}$ results in a representation where the action contains an ordinary 4D Einstein-Hilbert term, along with a scalar field $\phi$ derived from the scale factor $b$, and a classical matter action. In the Einstein frame representation of the theory, the scale factor $b(x)$ is no longer coupled to the Ricci scalar, but now becomes coupled to the matter sector of the theory[@Dicke62],[@FMbook].
Passing to the flat 4-dimensional spacetime limit with nonrelativistic matter, the 4D action can be recast in the form of a low energy theory for nonrelativistic matter, along with a relativistic scalar field with a potential $V\sim\Lambda b^{-1}$. The equations of motion for the system are obtained, and explicit, exact solutions for the scalar field are obtainable in certain cases. The classical matter particle has an attendant variable mass $m(\vec{x})$ in a flat spacetime.
A quantization of the Hamiltonian then yields a Schrödinger-like equation for the system. We use the quantization method described above, where the kinetic energy operator is given by (\[0\]),$$\hat{H}\psi(\vec{x})=\left[ \hat{T}+\mathcal{U}_{0}(\vec{x})\right]
\psi(\vec{x})=E\psi(\vec{x});\ \ \ \ \ \mathcal{U}_{0}(\vec{x})=\frac{1}{2}m(\vec{x})U_{0}(\vec{x})\label{4}$$
which is obtained from a quantization procedure for a classical nonrelativistic particle characterized by a Lagrangian in a flat space representation of the form given by (\[3\]). An application of this method using a particular solution of the dilaton field equation is seen to produce the system of (\[a1\]) and (\[a2\]), generalized to two or three spatial dimensions, with $x\rightarrow r=\sqrt{x^{2}+y^{2}}$ being a radial coordinate. In addition, there is a nontrivial dilatonic field configuration which couples to the particle oscillator. This problem has been examined previously, but without any connection of the position dependent mass to a dilaton field. In this example, the dilaton field has a structure resembling a global cosmic string. Another difference is that the mass function $m(\vec
{x})$ is not an arbitrary input function, but is determined by the field equation for the dilaton scalar $\phi$.
The model
=========
The 5D action and dimensional reduction
---------------------------------------
We begin by assuming a 5D dimensional spacetime equipped with a metric$$d\tilde{s}_{D}^{2}=\tilde{g}_{MN}(x^{\mu},y)dx^{M}dx^{N}=\tilde{g}_{\mu\nu
}(x^{\alpha})dx^{\mu}dx^{\nu}-b^{2}(x^{\alpha})dy^{2}\label{5}$$
where the 4-dimensional metric $\tilde{g}_{\mu\nu}(x)$ has a negative signature $(+,-,-,-)$. We use $\mu,\nu=0,1,2,3$ for the ordinary spacetime indices and $M,N=0,1,2,3,5$. Absolute values of metric determinants are denoted by $\tilde{g}_{5}=|\det\tilde{g}_{MN}|$, and $\tilde{g}=|\det\tilde
{g}_{\mu\nu}|$, so that $\tilde{g}_{5}=\tilde{g}b^{2}$. The action for the 5D theory that includes gravitation and matter is$$S=\int d^{5}x\sqrt{\tilde{g}_{5}}\left\{ \frac{1}{2\kappa_{5}^{2}}\left[
\tilde{R}_{5}[\tilde{g}_{MN}]-2\Lambda\right] \right\} +S_{m}[\tilde{g}_{MN},\cdot\cdot\cdot]\label{6}$$
where $\Lambda$ is a cosmological constant, and $S_{m}$ is the matter action which depends upon particle velocities and metric $\tilde{g}_{MN}$. For instance, the matter action for a free classical particle is given by$$S_{m}=-\int m_{0}d\tilde{s}_{5}=-\int m_{0}\sqrt{\tilde{g}_{MN}\tilde{u}^{M}\tilde{u}^{N}}d\tilde{s}_{5}\label{7}$$
where $m_{0}$ is a constant mass parameter in the $D$-dimensional theory, and $\tilde{u}^{M}=dx^{M}/d\tilde{s}_{5}$, and we have $\sqrt{\tilde{g}_{MN}\tilde{u}^{M}\tilde{u}^{N}}=1$ when evaluated on the particle worldline where $d\tilde{s}_{5}^{2}=\tilde{g}_{MN}dx^{M}dx^{N}$ holds. We have a 5D gravitational constant denoted by $G_{5}$ with $\kappa_{5}^{2}=8\pi G_{5}$. It is related to the 4D gravitational constant $G$ by $G_{D}=V_{y}G$, and hence $\kappa_{D}^{2}=V_{y}\kappa^{2}$, where $V_{y}$ is the coordinate volume of internal space, $V_{y}=\int dy=2\pi
R_{0}$. We assume that particle trajectories and fields are $y$ independent, i.e., there are no Kaluza-Klein (KK) modes. Specifically, we assume that the classical particle’s trajectory is confined to the 4D spacetime, with $dy=0$ along the particle’s path. Therefore, along the particle’s trajectory we have $d\tilde{s}_{5}^{2}|_{\text{path}}=d\tilde{s}^{2}=\tilde{g}_{\mu\nu}(x)dx^{\mu}dx^{\nu}$ with $\tilde{u}^{5}=0$.
We dimensionally reduce the 5D theory to an effective 4D theory by performing an integration over the internal space with $d^{5}x=d^{4}xdy$ and using $\sqrt{\tilde{g}_{5}}=\sqrt{\tilde{g}}b(x)$. The 5D Ricci scalar $\tilde{R}[\tilde{g}_{MN}]=\tilde{g}^{MN}\tilde{R}_{MN}[\tilde{g}_{MN}]$, with $\tilde{R}_{MN}$ the Ricci tensor, can be broken into a 4D Ricci scalar $R[\tilde{g}_{\mu\nu}]$ plus terms involving the scale factor $b(x)$ and its derivatives. The result is (see, e.g.,[@bubbles],[@CGHW],[@MorrisGRG])$$\begin{array}
[c]{ll}S= & \int d^{4}x\sqrt{\tilde{g}}\{\dfrac{1}{2\kappa^{2}}\Big[b\tilde{R}[\tilde{g}_{\mu\nu}]+2\tilde{g}^{\mu\nu}\tilde{\nabla}_{\mu}\tilde{\nabla
}_{\nu}b-b\Lambda\Big]\}+S_{m}\end{array}
\label{8}$$
where $\tilde{\nabla}_{\mu}$ represents a covariant derivative with respect to the metric $\tilde{g}_{\mu\nu}$. The matter action for a particle of mass $m_{0}$ can be written as$$S_{m}=-\int m_{0}\sqrt{\tilde{g}_{\mu\nu}\tilde{u}^{\mu}\tilde{u}^{\nu}}d\tilde{s}+S_{int} \label{9}$$
where $S_{int}$ represents an action describing the interaction of the particle with any nongravitational sources in its environment.
Dropping a total divergence in the action, the action can be rewritten as$$S=\int d^{4}x\sqrt{\tilde{g}}\ b\left\{ \frac{1}{2\kappa^{2}}\tilde{R}[\tilde{g}_{\mu\nu}]-\frac{\Lambda}{\kappa^{2}}\right\} +S_{m}\label{11}$$
Notice that the action of (\[11\]) contains a scalar field $b$ that is nonminimally coupled to the Ricci scalar as appears in the term $b\tilde
{R}[\tilde{g}_{\mu\nu}]$, so that we have a 4D Jordan frame representation of the theory. We can perform a conformal transformation of the metric in order to recast the theory in the Einstein frame representation, where the scalar field decouples from the curvature scalar, but then becomes coupled to the matter sector[@Dicke62].
Conformal transformation to the Einstein frame
----------------------------------------------
Let us now define the Einstein frame metric $g_{\mu\nu}$ in terms of the Jordan frame metric $\tilde{g}_{\mu\nu}$:$$g_{\mu\nu}=b\tilde{g}_{\mu\nu},\ \ \ g^{\mu\nu}=b^{-1}\tilde{g}^{\mu\nu
},\ \ \ \sqrt{g}=b^{2}\sqrt{\tilde{g}}\label{12}$$
The action in the Einstein frame representation now takes the form[@CGHW],[@scatter]$$S=\int d^{4}x\sqrt{g}\left\{ \frac{1}{2\kappa^{2}}\left[ R[g_{\mu\nu}]+\frac{3}{2}b^{-2}g^{\mu\nu}(\nabla_{\mu}b)(\nabla_{\nu}b)\right]
-b^{-1}\frac{\Lambda}{\kappa^{2}}\right\} +S_{m}[b^{-1}g_{\mu\nu},\cdot
\cdot\cdot]\label{13}$$
Now define the scalar field $\phi(x)$ by$$b=e^{a\phi},\ \ \ \ \ \phi=\frac{1}{a}\ln b,\ \ \ \ \ a=\sqrt{\frac{2}{3}}\ \kappa\label{14}$$
where $a$ is a constant that is determined by requiring the kinetic term in the action to take the canonical form $\frac{1}{2}g^{\mu\nu}\nabla_{\mu}\phi\nabla_{\nu}\phi$. Then the scalar field $\phi$ is explicitly related to the extra dimensional scale factor $b$ by$$\phi(x)=\sqrt{\frac{3}{2}}\frac{1}{\kappa}\ln b(x),\ \ \ \ \ b(x)=\exp\left[
\sqrt{\frac{2}{3}}\ \kappa\phi(x)\right] \label{16}$$
and the action takes the form
\[17\]$$\begin{aligned}
S & =\int d^{4}x\sqrt{g}\left\{ \frac{1}{2\kappa^{2}}R[g_{\mu\nu}]+\frac
{1}{2}g^{\mu\nu}\nabla_{\mu}\phi\nabla_{\nu}\phi-e^{-a\phi}\frac{\Lambda
}{\kappa^{2}}\right\} +S_{m}[e^{-a\phi}g_{\mu\nu},\cdot\cdot\cdot
]\label{17a}\\
& =S_{R}+S_{\phi}+S_{m}\label{17b}$$
with $e^{-a\phi}=b^{-1}$, and we define
$$S_{R}=\int d^{4}x\sqrt{g}\frac{1}{2\kappa^{2}}R[g_{\mu\nu}],\ \ \ \ \ S_{\phi
}=\int d^{4}x\sqrt{g}\left[ \frac{1}{2}\partial^{\mu}\phi\partial_{\mu}\phi-V(\phi)\right] ,\ \ \ \ V(\phi)=\frac{\Lambda}{\kappa^{2}}e^{-a\phi
}\label{18}$$
The scalar field $\phi$ has canonical mass dimension 1, and $\kappa\phi$ is dimensionless.
The Einstein frame action contains an Einstein-Hilbert term for gravity, along with an action for the scalar field $\phi$. In addition, there is a matter action where matter fields have an anomalous coupling to the scalar field. In passing from the Jordan frame to the Einstein frame, the coupling of the scalar to the gravitational sector has been shifted to a coupling of the scalar to the matter sector.
Nonrelativistic matter
======================
Before obtaining an appropriate Schrödinger equation for a nonrelativistic matter particle, we first focus on the nonrelativistic limit of the action for a classical particle. The matter action given by (\[9\]) can be rewritten in the Einstein frame. The free particle portion of the action $S_{m}=S_{\text{free}}+S_{int}$ is given by$$S_{\text{free}}=-\int m_{0}d\tilde{s}=-\int m_{0}b^{-1/2}ds=-\int mds=-\int
m_{0}b^{-1/2}\sqrt{g_{\mu\nu}u^{\mu}u^{\nu}}ds\label{19}$$
This shows that the particle mass in the Einstein frame representation is given by[@Dicke62]$$m(x)=m_{0}b^{-1/2}\label{20}$$
Furthermore, in the nonrelativistic limit $ds=dt$, $u^{i}\ll u^{0}=1$ and we identify the nonrelativistic portion of the classical Lagrangian as$$L_{\text{free}}=-m_{0}b^{-1/2}(g_{\mu\nu}u^{\mu}u^{\nu})^{1/2}=-m_{0}b^{-1/2}(u_{0}u^{0}+g_{ij}u^{i}u^{j})^{1/2}=m_{0}b^{-1/2}\left[ \frac{1}{2}(\vec{u}\cdot\vec{u})-1\right] \label{21}$$
where use has been made of $-g_{ij}u^{i}u^{j}=\vec{u}\cdot\vec{u}=|\vec
{u}|^{2}\ll1$.
Next, we write the nonrelativistic limit of the interaction term in $S_{m}$ as$$S_{int}=\int L_{int}d\tilde{s}=-\int U(\vec{x})d\tilde{s}=-\int U(\vec
{x})b^{-1/2}ds\approx-\int b^{-1/2}U(\vec{x})dt\label{22}$$
and $L_{int}=b^{-1/2}U(\vec{x})$, where $U(\vec{x})$ is an arbitrary potential energy function defined in the Jordan frame, describing interactions of the particle with nongravitational forces. The nonrelativistic limit of the matter action is then$$S_{m}=\int Ldt=\int dtb^{-1/2}\left\{ \left[ \frac{1}{2}m_{0}\vec{u}\cdot\vec{u}\right] -\left[ U(\vec{x})+m_{0}\right] \right\} \label{23}$$
The potential $U(\vec{x})$ can be redefined to absorb the constant $m_{0}$, i.e., $U(\vec{x})+m_{0}\rightarrow U(\vec{x})$, and we can therefore write the Lagrangian for the nonrelativistic particle in the Einstein frame representation as$$L=b^{-1/2}\left[ \frac{1}{2}m_{0}\vec{u}\cdot\vec{u}-U(\vec{x})\right]
\equiv b^{-1/2}L_{0}\label{24}$$
where $L_{0}\equiv L|_{b=1}$. If we define $U_{0}(\vec{x})=2U(\vec{x})/m_{0}$, then (\[24\]) takes the form, in cartesian coordinates,$$L=\frac{1}{2}m(x)\left[ \delta_{ij}u^{i}u^{j}-U_{0}(\vec{x})\right]
\label{25}$$
which coincides with that in (\[3\]), in the limit of a flat spacetime where $g_{00}=1$, $g_{ij}=-\delta_{ij}$ (for cartesian coordinates).
The flat space nonrelativistic theory
=====================================
We want to focus on a laboratory type of setting for a nonrelativistic particle, i.e., a setting where the explicit curvature of spacetime can be ignored. In other words, we consider a small enough region of space where it becomes reasonable to approximate the spacetime as Minkowskian, with $g_{\mu\nu}=\eta_{\mu\nu}$, where $\eta_{\mu\nu}$ is the flat space Minkowski metric, appropriately expressed in terms of the coordinate system used. (For cartesian coordinates we have $\eta_{00}=\eta^{00}=1$, $\eta_{ij}=\eta
^{ij}=-\delta_{ij}=-\delta^{ij}$.) In this approximation the curvature action $S_{R}$ given by (\[18\]) is set to zero, and the action $S$ of (\[17\]) for the system reduces to
\[26\]$$\begin{aligned}
S & =S_{\phi}+S_{m}=\int d^{4}x\sqrt{\eta}\left[ \frac{1}{2}\eta^{\mu\nu
}\partial_{\mu}\phi\partial_{\nu}\phi-b^{-1}\frac{\Lambda}{\kappa^{2}}\right]
+S_{m}\label{26a}\\
& =\int d^{4}x\sqrt{\eta}\left[ \frac{1}{2}\eta^{\mu\nu}\partial_{\mu}\phi\partial_{\nu}\phi-e^{-a\phi}\frac{\Lambda}{\kappa^{2}}\right] +\int
dt\ e^{-a\phi/2}\left[ \frac{1}{2}m_{0}(-\eta_{ij})u^{i}u^{j}-U(\vec
{x})\right] \label{26b}$$
where (\[24\]) has been used for the matter Lagrangian, and the metric $\eta_{\mu\nu}$ has been left explicitly in the expressions, with a reminder that the metric is one for flat spacetime, with an arbitrary choice for cartesian or curvilinear coordinates. (We have $\eta_{00}=1$, and $\eta_{ij}$ is just the negative of the metric for a Euclidean space, with $-\eta_{ij}>0$ for $i\neq j$, and $\eta=|\det\eta_{\mu\nu}|$.) The parameterization $b=e^{a\phi}$, as given by (\[14\]), has also been used. A lagrangian density $\mathcal{L}_{m}$ for the matter particle can be defined by writing $\mathcal{L}_{m}=\sqrt{\eta}L\delta^{(3)}(\vec{x}-\vec{x}_{p})$ where $\vec
{x}_{p}$ locates the instantaneous position of the particle. Then (\[26\]) can be expressed as
$$\begin{aligned}
S & =\int d^{4}x\sqrt{\eta}\left[ \frac{1}{2}\eta^{\mu\nu}\partial_{\mu}\phi\partial_{\nu}\phi-e^{-a\phi}\frac{\Lambda}{\kappa^{2}}\right]
\nonumber\\
& +\int d^{4}x\sqrt{\eta}e^{-a\phi/2}\left[ \frac{1}{2}m_{0}(-\eta
_{ij})u^{i}u^{j}-U(\vec{x})\right] \delta^{(3)}(\vec{x}-\vec{x}_{p})\label{27}$$
From this action the equation of motion for the field $\phi$ can be obtained:$$\square\phi+\frac{\partial V(\phi)}{\partial\phi}-\sigma=0;\ \ \ \ \ V(\phi
)=e^{-a\phi}\frac{\Lambda}{\kappa^{2}},\ \ \ \ \ \sigma=\frac{\partial
\mathcal{L}_{m}}{\partial\phi}=\left( -\frac{a}{2}\right) \sqrt{\eta}L\delta^{(3)}(\vec{x}-\vec{x}_{p})\label{28}$$
where $\square=\nabla_{\mu}\partial^{\mu}=\partial_{t}^{2}-\nabla^{2}$. (The field $\phi$ is referred to as a 4D dilaton field.) Since $\mathcal{L}_{m}\propto\delta
^{(3)}(\vec{x}-\vec{x}_{p})$, we take the $\sigma$ term to vanish at all $\vec{x}\neq\vec{x}_{p}$. So, at least as a first approximation, we set $\sigma=0$ in the equation of motion in order to obtain a background solution for $\phi(x)$ that can couple to the matter particle in $L$. The equation of motion for $\phi$ therefore reads$$\square\phi-a\frac{\Lambda}{\kappa^{2}}e^{-a\phi}=0\label{29}$$
Here we focus attention on static solutions of (\[29\]), that satisfy$$\nabla^{2}\phi=-a\frac{\Lambda}{\kappa^{2}}e^{-a\phi}\label{30}$$
Solutions to this equation can be found for certain specific cases.
As an example, we consider $\phi(x,y)$ to be a function of the two cartesian coordinates $x$ and $y$, and (\[30\]) reads as$$(\partial_{x}^{2}+\partial_{y}^{2})\phi=-a\frac{\Lambda}{\kappa^{2}}e^{-a\phi
}=2\tilde{\kappa}\frac{\Lambda}{\kappa^{2}}e^{2\tilde{\kappa}\phi}\label{31}$$
where $\tilde{\kappa}\equiv-\frac{1}{2}a$. Then (\[31\]) is recognized as the 2D Liouville differential equation[@Liouville],[@Crowdy] (also see, for example,[@Gibbons] and [@DHoker]). Upon defining $\zeta
=x+iy$, the solutions are given by[@Gibbons],[@Crowdy]
$$\begin{aligned}
b^{-1} & =e^{2\tilde{\kappa}\phi}=\frac{2\kappa^{2}}{\tilde{\kappa}^{2}|\Lambda|}\frac{|f^{\prime}(\zeta)|^{2}}{\left[ |f(\zeta)|^{2}+1\right]
^{2}},\ \ \ (\Lambda=-|\Lambda|<0)\label{32a}\\
b^{-1} & =e^{2\tilde{\kappa}\phi}=\frac{2\kappa^{2}}{\tilde{\kappa}^{2}|\Lambda|}\frac{|f^{\prime}(\zeta)|^{2}}{\left[ |f(\zeta)|^{2}-1\right]
^{2}},\ \ \ (\Lambda=+|\Lambda|>0)\label{32b}$$
where $f(\zeta)$ is a holomorphic function of $\zeta$, and $f^{\prime}(\zeta)=df/d\zeta$. The matter coupling function $b^{-\frac{1}{2}}=e^{\tilde{\kappa}\phi}$ that appears in the matter particle lagrangian $L$ is
$$b^{-\frac{1}{2}}=e^{\tilde{\kappa}\phi}=\frac{\sqrt{2}\kappa}{|\tilde{\kappa
}|\sqrt{|\Lambda|}}\frac{|f^{\prime}(\zeta)|}{\left[ |f(\zeta)|^{2}\pm1\right] },\ \ (\Lambda=\mp|\Lambda|)\label{33}$$
Different choices for the function $f(\zeta)$ give different mathematical solutions to Liouville’s equation. (See, e.g.,[@Gibbons],[@MorrisPLB] for some applications to dilaton gravity and low energy string theory.)
Quantization
============
The classical lagrangian for a particle in the flat space nonrelativistic limit is given by (\[24\]), for instance. In terms of cartesian coordinates, we can write the lagrangian as$$L=b^{-1/2}\left[ \frac{1}{2}m_{0}\delta_{ij}u^{i}u^{j}-U(\vec{x})\right]
=T-\mathcal{U};\ \ \ \ \ \mathcal{U}=b^{-1/2}U(\vec{x})\label{34}$$
(For cartesian coordinates the spatial metric, with positive signature is $-g_{ij}=\delta_{ij}$, but one could use other coordinates with an appropriate metric $\delta_{ij}\rightarrow\gamma_{ij}$.)
As noted in (\[20\]), the effective mass of the particle in the Einstein frame representation is$$m(\vec{x})=m_{0}b^{-1/2}=m_{0}e^{-a\phi/2}\label{35}$$
where we assume that $b=b(\vec{x})$ is time independent. The classical kinetic energy is$$T=\frac{1}{2}m_{0}b^{-1/2}\delta_{ij}u^{i}u^{j}=\frac{1}{2}m(\vec{x})\delta_{ij}u^{i}u^{j}=\frac{1}{2}m(\vec{x})\vec{u}\cdot\vec{u}\label{36}$$
The canonically conjugate momentum is
\[37\]$$\begin{aligned}
p^{i} & =p_{i}=\frac{\partial L}{\partial u^{i}}=m_{0}b^{-1/2}u^{i}=m(\vec
{x})u^{i};\label{37a}\\
u^{i} & =\frac{p^{i}}{m(\vec{x})}=b^{1/2}\frac{p^{i}}{m_{0}}\label{37b}$$
Therefore, in terms of the canonical momentum, the kinetic energy is
$$T=b^{1/2}\delta_{ij}\frac{p^{i}p^{j}}{2m_{0}}=b^{1/2}\frac{\vec{p}\cdot\vec
{p}}{2m_{0}}\label{38}$$
We can now implement the quantization procedure described in the Introduction, and used in previous works[@cari],[@refe],[@axel1],[@epjp]. We write $T$ in a symmetrical form$$T=\frac{1}{2m_{0}}(b^{1/4}\vec{p})\cdot(b^{1/4}\vec{p})\label{39}$$
and then make the canonical quantum replacement $\vec{p}\rightarrow
-i\hbar\nabla$ in order to obtain the kinetic part of the quantum Hamiltonian $\hat{H}$:$$T\rightarrow\hat{T}=-\frac{\hbar^{2}}{2m_{0}}(b^{1/4}\nabla)\cdot
(b^{1/4}\nabla)=-\frac{\hbar^{2}}{2m_{0}}\left[ b^{1/2}\nabla^{2}+b^{1/4}(\nabla b^{1/4})\cdot\nabla\right] \label{40}$$
The classical Hamiltonian$$H=T+b^{-1/2}U(\vec{x})=T+\mathcal{U}(\vec{x})\label{41}$$
is then replaced by a quantum Hamiltonian$$\hat{H}=\hat{T}+\mathcal{U}(\vec{x})=\hat{T}+b^{-1/2}U(\vec{x})\label{42}$$
where $\hat{T}$ is given by (\[40\]).
Application: A dilaton string background
========================================
The nonrelativistic quantum particle
------------------------------------
As an example we consider a particle subjected to harmonic oscillation in the $x$-$y$ plane with translation invariance in the $z$ direction, or a particle subject to oscillation in a two dimensional space of the $x$-$y$ plane. We employ the quantization procedure described in the previous section. Consider now a solution of the 2D Liouville equation for which $f(\zeta
)=A\zeta$, where $\zeta=x+iy$, and hence $f^{\prime}(\zeta)=A$, $|f(\zeta
)|^{2}=A^{2}(x^{2}+y^{2})=A^{2}r^{2}$. Then (for $\Lambda=-|\Lambda|$) (\[33\]) gives a solution (see, for example, [@Gibbons] and [@MorrisPLB])$$b^{-1/2}=\frac{C}{(1+A^{2}r^{2})},\ \ \ \ C=\frac{\sqrt{2}\kappa A}{|\tilde{\kappa}|\sqrt{|\Lambda|}};\ \ \ \ \ b^{1/2}=\frac{(1+A^{2}r^{2})}{C};\ \ \ \ \ b^{1/4}=\frac{(1+A^{2}r^{2})^{1/2}}{\sqrt{C}}\label{43}$$
Using$$(\partial_{r}b^{1/4})=\frac{(1+A^{2}r^{2})^{-1/2}}{\sqrt{C}}\cdot
A^{2}r\label{44}$$
(\[40\]) yields$$\hat{T}=-\frac{\hbar^{2}}{2m_{0}}\left[ \frac{(1+A^{2}r^{2})}{C}\nabla
^{2}+\frac{A^{2}r}{C}\partial_{r}\right] \label{45}$$
Interestingly, this is the same form of the kinetic Hamiltonian that has been studied by various authors for a nonlinear quantum oscillator with position dependent mass[@cari],[@refe],[@axel1],[@epjp]. In fact, if we choose a potential $U(r)=\frac{1}{2}Kr^{2}$, we have a potential term in the Hamiltonian $\hat{H}$ given by $$\mathcal{U}(r)=b^{-1/2}U(r)=\frac{CK}{2}\frac{r^{2}}{(1+A^{2}r^{2})}\label{47}$$
Upon choosing $C=1$, $A^{2}=\lambda>0$, and $K=m_{0}\alpha^{2}$, the Hamiltonian is [@note]$$\hat{H}=-\frac{\hbar^{2}}{2m_{0}}\left[ (1+\lambda r^{2})\nabla^{2}+\lambda
r\partial_{r}\right] +\frac{1}{2}\frac{m_{0}\alpha^{2}r^{2}}{(1+\lambda
r^{2})} \label{48}$$
This system and related systems have been studied extensively for the cases of 1, 2, 3, and $n$ dimensions of space[@cari],[@refa],[@refe],[@axel1],[@jmp],[@epjp]. See, e.g., [@cari] and [@axel1]-[@epjp] for exact expressions for the normalizable, orthogonal eigenfunctions, along with the spectrum of energy eigenvalues for a class of Hamiltonians of this type. In those works, however, the mass function $m(\vec{x})$ is an assumed input function, whereas here it emerges as a solution of the 2D Liouville equation associated with the dilaton field $\phi(r)$.
Previous studies of this system have focused largely on its mathematical properties, with some mention that such systems may achieve physical realizations in condensed matter settings, as pointed out in[@cari], for instance. Here, however, we see this type of system being realized in the context of an inhomogeneous compactification of extra space dimensions, resulting in a quantum particle in ordinary space having an effective position dependent mass, whose variation depends upon a variation of the dilaton field, or equivalently, a variation in the size of the extra dimensions. The effective mass is$$m(r)=m_{0}b^{-1/2}(r)=\frac{m_{0}C}{(1+A^{2}r^{2})}\label{49}$$
in an effective potential$$\mathcal{U}(r)=b^{-1/2}(r)U(r)\propto\frac{r^{2}}{(1+A^{2}r^{2})}\label{50}$$
There is a discrete spectrum of bound states[@epjp], where we could consider nonrelativistic particles as becoming trapped within the dilatonic string-like core described by the dilaton coupling function $b^{-1/2}$, for which the size of the extra dimensions is$$b(r)\sim(1+A^{2}r^{2})^{2}\label{51}$$
So, the coupling function $b^{-1/2}=e^{\tilde{\kappa}\phi}$ has a maximum in the string core where $b\sim1$, and outside the core the coupling function $b^{-1/2}(r)$ decreases, and the extra dimensional scale factor $b(r)$ increases (i.e., the extra dimensions get bigger). Bound state particles are then localized within the dilatonic string core, where the extra dimensions are small, and the effective particle mass is large.
The dilaton string
------------------
We choose a negative cosmological constant, $\Lambda=-|\Lambda|$, so that from (\[28\]) the dilaton string has a potential given by$$V(\phi)=-\frac{|\Lambda|}{\kappa^{2}}e^{-a\phi}=-\frac{|\Lambda|}{\kappa^{2}}e^{2\tilde{\kappa}\phi}=-\frac{|\Lambda|}{\kappa^{2}}b^{-1}\label{52}$$
The energy density of the scalar field $\phi$ is$$\mathcal{H}=\mathcal{H}_{\text{kin}}+\mathcal{H}_{\text{pot}}=\frac{1}{2}(\partial_{r}\phi)^{2}+V(\phi) \label{53}$$
with $r=\sqrt{x^{2}+y^{2}}$ being the radial distance from the center of the vortex (2D space dimensions) or the center of the string (which is centered on the $z$-axis in a 3D space). We now adopt the settings $C=1$ and $A^{2}=\lambda>0$. We then have$$b=e^{a\phi}=(\lambda r^{2}+1)^{2},\ \ \phi=-\frac{1}{\tilde{\kappa}}\ln(\lambda r^{2}+1)=-\frac{1}{\tilde{\kappa}}\ln(\xi+1),\ \ \ \ \xi
\equiv\lambda r^{2}\label{54}$$
where $\xi=\lambda r^{2}$ is a dimensionless distance parameter. The energy per unit length of the string i.e., the string tension, is (assuming, for simplicity, $z$ independence) $$\mu=\mu_{\text{kin}}+\mu_{\text{pot}}=2\pi\int_{0}^{r_{C}}(\mathcal{H}_{\text{kin}}+\mathcal{H}_{\text{pot}})\ rdr=2\pi\int_{0}^{r_{C}}\mathcal{H}(r)\ rdr=\frac{\pi}{\lambda}\int_{0}^{\xi_{C}}\mathcal{H}(\xi
)d\xi\label{55}$$
where $r_{C}$ and $\xi_{C}$ are large distance cutoffs, as the tension diverges logarithmically, as with a global cosmic string[@Vilenkin],[@VSbook]. (This dilaton string resembles the string studied in [@Gibbons] and [@MorrisPLB], with the cosmological constant replacing a constant magnetic $H$ field.) Using (\[52\])-(\[55\]), some calculation yields$$\frac{\mu}{2\pi}=\frac{1}{\tilde{\kappa}^{2}}\left[ \ln(\xi_{C}+1)+\frac
{1}{(\xi_{C}+1)}-1\right] -\frac{|\Lambda|}{2\lambda\kappa^{2}}\left[
1-\frac{1}{(\xi_{C}+1)}\right] \label{56}$$
For $\xi_{C}\gg1$ ($r_{C}\gg1/\sqrt{\lambda}$) this simplifies to$$\frac{\mu}{2\pi}\approx\frac{1}{\tilde{\kappa}^{2}}\left( \ln\xi
_{C}-1\right) -\frac{|\Lambda|}{2\lambda\kappa^{2}} \label{57}$$
We note that since the potential is negative, but the kinetic term is positive, one can find$$\Big|\frac{\mathcal{H}_{\text{kin}}}{\mathcal{H}_{\text{pot}}}\Big|=\frac
{\frac{1}{2}(\partial_{r}\phi)^{2}}{|V(\phi)|}=\frac{2\lambda^{2}\kappa^{2}}{|\Lambda|\tilde{\kappa}^{2}}r^{2} \label{58}$$
so that the energy density becomes negative within a certain critical radius:$$\mathcal{H}\leq0\ \ \ \ \text{for\ \ \ \ }r\leq r_{crit}=\frac{|\Lambda
|^{1/2}|\tilde{\kappa}|}{\sqrt{2}\lambda\kappa},\ \ \ \ \ \ \xi_{crit}=\frac{|\Lambda|\tilde{\kappa}^{2}}{2\lambda\kappa^{2}} \label{59}$$
so that inside this portion of the string core the weak energy condition is violated. However, the string tension is positive, $\mu>0$, provided that $\xi_{C}\gg1$, and the condition$$\ln(\xi_{C}+1)>1+\xi_{crit} \label{60}$$
is satisfied.
Summary
=======
Beginning with a classical particle propagating in a four dimensional subspace of a five dimensional spacetime, where the internal space dimension gets inhomogeneously, toroidally, compactified, we arrive at an effective 4 dimensional theory for a classical particle. When a conformal transformation is made from the resulting 4D Jordan frame to a 4D Einstein frame, the particle acquires a position dependent mass. The classical nonrelativistic system in a flat Einstein frame spacetime can be quantized using existing quantization procedures. The result is the emergence of a quite general class of quantum systems that have position dependent mass and somewhat arbitrary potentials. Systems of this type, e.g., the nonlinear oscillator with a PDM, emerging from different physical circumstances, have been studied already. Here, we show another way in which this quantum system and generalizations of it can arise. It is also seen that within the type of scenario considered here there is a somewhat different physical interpretation involving a particle interacting with a dilaton field, which is associated with the size of the extra dimension.
The hope is that a possible physical realization of such systems, in the context of higher dimensional physics, will provide motivation for their further study, along with new physical interpretations of results materializing from these systems. We have presented an application of this formalism to obtain a previously studied system describing a nonlinear quantum oscillator with a position dependent mass. In the scenario presented here, however, the effective position dependent particle mass depends explicitly upon the scalar dilaton field (i.e., the size of the extra dimension), which itself may have nontrivial, interesting structure.
[99]{}
P.M. Mathews and M. Lakshmanan, $"$On a unique nonlinear oscillator$"$, Quart. Appl. Math. 32 (1974), 215-218
J.F. Cariñena, M.F. Rañada and M. Santander, $"$A quantum exactly-solvable nonlinear oscillator with quasi-harmonic behaviour$"$, Ann. Phys. 322 (2007), 434-459
J.F. Cariñena, M.F. Rañada, M. Santander and M. Senthilvelan, $"$A non-linear oscillator with quasi-harmonic behaviour: two- and n-dimensional oscillators$"$, Nonlinearity 17 (2004), 1941-1963
A. Ballesteros, A. Enciso, F.J. Herranz, O. Ragnisco and D. Riglioni, $"$Quantum mechanics on spaces of nonconstant curvature: The oscillator problem and superintegrability$"$, Ann. Phys. 326 (2011), 2053-2073
A. Schulze-Halberg and J.R. Morris, Special function solutions of a spectral problem for a nonlinear quantum oscillator, J. Phys. A 45 (2012), 305301 (9pp)
A. Schulze-Halberg and J.R. Morris, An exactly solvable three-dimensional nonlinear quantum oscillator, J. Math. Phys. 54, 112107 (2013)
A. Schulze-Halberg and J.R. Morris, Higher-dimensional realization of a nonlinear, one-parameter quantum oscillator, Eur. Phys. J. Plus 128, (2013)
J.F. Cariñena, M.F. Rañada and M. Santander, $"$The quantum harmonic oscillator on the sphere and the hyperbolic plane$"$, Ann. Phys. 322 (2007), 2249-2278
J.F. Cariñena, M.F. Rañada and M. Santander, $"$The quantum free particle on spherical and hyperbolic spaces: A curvature dependent approach. II$"$, J. Math. Phys. 53 (2012), 102109 (19pp)
K. Young, Position-dependent effective mass for inhomogeneous semiconductors, Phys. Rev. B39, 434-441 (1989)
A. Davidson and E.I. Guendelman, Gravitational bags, Phys. Rev. D34 (1986) 3264-3267
J.R. Morris, Domain bubbles of extra dimensions, Phys. Rev. D67 (2003) 025005 {e-Print: hep-th/0211175}
N. De Leon and J.R. Morris, Reflection and transmission at dimensional boundaries, Phys. Rev. D74 (2006) 045033 {e-Print: hep-th/0608113}
S. Carroll, J. Geddes, M.B. Hoffman, and R.M. Wald, Classical stabilization of homogeneous extra dimensions, Phys. Rev. D66 (2002) 024036 {e-Print: hep-th/0110149}
J.R. Morris, Dilatonic effects near naked singularities, Gen. Rel. Grav. 44 (2012) 437-448 {e-Print: arXiv:1111.0912 \[gr-qc\]}
R.H. Dicke, Mach’s principle and invariance under transformation of units, Phys. Rev. 125 (1962) 2163-2167
Y. Fugii and K. Maeda, *The Scalar-Tensor Theory of Gravitation* (Cambridge University Press, 2003)
J. Liouville, J. Math 18, (1853) 71
D.G. Crowdy, General solutions to the 2D Liouville equation, Int. J. Engng Sci. 35 (1997) 141-149
G.W. Gibbons and C.G. Wells, Flux confinement in dilatonic cosmic strings, Class. Quant. Grav. 11 (1994) 2499-2506 {e-Print: hep-th/9312014}
E.D. D’Hoker and R. Jackiw, Liouville field theory, Phys. Rev. D26 (1982) 3517
J.R. Morris, Generalized dilaton-Maxwell cosmic string and wall solutions, Phys. Lett. B641 (2006) 1-5 {e-Print: hep-th/0608101}
Our choice $C=1$ fixes the ratio $\lambda/|\Lambda|$, but we do not concern ourselves with this. This restriction could be obviated with a redefined constant $K=Cm_{0}\alpha^{2}$ along with a redefinition of the mass parameter, $Cm_{0}\rightarrow m_{0}$.
A. Vilenkin, Phys. Rep. 121, 263 (1985)
A. Vilenkin and E.P.S. Shellard, *Cosmic Strings and Other Topological Defects*, (Cambridge University Press, Cambridge, 1994)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Irtiza Hasan
- 'Shengcai Liao[^1]'
- Jinpeng Li
- Saad Ullah Akram
- Ling Shao
bibliography:
- 'main.bib'
title: 'Pedestrian Detection: The Elephant In The Room'
---
[^1]: Corresponding author
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In any connected non-compact semi-simple Lie group without factors locally isomorphic to $\operatorname{SL}_2({\mathbb R})$, there can be only finitely many lattices (up to isomorphism) of a given covolume. We show that there exist arbitrarily large families of pairwise non-isomorphic arithmetic lattices of the same covolume. We construct these lattices with the help of Bruhat-Tits theory, using Prasad’s volume formula to control their covolumes.'
address: |
Section de mathématiques\
Université de Genève\
2–4 rue du Lièvre, Case postale 64\
CH-1211 Genève 4\
SWITZERLAND
author:
- Vincent Emery
bibliography:
- '/home/vincent/Math/mes-textes/emery-bib.bib'
title: Arbitrarily large families of spaces of the same volume
---
[^1]
Introduction {#sec:intro}
============
Let ${\mathcal G}$ be a connected semi-simple real Lie group without compact factors. For simplicity we will suppose that ${\mathcal G}$ is adjoint (i.e., with trivial center), though this is not a major restriction in this article. Any choice of a Haar measure $\mu$ on ${\mathcal G}$ assigns a *covolume* $\mu(\Gamma {\backslash}{\mathcal G}) \in {\mathbb R}_{> 0}$ to each lattice $\Gamma$ in ${\mathcal G}$. Wang’s theorem [@Wang72] asserts that there exist only finitely many irreducible lattices (up to conjugation) of bounded covolumes in ${\mathcal G}$ unless ${\mathcal G}$ is isomorphic to $\operatorname{PSL}_2({\mathbb R})$ or $\operatorname{PSL}_2({\mathbb C})$. In particular, there exist only finitely many irreducible lattices in ${\mathcal G}$ of a given covolume. For ${\mathcal G}$ isomorphic to $\operatorname{PSL}_2({\mathbb C})$ this property is still true, as follows from the work of Thurston and Jørgensen [@Thur80 Ch. 6]. In this paper we prove that the number of lattices in ${\mathcal G}$ of the same covolume can be arbitrarily large. In most cases, arbitrarily large families of lattices of equal covolume appear in the commensurability class of any arithmetic lattice of ${\mathcal G}$. This is the content of the following theorem. The symbol ${\mathfrak{g}_{\mathbb C}}$ denotes the complexification of the Lie algebra of ${\mathcal G}$.
\[thm:general\]
Let ${\mathcal G}$ be a connected adjoint semi-simple real Lie group without compact factors. We suppose that ${\mathfrak{g}_{\mathbb C}}$ has a simple factor that is not of type ${\mathrm{A}}_1$, ${\mathrm{A}}_2$ or ${\mathrm{A}}_3$. Let $\Gamma$ be an arithmetic lattice in ${\mathcal G}$. Then, for every $m \in {\mathbb N}$, there exist a family of $m$ lattices commensurable to $\Gamma$ that are pairwise non-isomorphic and have the same covolume in ${\mathcal G}$. These lattices can be chosen torsion-free.
Every arithmetic lattice $\Gamma \subset {\mathcal G}$ is constructed with the help of some algebraic group ${\mathrm G}$ defined over a number field $k$ (see Section \[ss:arithmetic-subgroup\]). To prove Theorem \[thm:general\], we use Bruhat-Tits theory to construct families of arithmetic subgroups in ${\mathrm G}(k)$ that are non-conjugate, and have equal covolume. By strong (Mostow) rigidity one obtains the analogous result with “pairwise non-conjugate” replaced with “pairwise non-isomorphic”. To control the covolume we use some computations that appear in Prasad’s volume formula [@Pra89]. To ensure that the subgroups constructed are not conjugate we need to exhibit parahoric subgroups in ${\mathrm G}(k_v)$ (where $k_v$ is a non-archimedean completion of $k$) that are not conjugate but of the same volume. This can be easily achieved when ${\mathrm G}$ is not of type ${\mathrm{A}}_n$ and is split over $k_v$. When ${\mathrm G}$ is of type ${\mathrm{A}}_n$ the Bruhat-Tits building of a split ${\mathrm G}(k_v)$ has more symmetries, and the argument must be slightly adapted. In particular, there we need the assumption $n \ge 4$, which explains the excluded cases in the statement of Theorem \[thm:general\]. The simple Lie groups excluded are listed in Table \[tab:exceptions\].
-------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------
Type ${\mathrm{A}}_1$: $\operatorname{PSL}_2({\mathbb R})$ and $\operatorname{PSL}_2({\mathbb C})$;
\[3pt\] Type ${\mathrm{A}}_2$: $\operatorname{PSL}_3({\mathbb R})$, $\operatorname{PSL}_3({\mathbb C})$ and $\operatorname{PU}(2,1)$;
\[3pt\] Type ${\mathrm{A}}_3$: $\operatorname{PSL}_4({\mathbb R})$, $\operatorname{PSL}_4({\mathbb C})$, $\operatorname{PSO}(5,1)$, $\operatorname{PU}(3,1)$ and $\operatorname{PU}(2,2)$.
-------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------
: Simple Lie groups not covered in Theorem \[thm:general\][]{data-label="tab:exceptions"}
\[eq:exceptions\]
For the Lie groups of type ${\mathrm{A}}_2$ and ${\mathrm{A}}_3$ we can use algebraic groups that are outer forms (type $^2{\mathrm{A}}_2$ and $^2{\mathrm{A}}_3$) to show the existence of arbitrarily large families of arithmetic lattices of the same covolume. In contrast with Theorem \[thm:general\], now each family corresponds to a different commensurability class.
\[thm:exceptions\]
Let ${\mathcal G}$ be a connected adjoint semi-simple Lie group without compact factors. We suppose that ${\mathfrak{g}_{\mathbb C}}$ contains only factors of type ${\mathrm{A}}_2$ (resp. only factors of type ${\mathrm{A}}_3$). Let $m \in {\mathbb N}$. Then there exists a family $\left\{ \Gamma_1, \dots, \Gamma_m \right\}$ of irreducible arithmetic lattices in ${\mathcal G}$ such that for $i,j \in
\left\{ 1,\dots m \right\}$:
1. $\Gamma_i$ is commensurable to $\Gamma_j$;
2. $\Gamma_i$ and $\Gamma_j$ have the same covolume in ${\mathcal G}$;
3. if $i \neq j$, then $\Gamma_i$ and $\Gamma_j$ are not isomorphic.
The lattices $\left\{ \Gamma_i
\right\}$ can be chosen torsion-free. Moreover, they can be chosen cocompact. They can be chosen non-cocompact unless there are no such lattices in ${\mathcal G}$.
It follows from Margulis’ arithmeticity theorem that irreducible lattices can only exist in a Lie group ${\mathcal G}$ that is isotypic (i.e., for which all the simple factors of ${\mathfrak{g}_{\mathbb C}}$ have the same type), so that the assumptions in Theorem \[thm:exceptions\] are minimal. The existence of irreducible cocompact lattices in any isotypic ${\mathcal G}$ was proved by Borel and Harder [@BorHar78]. Non-compact irreducible quotients of ${\mathcal G}$ do not always exist. For example there is no such quotient of $\operatorname{PU}(3,1) \times \operatorname{PSO}(5,1)$ (this example is detailed in [@WittMorr08 Prop. (15.31)]). A general criterion for the existence of non-cocompact arithmetic lattices appears in the work of Prasad-Rapinchuk [@PraRap06], where the authors extend the results of [@BorHar78]. The proof of Theorem \[thm:exceptions\] uses these existence results.
By Wang’s theorem, it is clear that the covolume common to the lattices of a family grows with the size of the family. Even though in this article we focus on qualitative results, we note that the proofs of Theorems \[thm:general\]–\[thm:exceptions\] could be used to obtain quantitative results on the growth of the covolume with the size of the family.
We now discuss the geometric significance of our results. Let ${X}$ be the symmetric space associated with ${\mathcal G}$, that is ${X}=
{\mathcal G}/K$ for a maximal compact subgroup $K \subset {\mathcal G}$. This class of spaces includes the *hyperbolic $n$-space* ${\mathcal H}^n$; we have that ${\mathcal H}^2$ is associated with ${\mathcal G}= \operatorname{PSL}_2({\mathbb R})$, and ${\mathcal H}^3$ with ${\mathcal G}=
\operatorname{PSL}_2({\mathbb C})$. For a torsion-free irreducible lattice $\Gamma \subset
{\mathcal G}$, the locally symmetric space $\Gamma {\backslash}{X}$ will be called *an ${X}$-manifold* (in particular it is irreducible and has finite volume). The following result follows directly from Theorems \[thm:general\]–\[thm:exceptions\] and the existence of cocompact arithmetic lattices in ${\mathcal G}$ (see for instance [@PraRap06 Theorem 1]).
\[cor:same-volume\]
Let $X$ be a Riemannian symmetric space of non-compact type that contains no factor isometric to ${\mathcal H}^2$ or ${\mathcal H}^3$, and suppose that irreducible quotients of $X$ do exist. Then there exist arbitrarily large families of pairwise non-isometric commensurable compact $X$-manifolds having the same volume. The analogue statement with non-compact $X$-manifolds is true unless all $X$-manifolds are compact.
The result for $X = {\mathcal H}^3$ was proved by Wielenberg for the case of non-compact manifolds [@Wiel81], and later by Apasanov-Gutsul for compact manifolds [@ApaGut92]. For $X = {\mathcal H}^4$ the result with non-compact manifolds was proved by Ivanšić in his thesis [@Ivan99]. All these results are obtained by geometric methods. In [@Zimm94] Zimmerman gave a new proof for $X = {\mathcal H}^3$ by exhibiting examples of ${\mathcal H}^3$-manifolds $M$ with first Betti number $\beta_1$ at least $2$, and showing that this property implies the existence of arbitrarily large families of covering spaces of $M$ of same degree. In [@Lubo96] Lubotzky showed that there exist (many) hyperbolic manifolds with $\beta_1 \ge 2$ in every dimension. Thus for all $X = {\mathcal H}^n$ we have a proof of Corollary \[cor:same-volume\] by Zimmerman’s method. Since super-rigidity implies that $H^1(\Gamma {\backslash}X,{\mathbb R}) = 0$ for irreducible lattices $\Gamma$ in ${\mathcal G}$ with ${\mathbb R}\mathrm{-rank}({\mathcal G}) \ge 2$, the same approach cannot be used to prove the result in this situation. Conversely, it does not seem that our method can be adapted to include the case of ${\mathcal H}^2$ and ${\mathcal H}^3$.
Very recently, Aka constructed non-isomorphic arithmetic lattices that have isomorphic profinite completions [@Aka]. In particular, his construction gives arbitrarily large families of lattices of equal covolume in the Lie group $\operatorname{SL}_n({\mathbb C})$, for any $n \ge 3$.
Acknowledgements {#acknowledgements .unnumbered}
----------------
I would like to thank Misha Belolipetsky, Pierre de la Harpe, Gopal Prasad and Matthew Stover for helpful comments on an early version of this paper. I also thank Menny Aka for a helpful correspondence on his preprint.
Arithmetic lattices {#sec:arithmetic-lattices}
===================
We can obviously reduce the proof of Theorem \[thm:general\] to the case of an irreducible $\Gamma$. Then, like in Theorem \[thm:exceptions\], ${\mathcal G}$ is supposed to be isotypic.
{#ss:arithmetic-subgroup}
For generalities on arithmetic groups we refer the reader to [@Zimmer84] and [@PlaRap94]. We briefly explain here how irreducible arithmetic lattices in ${\mathcal G}$ are obtained. Let $k$ be a number field with ring of integers ${\mathcal{O}}$. Let ${\mathrm G}$ be an absolutely simple simply connected algebraic group defined over $k$. We denote by ${\overline{{\mathrm G}}}$ the adjoint group of ${\mathrm G}$, i.e., the $k$-group defined as ${\mathrm G}$ modulo its center, and by $\pi : {\mathrm G}\to {\overline{{\mathrm G}}}$ the natural isogeny. Let ${\mathscr{S}}$ be the set of archimedean places $v$ of $k$ such that ${\mathrm G}(k_v)$ is non-compact. We denote by ${\mathrm G}_{\mathscr{S}}$ the product $\prod_{v \in {\mathscr{S}}} {\mathrm G}(k_v)$, and similarly for ${\overline{{\mathrm G}}}_{\mathscr{S}}$. Note that ${\mathrm G}_{\mathscr{S}}$ is connected. For any matrix realization of ${\mathrm G}$, the group ${\mathrm G}({\mathcal{O}})$ is an irreducible lattice in ${\mathrm G}_{\mathscr{S}}$. Suppose that the connected component $({\overline{{\mathrm G}}}_{\mathscr{S}})^\circ$ of ${{\overline{{\mathrm G}}}_{\mathscr{S}}}$ is isomorphic to ${\mathcal G}$. Then $\pi$ extends to a surjective map ${\pi_{\mathscr{S}}}:
{\mathrm G}_{\mathscr{S}}\to {\mathcal G}$. An irreducible lattice in ${\mathcal G}$ is called *an arithmetic lattice* if it is commensurable with a subgroup of the form ${\pi_{\mathscr{S}}}({\mathrm G}({\mathcal{O}}))$ for some $k$-group ${\mathrm G}$ as above.
In the following ${\mathrm G}$ will always be a $k$-group as above, which determines a commensurability class of arithmetic lattices in ${\mathcal G}$.
{#ss:adelic-group}
We denote by ${{V}_{\mathrm{f}}}$ the set of finite places of $k$, and by ${\mathbb{A}_\mathrm{f}}$ the ring of finite adèles of $k$. For each $v \in {{V}_{\mathrm{f}}}$ we consider $k_v$ the completion of $k$ with respect to $v$, and ${\mathcal{O}}_v \subset k_v$ its associated valuation ring. A collection ${P}=
({P}_v)_{v \in {{V}_{\mathrm{f}}}}$ of compact subgroups ${P}_v \subset {\mathrm G}(k_v)$ is called *coherent* if the product ${\mathcal{K}}_{P}= \prod_{v \in {{V}_{\mathrm{f}}}} {P}_v$ is open in the adelic group ${\mathrm G}({\mathbb{A}_\mathrm{f}})$ (see [@PlaRap94 Ch. 6] for information on adelic groups). For example, for any matrix realization of ${\mathrm G}$, the collection $({\mathrm G}({\mathcal{O}}_v))_{v \in {{V}_{\mathrm{f}}}}$ is coherent. For a coherent collection ${P}= ({P}_v)$, the group $$\begin{aligned}
{\Lambda}_{P}&=& {\mathrm G}(k) \cap \prod_{v \in {{V}_{\mathrm{f}}}} {P}_v,
\label{eq:PA_P}\end{aligned}$$ where ${\mathrm G}(k)$ is seen diagonally embedded into ${\mathrm G}({\mathbb{A}_\mathrm{f}})$, is an arithmetic subgroup of ${\mathrm G}(k)$ (and thus an arithmetic lattice in ${\mathrm G}_{\mathscr{S}}$). This follows from the equality ${\mathrm G}({\mathcal{O}}) = {\mathrm G}(k) \cap \prod_v
{\mathrm G}({\mathcal{O}}_v)$ together with the inequality $$\begin{aligned}
[{\Lambda}_{P}: {\Lambda}_{{P}'}] &\le& [{\mathcal{K}}_{P}: {\mathcal{K}}_{{P}'}],
\label{eq:inequality-global-local}\end{aligned}$$ valid for any two coherent collections ${P}$ and ${P}'$ with ${P}'_v \subset {P}_v$ for each $v\in {{V}_{\mathrm{f}}}$. Since ${\mathrm G}$ is simply connected, strong approximation holds [@PlaRap94 Theorem 7.12] and it follows that is in fact an equality. We put this (known) result in the following lemma.
\[lem:index-by-strong-approx\]
Let ${P}= ({P}_v)_{v \in {{V}_{\mathrm{f}}}}$ and ${P}' = ({P}'_v)_{v \in {{V}_{\mathrm{f}}}}$ be two coherent collections of compact subgroups such that ${P}'_v
\subset {P}_v \subset {\mathrm G}(k_v)$ for all $v \in {{V}_{\mathrm{f}}}$. Then
$$\begin{aligned}
[{\Lambda}_{P}: {\Lambda}_{{P}'}] &=& \prod_{v \in {{V}_{\mathrm{f}}}} [{P}_v:{P}_v'].
\label{eq:index-by-strong-approx}
\end{aligned}$$
{#ss:non-conjugated}
For every field extension $L|k$ with algebraic closure ${\overline{L}}$, the group of $L$-points given by ${\overline{{\mathrm G}}}(L)$ is identified with the inner automorphisms of ${\mathrm G}$ that are defined over $L$. Note that in general ${\overline{{\mathrm G}}}(L)$ is larger than the image of ${\mathrm G}(L)$ in ${\overline{{\mathrm G}}}({\overline{L}})$.
\[lem:non-conjugated\]
Let ${P}$ and ${P}'$ be two coherent collections of compact subgroups ${P}_v, {P}_v' \subset {\mathrm G}(k_v)$. Suppose that there exist a place $w \in {{V}_{\mathrm{f}}}$ such that ${P}_{w}$ and ${P}_{w}'$ are not conjugate by the action of ${\overline{{\mathrm G}}}(k_{w})$. Moreover, we suppose that ${P}_{w}$ and ${P}_{w}'$ contain the center of ${\mathrm G}(k_{w})$. Then ${\pi_{\mathscr{S}}}({\Lambda}_{P})$ and ${\pi_{\mathscr{S}}}({\Lambda}_{{P}'})$ are not conjugate in ${\mathcal G}$.
Let ${\mathrm C}$ be the center of ${\mathrm G}$. We may assume that each ${P}_v$ (resp. ${P}'_v$) contains the center ${\mathrm C}(k_v)$. If not replace ${P}_v$ by ${\mathrm C}(k_v) \cdot
{P}_v$; the image ${\pi_{\mathscr{S}}}({\Lambda}_{P})$ does not change with this modification, and the hypothesis at $w$ is kept.
Suppose that ${\pi_{\mathscr{S}}}({\Lambda}_{P})$ and ${\pi_{\mathscr{S}}}({\Lambda}_{{P}'})$ are conjugate in ${\mathcal G}$. Then ${\Lambda}_{{P}}$ and ${\Lambda}_{{P}'}$ are conjugate under the action of ${\mathcal G}\cong
({\overline{{\mathrm G}}}_{\mathscr{S}})^\circ$. Since arithmetic subgroups of ${\mathrm G}$ are Zariski-dense, we have more precisely that ${\Lambda}_{P}$ and ${\Lambda}_{{P}'}$ are conjugate by an element $g \in {\overline{{\mathrm G}}}(k)$. By strong approximation the closure of ${\Lambda}_{P}$ (resp. ${\Lambda}_{{P}'}$) in ${\mathrm G}(k_w)$ is ${P}_w$ (resp. ${P}'_w$), and it follows that $g$ conjugates ${P}_w$ and ${P}'_w$.
Parahoric subgroups and volume {#sec:parahorics}
==============================
In the following we assume that the reader has some knowledge of Bruhat-Tits theory. All the facts we need can be found in Tits’ survey [@Tits79]. See [@PlaRap94 §3.4] for a more elementary introduction.
{#ss:parahorics}
Let $v \in {{V}_{\mathrm{f}}}$. A *parahoric subgroup* of ${\mathrm G}(k_v)$, a certain kind of compact open subgroup of ${\mathrm G}(k_v)$, is by definition the stabilizer of a simplex in the Bruhat-Tits building attached to ${\mathrm G}(k_v)$. There are a finite number of conjugacy classes of parahoric subgroups in ${\mathrm G}(k_v)$; these conjugacy classes in ${\mathrm G}(k_v)$ correspond canonically to proper subsets of the local Dynkin diagram $\Delta_v$ of ${\mathrm G}(k_v)$. If ${P}_v \subset {\mathrm G}(k_v)$ is a parahoric subgroup, we denote by $\tau({P}_v) \subset \Delta_v$ its associated subset, and we call it the *type* of ${P}_v$. Two parahoric subgroups ${P}_v$ and ${P}_v'$ can be conjugate by an element of ${\overline{{\mathrm G}}}(k_v)$ only if there is an automorphism of $\Delta_v$ that sends $\tau({P}_v)$ to $\tau({P}_v')$.
{#ss:residual-group-scheme}
Let us denote by ${\mathfrak f}_v$ the residual field of $k_v$. To each parahoric subgroup ${P}_v \subset {\mathrm G}(k_v)$, a smooth affine group scheme over ${\mathcal{O}}_v$ is associated in a canonical way [@Tits79 §3.4.1]. By reduction modulo $v$, this determines in turn an algebraic group over ${\mathfrak f}_v$. Its maximal reductive quotient is a ${\mathfrak f}_v$-group that will be denoted by the symbol ${\overline M}_v$. The structure of ${\overline M}_v$ can be determined from $\tau({P}_v)$ and the local index of ${\mathrm G}(k_v)$ by the procedure described in [@Tits79 §3.5].
{#ss:lang-isogeny}
Let $({\overline M}_v,{\overline M}_v)$ be the commutator group of ${\overline M}_v$, and let ${R}({\overline M}_v)$ be the radical of ${\overline M}_v$. Both are defined over ${\mathfrak f}_v$, and we have (see [@Spring98 8.1.6]) $$\begin{aligned}
\label{eq:M_v-as-product-almost-direct}
{\overline M}_v &=& ({\overline M}_v,{\overline M}_v) \cdot {R}({\overline M}_v). \end{aligned}$$ The radical ${R}({\overline M}_v)$ is a central torus in ${\overline M}_v$, whose intersection with $({\overline M}_v,{\overline M}_v)$ is finite [@Spring98 7.3.1]. It follows that the product map $$\begin{aligned}
\label{eq:M_v-product-map}
({\overline M}_v,{\overline M}_v) \times {R}({\overline M}_v) &\to& {\overline M}_v\end{aligned}$$ is an isogeny. By applying Lang’s isogeny theorem [@PlaRap94 Prop. 6.3], we obtain that the order of ${\overline M}_v({\mathfrak f}_v)$ is given by the following: $$\begin{aligned}
\label{order-of-M_v-as-product}
|{\overline M}_v({\mathfrak f}_v)| &=& |({\overline M}_v,{\overline M}_v)({\mathfrak f}_v)| \cdot |{R}({\overline M}_v)({\mathfrak f}_v)|.\end{aligned}$$
\[thm:prasad-formula\]
Let $\mu$ be a Haar measure on ${\mathrm G}_{\mathscr{S}}$. Then there exists a constant ${c_{\mathrm G}}$ (depending on the algebraic group ${\mathrm G}$) such that for any coherent collection ${P}$ of parahoric subgroups ${P}_v \subset {\mathrm G}(k_v)$, we have $$\begin{aligned}
\mu(\Lambda_{P}{\backslash}{\mathrm G}_{\mathscr{S}}) &=& {c_{\mathrm G}}\prod_{v \in {{V}_{\mathrm{f}}}} \frac{|{\mathfrak f}_v|^{\left( t_v + \dim
{\overline M}_v \right)/2 }}{|{\overline M}_v({\mathfrak f}_v)|},
\end{aligned}$$ where for each $v \in {{V}_{\mathrm{f}}}$ the integer $t_v$ depends only on the $k_v$-structure of ${\mathrm G}$.
This theorem is a much weaker form of Prasad’s volume formula, given in [@Pra89 Theorem 3.7]. In fact, Prasad’s result explicitly gives the value of ${c_{\mathrm G}}$ for a natural normalization of the Haar measure $\mu$. Moreover, the integers $t_v$ are explicitly known. Since we want to prove qualitative results, we will not need more than the statement of Theorem \[thm:prasad-formula\].
Proof of Theorem \[thm:general\] {#sec:proof-1}
=================================
We now prove Theorem \[thm:general\], assuming that the group ${\mathcal G}$ is isotypic. Let $\Gamma \subset {\mathcal G}$ be an irreducible arithmetic lattice, with ${\mathrm G}$ and ${\overline{{\mathrm G}}}$ the associated $k$-groups as in Section \[ss:arithmetic-subgroup\]. We retain all notation introduced above.
{#ss:cebotarev}
The group ${\mathrm G}$ is quasi-split over $k_v$ for almost all places $v$ [@PlaRap94 Theorem 6.7]. Let us denote by ${T}$ the set of the places $v \in {{V}_{\mathrm{f}}}$ such ${\mathrm G}$ is not quasi-split over $k_v$. Let $\ell|k$ be the smallest Galois extension such that ${\mathrm G}$ is an inner form over $\ell$ (see for instance [@Spring98 Ch. 17], where this field is denoted by $E_\tau$). If $v \not
\in {T}$ is totally split in $\ell|k$, i.e., if $\ell \subset k_v$, then ${\mathrm G}$ is split over $k_v$. It follows from the Chebotarev density theorem that the set of places $v \not \in {T}$ that are totally split in $\ell|k$ is infinite. Let us denote this infinite subset of ${{V}_{\mathrm{f}}}$ by ${S}$.
{#ss:good-parahorics-general}
Let $v \in {S}$. The local Dynkin diagram $\Delta_v$ of ${\mathrm G}(k_v)$ can be found in [@Tits79 §4.2]. Let $n$ be the absolute rank of ${\mathcal G}$ (and of ${\mathrm G}$). We suppose first that ${\mathcal G}$ (and consequently ${\mathrm G}$ as well) is not of absolute type ${\mathrm{A}}_n$. Then there exist two vertices $\alpha_1,\alpha_2 \in \Delta_v$ such that $\alpha_1$ is hyperspecial and $\alpha_2$ is not. Let ${P^{(1)}}_v$ (resp. ${P^{(2)}}_v$) be a parahoric subgroup in ${\mathrm G}(k_v)$ of type $\tau({P^{(1)}}_v) = \left\{
\alpha_1 \right\}$ (resp. $\tau({P^{(2)}}_v) = \left\{ \alpha_2 \right\}$). Then ${P^{(1)}}_v$ and ${P^{(2)}}_v$ are not conjugate by the action of ${\overline{{\mathrm G}}}(k_v)$ (see Section \[ss:parahorics\]). Note also that these two groups, being parahoric subgroups, contain the center of ${\mathrm G}(k_v)$. We consider the subgroup ${\overline M}_v$ associated with ${P^{(1)}}_v$ (resp. associated with ${P^{(2)}}_v$). In both cases $i=1,2$ the radical ${R}({\overline M}_v)$ is a split torus of rank $n-1$ and the semi-simple part $({\overline M}_v,{\overline M}_v)$ is of type ${\mathrm{A}}_1$. From we see that the order of ${\overline M}_v({\mathfrak f}_v)$ is the same for ${P^{(1)}}_v$ and ${P^{(2)}}_v$.
If ${\mathrm G}$ is of type ${\mathrm{A}}_n$ then $\Delta_v$ is a cycle of $n+1$ vertices, all hyperspecial. The group ${\overline{{\mathrm G}}}(k_v)$ acts simply transitively by rotations on $\Delta_v$. Let us choose a labelling $\alpha_0, \dots, \alpha_n$ of the vertices that follows an orientation of $\Delta_v$. We now consider ${P^{(1)}}_v$ with $\tau({P^{(1)}}_v) = \left\{ \alpha_0,\alpha_2 \right\}$, and ${P^{(2)}}_v$ with $\tau({P^{(2)}}_v) =
\left\{ \alpha_0, \alpha_3 \right\}$. If $n\ge 4$ then no rotation of $\Delta_v$ sends $\tau({P^{(1)}}_v)$ to $\tau({P^{(2)}}_v)$, so that ${P^{(1)}}_v$ and ${P^{(2)}}_v$ are not conjugate by ${\overline{{\mathrm G}}}(k_v)$. Moreover, we can check as above that the order of ${\overline M}_v$ is the same for ${P^{(1)}}_v$ and ${P^{(2)}}_v$.
\[ss:good-parahorics-A\_n\]
{#ss:construct-families}
We consider a coherent collection ${P}$ of parahoric subgroups ${P}_v \subset
{\mathrm G}(k_v)$. Let $m \in {\mathbb N}$ and choose a finite subset ${S}_m \subset {S}$ of length $m$. For each $v \in {S}_m$ we replace $P_v$ by either ${P^{(1)}}_v$ or ${P^{(2)}}_v$, and consider the arithmetic subgroup in ${\mathrm G}(k)$ associated with this modified coherent collection. Thus we obtain $2^m$ different arithmetic subgroups in ${\mathrm G}(k)$, and by Lemma \[lem:non-conjugated\] their images in ${\mathcal G}$ are pairwise non-conjugate. But by Theorem \[thm:prasad-formula\] they all have the same covolume.
To obtain families of torsion-free lattices we make the following change. Let us choose two distinct places $v_1, v_2 \in {S}\setminus
{S}_m$, and for $i=1,2$ replace ${P}_{v_i}$ by its subgroup ${K}_i$ defined as the kernel of the reduction modulo $v_i$. We denote this modified coherent collection by ${P}'$. Let $p_i$ be the characteristic of ${\mathfrak f}_{v_i}$. Then ${K}_i$ is a pro-$p_i$-group [@PlaRap94 Lemma 3.8], and since $p_1 \neq p_2$ we have that ${K}_1 \cap {K}_2$ is torsion-free. Thus $\Lambda_{{P}'}$ is torsion-free. The above construction with the coherent collection ${P}'$ instead of ${P}$ now gives non-conjugate lattices in ${\mathcal G}$ that are torsion-free. Using Lemma \[lem:index-by-strong-approx\] we see that these sublattices also share the same covolume.
{#ss:strong-rigidity}
Let $\operatorname{Aut}({\mathcal G})$ be the automorphism group of ${\mathcal G}$. Then $\operatorname{Aut}({\mathcal G})/{\mathcal G}$ (where ${\mathcal G}$ acts on itself as inner automorphisms) is a group whose order is bounded by the symmetries of the Dynkin diagram of ${\mathcal G}$. In particular, it is a finite group. By letting $m$ tends to infinity, we have constructed arbitrarily large families of non-conjugate lattices in ${\mathcal G}$ of the same covolume. By considering each family modulo the equivalence induced by the action of $\operatorname{Aut}({\mathcal G})/{\mathcal G}$, we see that there exist arbitrarily large families of lattices that are not conjugate by $\operatorname{Aut}({\mathcal G})$. Since strong rigidity holds for all the lattices under consideration (see [@Zimmer84 §5.1] and the references given there), we get that these families consist of non-isomorphic lattices.
Proof of Theorem \[thm:exceptions\] {#sec:proof-2}
====================================
We now give the proof of Theorem \[thm:exceptions\]. Thus we suppose that ${\mathfrak{g}_{\mathbb C}}$ has only factors of type ${\mathrm{A}}_n$ (with $n = 2$ or $n = 3$). Let $m \in {\mathbb N}$.
{#ss:good-extension-ell}
Let $k$ be a number field that has as many complex places as there are simple factor of ${\mathcal G}$ isomorphic to $\operatorname{PSL}_{n+1}({\mathbb C})$. Let $\ell|k$ be a quadratic extension having one complex place for each factor of ${\mathcal G}$ that is projective unitary (i.e., of the form $\operatorname{PU}(p,q)$) or isomorphic to $\operatorname{PSL}_{n+1}({\mathbb C})$. Using approximation for $k$ (see [@Neuk99 Theorem (3.4)]) it is possible to choose $\alpha \in k$ such that $\ell = k(\sqrt{\alpha})$ is as above with the additional property that for the set ${R}\subset {{V}_{\mathrm{f}}}$ of ramified places in $\ell|k$ we have $2^{\# {R}} \ge m$.
{#ss:G-from-PraRap}
Let ${\mathrm G}_0$ be the quasi-split simply connected $k$-group of type ${\mathrm{A}}_n$ with splitting field $\ell$. By [@PraRap06 Theorem 1], there exists an inner form ${\mathrm G}$ of ${\mathrm G}_0$ such that ${\mathrm G}|k_v$ is quasi-split for all $v \in {R}$ and such that $({\overline{{\mathrm G}}}_{\mathscr{S}})^\circ \cong {\mathcal G}$. The group ${\mathrm G}$ can be chosen to be $k$-isotropic unless the condition (1) in [@PraRap06] is not satisfied at infinite places, in which case there is no isotropic $k$-group ${\mathrm G}$ with $({\overline{{\mathrm G}}}_{\mathscr{S}})^\circ
\cong {\mathcal G}$. We can always choose ${\mathrm G}$ to be anisotropic, by specifying in [@PraRap06 Theorem 1] that ${\mathrm G}$ is $k_v$-anisotropic at some $v
\in {{V}_{\mathrm{f}}}\setminus {R}$.
{#ss:end-pf-thm-2}
The local Dynkin diagram $\Delta_v$ of ${\mathrm G}(k_v)$ for $v \in {R}$ is shown in [@Tits79 §4.2]; it is named $\mbox{C--BC}_1$ for the type ${\mathrm{A}}_2$, and $\mbox{C--B}_2$ for ${\mathrm{A}}_3$ ($={\mathrm{D}}_3$). With this diagram at hand we can easily construct (similarly to Section \[ss:good-parahorics-general\]) a pair of non-conjugate parahoric subgroups of ${\mathrm G}(k_v)$ ($v \in {R}$) that have equal volume. Taking them as part of coherent collection we produce $m$ pairwise non-conjugate arithmetic subgroups that, by Theorem \[thm:prasad-formula\], are of the same covolume in ${\mathcal G}$. By Godement’s compactness criterion, these lattices are cocompact exactly when ${\mathrm G}$ is anisotropic. The last steps of the proof are verified exactly as in Sections \[ss:construct-families\]–\[ss:strong-rigidity\].
[^1]: Supported by Swiss National Science Foundation, Project number [PP00P2-128309/1]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article, we derive the fermionic formalism of Hamiltonians as well as corresponding excitation spectrums and states of Calogero-Sutherland(CS), Laughlin and Halperin systems, respectively. In addition, we study the triangular property of these Hamiltonians and prove the integrability in these three cases.'
address:
- 'Institute of Theoretical Physics, Chinese Academy of Sciences, Beijing, China 100190'
- 'State Key Laboratory of Theoretical Physics, Beijing, China 100190'
- 'Institute of Theoretical Physics, Beijing University of Technology, Beijing, China 100124'
- 'Department of Mathematics, Jilin University, Changchun, China 130012'
- 'School of Mathematics Sciences, Capital Normal University, Beijing, China 100048'
- |
Beijing Center for Mathematics and Information Interdisciplinary Sciences,\
Beijing, China 100048
author:
- 'Li-Qiang Cai'
- 'Li-Fang Wang'
- 'Jian-Feng Wu'
- Jie Yang
- Ming Yu
title: 'Fermionization, Triangularization and Integrability'
---
Calogero-Sutherland ,Jack polynomial ,CFT ,FQHE ,integrability ,fermionization ,triangularization
02.10.Ox ,02.30.Ik ,05.30.Pr ,05.45.Yv
Introduction and Results\[sec:Introduction\]
============================================
In the area of many body physics, fractional quantum Hall effects (FQHEs) and integrable models, are two fruitful and important classes. Many researchers believe these two are connected in a hundred and one ways. [@azuma1994explicit; @bergere2000composite; @bernevig2008model; @estienne2012conformal; @ha1994exact; @zhang1989effective; @zhang1992chern] Plenty of efforts have been dedicated to find out the intrinsic relationship.
In FQHEs, the Laughlin trial wavefunction reveals several remarkable properties of FQHEs at filling number $\nu=\frac{1}{2m+1},$ such as the fractional statistics as well as the topological orders. Later on a conformal field theory (CFT) realization was discovered which shows that the wavefunction is corresponding to a correlation function of certain vertex operators . Furthermore this idea is generalized to many other FQH states, e.g. Halperin state[@halperin1984statistics], Moore-Read state, and Read-Rezayi state[@read1999beyond], et.al[@bergholtz2008quantum; @gurarie1997haldane; @moore1991nonabelions; @wen1994chiral]. However, CFT is possibly not sufficient to drive the dynamics of the edge theory, since it only determines the behavior of the theory near critical point.[^1].
A more ambitious thinking is to find the Hamiltonian system behind the edge ground state. So far, there are two classes of Hamiltonian analysis for FQHE. One is the Chern-Simons approach, initiated by Zhang, Hansson and Kivelson in 1989[@zhang1989effective]. The other is the extended Hamiltonian theory, introduced by Murthy and Shankar in late 90’s[@murthy1999hamiltonian; @murthy2003hamiltonian; @shankar1997towards]. The later one contains Chern-Simons as its asymptotic theory.
In our study we try to approach the integrability problem in a different way. In fact, we are not meant to establish a unified Hamiltonian theory to solve the complicated many-body problem. Instead we are looking for the integrability behind FQHEs as well as the Hamiltonian expression of it. In order to do so we separate the excitations of FQHE into two simple classes: the perturbative class and nonperturbative one. The nonperturbative class dominates the states in Hilbert space, a.k.a. the basis, the perturbative class organizes those basis into physical states. So perturbations actually are provided as structure constants (or superposition coefficients). Interestingly, this idea is like in CFT, where correlation function is made by conformal block and structure constant (it encodes the multiplicity of the corresponding conformal block in the correlation function. ) Since the ground state should not change by perturbations it belongs to the nonperturbative class. Hence it describes a sort of wave without dissipation which implies that the ground state is a solitonic wave.
In this way we have related the FQHE theory to soliton theory, the other important area of many-body physics. The question now is to extract excitations from the solitonic wavefunction. The stable excitations from the soliton wavefunction, are those solutions of quantum mechanics equation for soliton wave[@das1989integrable]. In this quantum mechanics, the logarithmic of the soliton wavefunction is a scalar function, while its gradation, gives the effective “gauge” potential. Therefore, the Hamiltonian could be written as a Landau-Ginzberg pseudo-potential form.
Inspired by these observations and a previous work [@Wu:2011tt], we use the same method for Laughlin and Halperin states. Then we obtain complicated Hamiltonians with non-linear interactions. However, they are all exact solvable. The resolving strategy is as follows: firstly, we interpret the ground state as correlation function in CFT. Secondly, by Jastrow transformation we drop the contribution of ground state and obtain a relative simple Hamiltonian. Thirdly, the eigen-equation of the new Hamiltonian can be transformed into an operator equation acting on the coherent basis. Fourthly, it turns out that the operator formalism is exactly triangulated. Therefore we can extract the spectrum as well as the state in a recursive way. Finally, to analyze the integrability closely, we derive the fermionization for the bosonic theory. Hence the integrability is clearly determined by free fermions and the explicit triangularization.
We find, interestingly, the integrability behind Laughlin state, is the same as the famous Calogero-Sutherland model. Hence the excitations are those of Jack polynomials[@macdonald1995symmetric]. During last two decades, people claimed that ground states of some FQHEs have the same properties as those of Jack polynomials. For example the $(k,r,N)$-admissible representations (labeled by certain restricted Young diagrams) is related to the filling number $\nu=\frac{k}{r}$ FQHE ground state[@azuma1994explicit; @bernevig2008generalized; @bernevig2008properties; @estienne2010clustering; @estienne2010electron; @estienne2009relating; @feigin2002differential; @ha1994exact; @iso1995collective; @lee2014construction]. From our viewpoint the basic ingredients are Jack polynomials and additional restrictions, mostly from the fusion rule (which we do not explore in this article), will rule out some Jack polynomials systematically which results in the admissible representations.
The Halperin state, corresponding to the two-layer FQHE, shows a secret integrability dominated by also the triangularization, which says the number of boxes in Young diagram for the first layer always decreases while the one for the second layer increases and the total boxes of these two layers remain the same. The triangularization interaction, being triple in two kinds of bosonic operators, is quite complicated. It makes the explicit solution of the excitations slightly difficult. Nevertheless, since the triangularization is clear, we can give the explicit solution of the system in principle.
This article is organized as follows. In sec. 2, we review the famous Calogero-Sutherland model, its operator formalism, the CFT correspondence, the spectrum and eigenstates. In sec. 3 we obtain the fermionization of the CS theory followed by the fermionic triangularization and integrability. In sec. 4 and 5, we provide parallel analysis for Laughlin state and Halperin state. In sec. 6 we make a conclusion and discuss some further works.
The Calogero-Sutherland Model
=============================
We start our analysis from the famous Calogero-Sutherland(CS) model. It is an exact solvable model, describing $N$ interacting charged particles on a unit circle, with two-body interaction $$\begin{aligned}
H_{int}=\sum_{i<j} & \frac{\beta(\beta-1)}{\sin^{2}(x_{i}-x_{j})} & \,,\end{aligned}$$ in which $x_{i}$ defines the $i$-th particle’s position on the circle. For simplicity, we substitute $\beta=b^{2}$. Then CS Hamiltonian is written as $$H_{CS}=-\frac{1}{2}\sum_{i=1}^{N}\partial_{i}^{2}+\sum_{i<j}\frac{b^{2}(b^{2}-1)}{\sin^{2}(x_{i}-x_{j})}\,.$$
\[theo:isospectrum\]$H_{CS}$ is isospectral to another Hamiltonian $$\tilde{H}_{CS}=-\frac{1}{2}\sum_{i=1}^{N}(\partial_{i}+\partial_{i}\ln\prod_{j<k}\sin^{b^{2}}(x_{j}-x_{k}))(\partial_{i}-\partial_{i}\ln\prod_{r<s}\sin^{b^{2}}(x_{r}-x_{s})\,.$$ up to a universal shift of eigen-energy.
$\!\textbf{Proof of Theorem }$$\ref{theo:isospectrum}$: Defining the complex coordinate $z_{i}=e^{i2x_{i}}$, we have $\partial_{i}=2iz_{i}\partial_{z_{i}}$. Therefore $$\begin{aligned}
\partial_{i}\ln\prod_{j<k}\sin^{b^{2}}(x_{j}-x_{k}) & = & b^{2}\sum_{\substack { j\\i\neq j}}\cot(x_{i}-x_{j})\\
& = & ib^{2}\sum_{\substack { j\\i\neq j}}\frac{z_{i}+z_{j}}{z_{i}-z_{j}}\,,\end{aligned}$$ and the commutator $$\left[\partial_{i}, b^{2}\sum_{\substack{ k \\i\neq k}}\cot(x_{i}-x_{k})\right]=-b^{2}\sum_{\substack{ k \\k\neq i}}\frac{1}{\sin^{2}(x_{k}-x_{i})}.$$ We can rewrite the $\tilde{H}_{CS}$ as the following formula $$\begin{aligned}
\tilde{H}_{CS} & = & -\frac{1}{2}\sum_{i=1}^{N}\partial_{i}^{2}-b^{2}\sum_{i<j}\frac{1}{\sin^{2}(x_{i}-x_{j})}\\
& & +\frac{1}{2}b^{4}\sum_{i\neq j,i\neq k}\cot\left(x_{i}-x_{j}\right)\cot\left(x_{i}-x_{k}\right)\,.\end{aligned}$$ Using the identity $$\begin{aligned}
\sum_{\hbox{distinct \ }i, j, k}\cot\left(x_{i}-x_{j}\right)\cot\left(x_{i}-x_{k}\right)+{i, j, k}\text{{\ cyclic}}\\
=\sum_{\hbox{distinct\ } i, j, k}(-1)=-N(N-1)(N-2) & \,,\end{aligned}$$ and the $j=k$ contribution $$\sum_{i\neq j}\cot^{2}\left(x_{i}-x_{j}\right)=-\sum_{i\neq j}1+\sum_{i\neq j}\frac{1}{\sin^{2}(x_{i}-x_{j})}\,,$$ we now have the form of $\tilde{H}_{CS}$ as $$\begin{aligned}
\tilde{H}_{CS} & = & -\frac{1}{2}\sum_{i=1}^{N}\partial_{i}^{2}+b^{2}(b^{2}-1)\sum_{i<j}\frac{1}{\sin^{2}(x_{i}-x_{j})}-\frac{1}{6}b^4(N-1)N(N+1)\,,\nonumber \\
& = & H_{CS}-\frac{1}{6}b^4N(N-1)(N+1)\end{aligned}$$ so Theorem $\ref{theo:isospectrum}$ is proved. Q.E.D.
It is now nature to consider the $\tilde{H}_{CS}$ rather than $H_{CS}$ since the later one, when acting on the ground state, will have a large energy (proportional to $N^{3}$) contribution to the spectrum. The ground state of $\tilde{H}_{CS}$ is simply $$\Psi_{CS}=\prod_{i<j}\sin^{\beta}(x_{i}-x_{j}),\quad \tilde{H}_{CS}\Psi_{CS}=0\text{\,.}\label{eq:GroundCS}$$ To extract the spectrums as well as corresponding excitation states, we need to eliminate the contribution of ground state. It implies the Jacobi transformation $$2H'_{CS}=\Psi_{CS}^{-1}\tilde{H}_{CS}\Psi_{CS}\,.$$ In this way, we have $$\begin{aligned}
2H'_{CS} & = & -\frac{1}{2}\sum_{i}(\partial_{i}+2\partial_{i}\ln\Psi_{CS})\partial_{i}\nonumber \\
& = & -\frac{1}{2}\sum_{i}(2iz_{i}\partial_{z_{i}}+i2b^{2}\sum_{i\neq j}\frac{z_{i}+z_{j}}{z_{i}-z_{j}})(2iz_{i}\partial_{z_{i}})\nonumber \\
& = & 2\sum_{i}(z_{i}\partial_{z_{i}})^{2}+2b^{2}\sum_{i<j}\frac{z_{i}+z_{j}}{z_{i}-z_{j}}(z_{i}\partial_{z_{i}}-z_{j}\partial_{z_{j}})\,.\label{eq:complexHcs}\end{aligned}$$
Bosonic oscillator formalism of $H'_{CS}$
-----------------------------------------
The ground state as in ($\ref{eq:GroundCS}$) can be understood as a CFT correlation function, that is $$\Psi_{CS}({z_{i}})\simeq\langle k_{f}|\prod_{i=1}^{N}V_{b}(z_{i})|k_{in}\rangle\,,$$ with the vertex operator defined by $$V_{b}(z)\equiv :e^{b\phi(z)}:\,,$$ and the bosonic field has the standard mode expansion $$\begin{aligned}
\phi(z) & = & q_{0}+p_{0}\ln z+\sum_{n\neq0}\frac{a_{-n}}{n}z^{n}\,,\\
{}[a_{n},a_{m}] & = & n\delta_{n+m,0},\quad [p_{0},q_{0}]=1\,.\end{aligned}$$ We can show that briefly. The OPE of vertex operators reads $$V_{a}(z)V_{b}(w)=(z-w)^{ab}:V_{a}V_{b}(\frac{z+w}{2}):\,.$$ If we choose the initial (final) momentum of right (left) vacuum $k_{in}=\frac{b}{2}(1-N)$ ($k_{f}=k_{in}+Nb$) , the correlation function is charge neutral and gives the result $$\begin{aligned}
\langle k_{f}|\prod_{i=1}^{N}V_{b}(z_{i})|k_{in}\rangle & = & \prod_{i<j}(z_{i}-z_{j})^{b^{2}}\prod_{i}(z_{i})^{\frac{b^{2}}{2}(1-N)}\\
& = & \prod_{i<j}\left(\frac{z_{i}-z_{j}}{\sqrt{z_{i}z_{j}}}\right)^{b^{2}}\\
& = & \prod_{i<j}(2i\sin(x_{i}-x_{j}))^{b^{2}}\,.\end{aligned}$$ Therefore up to a constant factor, it is the ground state of CS model. The excitation state, in principle, will be a state in the Fock space of the conformal field theory, which in general is a polynomial of bosonic oscillators. It implies there are one-to-one correspondence from the excitation wavefunction to an oscillator polynomial. The basic relation is the coherent relation such that $$a_{n}\prod_{i}V_{b}^{-}(z_{i})|k_{in}\rangle=b\sum_{i}z_{i}^{n}\prod_{i}V_{b}^{-}(z_{i})|k_{in}\rangle\,.\label{eq:coherent-1}$$ It relates the bosonic oscillator mode $a_{m}$to a symmetric polynomial (or symmetric function if $N\rightarrow\infty$). If we define the excitation state as follows $$\begin{aligned}
\frac{1}{2}\tilde{H}_{CS}\Psi_{\lambda}^{\beta}({z_{i}}) & = & \Psi_{CS}H'_{CS}P_{\lambda}^{\beta}({z_{i}})=E_{\lambda}^{\beta}\Psi_{\lambda}^{\beta}({z_{i}})\\
\Psi_{\lambda}^{\beta}({z_{i}}) & = & \Psi_{CS}({z_{i}})P_{\lambda}^{\beta}({z_{i}})\\
& \simeq & \langle k_{f}|\prod_{i=1}^{N}V_{b}(z_{i})P_{\lambda}^{\beta}(a^{-})|k_{in}\rangle\,,\end{aligned}$$ then $$\langle k_{f}|P_{\lambda}^{\beta}(a^{+})H'_{CS}(a)\prod_{i=1}^{N}V_{b}^{-}(z_{i})|k_{in}\rangle=E_{\lambda}^{\beta}P_{\lambda}^{\beta}({z_{i}})\,.\label{eq:OpereqCS}$$ We have defined here the normal-ordered operator formalism of $H'_{CS}(a)\equiv H$ , such that
$$H|P_{\lambda}^{\beta}\rangle=E_{\lambda}|P_{\lambda}^{\beta}\rangle.$$ The next step is to translate the differential formalism Hamiltonian $H'_{CS}$ into operator formalism with the help of coherent relation (\[eq:coherent-1\]). We have the following relations $$\begin{aligned}
z_{i}\partial_{z_{i}}\bullet = b\sum_{n>0}a_{-n}z_{i}^{n}\bullet,& &
(z_{i}\partial_{z_{i}})^{2}\bullet = b^{2}\sum_{n,m>0}a_{-n}a_{-m}z_{i}^{n+m}\bullet
+b\sum_{n>0}na_{-n}z_{i}^{n}\bullet\\
\sum_{i<j}\frac{z_{i}+z_{j}}{z_{i}-z_{j}}(z_{i}\partial_{z_{i}}-z_{j}\partial_{z_{j}})\bullet & = & b\sum_{i<j,n>0}a_{-n}\frac{z_{i}+z_{j}}{z_{i}-z_{j}}(z_{i}^{n}-z_{j}^{n})\bullet\\
=b\sum_{i<j,n>0}a_{-n}(z_{i}^{n}+2z_{i}^{n-1}z_{j} & + & \cdots2z_{i}z_{j}^{n-1}+z_{j}^{n})\bullet\\
=\left(b\sum_{n,m>0,i,j}a_{-n}z_{i}^{n-m}z_{j}^{m} \right.& + & \left. Nb\sum_{n>0,i=1}^N a_{-n}z_{i}^{n}-b\sum_{n>0,i=1}^N na_{-n}z_{i}^{n}\right)\bullet\end{aligned}$$ where $\bullet$ denotes $\prod_{j=1}^{N}V_{b}^{-}(z_{j})|k_{i}\rangle$. So we have the CS Hamiltonian in bosonic operator formalism $$\begin{aligned}
H=\sum_{n,m>0}b(a_{-n}a_{-m}a_{n+m}+a_{-n-m}a_{n}a_{m})\\
+(1-b^{2})\sum_{n>0}na_{-n}a_{n}+b^{2}N\sum_{n>0}a_{-n}a_{n}\,.\nonumber\end{aligned}$$ The last term involves the level of corresponding excitations, when $N\rightarrow\infty$, it overwhelms the excitation spectrum since it is much larger than other contributions in $H$. In our analysis, we treat it as the background and we ignore this term. Besides, if we set $$\tilde{a}_{-n}=\frac{a_{-n}}{b},\,\,\tilde{a}_{n}=a_{n}b\,,{\text for}\,\,n>0$$ then we rewrite $H$ as $$\begin{aligned}
H & = & \sum_{n,m>0}b(b\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}+\frac{1}{b}\tilde{a}_{-n-m}\tilde{a}_{n}\tilde{a}_{m})\label{deformcs}\\
& + & (1-b^{2})\sum_{n>0}n\tilde{a}_{-n}\tilde{a}_{n}\nonumber \\
& = & \sum_{n,m>0}(\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}+\tilde{a}_{-n-m}\tilde{a}_{n}\tilde{a}_{m})\nonumber \\
& + & (1-b^{2})\left(\sum_{n>0}n\tilde{a}_{-n}\tilde{a}_{n}-\sum_{n,m>0}\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}\right)\,.\nonumber\end{aligned}$$ It is easy to see that $\tilde{a}$ still holds the Heisenberg algebra so that the fermionization is exact. We split the Hamiltonian into free part (the first line of last equality of (\[deformcs\])), which is the same as free fermions, and the interacting part (the second line of last equality of (\[deformcs\])).
Eigenstate and spectrum
-----------------------
The CS model is exactly solvable. To see that, we first classify the Fock space expanded by bosons by its level $\mathcal{N}=\sum_{n>0}a_{-n}a_{n}$ such that an arbitrary state $$|n_{1},n_{2},\cdots,n_{l}\rangle=a_{-n_{1}}a_{-n_{2}}\cdots\, a_{-n_{l}}|0\rangle\,,\label{eq:Fockstate}$$ has a level $$\mathcal{N}|n_{1},n_{2},\cdots\, n_{l}\rangle=\sum_{i=1}^{l}n_{i}|n_{1},n_{2},\cdots\, n_{l}\rangle\,.$$ In this classification, there are $P(k)$, the partition number of $k$, states at a given level $k$. It is easy to check that $H$ commutes with $\mathcal{N}$. Hence they can have common eigenstates. Therefore, the eigenstate of $H$ can be obtained by a superposition of states like (\[eq:Fockstate\]). A closer observation shows that the Hamiltonian $H$ acting on a state (\[eq:Fockstate\]) at level $k$ by certain times will definitely generate the lowest state $|1^{k}\rangle\equiv(a_{-1})^{k}|0\rangle$. We can choose the coefficients of the Fock state $|1^{k}\rangle$ of all states at level $k$ to be the same and equal to $b^{-k}$.[^2] By this choice, we have removed the irrelevant c-number common factor of each eigenstate. For example, at level 4, we assume an eigenstate has the following formalism $$|P_{\lambda}^{\beta}\rangle=b^{-4}((a_{-1})^{4}+\alpha_{1}a_{-2}(a_{-1})^{2}+\alpha_{2}a_{-2}^{2}+\alpha_{3}a_{-3}a_{-1}+\alpha_{4}a_{-4})|0\rangle\,,$$ Thus there are $P(4)-1=5-1=4$ unknown coefficients and also the eigen-energy $E_{\lambda}^{\beta}$ is not known. However, compare all the coefficients of the eigen-equation $$H|P_{\lambda}^{\beta}\rangle=E_{\lambda}^{\beta}|P_{\lambda}^{\beta}\rangle\,,$$ we have in total 5 independent equations. They in turn determine the eigenstate completely. The generalization to level $k$ is then straightforward.
However, this method does not provide a clear relation between the eigenstate and the Young diagram underlining the theory. In general, one can define by hand a sequence of eigen-energies at a given level so that each state is uniquely related to a Young diagram. But the reason is weak and unnatural. However, it is quite natural to see the Young diagram from the fermionic picture, which we will explore in next section.
Fermionization
==============
Fermionization of free term
---------------------------
We now rewrite the CS Hamiltonian as $H\equiv H_0 + H_{int}$, here $$\label{eq:freefermion}
H_0 = \sum_{n,m>0}(\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}+\tilde{a}_{-n-m}\tilde{a}_{n}\tilde{a}_{m})$$ is the free part, while $$\begin{aligned}
\label{eq:intpart}
H_{int} = (1-b^{2})\left(\sum_{n>0}n\tilde{a}_{-n}\tilde{a}_{n}-\sum_{n,m>0}\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}\right)\end{aligned}$$ is the interaction part which could be separated into two components $$\begin{aligned}
H_int &=&H_1+H_2, H_1 = (1-b^{2})(-\sum_{n,m>0}\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m})\\
H_2&=&(1-b^{2})\sum_{n}n\tilde{a}_{-n}\tilde{a}_{n} \end{aligned}$$ for further convenience. Now we want to fermionize the deformed bosonic Hamiltonian $H$ by introducing $$\tilde{a}_{n}=\sum_{r\in\mathbb{Z}+\frac{1}{2}}:\psi_{n-r}\psi_{r}^{*}:\,,$$ and also the free Virasoro generator $$\begin{aligned}
\tilde{T}(z) & = & \frac{1}{2}(\partial_{z}\tilde{\phi}(z))^{2}=-\frac{1}{2}[\psi\partial\psi^{*}+\psi^{*}\partial\psi](z)\\
\tilde{L}_{n} & = & \frac{1}{2}\sum_{m}:\tilde{a}_{n-m}\tilde{a}_{m}:=\sum_{r>0,r\in\mathbb{Z}+\frac{1}{2}}(r+\frac{n}{2}):\psi_{-r}\psi_{n+r}^{*}:\end{aligned}$$ Firstly, let us consider the free part $H_0$. Notice that, $$\begin{aligned}
\sum_{n>0}\tilde{L}_{-n}\tilde{a}_{n} & = & \sum_{n,m>0}\tilde{a}_{-n-m}\tilde{a}_{m}\tilde{a}_{n}+\frac{1}{2}\sum_{n>m>0}\tilde{a}_{-n+m}\tilde{a}_{-m}\tilde{a}_{n}\\
\sum_{n>0}\tilde{a}_{-n}\tilde{L}_{n} & = & \frac{1}{2}\sum_{n>m>0}\tilde{a}_{-n}\tilde{a}_{n-m}\tilde{a}_{m}+\sum_{n,m>0}\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}\,.\end{aligned}$$ It gives rise to $$\begin{aligned}
H_{0}=\sum_{n,m>0}(\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}+\tilde{a}_{-n-m}\tilde{a}_{n}\tilde{a}_{m})=\frac{2}{3}\left(\sum_{n>0}\tilde{L}_{-n}\tilde{a}_{n}+\tilde{a}_{-n}\tilde{L}_{n}\right)\\
\sum_{n,m>0}\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}=\frac{2}{3}\left(\sum_{n>0}[2\tilde{a}_{-n}\tilde{L}_{n}-\tilde{L}_{-n}\tilde{a}_{n}]\right)\,.\end{aligned}$$ Notice that the $H_{0}$ is just the zero mode of the OPE of $$\frac{1}{2\pi i}\oint\frac{dz}{z-w}\tilde{T}(z)\partial_{w}\tilde{\phi}(w)\,.$$ In fermionic representation, we have $$\begin{aligned}
H_{0}(w) & = & \frac{2}{3}\oint\frac{dz}{2\pi i(z-w)}\left(-\frac{1}{2}[\psi\partial_{z}\psi^{*}+\psi^{*}\partial_{z}\psi])(z)[\psi\psi^{*}](w)\right)\\
& = & \frac{2}{3}\oint\frac{dz}{2\pi i(z-w)}\left\{ \frac{1}{2}\left(\frac{1}{(z-w)^{3}}+\frac{\psi(z)\psi^{*}(w)}{(z-w)^{2}}-\frac{\partial_{z}\psi^{*}(z)\psi(w)}{z-w}\right)\right.\nonumber \\
& & +\left.\frac{1}{2}\left(-\frac{1}{(z-w)^{3}}-\frac{\psi^{*}(z)\psi(w)}{(z-w)^{2}}+\frac{\partial_{z}\psi(z)\psi^{*}(w)}{z-w}\right)\right\} \nonumber \\
& = & \frac{1}{3}\oint\frac{dz}{2\pi i(z-w)}\left\{ \frac{\psi(z)\psi^{*}(w)-\psi^{*}(z)\psi(w)}{(z-w)^{2}}\right.\nonumber \\
& & +\left.\left(\frac{\partial_{z}\psi(z)\psi^{*}(w)-\partial_{z}\psi^{*}(z)\psi(w)}{z-w}\right)\right\} \nonumber \\
& = & \frac{1}{2}\left([(\partial_{w})^{2}\psi(w)]\psi^{*}(w)-[(\partial_{w})^{2}\psi^{*}(w)]\psi(w)\right)\,.\nonumber\end{aligned}$$ The operator formalism $H_{0}$ is $$\begin{aligned}
H_{0} & = & \frac{1}{2\pi i}\oint w^{2}dwH_{0}(w)\\
& = & \frac{1}{2}\left([(\partial_{w})^{2}\psi(w)]\psi^{*}(w)-[(\partial_{w})^{2}\psi^{*}(w)]\psi(w)\right)\nonumber \\
& = & \frac{1}{2}\sum_{r\in\mathbb{Z}+\frac{1}{2}}:\psi_{-r}\psi_{r}^{*}:\left((-r-\frac{1}{2})(-r-\frac{3}{2})+(r-\frac{1}{2})(r-\frac{3}{2})\right)\nonumber \\
& = & \sum_{r>0,r\in\mathbb{Z}+\frac{1}{2}}\left(r^{2}+\frac{3}{4}\right)(\psi_{-r}\psi_{r}^{*}-\psi_{-r}^{*}\psi_{r})\,.\nonumber\end{aligned}$$
Fermionization of triple bosons term
------------------------------------
Now let us consider the term $$\begin{aligned}
& & (1-b^{2})(-\sum_{n,m>0}\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m})\\
& & =\frac{2}{3}(1-b^{2})\sum_{n}[-2\tilde{a}_{-n}\tilde{L}_{n}+\tilde{L}_{-n}\tilde{a}_{n}]\\
& & =\frac{2}{3}(1-b^{2})\left(\sum_{n}{{\diamond}\atop {\diamond}}\tilde{L}_{-n}\tilde{a}_{n}-2\tilde{a}_{-n}\tilde{L}_{n}{{\diamond}\atop {\diamond}}+{\text{Contractions}}\right)\,.\end{aligned}$$ In formulation of $\tilde{a}$ and $\tilde{L}$, it is easier to calculate the fermionic expression. We have $$\begin{aligned}
{{\diamond}\atop {\diamond}}\tilde{L}_{-n}\tilde{a}_{n}-2\tilde{a}_{-n}\tilde{L}_{n}{{\diamond}\atop {\diamond}} & = & \sum_{\begin{subarray}{c}
r\in\mathbb{Z}+\frac{1}{2}\\
s\in\mathbb{Z}+\frac{1}{2}
\end{subarray}}\left((r-\frac{n}{2}):\psi_{-r}\psi_{-n+r}^{*}\psi_{-s}\psi_{n+s}^{*}:\right.\\
& & -\left.(2s+n):\psi_{-r}\psi_{-n+r}^{*}\psi_{-s}\psi_{n+s}:\right)\nonumber \\
& = & \sum_{\begin{subarray}{c}
r,s,k,l\in\mathbb{Z}+\frac{1}{2}\\
r+s<0,l+k>0\\
r+s+k+l=0
\end{subarray}}:\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:\left(\frac{1}{2}(s-r)+k-l\right)\,.\nonumber\end{aligned}$$ To calculate the contractions, we need to calculate the commutator $$\begin{aligned}
[\tilde{L}_{-n},\psi_{-k}] & = & \sum_{r}\left[(r-\frac{n}{2}):\psi_{-r}\psi_{-n+r}^{*}:\,,\,\,\psi_{-k}\right]\\
& = & \left(\theta(k>0)\sum_{r>0}\delta_{r-n,k}\psi_{-r}\right.\nonumber \\
& & +\left.\theta(k<0)\sum_{r<0}\delta_{r-n,k}\psi_{-r}\right)(r-\frac{n}{2})\nonumber \\
& = & (k+\frac{n}{2})\psi_{-n-k}\nonumber \\
\,[\tilde{L}_{-n},\psi_{n+k}^{*}] & = & \sum_{r}\left[(r-\frac{n}{2}):\psi_{-r}\psi_{-n+r}^{*}:\,,\,\,\psi_{n+k}^{*}\right]\\
& = & \left(-\theta(n+k>0)\sum_{r>0}\delta_{r,n+k}\psi_{-n+r}^{*}\right.\nonumber \\
& & -\left.\theta(n+k<0)\sum_{r<0}\delta_{r,n+k}\psi_{-n+r}^{*}\right)(r-\frac{n}{2})\nonumber \\
& = & -(k+\frac{n}{2})\psi_{k}^{*}\nonumber \\
\,[\psi_{-k},\tilde{L}_{n}] & = & (-k+\frac{n}{2})\psi_{n-k}\\
\,[\psi_{-n+k}^{*},\tilde{L}_{n}] & = & (k-\frac{n}{2})\psi_{k}^{*}\,.\end{aligned}$$ Hence we have the contraction as follows $$\begin{aligned}
{\text{Contractions}} & = & \left(\sum_{n\in\mathbb{Z},k>0}:[\tilde{L}_{-n},\psi_{-k}]\psi_{n+k}^{*}:+\sum_{n\in\mathbb{Z},n+k<0}:\psi_{-k}[\tilde{L}_{-n},\,\psi_{n+k}^{*}]:\right.\nonumber \\
& & -2\left.\left(\sum_{n\in\mathbb{Z},n-k<0}:\psi_{-k}[\psi_{-n+k}^{*},\tilde{L}_{n}]:+\sum_{n\in\mathbb{Z},k<0}:[\psi_{-k},\,\tilde{L}_{n}]\psi_{-n+k}^{*}:\right)\right)\nonumber \\
& = & \left(\sum_{k>0}(k+\frac{n}{2}):\psi_{-n-k}\psi_{n+k}^{*}:-\sum_{n+k<0}(k+\frac{n}{2}):\psi_{-k}\psi_{k}^{*}:\right.\nonumber \\
& & -2\left.\sum_{n<k}(k-\frac{n}{2}):\psi_{-k}\psi_{k}^{*}:+2\sum_{k<0}(k-\frac{n}{2}):\psi_{n-k}\psi_{-n+k}^{*}:\right)\nonumber \\
& = & \sum_{k>n}(k-\frac{n}{2}):\psi_{-k}\psi_{k}:-\sum_{k<-n}(k+\frac{n}{2}):\psi_{-k}\psi_{k}^{*}:\nonumber \\
& & +\sum_{k<-n}(2k+n):\psi_{-k}\psi_{k}^{*}:-\sum_{k>n}(2k-n):\psi_{-k}\psi_{k}^{*}:\nonumber \\
& = & \left(\sum_{k>n}(\frac{n}{2}-k)+\sum_{k<-n}(k+\frac{n}{2})\right):\psi_{-k}\psi_{k}^{*}:\nonumber \\
& = & \sum_{0<n<k}(\frac{n}{2}-k)(\psi_{-k}\psi_{k}^{*}-\psi_{-k}^{*}\psi_{k})\\
& = & \sum_{k>0}\left((-k)(k-\frac{1}{2})+\frac{1}{4}(k+\frac{1}{2})(k-\frac{1}{2})\right)(\psi_{-k}\psi_{k}^{*}-\psi_{-k}^{*}\psi_{k})\nonumber \\
& = & -\frac{1}{16}\sum_{k>0,k\in\mathbb{Z}+\frac{1}{2}}(6k-1)(2k-1)(\psi_{-k}\psi_{k}^{*}-\psi_{-k}^{*}\psi_{k})\end{aligned}$$ Now the whole $H_{1}$ is $$\begin{aligned}
H_{1} & = & \frac{2}{3}(1-b^{2})\left(\sum_{\begin{subarray}{c}
r,s,k,l\in\mathbb{Z}+\frac{1}{2}\\
r+s<0,l+k>0\\
r+s+k+l=0
\end{subarray}}:\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:\left(\frac{1}{2}(s-r)+k-l\right)\right.\\
& & -\left.\frac{1}{16}\sum_{k>0,k\in\mathbb{Z}+\frac{1}{2}}(6k-1)(2k-1)(\psi_{-k}\psi_{k}^{*}-\psi_{-k}^{*}\psi_{k})\right)\nonumber\end{aligned}$$
Fermionization of double bosons term
------------------------------------
The last term in the Hamiltonian is the term $$H_{2}=(1-b^{2})\sum_{n}n\tilde{a}_{-n}\tilde{a}_{n}\,,$$ fermionization leads to an expression $$\begin{aligned}
H_{2} & = & (1-b^{2})\left(\sum_{r,k\in\mathbb{Z}+\frac{1}{2}}n:\psi_{k}\psi_{-n-k}^{*}\psi_{r}\psi_{n-r}^{*}:+\text{Contractions}\right)\label{H2}\\
& = & (1-b^{2})\left(\sum_{\begin{subarray}{c}
r+s<0,k+l>0\\
r+s+k+l=0\\
r,s,k,l\in\mathbb{Z}+\frac{1}{2}
\end{subarray}}(k+l):\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:+\text{Contractions}\right)\,.\nonumber\end{aligned}$$ The contractions in (\[H2\]) now can be calculated as $$\begin{aligned}
\text{Contractions} & = & \sum_{\begin{subarray}{c}
n>0\\
k,l>0
\end{subarray}}n\left((\psi_{-k}\psi_{k-n}^{*}-\psi_{-n-k}^{*}\psi_{k})(\psi_{-l}\psi_{n+l}^{*}-\psi_{n-l}^{*}\psi_{l})\right)\nonumber \\
& & -\sum_{r,k\in\mathbb{Z}+\frac{1}{2}}n:\psi_{k}\psi_{-n-k}^{*}\psi_{r}\psi_{n-r}^{*}:\nonumber \\
& = & \sum_{k>l>0}(k-l)\psi_{-k}\psi_{k}^{*}+\sum_{0<k<l}(l-k)\psi_{-l}^{*}\psi_{l}\nonumber \\
& = & \sum_{k>0}(\frac{(k-\frac{1}{2})(k+\frac{1}{2})}{2})(\psi_{-k}\psi_{k}^{*}+\psi_{-k}^{*}\psi_{k})\nonumber \\
& = & \sum_{k>0}\frac{1}{2}(k^{2}-\frac{1}{4})(\psi_{-k}\psi_{k}^{*}+\psi_{-k}^{*}\psi_{k})\,.\end{aligned}$$
Full expression
---------------
Combining all the expressions, we obtain the total $H$ $$\begin{aligned}
\label{eq:full_Hamiltonian}
H & = & H_{0}+H_{1}+H_{2}=\sum_{k>0,k\in\mathbb{Z}+\frac{1}{2}}\left(k^{2}+\frac{3}{4}\right)(\psi_{-k}\psi_{k}^{*}-\psi_{-k}^{*}\psi_{k})\\
& & +\frac{2}{3}(1-b^{2})\left(\sum_{\begin{subarray}{c}
r,s,k,l\in\mathbb{Z}+\frac{1}{2}\\
r+s<0,l+k>0\\
r+s+k+l=0
\end{subarray}}:\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:\left(\frac{1}{2}(s-r)+k-l\right)\right.\nonumber \\
& & -\left.\frac{1}{16}\sum_{k>0,k\in\mathbb{Z}+\frac{1}{2}}(6k-1)(2k-1)(\psi_{-k}\psi_{k}^{*}-\psi_{-k}^{*}\psi_{k})\right)\nonumber \\
& & +(1-b^{2})\left(\sum_{\begin{subarray}{c}
r+s<0,k+l>0\\
r+s+k+l=0\\
r,s,k,l\in\mathbb{Z}+\frac{1}{2}
\end{subarray}}(k+l):\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:\right.\nonumber \\
& & +\left.\sum_{k>0}\frac{1}{2}(k^{2}-\frac{1}{4})(\psi_{-k}\psi_{k}^{*}+\psi_{-k}^{*}\psi_{k})\right)\nonumber \\
& = & H_{0}+(1-b^{2})H_{d}+(1-b^{2})H_{t}\,.\nonumber\end{aligned}$$ Here we introduce $$\begin{aligned}
H_{d} & = & \sum_{k>0,k\in\mathbb{Z}+\frac{1}{2}}\frac{1}{3}(k-\frac{1}{2})\psi_{-k}\psi_{k}^{*}+(k-\frac{1}{2})(k+\frac{1}{6})\psi_{-k}^{*}\psi_{k}\\
& & +\sum_{\begin{subarray}{c}
k+l>0\\
k,l\in\mathbb{Z}+\frac{1}{2}
\end{subarray}}\frac{2}{3}(2k+l):\psi_{-l}\psi_{-k}^{*}\psi_{k}\psi_{l}^{*}:\,,\nonumber \\
H_{t} & = & \sum_{\begin{subarray}{c}
k+l>0\\
s,k,l\in\mathbb{Z}+\frac{1}{2}
\end{subarray}}(2k+\frac{2}{3}(s+l)):\psi_{-s-k-l}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:\end{aligned}$$
Fermionic Triangularization
---------------------------
### $H_{d}$ shifts eigen-energy
Now it is clear that $$H_0^\beta\equiv H_{0}+(1-b^{2})H_{d}$$ preserves the Schur state, but $H_{d}$ changes the eigen-energy. The aim here is to prove the eigen-energy $$E_{\lambda}^{\beta}=E_{\lambda}^{1}+(1-b^{2})\sum_{i}^{\ell(\lambda^{t})}(\lambda_{i}^{t})^{2}\, ,$$ where $$E_{\lambda}^{1}=\sum_{i}^{d(\lambda)}(\lambda_{i}-i+\frac{1}{2})^{2}-(\lambda_{i}^{t}-i+\frac{1}{2})^{2}$$ is the eigen-energy of free fermions excitation labeled by Young diagram $\lambda$. For simplicity, we define $$\begin{aligned}
n_{i}(\lambda)=\lambda_{i}-i+\frac{1}{2},\,\, m_{i}(\lambda)=\lambda_{i}^{t}-i+\frac{1}{2}\,.\end{aligned}$$
The eigen-energy of $H_{d}$ related to Schur state $\lambda$ is $$E_{\lambda}^{d}=\sum_{i}^{\lambda_{1}}(\lambda_{i}^{t})^{2}\,.$$ \[Theo:Hdenergy\]
![The (2,2,1,1) Young diagram[]{data-label="fig:Fig1"}](Fig1.eps "fig:"){width=".2\textwidth"}\
Before proving this theorem, we now consider an example $\lambda=\{2,2,1,1\}$ as shown in Fig. \[fig:Fig1\]. The corresponding Schur state is $$|\lambda\rangle=(-)\psi_{-3/2}\psi_{-7/2}^{*}\psi_{-1/2}\psi_{-1/2}^{*}|vac\rangle\,.$$ From the formalism of $H_{d}$, the energy eigenvalue is contributed by the following terms $$\begin{aligned}
& & \frac{2}{3}\left(\frac{3}{2}-1\right)\psi_{-3/2}\psi_{3/2}^{*}\,,\,\left((\frac{7}{2}-\frac{1}{2})(\frac{7}{2}+\frac{1}{6})\right)\psi_{-7/2}^{*}\psi_{7/2}\\
& & \frac{2}{3}\left\{ \left(7+\frac{3}{2}\right)\psi_{-3/2}\psi_{-7/2}^{*}\psi_{7/2}\psi_{3/2}^{*}+\left(1+\frac{1}{2}\right)\psi_{-1/2}\psi_{-1/2}^{*}\psi_{1/2}\psi_{-1/2}^{*}\right.\\
& & \hspace{0.5cm}+\left(7+\frac{1}{2}\right)\psi_{-1/2}\psi_{-7/2}^{*}\psi_{7/2}\psi_{1/2}^{*}+\left(1+\frac{3}{2}\right)\psi_{-3/2}\psi_{-1/2}^{*}\psi_{1/2}\psi_{3/2}^{*}\\
& & \left.\hspace{0.5cm}+\left(1-\frac{3}{2}\right)\psi_{-3/2}\psi_{-1/2}\psi_{-1/2}^{*}\psi_{3/2}^{*}-\left(7-\frac{1}{2}\right)\psi_{-1/2}^{*}\psi_{-7/2}^{*}\psi_{7/2}\psi_{1/2}\right\} \,.\end{aligned}$$ Combine all of them, we have the eigenvalue of $H_{d}$ $$E_{\lambda}^{d}=\frac{1}{3}+11+\frac{2}{3}(10+10-7)=20=16+4\,.$$ [ **Proof of Theorem \[Theo:Hdenergy\]**]{}
For a generic Schur state, the proof of Theorem \[Theo:Hdenergy\] is following. A generic Schur state is denoted by $$|\lambda\rangle=(-)^{\sum_{i}^{d(\lambda)}\left(m_{i}-\frac{1}{2}\right)}\prod_{i}^{d(\lambda)}\psi_{-n_{i}}\psi_{-m_{i}}^{*}|vac\rangle\,.$$ Acting on it, the first two terms of $H_{d}$ contribute $$\begin{aligned}
\sum_{i}^{d(\lambda)}\frac{1}{3}(n_{i}-\frac{1}{2})+\left(m_{i}-\frac{1}{2}\right)\left(m_{i}+\frac{1}{6}\right)\,.\end{aligned}$$ The third term contains $$\left(\begin{array}{c}
2d(\lambda)\\
2
\end{array}\right)=d(\lambda)(2d(\lambda)-1)$$ terms. We can have three independent picking strategies. Among them there are $d(\lambda)^{2}$ terms picked from a pair of $\psi$ and $\psi^{*}$ which we call $\psi\psi^{*}$-strategy. Other terms are from two fermionic modes picked from either the set of all $\psi$’s or $\psi^{*}$’s (the $\psi\psi$-strategy and $\psi^{*}\psi^{*}$-strategy).
We first consider the $\psi\psi^{*}$ -strategy. The contribution is $$E_{\psi\psi^{*}}=\frac{2}{3}\sum_{i,j=1}^{d(\lambda)}(2m_{i}+n_{j})\,.$$ The $\psi\psi$ strategy has a signature contribution (-1). Its contribution to the total energy is $$E_{\psi\psi}=\frac{2}{3}\sum_{i>j}^{d(\lambda)}(2n_{i}-n_{j})\,.$$ Similarly, the $\psi^{*}\psi^{*}$ -strategy gives rise to $$E_{\psi^{*}\psi^{*}}=\frac{2}{3}\sum_{i<j}^{d(\lambda)}(-2m_{i}+m_{j})\,.$$ The summation conditions $i>j$ or $i<j$ reflects the condition $k+l>0$. Thus we have translated the Theorem \[Theo:Hdenergy\] into the form that $$\begin{aligned}
& & \sum_{i}^{\ell(\lambda)}(\lambda_{i}^{t})^{2}=\sum_{i}^{d(\lambda)}\frac{1}{3}(n_{i}-\frac{1}{2})+\left(m_{i}-\frac{1}{2}\right)\left(m_{i}+\frac{1}{6}\right)\label{imptcombine}\\
& & +\frac{2}{3}\left(\sum_{i,j=1}^{d(\lambda)}(2m_{i}+n_{j})+\sum_{i>j}^{d(\lambda)}(2n_{i}-n_{j})+\sum_{i<j}^{d(\lambda)}(-2m_{i}+m_{j})\right)\,.\nonumber\end{aligned}$$ We use mathematical induction to prove it here. A direct proof can be found in Appendix.
Suppose the relation holds for a Young diagram $\lambda$, with $|\lambda|=n$. We will prove it holds for adding one more box to $\lambda$ under or right to the certain position $(i,j)$. These two different cases are following.
**Case I**: $\lambda_{j}^{t}\geq j$ $ \lambda^t_{j-1}> i$ and the box is attached just under the $(i,j)$ box, so that $$m_{j}\rightarrow m_{j}+1\,.$$
The change of R.H.S of (\[imptcombine\]) will be $$\begin{aligned}
\Delta_{RHS} & = & \left(2m_{j}+\frac{2}{3}\right)+\frac{2}{3}\left(2d(\lambda)+j-1-2(d(\lambda)-j)\right)\\
& = & 2m_{j}+2j=2\lambda_{j}^{t}+1=\Delta_{LHS}\,.\nonumber\end{aligned}$$
**Case II**: $\lambda_{i}\geq i$ $\lambda_{i-1}>j$, the box is attached to the right of the box $(i,j)$, so that $$n_{i}\rightarrow n_{i}+1\,.$$
The change of R.H.S of (\[imptcombine\]) will be $$\begin{aligned}
\Delta_{RHS} & = & \frac{1}{3}+\frac{2}{3}(3i-2)=2i-1\,.\end{aligned}$$ Notice that the $j+1$-th column has $\lambda_{j+1}^{t}=i-1$, so the change of L.H.S of (\[imptcombine\]) is $$\Delta_{L.H.S}=i^{2}-(i-1)^{2}=2i-1=\Delta_{R.H.S}\,.$$ Now we conclude that the identity (\[imptcombine\]) holds for any Young diagram and the Theorem \[Theo:Hdenergy\] is proved. Q.E.D.
### $H_{t}$ squeezes the state
As argued previously, the $H_t$ always squeezes the original Young diagram $\lambda$ for a given Schur state to some “thinner” $\lambda^{\prime}$’s. To be precise after its action $$\lambda^{\prime}<\lambda \quad \hbox{meaning} \quad \sum_{i=1}^{j}\lambda_{i}^{\prime}<\sum_{i=1}^{j}\lambda_{i},\,\text{for }\, j=1,2,\cdots\,.\label{squeeze}$$ To see this triangular property in a more transparent form, we first reduce the summation area from $r+s<0,k+l>0$ to $k+l>0,k+s<0,k>r$ and $k<r$. The $k<r$ case can be obtained from re-labeling the index ($r\leftrightarrow k$) and exchanging $\psi_{r}$ and $\psi_{k}$ $$\begin{aligned}
\frac{1}{1-b^{2}}H_{t} & = & \sum_{\begin{subarray}{c}
k+l>0, r+s+k+l=0\\
s,k,l\in\mathbb{Z}+\frac{1}{2}
\end{subarray}}\left(\frac{2}{3}(2k-r)\right):\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:\\
& = & \sum_{\begin{subarray}{c}
k+l>0,k+s<0\\
k>r,r+s+k+l=0
\end{subarray}}\left(\frac{2}{3}(2k-r-2r+k)\right):\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:\\
& = & \sum_{\begin{subarray}{c}
k+l>0,k+s<0\\
k>r,r+s+k+l=0
\end{subarray}}2(k-r):\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}: \, .\end{aligned}$$ $H_{t}$ is simplified and the summation area is decomposed into five cases $$\begin{aligned}
\frac{1}{1-b^{2}}H_{t} & = & 2\sum_{\begin{subarray}{c}
k+l>0,k+s<0\\
k>r,k+l+r+s=0
\end{subarray}}(k-r):\psi_{r}\psi_{s}^{*}\psi_{k}\psi_{l}^{*}:\label{eq:triangularH}\\
& = & 2\sum_{n=1,k>r>0}^{n=s-1/2}(k-r)\psi_{-k-n}^{*}\psi_{-r+n}^{*}\psi_{r}\psi_{k}\nonumber \\
& & +2\sum_{n=1,k>r>0}^{n=[(k-r-1)/2]}(r-k+2n)\psi_{-k+n}\psi_{-r-n}\psi_{r}^{*}\psi_{k}^{*}\nonumber \\
& & +2\sum_{n=1,k,r>0}^{n=r-1/2}(k+r-n)\psi_{-r+n}\psi_{-k-n}^{*}\psi_{k}\psi_{r}^{*}\\
& & +2\sum_{k>r>0,s>0}(r-k)\psi_{-r-s-k}^{*}\psi_{r}\psi_{k}\psi_{s}^{*}\nonumber \\
& & +2\sum_{k>r>0,s>0}(k-r)\psi_{-k}\psi_{-s}^{*}\psi_{-r}\psi_{r+s+k}^{*}\,.\nonumber\end{aligned}$$ Acting on a Schur state corresponding to a Young diagram $H_t$ gives rise to the following five processes.
**Case I:**The first line of the second equality in (\[eq:triangularH\]) annihilates two columns of the corresponding Young diagram of length $r$ and $k$ (k>r) and creates two new columns of length $k+n$ and $r-n$, which makes long column longer and short column shorter simultaneously. So it squeezes the Young diagram.
**Case II:** The second line annihilates two rows of length $r$ and $k$ (k>r) and generates two new rows of length $k-n$ and $r+n$, which makes long row shorter and short row longer, but not longer than the new shorter row, also, it squeezes the Young diagram.
**Case III:** The third line annihilates one row of length $k$ and one column $r$ and generates a shorter row (length $r-n$) and a longer column (length $k+n$).
**Case IV:** The fourth line annihilates two column of length $r$ and $k$ (k>r) and one row of length $s$ and generates a single column of length $r+k+s$ .
**Case V:** The fifth line annihilates a long row of length $r+k+s$ into three short columns of lengths $r$, $k$, $s$ respectively .
From the above analysis, we conclude that the $H_{t}$, when acting on a Schur state, will generate series of squeezed states. We call this property the fermionic triangularization.
Integrability
-------------
The triangular property means the eigenstate can be understood as $|P_{\lambda}(\psi\psi^{*})\rangle$, with $$P_{\lambda}=s_{\lambda}+\sum_{\mu<\lambda}c_{\lambda,\mu}s_{\mu}\,.$$ This ansatz is due to the eigenstate of bosonic $H$ is a function of bosonic oscillators $a_{-n}$’s. When $a_{-n}$ acts on a coherent basis, the eigenstate will be a power sum symmetric function $$\begin{aligned}
& & \tilde{a}_{-n}\exp\left(\sum_{i,n}b\frac{a_{n}}{-n}z_{i}^{n}\right)|0\rangle=\exp\left(\sum_{i,n}b\frac{a_{n}}{-n}z_{i}^{n}\right)\sum_{i}z_{i}^{n}|0\rangle\nonumber \\
& & \Rightarrow\tilde{a}_{-n}\simeq\sum_{i}z_{i}^{n}=p_{n}(z_{i})\,.\end{aligned}$$ It in turn determines the eigenstate itself as a symmetric function. The eigenvalue of $H$ is $$\begin{aligned}
H|P_{\lambda}(\psi\psi^{*})\rangle=E_{\lambda}^{\beta}|P_{\lambda}(\psi\psi^{*})\rangle\,.\label{eigeneqn}\end{aligned}$$ Let us consider the leading Schur function $s_{\lambda}$. Because $H_{t}$ changes basis, the eigenvalue should come from the action of $H_{0}^{\beta}$ on $$|s_{\lambda}\rangle=s_{\lambda}(\psi\psi^{*})|vac\rangle\,.$$
The equation (\[eigeneqn\]), has an explicit solution (up to a constant similarity transformation) that $$|P_{\lambda}\rangle \propto R(E)|s_{\lambda}\rangle\,,$$ where $$R(E) = \frac{1}{1-\frac{1-b^{2}}{E_{\lambda}^{\beta}-H_{0}^{\beta}}H_{t}}\,.\label{explicitsol}$$ This solution can be derived as follows. We first rewrite $H$ as $$H=E-(E-H_{0}^{\beta})\left(1-\frac{1-b^{2}}{E-H_{0}^{\beta}}H_{t}\right)\,\label{decompotrick}$$ then $$\begin{aligned}
H|P_{\lambda}\rangle & = & \left(E_{\lambda}^{\beta}-(E_{\lambda}^{\beta}-H_{0}^{\beta})\left(1-\frac{1-b^{2}}{E_{\lambda}^{\beta}-H_{0}^{\beta}}H_{t}\right)\right)R(E)|s_{\lambda}\rangle\\
& = & E_{\lambda}^{\beta}R(E)|s_{\lambda}\rangle-(E_{\lambda}^{\beta}-H_{0}^{\beta})|s_{\lambda}\rangle=E_{\lambda}^{\beta}|P_{\lambda}\rangle\,.\nonumber\end{aligned}$$
Notice that till now we have used the deformed bosonic oscillators $\tilde{a}$. It is not convenient when we consider the standard symmetric function formulae. We need to introduce a similarity transformation that transforms $\tilde{a}$’s back to $a$’s.\
$$D=\exp\bigl(-\log(b)(\tilde{q}\tilde{a}_{0}+\sum_{n>0}\tilde{a}_{-n}\tilde{a}_{n}/n)\bigr)\,,$$ or equivalently, we have a fermionic formalism $$D_{\pm}=b^{\pm\frac{1}{2}\sum_{r>0}(\psi_{-r}\psi_{r}^{*}+\psi_{-r}^{*}\psi_{r})}\,,$$ where $\pm$ means acting on bra(left) or ket(right) state respectively.
Laughlin state and its Hamiltonian
==================================
The Laughlin state is defined as $$\begin{aligned}
\Psi_{L}(\{z_{i}\})=\prod_{i<j}(z_{i}-z_{j})^{b^{2}}\exp\left(-\sum_{i}\frac{|z_{i}^{2}|}{4\ell}\right)\,,\end{aligned}$$ where $b^{-2}=\nu\in\mathbb{Z}$ is the filling fraction. In the formula $$\ell=\sqrt{\frac{\hbar}{eB}}$$ is the basic magnetic length scale. Usually it is normalized to be 1. $e$ stands for the electron charge and $B$ is the external magnetic field strength. The Gaussian factor $$\exp\left(-\sum_{i}\frac{|z_{i}^{2}|}{4\ell}\right)\,,$$ will be ignored later on since it will bring in excessive clutters.
The Laughlin wavefunction can be understood as follows. Consider there are $N$ quasi-particles containing in the interior of a disk of area $A_{N}=2N\pi b^{2}$. The ground state of this system is a correlation function of these $N$ free quasi-particles in a background magnetic field, which in general makes the total charge to be zero, that is, a neutral correlation function. We define the vertex operator for a quasi-particle $$\begin{aligned}
& & V_{b}(z)=e^{b\phi(z)}\,.\,\end{aligned}$$ The density $\rho_{0}$ of the $\phi$ field on the disk, for a ground state, should be uniform anywhere. Otherwise a density wave will be excited. It is simply $$\begin{aligned}
\rho_{0} & = & \frac{N}{2\pi Nb^{2}}=\frac{1}{2\pi b^{2}}\,.\end{aligned}$$ Thus the background charge is $$S=e^{-b\int d^{2}w\rho_{0}\phi(w)}\,.$$ Without this factor, the correlation function will vanish. Now the correlation function is written as $$\begin{aligned}
\langle0\prod_{i=1}^{N}V_{b}(z_{i})S|0\rangle & = & \Bigg\langle\prod_{i=1}^{N}e^{b\phi(z)}\exp\bigg(-b\int d^{2}w\rho_{0}\phi(w)\bigg)\Bigg\rangle\nonumber \\
& = & \prod_{i<j}^{N}(z_{i}-z_{j})^{b^{2}}\equiv\tilde{\Psi}_{L}(\{z_{i}\})\end{aligned}$$ Here we have defined the so-called reduced wave function $\tilde{\Psi}_{L}(\{z_{i}\})$ without the Gaussian factor. We now can consider the polynomial excitations of this ground state and moreover the behind integrability of this system.
From ground state to the Hamiltonian
------------------------------------
As we have done in the case of Calogero-Sutherland model, we now propose a Hamiltonian exactly has this Laughlin state as its ground state, that is, $$\begin{aligned}
H_{L} & = & \sum_{i}(\partial_{i}+\partial_{i}(\ln\tilde{\Psi}\{z_{i}\}))(\partial_{i}-\partial_{i} (\ln\tilde{\Psi}\{z_{i}\})).\end{aligned}$$ Here $\partial_{i}=\partial/\partial x_{i}\equiv z_{i}\partial_{z_{i}}$, or equivalently, $z_{i}=e^{x_{i}}$. By separating out the ground state contribution as follows, we have
$$\begin{aligned}
\tilde{{H}} & = & \tilde{(\Psi}\{z_{i}\})^{-1}H_{L}\tilde{\Psi}\{z_{i}\}\nonumber \\
& = & \sum_{i}\left(z_{i}\partial_{z_{i}}+2z_{i}\partial_{z_{i}}(\ln\tilde{\Psi}\{z_{i}\})z_{i}\partial_{z_{i}}\right)\label{eq:Htilde}\\
& = & \sum_{i}(z_{i}\partial_{z_{i}})^{2}-2b^{2}\sum_{i<j}\frac{z_{i}^{2}\partial_{z_{i}}}{z_{i}-z_{j}}\nonumber \\
& = & \sum_{i}(z_{i}\partial_{z_{i}})^{2}-b^{2}\sum_{i<j}\frac{z_{i}^{2}\partial_{z_{i}}-z_{j}^{2}\partial_{z_{j}}}{z_{i}-z_{j}}\,.\nonumber\end{aligned}$$
When acting on the normal ordered vertex operators, we have $$\begin{aligned}
z_{i}\partial_{z_{i}}\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle & = & b\sum_{n}a_{-n}z_{i}^{n}\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle\,,\nonumber \\
(z_{i}\partial_{z_{i}})^{2}\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle & = & b^{2}\sum_{n,m}a_{-n}a_{-m}z_{i}^{n+m}\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle\nonumber \\
& + & b\sum_{n}na_{-n}z_{i}^{n}\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle\,,\nonumber \\
a_{k}\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle & = & \sum_{i}bz_{i}^{n}\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle\,,\label{eq:coherent}\\
\frac{z_{i}^{2}\partial_{z_{i}}-z_{j}^{2}\partial_{z_{j}}}{z_{i}-z_{j}}\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle\nonumber \\
=\frac{b\sum_{n}a_{-n}\left(z_{i}^{n+1}-z_{j}^{n+1}\right)}{z_{i}-z_{j}} & & \prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle\nonumber \\
=b\sum_{n}a_{-n}(z_{i}^{n}+z_{i}^{n-1}z_{j}+ & \cdots & +z_{i}z_{j}^{n-1}+z_{j}^{n})\prod_{j}^{N}\exp\left(b\sum_{n,j}\frac{a_{-n}}{n}z_{j}^{n}\right)|0\rangle\,,\label{eq:interactionterm}\end{aligned}$$ and substituting ($\ref{eq:coherent}$) into ($\ref{eq:Htilde}$), we have the free part of ($\ref{eq:Htilde}$) as $$\tilde{{H}_{0}}=\sum_{n>0}na_{-n}a_{n}+b\sum_{n,m>0}a_{-n}a_{-m}a_{n+m}$$ We need to be careful with the interaction term ($\ref{eq:interactionterm}$). The summation on $i,\, j$ is not arbitrary. To turn this term into bosonic operator formalism, we have to do a trick as follow. Firstly, we change the summation from $i<j$ to $i\neq j$. It gives rise to a simple $\frac{1}{2}$ factor. Secondly, we add $i=j$ terms into the summation and then finally subtract these terms. Follow these steps, we have $$\begin{aligned}
b^{3}\sum_{i<j}\,\sum_{n>0}a_{-n}(z_{i}^{n}+z_{i}^{n-1}z_{j}+&\cdots&+z_{i}z_{j}^{n-1}+z_{j}^{n}) = \frac{b}{2}\sum_{n,m>0}a_{-n-m}a_{n}a_{m}\nonumber \\
-\frac{1}{2}b^{2}\sum_{n>0}na_{-n}a_{n} & + & b^{2}(N-1)\sum_{n>0}a_{-n}a_{n\,,}\label{eq:interoper}\end{aligned}$$ The last term of ($\ref{eq:interoper}$) is an irrelevant total energy of free quasi-particles. When acting on a level $n$ state, it gives rise to an eigen-energy $$E_{ir}|n\rangle=b^{2}(N-1)n|n\rangle\,$$ as expected which describes a system of $N-1$ copies of non-interacting bosonic oscillators with the frequency $b^{2}$. Therefore we ignore its contribution and now we obtain the operator formalism of $H_{L}$
$$H_{L}=\sum_{n}\left(1-\frac{b^{2}}{2}\right)na_{-n}a_{n}+\sum_{n,m}b\left(a_{-n}a_{-m}a_{n+m}+\frac{1}{2}a_{-n-m}a_{n}a_{m}\right)\,.\label{eq:HLaughlin}$$
Fermionization of $H_{L}$
-------------------------
Define the deformed bosonic modes as $$\begin{aligned}
2b^{-1}\tilde{a}_{n} & = & a_{n}\, , \quad \frac{1}{2}b\tilde{a}_{-n}=a_{-n\,,}\label{eq:deformLau}\end{aligned}$$ so that the triple-$a$ terms in ($\ref{eq:HLaughlin}$) now becomes $$\begin{aligned}
&&\sum_{n,m>0}\left(\frac{1}{2}b^{2}\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}+\tilde{a}_{-n-m}\tilde{a}_{n}\tilde{a}_{m}\right)\\
&=&\sum_{n,m>0}\left(\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}+\tilde{a}_{-n-m}\tilde{a}_{n}\tilde{a}_{m}\right) - (1-\frac{b^{2}}{2})\tilde{a}_{-n}\tilde{a}_{-m}\tilde{a}_{n+m}\,.\end{aligned}$$ It is straightforward to write down the fermionic formalism as the case of the Calogero-Sutherland model in previous section. Hence we get $$H_{L}=H_{0}+(1-\frac{b^{2}}{2})H_{d}+(1-\frac{b^{2}}{2})H_{t}\equiv H_{l}+(1-\frac{b^{2}}{2})H_{t}\,.\label{eq:LaughlinHamilton}$$ The eigen-energy for this Hamiltonian is $$E_{\lambda}^{Lau}=\sum_{i=1}^{\lambda_{1}^{t}}(\lambda_{i})^{2}-\frac{b^{2}}{2}\sum_{i=1}^{\lambda_{1}}(\lambda_{i}^{t})^{2}\,.\label{eq:eigenLau}$$ The corresponding excitation state is $$|P_{\lambda}^{Lau}\rangle=D^{Lau}\frac{1}{1-\frac{1-b^{2}/2}{E_{\lambda}^{Lau}-H_{l}}H_{t}}s_{\lambda}(\psi\psi^{*})|vac\rangle\,,$$ where $D^{Lau}$ is the similarity transformation related to ($\ref{eq:deformLau}$). $$D^{Lau}=\exp\bigl(-\log(b/2)(\tilde{q}\tilde{a}_{0}+\sum_{n>0}\tilde{a}_{-n}\tilde{a}_{n}/n)\bigr)\,,$$ and the fermionic expression: $$D_{\pm}^{Lau}=\left(\frac{b}{2}\right)^{\pm\sum_{r>0}(\psi_{-r}\psi_{r}^{*}+\psi_{-r}^{*}\psi_{r})}\,.$$ Notice that the polynomial is not a new polynomial, it is a Jack polynomial with the parameter $\beta=\frac{b^{2}}{2}\,.$
Since the integrability of the Laughlin theory is exactly the same as that of CS model we ignore its analysis here.
Halperin State and Two-layer System
===================================
Now we can consider the two-layer system. The corresponding ground state wavefunction is the Halperin state, which reads $$\tilde{\Psi}_{H}\left({z_{i}},{w_{j}}\right)=\prod_{i<j}^{N}(z_{i}-z_{j})^{p}\prod_{m<n}^{M}(w_{m}-w_{n})^{q}\prod_{i,m}^{N,M}(z_{i}-w_{m})^{r}\,.\label{eq:Halperin}$$ The complexity of this wave-function lies in that it involves an interaction between two layers. Let us first define the bosonic fields $$\begin{aligned}
\phi^{1}(z) & = & q_{0}^{1}+a_{0}^{1}\ln z+\sum_{n\neq0}\frac{a_{-n}^{1}}{n}z^{n}\,,\\
\phi^{2}(w) & = & q_{0}^{2}+a_{0}^{2}\ln w+\sum_{n\neq0}\frac{a_{-n}^{2}}{n}w^{n}\,,\end{aligned}$$ with commutation relation $$[a_{n}^{I}\,,\, a_{m}^{J}]=n\delta^{IJ}\delta_{n+m,0}\,,\, n,m\in\mathbb{Z},\, I,J\in{1,2}\,.$$ The related coordinate system is defined as $$\begin{aligned}
Z_{I} & = & z_{I},\,\mbox{if}\,\ I\leq N\,,\\
Z_{I} & = & w_{I-N}\,,\, if\ \, N<I\leq N+M\,,\\
\partial_{I} & = & Z_{I}\partial_{Z_{I}}\,,\end{aligned}$$ we can now write down the Hamiltonian of this system as $$\begin{aligned}
H^{Hal} & = & \sum_{I}(\partial_{I}-2\partial_{I}(\ln\tilde{\Psi}_{H}(Z_{I}))\partial_{I}\\
& = & \sum_{i=1}^{N}\left(z_{i}\partial_{z_{i}}-2z_{i}\partial_{z_{i}}\left(\ln\prod_{i<j}(z_{i}-z_{j})^{p}\right)\right.\\
& & +\left.2z_{i}\partial_{z_{i}}\left(\ln\prod_{i,m}(z_{i}-w_{m})^{r}\right)\right)z_{i}\partial_{z_{i}}\\
& + & \sum_{m=1}^{M}\left(w_{m}\partial_{w_{m}}-2w_{m}\partial_{w_{m}}\left(\ln\prod_{m<n}(w_{m}-w_{n})^{q}\right)\right.\\
& & +\left.2w_{m}\partial_{w_{m}}\left(\ln\prod_{i,m}(z_{i}-w_{m})^{r}\right)\right)w_{m}\partial_{w_{m}}\\
& = & H_{L}(p)+H_{L}(q)+2r\sum_{m,i}\frac{z_{i}^{2}\partial_{z_{i}}}{z_{i}-w_{m}}-2r\sum_{m,i}\frac{w_{m}^{2}\partial_{w_{m}}}{z_{i}-w_{m}}\,,\\
& = & H_{L}(p)+H_{L}(q)+2r\sum_{n\geq0,m,i}\left(\frac{w_{m}}{z_{i}}\right)^{n}z_{i}\partial_{z_{i}}-2r\sum_{n>0,m,i}\left(\frac{w_{m}}{z_{i}}\right)^{n}w_{m}\partial_{w_{m}}\,\,,\end{aligned}$$ where $H_{L}(p)$ is the Laughlin Hamiltonian defined before in which $b^{2}\rightarrow p\,.$ The vertex operators in this system are $$V_{\sqrt{p}}^{1}(z)=e^{\sqrt{p}\phi^{1}(z)}\,,\: V_{\sqrt{q}}^{2}(w)=e^{\sqrt{q}\phi^{2}(w)}\,.$$ The differential operators relate to operators and power-sum polynomials are defined as follows $$\begin{aligned}
z_{i}\partial_{z_{i}}\prod_{i=1}^{N}V_{\sqrt{p}}^{1,-}(z_{i}) & = & \sum_{n>0}\sqrt{p}a_{-n}^{1}z_{i}^{n}\prod_{i=1}^{N}V_{\sqrt{p}}^{1,-}(z_{i})\,,\\
w_{m}\partial_{w_{m}}\prod_{j=1}^{M}V_{\sqrt{q}}^{2,-}(w_{j}) & = & \sum_{n>0}\sqrt{q}a_{-n}^{2}w_{j}^{n}\prod_{n=1}^{M}V_{\sqrt{q}}^{2,-}(w_{j})\,,\\
a_{n}^{1}\prod_{i=1}^{N}V_{\sqrt{p}}^{1,-}(z_{i})|0\rangle & = & \sqrt{p}\sum_{i}z_{i}^{n}\prod_{i=1}^{N}V_{\sqrt{p}}^{1,-}(z_{i})|0\rangle\\
a_{n}^{2}\prod_{j=1}^{N}V_{\sqrt{p}}^{2,-}(w_{j})|0\rangle & = & \sqrt{q}\sum_{i}w_{j}^{n}\prod_{i=1}^{N}V_{\sqrt{p}}^{1,-}(w_{j})|0\rangle\end{aligned}$$ Therefore we have $$\begin{aligned}
2r\sum_{n\geq0,m,i}\left(\frac{w_{m}}{z_{i}}\right)^{n}z_{i}\partial_{z_{i}} & \rightarrow & 2r\frac{1}{\sqrt{q}}\sum_{\begin{array}[t]{c}
n\geq0\\
m>0
\end{array}}a_{-n}^{1}a_{n-m}^{1}a_{m}^{2}\,,\\
2r\sum_{n>0,m,i}\left(\frac{w_{m}}{z_{i}}\right)^{n}w_{m}\partial_{w_{m}} & \rightarrow & 2r\frac{1}{\sqrt{p}}\sum_{\begin{array}[t]{c}
n>0\\
m>0
\end{array}}a_{-m}^{1}a_{-n}^{2}a_{n+m}^{2}\,.\end{aligned}$$ It is now easy to write down the bosonic operator formalism, we have $$\begin{aligned}
H^{Hal} & = & H_{L}(p,a^{1})+H_{L}(q,a^{2})+H_{int}\,,\\
H_{int} & = & 2r\left(\frac{1}{\sqrt{q}}\sum_{\begin{array}[t]{c}
n\geq0\\
m>0
\end{array}}a_{-n}^{1}a_{n-m}^{1}a_{m}^{2}-\frac{1}{\sqrt{p}}\sum_{\begin{array}[t]{c}
n>0\\
m>0
\end{array}}a_{-m}^{1}a_{-n}^{2}a_{n+m}^{2}\right)\\
& = & r\left(\frac{1}{\sqrt{q}}\sum_{m>0}L_{-m}^{1}a_{m}^{2}+\sum_{n,m>0}(\frac{1}{\sqrt{q}}a_{-n}^{1}-\frac{2}{\sqrt{p}}a_{-n}^{2})a_{-m}^{1}a_{n+m}^{2}\right)\,.\end{aligned}$$ Notice that the $H_{int}$ is a fascinating interaction. It is always a triangular term such that it subtracts boxes in Young diagram $\mu$ on layer-2 and adds the same number of boxes into Young diagram $\lambda$ on layer-1. The Hilbert space of this Hamiltonian is expanded by coupled bi-Jack functions, that is, $$|\Omega_{\lambda,\mu}^{0}\rangle=|P_{\lambda}^{1}(p/2)\rangle\otimes|P_{\mu}^{2}(q/2)\rangle\,.$$ It is not an eigenstate of the Halperin Hamiltonian, but only a highest weight state of the eigenstate, which can be obtained as following formula $$|\Omega_{\lambda,\mu}^{r}\rangle=\frac{1}{1-\frac{1}{E_{\lambda,\mu}-H_{L}(p)-H_{L}(q)}H_{int}}|\Omega_{\lambda,\mu}^{0}\rangle\,,$$ where the eigen-energy is written as $$E_{\lambda,\mu}=\sum_{i}^{\lambda_{1}^{t}}(\lambda_{i})^{2}+\sum_{j}^{\mu_{1}^{t}}(\mu_{j})^{2}-\frac{p}{2}\sum_{k}^{\lambda_{1}}(\lambda_{k}^{t})^{2}-\frac{q}{2}\sum_{l}^{\mu_{1}}(\mu_{k}^{t})^{2}\,.$$ The fermionization of the Hamiltonian as in ($\ref{eq:Halperin}$) is not hard. Let us first consider the first term in $H_{int}$ . It is $$\begin{aligned}
H_{int}^{1} & = & \frac{2r}{\sqrt{q}}\sum_{\begin{array}[t]{c}
n\geq0\\
m>0
\end{array}}a_{-n}^{1}a_{n-m}^{1}a_{m}^{2}\\
& = & \frac{2r}{\sqrt{q}}\left(\sum_{\begin{array}[t]{c}
r,s,k,l,u\in\mathbb{Z}+\frac{1}{2}\\
r+s+k+l=-m<0
\end{array}}:\psi_{r}^{1}\psi_{s}^{1*}\psi_{k}^{1}\psi_{l}^{1*}\psi_{-u}^{2}\psi_{m+u}^{2*}:\right.\\
& & \hspace{1em}\hspace{1em}+\left.\sum_{u\in\mathbb{Z}+\frac{1}{2},m>0}:\psi_{-u}^{2}\psi_{m+u}^{2*}:(\text{Contractions})\right)\,,\end{aligned}$$ where the contractions is calculated as $$\begin{aligned}
\text{Contractions} & = & \sum_{\begin{array}[t]{c}
\begin{array}[t]{c}
n,r>0\end{array}\end{array}}:(\psi_{-r}^{1}\psi_{-n+r}^{1*}-\psi_{-r-n}^{1*}\psi_{r}^{1})(\psi_{-s}^{1}\psi_{n-m+s}^{1*}-\psi_{n-m-s}^{1*}\psi_{s}^{1}):\,,\\
& = & \sum_{r>0}(\psi_{-r}^{1}\psi_{r-m}^{1*}+\psi_{-m-s}^{1*}\psi_{s}^{1})\,.\end{aligned}$$ Therefore $$\begin{aligned}
H_{int}^{1} & = & \frac{2r}{\sqrt{q}}\left(\sum_{\begin{array}[t]{c}
r,s,k,l,u\in\mathbb{Z}+\frac{1}{2}\\
r+s+k+l=-m<0
\end{array}}:\psi_{r}^{1}\psi_{s}^{1*}\psi_{k}^{1}\psi_{l}^{1*}\psi_{-u}^{2}\psi_{m+u}^{2*}:\right.\\
& & \hspace{1em}\hspace{1em}+\left.\sum_{\begin{array}[t]{c}
u>0\\
r,m>0
\end{array}}(\psi_{-u}^{2}\psi_{m+u}^{2*}-\psi_{m-u}^{2*}\psi_{u}^{2})(\psi_{-r}^{1}\psi_{r-m}^{1*}+\psi_{-m-r}^{1*}\psi_{r}^{1})\right)\,.\end{aligned}$$ Similarly, we have $$\begin{aligned}
H_{int}^{2} & = & -\frac{2r}{\sqrt{p}}\sum_{\begin{array}[t]{c}
n,m>0\end{array}}a_{-m}^{1}a_{-n}^{2}a_{n+m}^{2}\\
& = & -\frac{2r}{\sqrt{p}}\left(\sum_{\begin{array}[t]{c}
r,s,k,l,u\in\mathbb{Z}+\frac{1}{2}\\
r+s+k+l=m>0
\end{array}}:\psi_{r}^{2}\psi_{s}^{2*}\psi_{k}^{2}\psi_{l}^{2*}\psi_{-u}^{1}\psi_{-m+u}^{1*}:\right.\\
& & \hspace{1em}\hspace{1em}+\left.\sum_{u\in\mathbb{Z}+\frac{1}{2},m>0}:\psi_{-u}^{1}\psi_{-m+u}^{1*}:(\text{Contractions})\right)\\
& & =-\frac{2r}{\sqrt{p}}\left(\sum_{\begin{array}[t]{c}
r,s,k,l,u\in\mathbb{Z}+\frac{1}{2}\\
r+s+k+l=m>0
\end{array}}:\psi_{r}^{2}\psi_{s}^{2*}\psi_{k}^{2}\psi_{l}^{2*}\psi_{-u}^{1}\psi_{-m+u}^{1*}:\right.\\
& & \hspace{1em}\hspace{1em}+\left.\sum_{\begin{array}[t]{c}
u,m>0\\
r>0
\end{array}}(\psi_{-u}^{1}\psi_{-m+u}^{1*}-\psi_{-m-u}^{1*}\psi_{u}^{1})(\psi_{-r}^{2}\psi_{r+m}^{2*}+\psi_{m-r}^{2*}\psi_{r}^{2})\right)\,.\end{aligned}$$ Though this fermionic formalism is not effective in calculating the eigenstate, it plays an important role in deriving the behind tau-function and its Hirota integrability of this system. Actually, the fermionic formalism completely defines the fermionic orbit of a generator of $GL(\infty)$ which in turn determine the tau-function of this theory. We are working in detail in this direction.
Similarity transformation
-------------------------
As explained before, we need to do a similarity transformation to recover the deformed operator formalism of eigenstates to a standard operator formalism. It is easy to write down this similarity transformation, that is $$D^{Hal}=D^{Lau}(b^{2}\rightarrow p,\, a\rightarrow a^{1})D^{Lau}(b^{2}\rightarrow q,\, a\rightarrow a^{2})\,.$$
Conclusions and Future Works
============================
In conclusion, we introduce a systematic way to extract the integrability of several models. For CS model, we express the Hamiltonian in bosonic as well as fermionic representations. The eigenstate and eigenvalue are obtained explicitly. The construction of Jack state, in the fermionic representation, is highly involved in the fermionic triangularization of fermionic Hamiltonian of the CS model. For Laughlin and Halperin states, which can be seen as solitonic wavefunction, we construct the corresponding Hamiltonians in the same manner as CS model. We obtain their bosonic and fermionic representations. The explicit solutions, e.g. excitations and eigenvalues, are exactly solved. The integrability of Laughlin state, is the same as that of CS model, while for Halperin state, the integrability is determined also by the triangularization nature of the Hamiltonian.
There are several problems worthy of exploring in the future. Firstly, though the integrability in this article are inherited from free fermions, it is not clear to the authors that how to construct the integrable hierarchies. In soliton theory, the integrable hierarchy can be determined by the Lax operators. The Lax method is not the expected one for solving the problem since there are in general integration operators (the pseudo-differential operators) additional to usual differential operators. A possible solution may be the inverse scattering method, which will relate the integrable hierarchy to the inverse scattering equation[@das1989integrable]. This hierarchy tells us how integral of motions can be constructed by a recursive relation. A higher level integral of motion determines a refiner excitation structure of the model[@jimbo1983solitons]. Secondly, for FQHEs, people believe special Jack polynomial may be related to certain FQHE wavefunction. It still remains mysterious to us what kind of constraint leads to a truncation of the fusion rule of Jack polynomials. Thirdly, we expect a direct generalization of our analysis to Haldane-Shastry model[@haldane1988exact; @shastry1988exact; @talstra1995integrals], or the spin CS model, we are working on that.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors are grateful to Morningside Center of Chinese Academy of Sciences and Kavli Institute for Theoretical Physics China at the Chinese Academy of Sciences for providing excellent research environment and financial support to our seminar in mathematical physics. Currently Jie Yang is a visiting scholar in the Department of Physics at University of California, Berkeley and is grateful for the hospitality during the 2014-2015 academic year. The project is partially supported by National Natural Science Foundation of China (No. 11401400), Returned Oversea Students Fund in Beijing, and Specialized Research Fund for the Doctoral Program of Higher Education (No. 20121108120005).
Appendix {#appendix .unnumbered}
========
We will denote a partition by its parts $(\lambda_1, \lambda_2, \cdots, \lambda_{\lambda^t_1})$ and the Frobenius notation $(\alpha_1, \cdots, \alpha_d|\beta_1, \cdots, \beta_d) $ as well where $d$ is the diagonal length of $\lambda$. With this notation $n_i$ and $m_i$ in theorem 4 are related with $\alpha_i$ and $\beta_i$ in such a way that $n_i=\alpha_i+1/2$ and $m_i=\beta_i+1/2$. Theorem 4 is then written as $$\begin{aligned}
\sum_{i=1}^{\lambda_1}(\lambda^t_i)^2&=& \sum_{i=1}^d\left[\frac{\alpha_i}{3}+\beta_i^2+\frac{2\beta_i}{3}\right]\\
&&\quad +\frac{2}{3}d\left[\sum_{i=1}^d(\alpha_i+2\beta_i+\frac{3}{2})
+\sum_{i=1}^d2\alpha_i(i-1)\right.\\&&\left.-\sum_{i=1}^d(d-i)\alpha_i+\sum_{i=1}^d\beta_i(i-1)-2\sum_{i=1}^d\beta_i(d-i)\right]\\
&=&\sum_{i=1}^d(\beta_i^2+2i\beta_i+2i\alpha_i-\alpha_i)+d^2.\end{aligned}$$ To prove this theorem we need two preliminary steps.
Step 1: $$\label{eq:kappa}
2[n(\lambda^t)-n(\lambda)]=\sum_{i=1}^d \alpha_i(\alpha_i+1)-\beta_i(\beta_i+1).$$ It is obvious to prove the first step by using two very useful identities among several multi-number sets, namely $$\{\beta_i, (i\leq d)\}=\{0, 1, \cdots, \lambda_1^t-1\}-\{i-\lambda_i-1, (d+1\leq i\leq \lambda_1^t) \},$$ and similarly $$\{\alpha_i, (i\leq d)\}=\{0, 1, \cdots, \lambda_1-1\}-\{i-\lambda^t_i-1, (d+1\leq i\leq \lambda_1)\}.$$ These two identities are also very useful in proving step 2. Now let us compute $$\begin{aligned}
2[n(\lambda^t)-n(\lambda)]&=&\sum_{i=1}^{\lambda_1^t} \lambda_i(\lambda_i-2i+1)\\
&=& \sum_{i=1}^{\lambda_1^t} (\lambda_i-i)^2+|\lambda|-\sum_{i=1}^{\lambda_1^t}i^2\\
&=& \sum_{i=1}^d \alpha_i^2 +\sum_{i=d+1}^{\lambda_1^t} (-\lambda_i+i-1+1)^2 +|\lambda|
-\sum_{i=1}^{\lambda_1^t}i^2.\end{aligned}$$ We apply one of the two identities. Therefore one term becomes $$\sum_{i=d+1}^{\lambda_1^t} (-\lambda_i+i-1+1)^2 =\sum_{i=0}^{\lambda^t_1-1}(i+1)^2-\sum_{i=1}^d (\beta_i+1)^2$$ and since $$|\lambda|=\sum_{i=1}^d(\alpha_i+\beta_i+1),$$ combining all terms together we obtain the result in (\[eq:kappa\]).
Step 2:
$$\label{eq:hook}
2[n(\lambda^t)+n(\lambda)]=\sum_{i=1}^d (\alpha_i^2+\beta_i^2+4 i\alpha_i+4 i\beta_i -3\alpha_i-3\beta_i)+2d(d-1).$$
To prove (\[eq:hook\]), we recall a formula in Macdonald’s book [@macdonald1995symmetric] $$\sum_{x\in\lambda}h(x)=\sum_{(i, j)\in\lambda}(\lambda_i-i+\lambda^t_j-j+1)=n(\lambda^t)+n(\lambda)+|\lambda|.$$ Hence $$2[n(\lambda^t)+n(\lambda)]=2\sum_{(i, j)\in\lambda} (\lambda_i -i +\lambda^t_j-j).$$ Now let us compute $$\begin{aligned}
\sum_{(i, j)\in\lambda}\lambda_i-i+\lambda_j^t-j&=&\sum_{i=1}^{\lambda^t_1}\sum_{j=1}^{\lambda_i} \lambda_i-i+\lambda_j^t-j \\
&=& \left[\sum_{i=1}^d\sum_{j=1}^d+\sum_{i=1}^d\sum_{j=d+1}^{\lambda_i}+\sum_{j=1}^d\sum_{i=d+1}^{\lambda^t_j}\right] (\lambda_i-i+\lambda_j^t-j).\end{aligned}$$ The regions of the summation are shown in Fig. \[fig:regions\].
![The three regions of a Young diagram are separated by solid lines, thick lines, and dashed lines. []{data-label="fig:regions"}](Young_923.eps){width="50.00000%"}
The square region surrounded by solid lines is simply $$\sum_{i=1}^d\sum_{j=1}^d (\lambda_i-i+\lambda_j^t-j)
=d\sum_{i=1}^d (\alpha_i+\beta_i).$$ Now let us compute the sum for the region surrounded by the thick lines. $$\sum_{i=1}^d\sum_{j=d+1}^{\lambda_i}(\lambda_i-i)=\sum_{i=1}^d(\lambda_i-i)(\lambda_i-d)=\sum_{i=1}^d(\lambda_i^2-i\lambda_i-d\lambda_i+id).$$ We obtain $$\sum_{i=1}^d\sum_{j=d+1}^{\lambda_i}(\lambda_j^t-j+1-1)
=-\sum_{i=1}^d\left[\sum_{j=0}^{\lambda_i-i} j-(\sum_{j=i}^d\alpha_j)\right]-\sum_{i=1}^d(\lambda_i-d)$$ where we have used $$\{j-\lambda_j^t-1, (d+1\leq j\leq \lambda_i)\}=\{0, 1, 2, \cdots, \lambda_i-i\}-\{\alpha_j, (i\leq j\leq d)\}.$$ Therefore $$\begin{aligned}
\sum_{i=1}^d\sum_{j=d+1}^{\lambda_i}(\lambda_j^t-j+1-1)&=&-\sum_{i=1}^d\frac{(\lambda_i-i)(\lambda_i-i+1)}{2}+\sum_{i=1}^d i\alpha_i-\sum_{i=1}^d(\lambda_i-d)\\
&=&\sum_{i=1}^d\left[-\frac{\lambda_i^2}{2}+i\lambda_i-\frac{i(i-1)}{2}-\frac{\lambda_i}{2}+i\alpha_i-(\lambda_i-i+i-d)\right].\end{aligned}$$ Combining them we get $$\begin{aligned}
\sum_{i=1}^d\sum_{j=d+1}^{\lambda_i} (\lambda_i-i+\lambda_j^t-j) &=&\sum_{i=1}^d \left[\frac{(\lambda_i-i)^2}{2}+i(\lambda_i-i)+i\alpha_i-d(\lambda_i-i)-\frac{3}{2}(\lambda_i-i)-i+d\right]\\
&=&\sum_{i=1}^d\left[\frac{\alpha_i^2}{2}+2i\alpha_i-d\alpha_i-\frac{3\alpha_i}{2}+d-i \right].\end{aligned}$$ Similarly we can compute the region surrounded by the dashed lines and the result is just replacing $\alpha$ with $\beta$ in above formula. Hence we have $$\sum_{(i, j)\in\lambda} \lambda_i -i +\lambda^t_j-j= \sum_{i=1}^d\left[ \frac{\alpha_i^2}{2}+\frac{\beta_i^2}{2}+2i \alpha_i+2i\beta_i-\frac{3\alpha_i}{2}-\frac{3\beta_i}{2}+2(d-i)\right]$$ Therefore twice of it will give rise to (\[eq:hook\]).
Now let us compute $$4n(\lambda)=2[n(\lambda^t)+n(\lambda)]-2[n(\lambda^t)-n(\lambda)]=\sum_{i=1}^d(2\beta_i^2+4i\alpha_i+4i\beta_i-4\alpha_i-2\beta_i)+2d(d-1).$$ Since $$4n(\lambda)=2\sum_{i=1}^{\lambda_1}\lambda^t_i(\lambda^t_i-1)=2\sum_{i=1}^{\lambda_1}(\lambda^t_i)^2-2\sum_{i=1}^d\alpha_i-2\sum_{i=1}^d\beta_i-2d,$$ We obtain $$\sum_{i=1}^{\lambda_1}(\lambda^t_i)^2=\sum_{i=1}^d(\beta_i^2+2i\alpha_i-\alpha_i+2i\beta_i )+d^2.$$
References {#references .unnumbered}
==========
[10]{} H. Azuma and S. Iso. Explicit relation of the quantum hall effect and the calogero-sutherland model. , 331(1):107–113, 1994.
M. C. Bergere. Composite particles and the eigenstates of calogero–sutherland and ruijsenaars–schneider. , 41(11):7234–7251, 2000.
E. J. Bergholtz, T. H. Hansson, M. Hermanns, A. Karlhede, and S. Viefers. Quantum hall hierarchy wave functions: From conformal correlators to tao-thouless states. , 77(16):165325, 2008.
B. A. Bernevig and F. D. M. Haldane. Generalized clustering conditions of jack polynomials at negative jack parameter $\alpha$. , 77(18):184502, 2008.
B. A. Bernevig and F. D. M. Haldane. Model fractional quantum hall states and jack polynomials. , 100(24):246802, 2008.
B. A. Bernevig and F. D. M. Haldane. Properties of non-abelian fractional quantum hall states at filling $\nu$= k/r. , 101(24):246806, 2008.
A. Das. , volume 30. World Scientific, 1989.
B. Estienne, N. Regnault, and R. Santachiara. Clustering properties, jack polynomials and unitary conformal field theories. , 824(3):539–562, 2010.
B. Estienne, B. A. Bernevig, and R. Santachiara. Electron-quasihole duality and second-order differential equation for read-rezayi and jack wave functions. , 82(20):205307, 2010.
B. Estienne, V. Pasquier, R. Santachiara, and D. Serban. Conformal blocks in virasoro and w theories: duality and the calogero–sutherland model. , 860(3):377–420, 2012.
B. Estienne and R. Santachiara. Relating jack wavefunctions to$\backslash$ textrm $\{$WA$\}$ \_ $\{$k-1$\}$ theories. , 42(44):445209, 2009.
B. Feigin, M. Jimbo, T. Miwa, and Eugene Mukhin. A differential ideal of symmetric polynomials spanned by jack polynomials at r$\beta$=-(r= 1)/(k+ 1). , 2002(23):1223–1237, 2002.
V. Gurarie, M. Flohr, and C. Nayak. The haldane-rezayi quantum hall state and conformal field theory. , 498(3):513–538, 1997.
Z. N. C. Ha. Exact dynamical correlation functions of calogero-sutherland model and one-dimensional fractional statistics. , 73(12):1574, 1994.
F. D. M. Haldane. Exact jastrow-gutzwiller resonating-valence-bond ground state of the spin-(1/2 antiferromagnetic heisenberg chain with 1/r 2 exchange. , 60(7):635, 1988.
B. I. Halperin. Statistics of quasiparticles and the hierarchy of fractional quantized hall states. , 52(18):1583, 1984.
S. Iso and S. J. Rey. Collective field theory of the fractional quantum hall edge state and the calogero-sutherland model. , 352(1):111–116, 1995.
M. Jimbo and T. Miwa. Solitons and infinite dimensional lie algebras. , 19(3):943–1001, 1983.
K. H. Lee, Z. X. Hu, and X. Wan. Construction of edge states in fractional quantum hall systems by jack polynomials. , 89(16):165124, 2014.
I. G. Macdonald. . New York, 1995.
G. Moore and N. Read. Nonabelions in the fractional quantum hall effect. , 360(2):362–396, 1991.
G. Murthy and R. Shankar. Hamiltonian description of composite fermions: Calculation of gaps. , 59(19):12260, 1999.
G. Murthy and R. Shankar. Hamiltonian theories of the fractional quantum hall effect. , 75(4):1101, 2003.
N. Read and E. Rezayi. Beyond paired quantum hall states: parafermions and incompressible states in the first excited landau level. , 59(12):8084, 1999.
R. Shankar and G. Murthy. Towards a field theory of fractional quantum hall states. , 79(22):4437, 1997.
B. S. Shastry. Exact solution of an s= 1/2 heisenberg antiferromagnetic chain with long-ranged interactions. , 60(7):639, 1988.
J. C. Talstra and F. D. M. Haldane. Integrals of motion of the haldane-shastry model. , 28(8):2369, 1995.
X. G. Wen and Y. S. Wu. Chiral operator product algebra hidden in certain fractional quantum hall wave functions. , 419(3):455–479, 1994.
J. F. Wu and M. Yu. . : 1110.6720, 2011.
S. C. Zhang, T. H. Hansson, and S. Kivelson. Effective-field-theory model for the fractional quantum hall effect. , 62(1):82, 1989.
S. C. Zhang. The chern–simons–landau–ginzburg theory of the fractional quantum hall effect. , 6(01):25–58, 1992.
[^1]: In the viewpoint of integrable hierarchy, the CFT Hamiltonian $L_0$, is the second Hamiltonian(integral of motion) of the system. However, the finer structures, such as explored in present article and [@estienne2012conformal], are determined by the third or higher level Hamiltonians .
[^2]: It is a standard choice of a normalized Jack polynomial. While for a non-normalized Jack polynomial the coefficient can be 1.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[ We study separating function sets. We find some necessary and sufficient conditions for $C_p(X)$ or $C_p^2(X)$ to have a point-separating subspace that is a metric space with certain nice properties. One of the corollaries to our discussion is that for a zero-dimensional $X$, $C_p(X)$ has a discrete point-separating space if and only if $C_p^2(X)$ does. ]{}'
address: 'Facultad de Ciencias Físico-Matematicas, Benemérita Universidad Autonoma de Puebla, Apdo postal 1152, Puebla, Puebla CP 72000, Mexico'
author:
- Raushan Buzyakova
- Oleg Okunev
title: A Note On Separating Function Sets
---
Introduction {#S:introduction}
============
To start our discussion, recall, that $F\subset C^n_p(X)$ is [*point-separating*]{} if for any distinct $x,y\in X$ there exists $\langle f_1,...,f_n\rangle\in F$ such that $f_i(x)\not = f_i(y)$ for some $i\leq n$. In this paper we are concerned with the following general problem.
[**Problem.**]{} [*Let $P$ be a nice property. Describe “$C_p(X)$ (or $C_p^m(X)$) having a point-separating subspace with $P$” in terms of the topology of $X$, $X^n$, or $X^\omega$.* ]{}
In this study, $P$ is the property of being a discrete space, a countable union of discrete subspaces, a metric compactum, or a discrete group. We obtain two characterizations of spaces $X$ for which $C^2_p(X)$ has a discrete point-separation subspace (Theorems \[thm:criterion\] and \[thm:criterionGCH\], and \[thm:criterionsigma\]). One of the characterizations is consistent and may have a chance for a ZFC proof. We also characterize zero-dmensional $Z$ with point-separating discrete subspaces in $C_p(X)$ (Theorems \[thm:criterion0dim\] and \[thm:criterionGCH0dim\], and \[thm:criterionsigma0dim\]). Questions of similar nature are quite popular among topologists interested in $C_p$-theory and have been considered in many papers.
In notation and terminology we follow [@Eng]. All spaces under consideration are assumed Tychonoff and [*infinite*]{}. By $s(X)$ we denote the supremum of cardinalities of discrete subspaces of $X$. By $iw(X)$ we denote the smallest weight of a Tychonoff subtopology of $X$. When we say that $D$ is a discrete subspace of $X$, $D$ need not be closed in $X$. By $\sigma_X(x^*)$ we denote the subspace of $X^\omega$ that consists of all points that differ from $x^*$ by finitely many coordinates. Since $\sigma_X(x)$ and $\sigma_X(y)$ are obviously homeomorphic we may simply write $\sigma_X$ and, as usual, refer to it as [*$\sigma$-product of $X^\omega$*]{}. A standard neighborhood of $f$ in $C_p(X)$ is in form $U(x_1,...,x_n; B_1,...,B_n)=\{g\in C_p(X): g(x_i)\in B_i\}$, where $B_i$ is open interval with rational endpoints for each $i$. Zero-dimensionality is understood in the sense of dim.
Discrete Point-Separating Subspaces {#S:discretesubspaces}
===================================
Our first goal is to find a characterization of those infinite $X$ for which $C_p(X)$ or $C_p^2(X)$ has a point-separating discrete subspace. We achieve the goal within a wide class of spaces. We start with a few helping lemmas.
The following Lemma is almost identical to Proposition II.5.5 in [@Arh] but due to cofinality restrictions we have to prove it using a similar argument.
\[lem:functionspace0\] (version of [@Arh II.5.5]) Assume that $C^2_p(X)$ has a discrete subspace of size $\tau$. Then the following hold:
1. If $cf(\tau)>\omega$, then $s(X^n)\geq \tau$ for some $n\in \omega$.
2. If $cf(\tau)=\omega$, then $s(\sigma_X)\geq \tau$.
Since part (2) is an obvious consequence of part (1), we will prove part (1) only. Let $D\subset C^2_p(X)$ be a $\tau$-sized discrete subspace. For each $\langle f, g\rangle\in D$ fix standard neighborhoods $U_f = U_f(x^f_1,...,x^f_{n_f}, I_1^f, ...,I_{n_f}^f)$ and $V_g = V_g(y^g_1,...,y^g_{m_g}, J_1^g, ...,J_{m_g}^g)$ such that $U_f\times V_g$ contains $\langle f,g\rangle$ and misses $D\setminus \{\langle f,g\rangle\}$. Since $cf(\tau ) >\omega$ we can find $n^*,m^*\in \omega$, $\langle I_i: i\leq n^*\rangle$, $\langle J_i: i\leq m^*\rangle$, and a $\tau$-sized $D'\subset D$ such that $n_f=n^*$, $m_g=m^*$, $\langle I_i^f: i\leq n^*\rangle = \langle I_i: i\leq n^*\rangle$, and $\langle J_i^g: i\leq m^*\rangle = \langle J_i: i\leq m^*\rangle$ for each $\langle f,g\rangle\in D'$. We can now conclude that for any distinct $\langle f,g\rangle,\ \langle f',g'\rangle\in D'$, either $\langle x_i^f:i\leq n^*\rangle \not = \langle x^{f'}_i:i\leq n^*\rangle$ or $\langle y_i^g:i\leq m^*\rangle \not = \langle y^{g'}_i:i\leq m^*\rangle$. Therefore, the set $S=\{\langle x^f_i,...,x_{n^*}^f, y_i^g,...,y^g_{m^*} \rangle: \langle f,g\rangle \in D' \}$ is $\tau$-sized. To show that $S$ is a discrete subspace of $X^{n^*+m^*}$, for each $\langle f,g\rangle\in D'$, put $U_f = f^{-1}(I_1)\times ...\times f^{-1}(I_{n^*})$ and $V_g = g^{-1}(J_1)\times ...\times g^{-1}(J_{m^*})$. Clearly $U_f\times V_g$ is a neighborhood of $\langle x^f_i,...,x_{n^*}^f, y_i^g,...,y^g_{m^*} \rangle$ in $X^{n^*+m^*}$. Next, fix $\langle f',g'\rangle\in D'\setminus \{\langle f,g\rangle\}$. By the choice of our neighborhoods, we may assume that $f\not \in U_{f'}$. Therefore, there exists $i\leq n^*$ such that $f(x^{f'}_i)\not \in I_i$. Therefore, $x_i^{f'}\not \in f^{-1}(I_i)$, which implies that $\langle x^{f'}_i,...,x_{n^*}^{f'}, y_i^{g'},...,y^{g'}_{m^*} \rangle\not \in U_f\times V_g$.
Note that if $C_p(X)$ or $C^2_p(X)$ has a discrete point-separating subspace of an infinite size $\tau$ , then $\tau \geq iw(X)$. If in addition $cf(\tau)>\omega$, then, by Lemma \[lem:functionspace0\], $s(X^n)\geq \tau \geq iw(X)$ for some $n$. Thus, the following statement holds.
\[thm:functionspace\] Assume that $C_p(X)$ or $C_p^2(X)$ has a discrete point-separating subspace of size $\tau$ with $cf(\tau)>\omega$. Then $s(X^n)\geq \tau\geq iw(X)$ for some natural number $n$.
We are now ready to formulate and prove two necessary conditions for $C_p(X)$ and $C_p^2(X)$ to have a point-separating discrete subspace.
\[thm:functionspace2\] If $C^2_p(X)$ has a discrete point-separating subspace, then $s(\sigma_X)\geq iw(X)$.
Put $\tau = iw(X)$. If $\tau$ is countable, then $X$ has a countable network. Since $X$ is infinite, it contains an infinite countable subspace. Hence, $s(\sigma_X)\geq iw(X)$.
We now assume that $\tau$ is uncountable. By Theorem \[thm:functionspace\] we may assume that $cf(\tau) = \omega$. Fix a strictly increasing sequence of cardinals $\tau_n$ of uncountable cofinalities so that $\tau =\sum_n\tau_n$. Since any point-separating subset of $C^2_p(X)$ must have size at least $\tau$, there exists a discrete subset of cardinality $\tau_n$ in $C^2_p(X)$ for each $n$. By Lemma \[lem:functionspace0\], there exists a discrete subset $D_n$ in some finite power of $X$ for each $n$. Therefore, $s(\sigma_X)\geq \tau$.
In all future arguments, the cases when a discrete point-separating subspace is finite can be handled as in Theorem \[thm:functionspace2\] and will therefore not be considered. For our next observation we need Zenor’s theorem [@Zen] stating that if $s(X\times Y)\leq \tau\geq \omega$ then either $hl(X)\leq \tau$ or $hd(Y)\leq \tau$.
\[thm:functionspace3\] Assume Generalized Continuum Hypothesis. If $C^2_p(X)$ has a discrete point-separating subspace, then $s(X^n)\geq iw(X)$ for some $n\in \omega$.
Put $\tau = iw(X)$. By Theorem \[thm:functionspace\] we may assume that $\tau$ is an infinite cardinal of countable cofinality. Assume the contrary. Then $s(X^4) = \lambda < \tau$. By Zenor’s theorem, $hl(X^2)\leq \lambda$ or $d(X^2)\leq \lambda$. If the former is the case, then the off-diagonal part of $X^2$ can be covered by $\lambda$-many functionally closed boxes, which implies that $iw(X)< \tau$. If $d(X^2)\leq \lambda$, then by Generalized Continuum Hypothesis, $w(X^2)$ is at most $2^\lambda<\tau$. Since both cases lead to contradictions, the statement is proved.
The assumptions in Theorem \[thm:functionspace3\] prompts the following questions.
\[que:zfc\] Does Theorem \[thm:functionspace3\] hold in ZFC?
Note that if one can construct a space $X$ such that $s(X^n)=\omega_n$ for all natural numbers $n$ and $iw(X)=\omega_\omega$, then the answer to Question \[que:zfc\] is a “no”.
At this point one may wonder if our study is justified. In other words, are we studying a non-empty class? Let $X$ be an non-metrizable compact space such that $X^n$ is hereditary separable for each $n$. Such a space exists. A consistent example of such a space is Ivanov’s modification [@Iva] of Fedorchuk’s example [@Fed]. Since $X^n$ is hereditarily separable, by Lemma \[lem:functionspace0\], no discrete subspace of $C_p^2(X)$ or $C_p(X)$ is uncountable. Since $X$ is not submetrizable, we conclude that no countable subspace of $C_p^2(X)$ or $C_p(X)$ is point-separating. Let us summarize this observation as follows.
\[ex:sspace\] There exists a consistent example of a compactum $X$ such that neither $C_p(X)$ nor $C^2_p(X)$ has a discrete separating subspace.
The authors believe that in some models of ZFC, no such example may exist, meaning that any space may have a discrete in itself point-separating function set.
\[que:mainexample\] Is there a ZFC example of a space $X$ such that no discrete subspace of $C_p(X)$ ($C_p^n(X)$) is point-separating?
We will next reverse the statement of Theorem \[thm:functionspace\], which will bring us to the promised characterizations.
\[thm:Xn\] If $X^n$ has a discrete subspace of size $iw(X)$ for some natural number $n$, then $C^2_p(X)$ has a point-separating discrete subspace.
Let $n$ be the smallest that satisfies the hypothesis of the lemma and put $\tau = iw(X)$. By the choice of $n$ there exists a $\tau$-sized discrete subspace $D$ of $X^n$ with the following property:
$|\{x(i):i\leq n\}|=n$ for each $x\in D$.
Let $\mathcal T$ be a Tychonoff subtopology of the topology of $X$ of weight $\tau$. Fix a $\tau$-sized network $\mathcal N$ for $\langle X, \mathcal T\rangle$ that consists of functionally closed subsets. Let $\mathcal P$ be the set of all pairs $\langle A, B\rangle$ of disjoint elements of $\mathcal N$. Enumerate $\mathcal P$ and $D$ as $\{\langle A_\alpha, B_\alpha\rangle:\alpha<\tau\}$ and $\{d_\alpha:\alpha<\tau\}$. Since $D$ is a discrete subspace, for each $\alpha<\tau$ we can fix a functionally closed set $B^\alpha_1\times ...\times B^\alpha_n$ that contains $d_\alpha$ in its interior and misses $D\setminus \{d_\alpha\}$. By [*Property*]{}, we may assume that $B^\alpha_i\cap B^\alpha_j=\emptyset$ for distinct $i$ and $j$.
We will next construct our desired subspace $\{\langle f_\alpha, g_\alpha\rangle:\alpha<\tau\}$ of $C_p(X)$.
Let $S_\alpha$ be a functionally closed subset of $X$ such that $X\setminus S_\alpha$ can be written as a union of $L_\alpha$ and $R_\alpha$ so that the following hold.
1. $cl_X(L_\alpha)\cap cl_X(R_\alpha)\subset S_\alpha$;
2. $A_\alpha\subset L_\alpha$ and $B_\alpha\subset R_\alpha$;
3. $d_\alpha(i)\in L_\alpha$ if $d_\alpha(i)\not \in B_\alpha$, and $d_\alpha(i)\in R_\alpha$ if $d_\alpha(i)\in B_\alpha$.
Such an $S_\alpha$ exists since $A_\alpha$ and $B_\alpha$ are functionally separable sets and the coordinate set of $d_\alpha$ is finite. Let $f_{\alpha, l}: L_\alpha\cup S_\alpha\to [-1, 0]$ be any continuous function that has the following properties:
$f^{-1}_{\alpha, l}(\{0\})= S_\alpha$;
$(\{d_\alpha(i):i\leq n\}\cap L_\alpha )\subset f^{-1}_{\alpha, l}([-1,-1/3))\subset \bigcup
\{B^\alpha_i :d_\alpha(i)\in L_\alpha\}$.
Such a function exists because the coordinate set of $d_\alpha$ is finite and $S_\alpha$ is functionally closed and misses the coordinate set of $d_\alpha$. Let $f_{\alpha , r}: R_\alpha\cup S_\alpha\to [0, 1]$ be any continuous function that has the following properties:
$f^{-1}_{\alpha, r}(\{0\})= S_\alpha$;
$(\{d_\alpha(i):i\leq n\}\cap R_\alpha )\subset f^{-1}_{\alpha, r}((1/3, 1])\subset \bigcup
\{B^\alpha_i :d_\alpha(i)\in R_\alpha\}$.
Put $f_\alpha = f_{\alpha ,l} \cup f_{\alpha , r}$. By L1 and R1, $f_\alpha$ is a continuous function from $X$ to $\mathbb R$.
: Let $g_\alpha$ be any continuous function that maps $B_i^\alpha$ to $(i-1/3, i+1/3)$. This can be done since $B_i^\alpha$’s form a disjoint finite collection of functionally closed sets.
It remains to show that $F=\{\langle f_\alpha, g_\alpha\rangle:\alpha<\tau\}$ is a point-separating discrete subspace. To show that $F$ is point-separating, fix distinct $a,b$ in $X$. Since $\mathcal N$ is a network, there exist disjoint $A,B\in \mathcal N$ that contain $a$ and $b$, respectively. Then $\langle A,B\rangle=\langle A_\alpha, B_\alpha\rangle\in \mathcal P$. By the definition of $f_\alpha$, $f_\alpha(a)=f_{\alpha, l}(a)<0$ and $f_\alpha(b)=f_{\alpha,r}(b)>0$.
To show that $F$ is discrete in itself, fix $\alpha$. Put $$U_\alpha=\{f: f(d_\alpha(i)) <-1/3\ if \ d_\alpha(i)\in L_\alpha, f(d_\alpha(i))>1/3\ if \ d_\alpha(i)\in R_\alpha\}$$ $$V_\alpha = \{g: g(d_\alpha(i))\in (i-1/3,i+1/3)\}$$ Clearly, $U_\alpha\times V_\alpha$ is a neighborhood of $\langle f_\alpha, g_\alpha\rangle$. To show that this neighborhood misses the rest of $F$, fix $\beta\not = \alpha$. There exists $i\leq n$ such that $d_\alpha(i)\not \in B^\beta_i$. We have two possible cases.
This case’s assumption is that $d_\alpha(i)\not \in \bigcup_{j\leq n}B^\beta_j$. By L2 and R2 of the definition of $f_\beta$, we have $f_\beta((d_\alpha^i))\in (-1/3,1/3)$. Hence $f_\beta\not\in U_\alpha$. Therefore, $\langle f_\beta,g_\beta\rangle\not \in U_\alpha\times V_\alpha$.
Assume Case 1 does not take place. Then there exists $j\leq n$ such that $d_\alpha(i)\in B^\beta_j$. By the choice of $i$, we have $i\not = j$. Therefore, $g_\beta(d_\alpha(i))\not \in (i-1/3, i+1/3)$. Hence $g_\beta\not\in V_\alpha$. Therefore, $\langle f_\beta,g_\beta\rangle\not \in U_\alpha\times V_\alpha$.
Statements \[thm:functionspace\], \[thm:Xn\], and \[thm:functionspace3\] result in the following criteria.
\[thm:criterion\] Let a space $X$ have $iw(X)$ of uncountable cofinality. Then $C^2_p(X)$ has a point-separating discrete subspace if and only if $s(X^n)\geq iw(X)$ for some $n$.
\[thm:criterionGCH\] Assume Generalized Continuum Hypothesis. Then $C^2_p(X)$ has a point-separating discrete subspace if and only if $s(X^n)\geq iw(X)$ for some $n$.
Note that criteria \[thm:criterion\] and \[thm:criterionGCH\] would hold for $C_p(X)$ if we could prove Theorem \[thm:Xn\] for $C_p(X)$.
\[que:CpX\] Assume that $X^n$ has a discrete subspace of size $iw(X)$ for some natural number $n$. Is it true that $C_p(X)$ has a discrete point-separating set?
Using an argument somewhat similar to that of Theorem \[thm:Xn\] we will next show that Question \[que:CpX\] has an affirmative answer if we assume that $C$ is zero-dimensional.
\[thm:Xnzerodim\] Assume that $X$ is zero-dimensional. If $X^n$ has a discrete subspace of size $iw(X)$, then $C_p(X)$ has a point-separating discrete subspace.
Let $n$ be the smallest that satisfies the hypothesis of the lemma and put $\tau = iw(X)$. By the choice of $n$ there exists a $\tau$-sized discrete subspace $D$ of $X^n$ with the following property:
$|\{x(i):i\leq n\}|=n\ for\ each\ x\in D.$
Let $\mathcal T$ be a Tychonoff subtopology of the topology of $X$ of weight $\tau$. Due to zero-dimensionality of $X$ and the factorization theorem of Mardesic [@Mar], we may assume that $\mathcal T$ is zero-dimensional too. Fix a $\tau$-sized network $\mathcal N$ for $\langle X, \mathcal T\rangle$ that consists of clopen subsets. Let $\mathcal P$ be the set of all pairs $\langle A, B\rangle$ of disjoint elements of $\mathcal N$. Enumerate $\mathcal P$ and $D$ as $\{\langle A_\alpha, B_\alpha\rangle:\alpha<\tau\}$ and $\{d_\alpha:\alpha<\tau\}$. We will next construct our desired subspace in $C_p(X)$.
\] Since $D$ is a discrete subspace, we can fix a clopen box $U^\alpha_1\times ...\times U^\alpha_n$ that contains $d_\alpha$ and misses $D\setminus \{d_\alpha\}$. By [*Property*]{}, we may assume that $U^\alpha_i\cap U_j^\alpha=\emptyset$ if $i\not = j$. Since $A_\alpha$ and $B_\alpha$ are disjoint, we may assume that each $U_i$ meets at most one of the sets $A_\alpha$ and $B_\alpha$. Define $f_\alpha: X\to \{0,1,2,...,n+1\}$ by letting $f_\alpha(U_i)=\{i\}$, $f_\alpha(A_\alpha\setminus \bigcup_{i\leq n}U_i)=\{0\}$, and $f_\alpha (X\setminus (A_\alpha\cup U_1\cup ...\cup U_n) )= \{n+1\}$.
It remains to show that $F=\{f_\alpha:\alpha<\tau\}$ is a point-separating discrete subspace. To show that $F$ is point-separating, fix distinct $a,b$ in $X$. Since $\mathcal N$ is a network, there exist disjoint $A,B\in \mathcal N$ that contain $a$ and $b$, respectively. Then $\langle A,B\rangle=\langle A_\alpha, B_\alpha\rangle\in \mathcal P$. Since no $U_i^\alpha$ meets both $A_\alpha$ and $B_\alpha$ at the same time, $f_\alpha(A_\alpha)$ misses $f_\alpha(B_\alpha)$.
To show that $F$ is discrete in itself, fix $f_\alpha$ and put $V_\alpha = \{f: f(d_\alpha(i)) \in (i-1/3, i+1/3), i\leq n\}$. Next fix any $\beta\not = \alpha$. Then there exists $i\leq n$ such that $d_\alpha(i)\not \in U^\beta_i$. Therefore, $f_\beta(d_\alpha(i))\not \in (i-1/3,i+1/3)$. Hence, $f_\beta\not \in U_\alpha$.
Note that Theorems \[thm:criterion\] and \[thm:criterionGCH\] are now true for $C_p(X)$ provided $X$ is zero-dimensional. Let us state the new versions for reference.
\[thm:criterion0dim\] Let a zero-dimensional space $X$ have $iw(X)$ of uncountable cofinality. Then $C_p(X)$ has a point-separating discrete subspace if and only if $s(X^n)\geq iw(X)$ for some natural number $n$.
\[thm:criterionGCH0dim\] Assume Generalized Continuum Hypothesis. Let $X$ be zero-dimensional. Then $C_p(X)$ has a point-separating discrete subspace if and only if $s(X^n)\geq iw(X)$ for some $n$.
For our final characterization discussion we would like to extract a technical statement from the proof of Theorem \[thm:Xn\] and prove one helpful proposition.
\[lem:extract\] Assume that a finite power of $X$ has a discrete subspace of size $\lambda$. Let $\{\langle A_\alpha, B_\alpha\rangle:\alpha<\lambda\}$ be a family of pairs of functionally closed disjoint subsets of $X$. Then there exists a discrete subspace $F$ in $C^2_p(X)$ with the following property:
(\*) If $a\in A_\alpha$ and $b\in B_\alpha$ for some $\alpha<\lambda$, then $f(a)\not = f(b)$ for some $f\in F$.
\[pro:union\] Let $C^m_p(X)$ contain a point-separating subspace which is a countable union of discrete subspaces. Then $C^m_p(X)$ has a discrete point-separating subspace.
We will prove the statement for $m=2$. Let $D=\cup_nD_n$ be a point-separating set of $C^2_p(X)$, where each $D_n$ is a discrete subspace. For each $n$, fix a homeomorphism $h_n:\mathbb R\to (n,n+1)$. Put $E_n = \{\langle h_n\circ f, h_n\circ g\rangle: \langle f,g\rangle \in D_n\}$. Clearly, $E_n$ separates $x$ and $y$ if and only if $D_n$ does. Also, $E_n$ is a discrete subspace of $C^2_p(X)$. Since all functions in $(\cup_kE_k)\setminus E_n$ target $\mathbb R\setminus (n,n+1)$, we conclude that the closure of $(\cup_kE_k)\setminus E_n$ misses $E_n$. Terefore, $\cup_n E_n$ is a point-separating discrete subspace of $C^2_p(X)$.
\[thm:criterionsigma\] $C^2_p(X)$ has a discrete point-separating subspace if and only if $s(\sigma_X) \geq iw(X)$.
Necessity is done in Theorem \[thm:functionspace2\]. To prove sufficiency, put $\tau = iw(X)$. Let $\mathcal N$ be a $\tau$-sized family of functionally closed subsets of $X$ that is a network for some Tychonoff subtopology of $X$. Let $\mathcal P$ consist of all pairs of disjoint elements of $\mathcal N$. For each $n$ we can find a discrete subset $D_n$ of $\sigma_X$ that lives in a copy of some finite power of $X$ so that $\tau = \sum_n |D_n|$. Next represent $\mathcal P$ as $\bigcup \mathcal P_n$, where $|\mathcal P_n|=|D_n|$. Applying Lemma \[lem:extract\] to $\mathcal P_n$ and $D_n$ for each $n$, we find a point-separating subspace in $C^2_p(X)$ that is a countable union of discrete subspaces. By Proposition \[pro:union\], $C^2_p(X)$ contains a discrete point-separating subspace.
An argument identical to that of Theorem \[thm:criterionsigma\] leads to the following statement for $C_p(X)$.
\[thm:criterionsigma0dim\] Assume that $X$ is zero-dimensional. Then $C_p(X)$ has a discrete point-separating subspace if and only if $s(\sigma_X) \geq iw(X)$.
Theorems \[thm:criterionsigma\] and \[thm:criterionsigma0dim\] imply the following.
\[CpXCpX2\] Let $X$ be a zero-dimensional space. Then $C_p(X)$ has a point-separating discrete subspaces if and only if $C^2_p(X)$ does.
Note that the image of a point-separating family under a homeomorphism need not be point-separating. Indeed, $\{id_{[0,1]}\}$ is a point-separating subspace of $C_p([0,1])$, However, one can construct an automorphism on $C_p([0,1])$ that caries $\{ id_{[0,1]}\}$ to $\{\vec 0\}$ which is not point-separating. In connection with this observation, it would be interesting to know if having a discrete point-separating subspace is preserved by homeomorphisms among function spaces. The answer is affirmative and to show it we will use the fact [@Arh I,1,6] that if $C_p(X)$ and $C_p(Y)$ are homeomorphic then $iw(X)=iw(Y)$.
\[thm:tequiv0dim\] Let $X$ and $Y$ be $t$-equivalent. If $C^2_p(X)$ has a discrete point-separating subspace, then so does $C^2_p(Y)$.
Fix a homeomorphism $\phi: C^2_p(X)\to C^2_p(Y)$ and a discrete point-separating subspace $D$ of $C^2_p(X)$.
Assume, first, that $|D|$ is finite. Then $iw(X)=\omega$. Hence $iw(Y)=\omega$. Since $Y$ is infinite, it contains a an infinite countable subspace. By Theorem \[thm:criterionsigma\] , $C^2_p(Y)$ contains a discrete point-separating subspace.
We now assume that $|D|$ is infinite. Then $|D|\geq iw(X)$. Therefore, $|\phi (D)|\geq iw(Y)$. By Lemma \[lem:functionspace0\], $s(\sigma_Y)\geq |\phi(D)|\geq iw(Y)$. By Theorem \[thm:criterionsigma\] , $C^2_p(Y)$ contains a discrete point-separating subspace.
Repeating the argument of Theorem \[thm:tequiv0dim\], we obtain the following.
\[thm:tequiv\] Let $X$ and $Y$ be zero-dimensional and $t$-equivalent. If $C_p(X)$ has a discrete point-separating subspace, then so does $C_p(Y)$.
While being a discrete subspace is already a nice property, it would be interesting to know when $C_p(X)$ or its finite power has a discrete point-separating subspace which is in addition a subgroup. Note that any discrete subgroup is closed. In addition, $C_p(X)$ can be covered by countably many shifts of any neighborhood of zero-function. Therefore, any discrete subgroup of $C_p(X)$ is countable. These observations lead to the following question.
Let $X$ be a separable metric space. Is it true that $C_p(X)$ has a discrete point-separating subgroup?
It is worth noting that separable metric spaces have many pretty point-separating subspaces as backed up by the next two statement.
\[thm:01\_1\] $C_p(X)$ has a point-separating subset homeomorphic to $[0,1]$ if and only $X$ admits a continuous injection into $\mathbb R^\omega$.
To prove necessity, let $F\subset C_p(X)$ a point-separating family homeomorphic to $[0,1]$. Then any dense subset of $F$ is point-separating too. Therefore, $C_p(X)$ has a countable point-separating family. Therefore, $X$ continuously injects into $\mathbb R^\omega$.
To prove sufficiency we need the following claim.
[*Claim. $\mathbb R^\omega$ embeds into $C_p([0,1])$.*]{}
To prove the claim, note that $C_p(\omega) = \mathbb R^\omega$ embeds into $C_p(\mathbb R)$ since $\omega$ is closed in $\mathbb R$. By Gulko-Hmyleva theorem [@GH] that $\mathbb R$ and $[0,1]$ are $t$-equivalent, we conclude that, $\mathbb R^\omega$ embeds into $C_p([0,1])$. The claim is proved.
By Claim $X$ injects into $C_p([0,1])$. Let $F$ ne the image of such an injection. Due to homogeneity we may assume that the identity function is in $F$.Therefore, $F$ generates the topology of $[0,1]$. Consider the evaluation map the evaluation function $\Psi_F:[0,1]\to C_p(F)$. Since $F$ generates the topology of $[0,1]$ , we conclude that $\Psi_F([0,1])$ genberate the topology of $F$. If $h:X\to F$ is a continuous bijection then the map $H: C_p(F)\to C_p(X)$ is a continuous injection, where $H(f) =hf$. Clearly, $H(\Psi_F([0,1]))=[0,1]$ is point separating.
\[thm:01\_2\] Let $X$ be a separable metric space. Then $C_p(X)$ has a topology-generating subspace homeomorphic to $[0,1]$.
Embed $X$ into $C_p([0,1])$ so that the image $F$ contains the identity map. The evaluation function $\Psi_F:[0,1]\to C_p(F)$. Since $F$ generates the topology of $[0,1]$ and therefore $\Psi_F([0,1])$ generates the topology of $F$. Since $F=X$, we conclude that $[0,1]=\Psi_F([0,1])$ generates the topology of $F=X$.
Note that Theorem \[thm:01\_2\] cannot be reversed. Indeed, $[0,1]$ generates the topology of $C_p([0,1])$ but the latter is not metrizable.
We would like to finish with two problems that are naturally prompted by our study.
Characterize spaces $X$ for which $C_p(X)$ has a closed discrete point-separating subset.
Characterize spaces $X$ for which $C_p(X)$ has a (closed) discrete topology-generating subset.
At last, the unattained goal of the paper is left as the following question.
Assume that $C_p(X)$ has a discrete subspace of size $iw(X)$. Is it true that $C_p(X)$ has a discrete point-separating set?
[99]{}
A. Arhangelskii, [*Topological Function Spaces*]{}, Math. Appl., vol. 78, Kluwer Academic Publishers, Dordrecht, 1992.
R. Engelking, [*General Topology*]{}, PWN, Warszawa, 1977.
V. V. Fedorchuk, [*A compact having a cardinality of continuum with no convergent sequences*]{}, Math. Proc. Cambridge Phil. Soc. 81(1977), 177-181
S. Gulko and T. Hmyleva, [*Compactaness is not preserved by $t$-equivalence*]{}, Mat Zametki, vol 39, 6 (1986), 895-903.
A. V. Ivanov, [*On bicompacta with hereditary separable finite powers*]{}, (in Russian) DAN SSSR, 243 (1978), 1109-1112.
S. Mardesic, [*On covering dimension and inverse limits of compact spaces*]{}, Ill. J. of Math. 4 (1960), 278-291.
P. Zenor, [*Hereditary m-separability and the hereditary m-Lindelöf property in product spaces and function spaces*]{}, Fund. Math. 106 (1980), 175-80.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Here we analyse the spectroscopic information gathered at a number of single CrO$_{2}$ / Pb interfaces. We examine thin films requiring additional interfacial layers to generate long range spin triplet proximity effect superconductivity (CrO$_{2}$/TiO$_{2}$) or not (CrO$_{2}$/Al$_{2}$O$_{3}$). We analyse the data using two theoretical models and explore the use of a parameter free method to determine the agreement between the models and experimental observations, showing the necessary temperature range that would be required to make a definitive statement. The use of the excess current as a further tool to distinguish between models is also examined. Analysis of the spectra demonstrates that the temperature dependence of the normalised zero bias conductance is independent of the substrate onto which the films are grown. This result has important implications for the engineering of interfaces required for the long range spin triplet proximity effect.'
author:
- 'K.A. Yates'
- 'M.S. Anwar'
- 'J. Aarts'
- 'O. Conde'
- 'M. Eschrig'
- 'T. Löfwander'
- 'L.F. Cohen'
bibliography:
- 'cro2bib.bib'
date: Published online 7th October 2013
title: 'Andreev spectroscopy of CrO$_{2}$ thin films on TiO$_{2}$ and Al$_{2}$O$_{3}$'
---
Introduction
============
In 2006 Keizer et al., reported the observation of a supercurrent transported over nearly a micron through fully spin polarised [@Soulen; @Yates; @Lofwander] CrO$_{2}$ in an SFS Josephson junction [@Keizer]. The conventional superconducting proximity effect is expected to be very short in CrO$_{2}$ [@EschrigPT] and certainly far shorter than the distances observed by Keizer et al. [@Keizer]. The result though was explicable within the developing theory of the long range spin triplet proximity effect (LRSTPE) [@EschrigPT; @EschrigNP; @Bergeret]. In this theory, two components are required to transform spin singlet Cooper pairs from the superconductor into spin parallel, triplet pairs that can exist in the ferromagnet: Spin mixing is required to mix the singlet pair ($|\uparrow \downarrow \rangle -|\downarrow \uparrow \rangle$) into a triplet pair of the form ($|\uparrow \downarrow \rangle +|\downarrow \uparrow \rangle$) while a further spin flip (or spin transformation) process is needed to change that spin opposite triplet pair component into the long range spin parallel ($|\uparrow \uparrow \rangle$ and $|\downarrow \downarrow \rangle$) one. The theory predicted that magnetic inhomogeneity at the interface between a ferromagnet and a superconductor provided the key to both the spin mixing, defined by a spin mixing angle $\theta$ and spin transformation processes required for generation of the LRSTPE [@EschrigPT; @EschrigNP; @Bergeret].
Subsequent experiments have confirmed that the inhomogeneity requirement for LRSTPE generation can be satisfied by engineered multilayer contacts [@Khaire] or intrinsic magnetic inhomogeneity such as the spiral ordering found in the rare earth metal holmium [@Robinson; @Sosnin]. Recent results on CrO$_{2}$ show that the LRSTPE can be observed in CrO$_{2}$ grown on Al$_{2}$O$_{3}$ substrates using simple superconducting contacts [@AnwarPRB] but only through CrO$_{2}$ films grown onto TiO$_{2}$ substrates when engineered contacts incorporating a thin Ni layer are employed [@AnwarAPL]. Here we examine the spectroscopic information from single S/F, Pb/CrO$_{2}$ contacts using CrO$_{2}$ thin films grown on Al$_{2}$O$_{3}$ or TiO$_{2}$ substrates [@AnwarPRB; @AnwarAPL]. We do this in the context of the Mazin modified Blonder-Tinkham-Klapwijk (BTK) [@BTK; @Mazin] and the more recent LRSTPE spin mixing model (SMM) [@Lofwander; @Grein] and find that the latter suggests that CrO$_{2}$ has strong intrinsic spin mixing properties independent of the substrate on which it is grown.
The spectroscopic method we use is point contact spectroscopy (PCS). The technique has been used extensively to investigate the transport spin polarization of candidate materials for spintronics [@Soulen; @Bugoslavsky] using the modified BTK model. In this model, the spectra can be fitted using four parameters: the superconducting gap, $\Delta$, a measure of the interface scattering, $Z$ (a spin independent delta function parameter), the spin polarisation of the transport carriers, $P$ and either a spreading resistance, $r_{s}$, that captures the series resistance of the film as is used here [@Lofwander; @Woods], or a spectral broadening parameter that incorporates thermal and non-thermal smearing, $\omega$ [@Bugoslavsky]. Within the modified BTK model, the spins are treated equally as they cross the interface and no consideration is given for effects such as spin dependent scattering [@Xia]. It is clear though that for situations where there are strong spin mixing effects, the two spins will conduct differently across the interface. The SMM model has been proposed to account for these differences. In such cases, the conductance spectra observed by PCS will differ considerably from the predictions of the modified BTK model [@Lofwander; @Piano]. In reference [@Lofwander], two independent measures were proposed to distinguish between the two models: Using the variation of the zero bias conductance $G_{0}$ of a contact, and the evolution of the excess current $I_{ex}$ measured at large voltage ($eV >> \Delta$), normalised as $I_{ex}R_{n}/\Delta$ where $R_{n}$ is the normal state resistance, both as a function of temperature [@Lofwander].
Experimental
============
Films of 100nm thickness were grown onto TiO$_{2}$ substrates by chemical vapour deposition (CVD) as described in [@AnwarPRB; @AnwarAPL]. Films grown onto Al$_{2}$O$_{3}$ substrates were of varying thickness and are described in ref [@Yates; @Sousa]. Point contact measurements were taken using mechanically sharpened Pb tips (T$_{c}$ = 7.2K) and using a differential screw mechanism to slowly bring the tip into contact with the sample [@Bugoslavsky] in a dewar of liquid helium. A contact was established at low temperature, and spectra were taken at increasing temperatures until a temperature just greater than the T$_{c}$ of Pb was reached. The background conductance was found to be temperature dependent meaning that it was not possible to normalise the low temperature spectra using the spectrum taken at T $>$ T$_{c}$. In order to normalise the spectra, it was found necessary to use a fourth order polynomial curve to fit the experimental curves above the region $|V| \geq 5 $mV. This was performed at each temperature and the data was then divided by this polynomial curve. An example of the resulting normalised curve used in the fitting routine is shown in figure \[fig.1\]. Spectra were fitted with the SMM and the Mazin modified BTK models. In each case the value for the gap energy, $\Delta$, was fixed to be that of Pb (ie. $\Delta _{0}$ = 1.35meV). For the SMM case, the fitting parameters were the spin mixing angle, $\theta$, a measure of the interface scattering, $Z_{smm}$ and $r_{s}$. Following the method of ref [@Lofwander], the value of $P$ was assumed to be fixed to 100% while the misalignment angle, $\alpha$, was fixed at $\pi /2$ as described in references [@Lofwander; @Grein] and used in [@Lofwander]. For fitting to the Mazin modified BTK model, the fit parameters were the polarisation, $P$, $Z_{BTK}$ and $r_{s}$.
Results and Discussion
======================
Figure \[fig.2\] shows the temperature dependence of a contact made onto the CrO$_{2}$ thin film. All spectra showed a suppression of the V $<$ $\Delta$ conductance at T $<$ T$_{c}$ consistent with contacts onto highly spin polarised films. As the temperature increased, the suppression of the zero bias conductance reduced until the background spectra were obtained at T $\simeq$ 6.5K, indicating the superconducting critical temperature of Pb was suppressed at these interfaces. In common with previous studies, we will denote the local critical temperature of the contact as T$_{c}^{A}$. Note that a suppressed T$_{c}$, can result from at least one of two scenarios, either the tips used have been somewhat oxidized or possibly, that there is reasonably strong proximity effect. The latter would have to be taken into account in the fitting model used, as described for example in Strijkers et al.,[@Strijkers], but in our case no evidence for bulk and suppressed gap features (i.e. features associated with two superconducting energy gaps) are observed in the spectra and hence the analysis by Strijkers et al., cannnot be applied [@Strijkers].
![Spectrum F at 4.2K normalised by an order 4 polynomial fit (see text) with the BTK fit (red solid line) and the SMM fit (grey dashed line). Inset shows the IV characteristic at 4.2K (black line) and 7.3 K (grey line).[]{data-label="fig.1"}](finfig1.eps){width="8cm"}
![Point contact Andreev reflection spectra onto CrO$_{2}$ with a Pb tip at (from bottom) 4.2K, 5.1K, 6.0K, 6.4K, 7.3K.[]{data-label="fig.2"}](finfig2.eps){width="8cm"}
An important comparison between the SMM and Mazin modified BTK models is the behaviour of the zero bias conductance as a function of temperature [@Lofwander]. Figure 3 shows the zero bias conductance (normalised to the zero bias conductance at T $\geq$ T$_{c}^{A}$) as a function of temperature. The attraction of this comparison is that the zero bias conductance value is taken straight from the raw data and no fitting is involved. Data on two sets of films are presented; those grown on TiO$_{2}$ substrates and data from films grown on Al$_{2}$O$_{3}$ substrates used in a previous study [@Yates]. The films grown on TiO$_{2}$ have been shown to support LRSTPE [@AnwarSUST], and those grown on Al$_{2}$O$_{3}$ were found to exhibit very high transport spin polarisation values of $\simeq$ 90 % [@Yates; @Branford]. Both sets of data show a close to linear decrease of $G_{0}/G_{N}$ down to T/T$_{c}^{A} \simeq$ 0.6. It is interesting to note that a near linear dependence is expected in the SMM if P = 100% and $\theta \simeq \pi /2$ whereas the behaviour should show quasi-exponential behaviour in the modified BTK or in the SMM case with $\theta = 0$ (for zero non-thermal broadening and/or zero series resistance), the latter case means that the interface is no longer spin active and the BTK results are recovered as shown by the theoretically generated data also shown in the figure. In our case, the temperature window of our experiments is restricted, nevertheless the trends in the experimental curves suggest that the SMM with $\theta \simeq \pi/2$ fits the majority of data points for films on both types of substrate. Note that the data can be equally well explained within the BTK model if a polarisation significantly less than 100% and either a large non-thermal broadening (quite usual for point contact type experiments [@Bugoslavsky]) or a large series resistance $r_{s}$ is considered for all contacts.
![$G_{0}/G_{N}$ for contacts grown onto TiO$_{2}$ substrates (closed symbols) onto Al$_{2}$O$_{3}$ substrates (open symbols). Also shown is the predicted behaviour for $\theta = \pi /2$ (black) and $\theta$ = 0 (grey) for $Z$ = 0.1 (solid line), 0.26 (dotted line), 0.5 (thick dashed line), 1.0 (dashed line).[]{data-label="fig.3"}](Figure3widerlines.eps){width="8cm"}
In order to compare the parameters generated by fitting to each model, the lowest temperature spectra of data sets B (4.3K) and F (4.2K) (both on TiO$_{2}$) were each normalised by dividing by the polynomial background. The resulting fits to both models are shown in figure 1. On obtaining the fit parameters $P$, $Z_{BTK}$, $r_{s}$ (modified BTK) or $\theta$, $Z_{smm}$, $r_{s}$ (SMM), the $G_{0}/G_{N}$ curve was generated for each data set and model, figure 4(a-d). Note that for consistency, the data for sets B and F have been replotted in figure 4 as $G_{0}/G_{N\prime}$ where $G_{N\prime}$ is the ‘conductance’ at the zero bias point of the polynomial fit. The difference between these two normalisation methods is minimal as can be seen by comparing the data for set B to the high temperature normalised data in figure 4c. Following full fitting of the lowest temperature spectrum and generation of the $G_{0}/G_{N\prime}$ from these parameters, it can be seen that within our temperature window we are unable to differentiate between the models. However, data taken to a lower temperature or the employment of superconductor with a higher critical temperature would facilitate the comparison between models. As it stands, the SMM model assumes the films have 100% polarisation, which is not unreasonable given the fact they strongly support LRSTPE [@AnwarSUST], but this is not definitive proof. Spin polarised photoemission although restricted to the top few nanometers of the surface would be a useful additional characterisation tool. Fitting the data within the modified BTK model produces lower values of polarisation (coupled with a high value of $r_{s}$ for contact F.)
![$G_{0}/G_{N\prime}$ and the predicted behaviour based on the low temperature fits for the spin mixing model for contact (a) B and (b) F and the modified BTK model for contact (c) B and (d) F, fitting parameters are included in the figure. The open symbol data in (c) is the data for contact B normalised to T$>$T$_{c}$.[]{data-label="fig.4"}](finfig4.eps){width="8cm"}
Measurement of the excess current was suggested in ref [@Lofwander], as a tool for checking the validity of the fits either to the modified BTK or the SMM. In its simplest form, the deviation of the I-V characteristic from Ohmic behaviour gives the current deficit of the contact, $I_{ex}$. The magnitude of the deficit can then be plotted as $I_{ex}R_{n}/\Delta$ where $R_{n}$ is the normal state resistance. The value of $I_{ex}R_{n}/\Delta$ is predicted to vary as a function of $Z$ and $P$ in the modified BTK model and $Z$, $P$ and $\theta$ in the spin mixing model [@Lofwander]. The validity of the fit can be established by comparing the value of $I_{ex}R_{n}/\Delta$ at $Z_{fit}$ with the values for the other parameters extracted. For the contacts shown in figure 4(a-d), the excess current values ($I_{ex}R_{n}/\Delta$) predicted are, for contact B, -0.46 (SMM) and -0.39 (BTK) while for contact F, $I_{ex}R_{n}/\Delta \simeq$ -0.43 (SMM), -0.42 (BTK). An unfortunate complication in our data is that the I-V curves taken at T$>$T$_{c}$ are also non-Ohmic (see for example the conductance curve at T$>$T$_{c}$ in figure 2 and the IV curve in the inset to figure 1) and it is this non-Ohmic response that dominates the excess current evaluation. In order to account for this, two methods were applied to approximate $I_{ex}$, firstly the IV at T $<$ T$_{c}$ was subtracted from that just above T$_{c}$. The value of the $I_{ex}$ was then averaged over $|5-10|$mV and the standard deviation from this average was taken as the error. Secondly, an Ohmic response was taken away from the data at low temperature and compared with the same Ohmic response for data taken at T$>$T$_{c}$, with the final value for $I_{ex}R_{n}/\Delta$ again being taken as the average over $|5-10|$mV and the standard deviation as the error. Within the considerable error, the data set matches both predictions from BTK and SMM models. Therefore although in general this method may provide additional validation, for the particular case of CrO$_{2}$, where there is a strong temperature dependent background, it does not help distinguish between the models.
The main result is that there is no detectable difference in behaviour between films grown onto Al$_{2}$O$_{3}$ and those grown onto TiO$_{2}$ using a model independent method. It has been observed previously [@AnwarAPL] that films grown onto TiO$_{2}$ supported the LRSTPE when a magnetically inhomogeneous layer was inserted. It was suggested that the higher degree of magnetic homogeneity achieved for films grown on these substrates may have meant that some components needed to generate the LRSTPE were lacking [@AnwarAPL]. Stimulated by the interpretation within the SMM, we propose that spin mixing is the common ingredient in generating the LRSTPE, and that the spin flip process is furnished by the magnetic disorder (films on Al$_{2}$O$_{3}$) or an extra magnetic layer (Ni for the case of films on TiO$_{2}$) [@AnwarAPL]. It is interesting to speculate that also spin-orbit scattering might be used for this purpose.
In summary we have revisited the spectroscopic information obtained on S/F contacts between Pb and CrO$_{2}$. Although within the limits of our experimental temperature window we are unable to differentiate between models, we have set out the types of experiments that would need to be carried out in order to do so. No difference is found in the conductance spectra taken on CrO$_{2}$ films grown on different substrates despite their different behaviours in terms of the generation of the long range spin triplet proximity effect.
ME acknowledges support from EPSRC under grant reference EP/J010618/1, LFC and KY acknowledge support from EPSRC under grant reference EP/H040048/1.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The search for diffuse non-thermal inverse Compton (IC) emission from galaxy clusters at hard X-ray energies has been undertaken with many instruments, with most detections being either of low significance or controversial. Because all prior telescopes sensitive at $E > 10$ keV do not focus light and have degree-scale fields of view, their backgrounds are both high and difficult to characterize. The associated uncertainties result in lower sensitivity to IC emission and a greater chance of false detection. In this work, we present 266 ks observations of the Bullet cluster, which is detected in the energy range 3–30 keV. ’s unprecedented hard X-ray focusing capability largely eliminates confusion between diffuse IC and point sources; however, at the highest energies the background still dominates and must be well understood. To this end, we have developed a complete background model constructed of physically inspired components constrained by extragalactic survey field observations, the specific parameters of which are derived locally from data in non-source regions of target observations. Applying the background model to the Bullet cluster data, we find that the spectrum is well – but not perfectly – described as an isothermal plasma with $kT = 14.2 \pm 0.2$ keV. To slightly improve the fit, a second temperature component is added, which appears to account for lower temperature emission from the cool core, pushing the primary component to $kT \sim 15.3$ keV. We see no convincing need to invoke an IC component to describe the spectrum of the Bullet cluster, and instead argue that it is dominated at all energies by emission from purely thermal gas. The conservatively derived 90% upper limit on the IC flux of $1.1 \times 10^{-12}$ erg s$^{-1}$ cm$^{-2}$ (50–100 keV), implying a lower limit on $B \ga 0.2$ $\mu$G, is barely consistent with detected fluxes previously reported. In addition to discussing the possible origin of this discrepancy, we remark on the potential implications of this analysis for the prospects for detecting IC in galaxy clusters in the future.'
author:
- 'Daniel R. Wik, A. Hornstrup, S. Molendi, G. Madejski, F. A. Harrison, A. Zoglauer, B. W. Grefenstette, F. Gastaldello, K. K. Madsen, N. J. Westergaard, D. D. M. Ferreira, T. Kitaguchi, K. Pedersen, S. E. Boggs, F. E. Christensen, W. W. Craig, C. J. Hailey, D. Stern, W. W. Zhang,'
title: |
Observations of the Bullet Cluster:\
Constraints on Inverse Compton Emission
---
Introduction {#sec:intro}
============
A number of observations, mainly at radio frequencies, have established that relativistic particles and magnetic fields are part of the intracluster medium (ICM) of galaxy clusters [e.g., @GF04]. The large ($\sim$Mpc) scale, diffuse structures known as radio halos and relics are produced by relativistic electrons spiraling around $\sim$$\mu$G magnetic fields. The synchrotron emission is a product of both the particle and magnetic field energy densities, the latter of which is not well constrained globally from these or other observations. However, the electron population can be independently detected through inverse Compton (IC) scattering off of ubiquitous Cosmic Microwave Background (CMB) photons, which are up-scattered to X-ray energies and may be observable if the electron population is sufficiently intense [@Rep79]. For single electrons or populations with power law energy distributions, the ratio of IC to synchrotron flux gives a direct, unbiased measurement of the average magnetic field strength $B$ in the ICM of a cluster. The magnetic field plays a potentially important role in the dynamics and structure of the ICM, such as in sloshing cool cores where $B$ may be locally amplified so that the magnetic pressure is comparable to the thermal pressure [@ZML11]. Detections of IC emission, therefore, probe whether the non-thermal phase is energetically important or, particularly if the average magnetic field is large, it is sizable enough to affect the dynamics and structure of the thermal gas.
The quest for the detection of IC emission associated with galaxy clusters began with the launch of the first X-ray sensitive sounding rockets and satellites, although the origin of extended, $\sim$ keV X-rays from clusters was soon recognized to be thermal [e.g., @ST72; @MCD+76]. Even so, in clusters with radio halos or relics, IC emission [*must*]{} exist at some level, since the CMB is cosmological. Thermal X-ray photons are simply too numerous at $E \la 10$ keV for a reliable detection of the IC component; at higher energies, however, the bremsstrahlung continuum falls off exponentially, allowing the non-thermal IC emission to eventually dominate and produce “excess” flux in the spectrum. While the first IC searches with yielded only upper limits, and thus lower limits on the average strength of ICM magnetic fields, $B \ga 0.1 \mu$G [@Rep87; @RG88], the next generation of hard X-ray capable satellites – and – produced detections in several clusters, although mostly of marginal significance [for a review, see, e.g., @RNO+08]. The most recent observatories – and – however, have largely failed to confirm IC at similar levels [@Aje+09; @Aje+10; @Wik+12; @Ota12]. One exception is the Bullet cluster (a.k.a. 1E 0657-56, RX J0658-5557), although the detection significance of the non-thermal component is marginal in both the and data alone.
The observation of the Bullet cluster’s had X-ray emission was not very constraining, but the overall spectrum from the PCA and HEXTE instruments, fit jointly with MOS data, favored a non-thermal tail at not quite $3\sigma$ significance [@PML06]. A two-temperature model fit the data equally well, but the higher temperature component had a nearly unphysically high temperature ($\sim 50$ keV) for a large (10%) fraction of the total emissivity. In a similar analysis, the data were simultaneously fit with a spectrum from the BAT all sky survey, and the non-thermal component was confirmed at the $5\sigma$ confidence level [@Aje+10]. However, a two-temperature model technically did a better job of describing the spectra, although the secondary temperature component was very low (1.1 keV), causing the authors to reject this interpretation. While this low temperature component is certainly not physical, the fact that a model can fit the data so well when an extra component is added solely at low energies indicates that the non-thermal component is not being strongly driven by the BAT data. Further confirmation of an IC component in the Bullet cluster is clearly necessary to rule out a purely thermal description of the hard band emission and uphold the implied magnetic field strength of $\sim 0.16 \mu$G.
The intriguing evidence for a non-thermal excess at hard energies coupled with its smaller angular size makes the Bullet cluster an ideal galaxy cluster target for the X-ray observatory [@Har+13]. is the first focusing hard X-ray telescope with a bandpass between 3 and 80 keV and is the first telescope with the ability to focus X-rays in the hard X-ray band above 10 keV. It has an effective area at 30 keV of $2 \times 110$ cm$^{2}$ and imaging half power diameter (HPD) of $58\arcsec$. While the effective area is somewhat lower than that of previous instruments, the focusing capability vastly reduces the background level and its associated uncertainties. Whereas collimators onboard , , and have quite large, $\ga 1$fields of view (FOVs) that include substantial emission from cosmic X-ray background (CXB) sources, the equivalent region of the Bullet cluster within spans $\sim 100\times$ less solid angle on the sky. Also, for clusters that fit well within ’s $\sim 13\arcmin \times 13$FOV, simultaneous offset regions can be used to precisely characterize the background to an extent not possible with collimated instruments.
We describe the two observations and their generic processing in Section \[sec:obs\]. In Section \[sec:cal\], the modeling of the background and its systematics and the overall flux calibration are briefly described (see Appendices \[sec:appendixbgd\] and \[sec:appendixsim\] for details). We examine hard band images and the character of the global spectrum in Section \[sec:analy\]. Lastly, the implications of these results are discussed in Section \[sec:disc\]. We assume a flat cosmology with $\Omega_M = 0.23$ and $H_0 = 70$ km s$^{-1}$ Mpc$^{-1}$. Unless otherwise stated, all uncertainties are given at the 90% confidence level.
Observations and Standard Processing {#sec:obs}
====================================
The Bullet cluster was observed by in two epochs. The optical axis fell near the centroid of the large-scale X-ray emission in the first observation and near the western shock driven by the bullet subcluster in the second. The first pointing was carried out over a little under 3 days, 18–20 October 2012, for a total unfiltered exposure of 231 ks. For the second pointing, the Bullet cluster was observed for a slightly longer raw exposure of 287 ks from 1–4 November 2012. To filter the events, standard pipeline processing (HEASoft v6.13 and NuSTARDAS v1.1.1) was applied along with stricter criteria regarding passages through the South Atlantic Anomaly (SAA) and a “tentacle”-like region of higher activity near part of the SAA; in the call to the general processing routine that creates Level 2 data products, [nupipeline]{}, the following flags are included: [SAAMODE=STRICT]{} and [TENTACLE=yes]{}. These additional flags reduce the cleaned exposure time by $\la 10$% from what it would otherwise be, but also reduce background uncertainties. No strong fluctuations are present in light curves culled from the cleaned events, suggesting a stable background, so no further time periods were excluded.
From the cleaned event files, we directly extract images like those shown in Figure \[fig:rawimgs\] and light curves using [xselect]{}, create exposure maps using [nuexpomap]{}, and extract spectra and associated response matrix (RMF) and auxiliary response (ARF) files using [nuproducts]{}. The call to [nuproducts]{} includes [extended=yes]{}, most appropriate for extended sources, which weights the RMF and ARF based on the distribution of events within the extraction region, assuming that to be equivalent to the true extent of the source. Although the effective smoothing of the source due to the point spread function (PSF) is not folded in with the weighting, the relatively narrow full width at half maximum (FWHM) of $\sim$ 18lessens the impact of this omission. The response across a given detector is largely uniform, so the RMFs of the four detectors are simply averaged by the weighted fraction each detector contributes to a region. In addition to the mirror response, the ARF includes low energy absorption in the detectors (due to a CZT dead layer and platinum electrodes) and is also “corrected” to a canonical power law Crab spectrum of photon index 2.1 and normalization 9.7 photons s$^{-1}$ cm$^{-2}$ keV$^{-1}$ at 1 keV. The remaining products necessary to analyze the spectra – background spectra and a PSF-corrected flux calibration – are tailored for this analysis and described in Section \[sec:cal\].
[ccccccc]{} 700055002 & 104.63207 & -55.924552 & 231 & 126\
700056002 & 104.53211 & -55.919636 & 287 & 140
![Cleaned events projected in sky coordinates from 3–20 keV; pixels with no events are displayed white while pixels with 1 or $\geq 2$ events are displayed grey or black, respectively. Top row: ObsID 700055002 images; Bottom Row: ObsID 700056002 images. The left and right columns show the data from the A and B telescopes, respectively. \[fig:rawimgs\]](rawA0.ps "fig:"){width="5cm"} ![Cleaned events projected in sky coordinates from 3–20 keV; pixels with no events are displayed white while pixels with 1 or $\geq 2$ events are displayed grey or black, respectively. Top row: ObsID 700055002 images; Bottom Row: ObsID 700056002 images. The left and right columns show the data from the A and B telescopes, respectively. \[fig:rawimgs\]](rawB0.ps "fig:"){width="5cm"} ![Cleaned events projected in sky coordinates from 3–20 keV; pixels with no events are displayed white while pixels with 1 or $\geq 2$ events are displayed grey or black, respectively. Top row: ObsID 700055002 images; Bottom Row: ObsID 700056002 images. The left and right columns show the data from the A and B telescopes, respectively. \[fig:rawimgs\]](rawA1.ps "fig:"){width="5cm"} ![Cleaned events projected in sky coordinates from 3–20 keV; pixels with no events are displayed white while pixels with 1 or $\geq 2$ events are displayed grey or black, respectively. Top row: ObsID 700055002 images; Bottom Row: ObsID 700056002 images. The left and right columns show the data from the A and B telescopes, respectively. \[fig:rawimgs\]](rawB1.ps "fig:"){width="5cm"}
Background Modeling and Flux Calibration {#sec:cal}
========================================
One of ’s pioneering technologies, at least for an astrophysics X-ray mission, is the separation of its optics and focal plane modules by an open mast structure that was extended after launch. The telescope is thus open and subject to stray light, which dominates the background at low energies and creates a spatial gradient across the FOV. The stray light must be distinguished from the instrumental background, which varies from detector to detector but is otherwise spatially uniform, in order to use local background regions for any source region. Also, because the PSF scatters some emission outside our extraction region, we must estimate the fraction of the emission collected within the region by convolving the cluster’s true spatial distribution with the PSF, which varies with off-axis angle. Our solution to these challenges is outlined below.
Background {#sec:cal:bgd}
----------
As is typical, the background has both intrinsic and extrinsic components, which for vary in relative importance both spectrally, spatially, and somewhat temporally. For faint sources where the background is a significant fraction of the source counts, it is to some degree inappropriate to naively extract and rescale a spectrum from elsewhere in the FOV to use as a background. However, because the background components are reasonably well understood and stable, we can model its instantaneous composition from source-free regions and, using what we know about the spatial variations of each component, extrapolate that model to the source region. The physical origin of the background components are briefly described below; for details on the specific models and how the background is actually fit with them, see Appendix \[sec:appendixbgd\]. These components are all identified in the spectra shown in Figure \[fig:app:stack\], and it may benefit the reader to refer to it and the following section simultaneously.
### Components {#sec:cal:bgd:comp}
[**Internal**]{}: The radiation environment of ’s orbit leads to a roughly flat background across all channels. An underlying featureless continuum is produced primarily, but probably not entirely, by high energy gamma rays that either pass through the anti-coincidence shield and Compton scatter in the detector or scatter untriggered in the shield itself. The remainder of the internal background consists of various activation and fluorescence lines, which are mostly resolved and only dominate the background between 22-32 keV. Above these energies weaker lines are still present, but the continuum dominates. More details can be found in Appendix \[sec:appendixbgd:overview\] [**Aperture Stray Light**]{}: Because the space between the optics and focal plane benches is not fully baffled, a series of aperture stops protrude from the focal plane bench to block unfocused X-rays from striking the detectors (for a diagram of this geometry, see Figure \[fig:app:geometry\]). Due to technical implementation limitations, the aperture stop does not exclude 100% of the stray light, leaving a few degree window centered on each mirror module. Although the optics bench itself blocks much of the FOV, there remain lines of sight connecting every detector pixel, through the aperture stop, to regions of open sky. The amount of sky visible to any given pixel is location-dependent. Since the CXB is roughly uniform on large scales, the stray light from the CXB through the aperture stop (hereafter called the “Aperture" background) produces a smooth gradient across the detector plane that depends on the orientation of the detectors and the apparent position of the optics module. The CXB spectral shape is consistent with that found by previous missions, and we adopt the canonical [*HEAO-1 A2*]{} spectral model, valid from 3–60 keV [@Bol87]. Due to cosmic variance, the precise normalization for any given observation should be measured intrinsically (see Appendix \[sec:appendixsim\] for details).
[**Reflected and Scattered Stray Light**]{}: Besides direct exposure to sources of stray light, the open geometry of the spacecraft is susceptible to reflected and scattered X-rays from the entire sky. One possible reflecting surface – along with many other parts of the observatory, including the mast – is the backside of the aperture stops, which are clearly visible to the detectors. There are three potential sources of reflected emission: the CXB, the Earth, and the Sun. Because such a large fraction of the sky is visible to the backside of the aperture stops, they are capable of reflecting a contribution of 10-20% of the total unfocused (i.e., “Aperture") CXB emission despite their smaller solid angle and low reflectivity. Assuming the spectrum is unchanged and uniformly illuminates the detectors, this extra emission simply adds to that coming through the aperture stops. Emission from the Sun (“Solar"), and potentially the Earth’s albedo, is much softer and also much more variable. During episodes of high solar activity, the background below $E \sim$ 5–6 keV will be dominated by a $\sim 1$ keV thermal spectrum of solar abundance, but even during less active periods this component accounts for $\sim 40$% of the $E \la 5$ keV total. The “Solar” emission is only present when the satellite is illuminated by the Sun, so there is no doubt as to its origin. There are also some weak fluorescence lines from material elsewhere on the spacecraft, such as the mast, that contribute to the background, although their origin and contribution is still under active investigation.
[**Focused Cosmic Background**]{}: Unlike the above components, there always exists an inherent “background” from other unresolved foreground/background sources within the FOV that are not of primary scientific interest. While subdominant at all energies, the focused CXB (“fCXB") contributes noticeably below 15 keV – having roughly 10% the flux of the “Aperture" CXB – with a slightly softer spectrum than the “Aperture" CXB since it has been modulated by the mirror effective area, which begins to decline above 10 keV.
### Systematic Uncertainties {#sec:cal:bgd:sys}
Although we directly measure the contribution of each component, we do not do so with infinite precision or accuracy. Inaccurately estimated systematic offsets can easily lead to “detections,” especially when the associated precision of a component is overestimated. Faint spectral components, such as IC emission in galaxy clusters, fall into in this category since they tend to reside in background-dominated regimes. Therefore, we must have some sense of the systematic uncertainty intrinsic to the background, and as much as possible to each component of the background. For some components, like the internal background, the systematic uncertainty could in theory be arbitrarily close to 0%. In practice, of course, uncertainties of less than a few percent are difficult to achieve. Components with a cosmic origin, however, have systematic uncertainty floors due to their very natures. While these uncertainties are sometimes large, they may also be well known, as in the case of the CXB.
At higher ($E > 40$ keV) energies, where the internal background strongly dominates, performing the background fitting procedure outlined in Appendix \[sec:appendixsim\] on the first-pass ECDFS survey fields results in an accurate reconstruction of the background level with a standard deviation of $< 3$% after accounting for the effect of statistical fluctuations. Although the real uncertainty may be smaller, the large statistical uncertainties due to the shorter exposure time ($\sim 40$ ks/field) make it difficult to surmise with greater precision. We adopt a conservative uncertainty of 3% for the entire energy range. Because much of this regime is dominated by lines, whose normalizations have independent systematic uncertainties that are dwarfed by their statistical uncertainties, a global shift up or down maximizes this background’s impact on fits to the cluster spectrum.
The shape of the CXB spectrum has been well-measured by other missions [e.g., @TCC+10], and although it may vary on small scales, the larger scales relevant to are unlikely to exhibit noticeable deviations from the average spectrum. The overall normalization, however, depends critically on the number of more rare, brighter sources, which varies from one location to another on the sky. Because we have no way to exclude the brightest sources, even the variance on large scales (0.3–10 deg$^2$/pixel over a total solid angle of 37.2 deg$^2$) can be high. We can eliminate much of this uncertainty by directly measuring it in the non-source regions of an observation. This technique is especially powerful thanks to the strong correlation of the CXB normalization between source and non-source regions. Each CXB point source produces an aperture-shaped (circular aperture stop opening modulated by any fraction blocked by the optics bench) “plateau” of emission across the detectors, so many pixels “see" the same sources seen by other pixels, especially those nearby. However, background and source regions will not contain all the same CXB sources, so a residual uncertainty remains. Based on simulations of the $\log N$–$\log S$ from @KWK+07, we find that for the approximate location of the cluster on the detectors the residual systematic for the aperture CXB is 8% ($1\sigma$).
In principle, scattered and reflected X-rays (contributing at the lowest energies) should be nearly perfectly correlated between all pixels, even if their spatial distribution is not necessarily uniform. Because we do not know exactly where the scattering is taking place, we cannot predict the appearance of this emission like we can for the “Aperture" background. It does not appear to be flat; independent fits of spectra from the various detectors give different normalizations. Unfortunately, it is not yet feasible to empirically determine the shape any more finely than that at this time. Based on the same exercise used to constrain the internal value, we find a systematic uncertainty for the “Solar" component of 10% ($1\sigma$).
For the “fCXB" emission, we can apply a straightforward shorthand estimate of cosmic variance, consistent with the method used for the “Aperture" component but based on an empirical estimate of the variance. We assume a conservative point source detection threshold of $3 \times 10^{-13}$ ergs s$^{-1}$ cm$^{-2}$ (20–30 keV), below which individual sources would not be obvious embedded within the cluster emission (see Section \[sec:analy:imgs\] for details). The variance scales as $\sigma_{\rm CXB}/I_{\rm CXB} \propto \Omega^{-0.5} S_{\rm cut}^{0.25}$, where $\Omega$ is the solid angle on the sky and $S_{\rm cut}$ is the flux limit for excised point sources. For our elliptical source region, shown in Figure \[fig:imgs\], $\Omega = 31$ arcmin$^2$. We can estimate the variance in our observation relative to another measurement assuming a $\log N$–$\log S$ relation of $N(S) \propto S^{-1.5}$. Using the [*HEAO-1 A2*]{} estimate [@Sha83; @BMC00; @RGS+03] with $\Omega = 15.8 \, {\rm deg}^2$, $S_{\rm cut}$(20–30 keV) $= 2.1\times10^{-11}$ erg s$^{-1}$ cm$^{-2}$, and $\sigma_{CXB}/I_{CXB} = 2.8\%$ (1$\sigma$), we find a variance and thus systematic uncertainty of $\sim 42$% ($1\sigma$) for our extraction region.
Flux Calibration {#sec:cal:flux}
----------------
Since we want the total cluster flux, to first order we could simply use as large a region as possible and assume that includes all the emission. However, the PSF wings cause a fraction of the flux to get redistributed far from its true origin on the sky, which results in some emission being scattered beyond the FOV as defined by the detectors. Detector gaps also miss flux, and one just happens to fall across the brightest part of the Bullet during the second observation. These effects require careful correction so that the exposure across the field is accurate.
As mentioned in Section \[sec:obs\], the ARF for an extended source is created by averaging the vignetting function across the region, weighted by the distribution of events. Extended source ARFs are not additionally corrected for any source emission scattered out of the region through the wings of the PSF. To get a proper total flux for the Bullet cluster spectrum, we must take the PSF and estimate the fraction of the total emission captured inside the region. This task is not entirely trivial since not only does the PSF shape vary with off-axis angle, but the off-axis angle varies for any given position on the sky over the course of an observation. Normally one could neglect the variation in shape, as it only becomes a measurable effect for large ($\ga 3\arcmin$) off-axis angles. Because the placement of the cluster in the second observation results in large off-axis angles for its eastern parts, we include these minor adjustments to the PSF shape. Following @Nyn+13, we can construct composite or effective PSFs for our particular observations across the entire cluster, so that each position has an appropriate PSF associated with it. Now armed with a set of position-dependent PSFs (but not energy-dependent), the flux in the wings can be directly computed. Note that the PSF varies weakly as a function of energy; below $\sim 8$ keV, the FWHM is up to 10% broader than it is at higher energies, although the encircled energy fraction within a radius of $\ga 1\arcmin$ agrees to within a few percent at all energies. The latter behavior justifies our use of an energy-independent PSF.
Ideally, we would like to take the true flux distribution from the cluster and convolve it with the PSFs to estimate the redistributed fraction, but above $\sim$ 7 keV is the only telescope capable of making a reliable image. To estimate the fraction of the total flux in the 3–20 keV energy range within our spectral extraction region, we generate PSFs in a 25$\times$28 grid – each position separated by 1 FWHM of 18– and roughly fit them to the A and B telescope images. The extraction region encompasses 95% of the intrinsic flux from the cluster, and a net $\sim 5$% of that is scattered out of the region by the PSF. Thus, in terms of total cluster emission, our spectrum captures $\sim 90$% of the total 3–20 keV flux. When comparing to past observations, our quoted model normalizations and fluxes would then be 10% lower; however, the overall effective area given in the calibration used here is $\sim 15$% lower than that needed to match with XRT and EPIC fluxes, which means our fluxes should also be decreased 15% (this adjustment is present in later [CALDB]{} releases). Since these corrections roughly cancel out, and given the uncertain nature of absolute calibration between telescopes, we do not further adjust the normalizations and fluxes derived from model fits to the spectra.
Images and Spectra {#sec:analy}
==================
Images {#sec:analy:imgs}
------
Although the goal of this paper is to determine the character of the hardest emission in the cluster, we must first confirm that no reasonably bright point sources contaminate that emission. Unlike all previous observatories, ’s unprecedented spatial resolution at hard energies makes a task heretofore impossible as simple as examining the images.
The pipeline-filtered event files are sufficiently processed to produce images, which can be done in arbitrary energy bands by further filtering on the PHA column in, e.g., [xselect]{}. However, calibrated images also require exposure-correction and background-subtraction; the necessary images are generated from [nuexpomap]{} and [nuskybgd]{}, respectively. The latter is not part of the software distribution, but was developed independently as part of this work. (see Appendices \[sec:appendixbgd\] and \[sec:appendixsim\]). We create exposure maps at single energies for each band, which roughly correspond to the mean emission-weighted energy of the band. To mosaic the two epochs along with the data from both telescopes, we also need to correct for offsets due to the $\sim 5$uncertainty in the reconstructed astrometry. No obvious point sources appear within the FOV, so we estimate the necessary shifts using the global distribution of the cluster emission and find slight offsets of 0 to 3 pixels relative to the first epoch’s A telescope astrometry. Because the 2.46pixels significantly oversample the PSF, the final images are smoothed by 5 pixels, more consistent with the PSF’s FWHM of $\sim 18$.
Images in four energy bands (top: 3–8 keV, 8–15 keV; bottom: 15–30 keV, 30-40 keV) are presented in Figure \[fig:imgs\]. The white ellipse shows the extraction region for spectra discussed in Section \[sec:analy:spec\]. From 3–8 keV, the cluster resembles the or images blurred by the larger PSF, except that the “bullet” to the west is relatively de-emphasized since it is composed of cooler ($\la 7$ keV) gas than is the main subcluster ($kT \sim 14$ keV) and ’s response is more sensitive to harder emission in this band. Above 8 keV, the “bullet” essentially disappears, although the halo of shocked gas surrounding it is clearly visible. The cluster begins to approach the level of the background above 15 keV, and above 30 keV whatever detectable emission remains is highly background-dominated. While the overall morphology changes slightly with energy – a subject of a future paper – it does not deviate appreciably from what one would expect extrapolating a temperature map measured at energies $< 8$ keV, suggesting the origin of the $E > 8$ keV emission is also mostly, if not entirely, thermal as well.
Most critically, there is no indication of a background AGN whose emission could masquerade as the non-thermal emission we are searching for. The Bullet cluster is generally free of bright point sources; the contribution of obvious point source emission in the 0.8–4 keV band from a 0.5 Ms mosaic (courtesy M. Markevitch) is $\sim 0.9$% of the total cluster emission, or a flux of roughly $7 \times 10^{-14}$ ergs s$^{-1}$ cm$^{-2}$. Of course, considering these sources alone does not protect us from contamination by absorbed or very hard sources. While there are no bright point sources in our images, ’s large PSF makes it more difficult to distinguish point sources embedded within the diffuse cluster emission. We can estimate the approximate brightness of point sources that we would be able to identify visually by adding a fake source to the data and noting the flux above which the source becomes readily apparent. The resulting flux limit is likely higher than what might be achieved with [wavdetect]{} or some other point source identification method, but relying on simple visual inspection of images is straightforward and sufficient for these purposes. In images covering the entire relevant energy band (3–40 keV), point sources would be clearly identified within a radius of $\sim 1$ if they are $\ga 5$% of the total cluster flux and $\ga 1$% outside this radius, corresponding to flux limits of 2–9 $\times 10^{-13}$ ergs s$^{-1}$ cm$^{-2}$. At higher energies, the signal-to-noise rapidly declines; in the 20–30 keV band point sources only become obvious when they have $\ga 20$% of the cluster emission at those energies, or $\ga 3 \times 10^{-13}$ ergs s$^{-1}$ cm$^{-2}$. These limits are roughly comparable, with the 5% limit above translating to a flux in the 20–30 keV band of $\ga 2 \times 10^{-13}$ ergs s$^{-1}$ cm$^{-2}$, assuming a power law spectrum with a typical photon index of 1.8. Note that the entire FOV of is at least a factor of two smaller than the effective PSF of the HXD-PIN and BAT instruments, further reducing the comparative chance of a point source contaminating the hard X-ray spectrum.
{width="8cm"} {width="8cm"} {width="8cm"} {width="8cm"}
Spectrum {#sec:analy:spec}
--------
Figure \[fig:specraw\] displays the four raw spectra extracted with [nuproducts]{} from the elliptical region illustrated in Figure \[fig:imgs\]. The consistency between the four spectra demonstrate the very similar effective areas between the two telescopes, the shallow vignetting function below 20 keV (the primary difference between the two epochs is the off-axis angle of the cluster centroid), and the stability of the background (whose dominance coincides with the appearance of strong lines just above 20 keV). The detection of an excess above the thermal tail clearly depends critically on the reproducibility of the background. Via the procedure discussed in detail in Appendices \[sec:appendixbgd\] and \[sec:appendixsim\], we have an empirical model for what the background emission should be in this region, based on blank field observations, which has been fit to non-source regions from these observations. This is our best guess for the background spectrum of each observation and telescope, but it is only the most likely state of the background; the actual background may be somewhat different given systematic and statistical uncertainties. A proper background should mimic the statistical impact of the actual background, having both the same area and exposure time of the source region. Typical backgrounds are often taken from larger regions or longer exposures in order to minimize the impact of statistical fluctuations in background regions that could bias the background level in the source region. However, this procedure underestimates the background-subtracted error per channel since the true background suffers from larger statistical uncertainties, which is important wherever the background dominates.
One solution is to jointly fit the background and source data together, but this requires simultaneously fitting 20 spectra each with 3675 unbinnned channels, making it computationally challenging just to find a good fit let alone calculate errors on parameter values. To circumvent this difficulty, we separate the background and source modeling phases but attempt to retain a statistically appropriate treatment of the background. As described in Appendix \[sec:appendixsim\], a nominal background model is found for the source region for each epoch and telescope. We then simulate, for the same exposure time, a background for the region from this model using [fakeit]{} in [XSPEC]{}, including Poisson fluctuations. Of course, the resulting background spectra fail to incorporate any systematic offsets from the nominal model, and statistical fluctuations introduced to the spectrum have the potential to bias fit parameters as well.
Although not the only path forward, we choose to simulate many realizations of the four backgrounds, fitting the spectra with each set. This procedure naturally allows systematic uncertainties to be incorporated as well, since the several background component model normalizations can be randomly varied to reflect those uncertainties. Each background thus represents a possible version of the true background, ideally in proportion to the likelihood that it matches the true background. A similar approach was taken to incorporate background systematic uncertainties in @MGE+11. We assume Gaussian fluctuations about the normalizations of each component with magnitudes given in Section \[sec:cal:bgd:sys\]; Appendix \[sec:appendixsim\] outlines the specific methodology in detail.
For continuum-driven fits on data binned to just above the Gaussian limit (25-30 counts/bin), the $\chi^2$ statistic is known to be biased, especially for fits using a large number of bins [@LM07; @HLB09]. Briefly, the weights $w$ on bins with negative fluctuations are overestimated while bins with positive fluctuations are underestimated, since $w = 1 / \sqrt{N}$, so the $\chi^2$ statistic drives the global best-fit model below the data. The model derived from fits to the background spectra, for example, are biased by 2–3% when $\chi^2$ is used as the fit statistic. To avoid this and similar issues with fitting the Bullet cluster spectra, we use the [XSPEC]{} command [statistic cstat]{}, which applies the W statistic, a Cash-like statistic appropriate for fits with unmodeled background spectra. Bins with no counts have a tendency to confuse the implementation of this statistic in [XSPEC]{}, so we also group the spectra such that there are at least 3 counts in each bin in [*both*]{} data and background spectra.
### Models {#sec:analy:spec:models}
Armed with a reliable way to deal with the background, we can confidently evaluate the nature of the hardest detectable emission from the Bullet cluster. A strong motivation for these observations was to confirm and better characterize the non-thermal component claimed in @Aje+10. In clusters with radio halos and relics, such as the Bullet cluster, IC emission – the only diffuse interpretation for a non-thermal tail – must be present at some level. If the IC emission is bright and begins to dominate the spectrum over the thermal emission at a low enough energy, then the spectrum will be trivial to model. The characterization of a weaker IC component, however, depends much more on the model employed to discriminate it from the thermal emission. Of course, our spectrum falls within the latter regime. The range of models considered is somewhat restricted, but appropriate for the data, consisting of single temperature (1T), two temperature (2T), and single temperature plus power law (T+IC) components. The thermal components are calculated using the version of the @Kaa92 plasma code implemented in [XSPEC]{}.
The 1T model provides the simplest possible description of the spectrum. Emission dominated by isothermal or nearly isothermal gas will be satisfactorily characterized with a single temperature component, since ’s 0.4 keV FWHM resolution does not allow us to easily separate the K$\alpha$ line complexes near He-like and H-like Fe at 6.7 and 6.9 keV, respectively. Therefore, we are entirely reliant on the shape of the largely featureless continuum to discern multi-temperature gas. Given our broad bandpass (3–30 keV), the 1T model is unlikely to account for all the truly thermal emission. From and , spatially resolved spectroscopy clearly demonstrates that the ICM contains gas spanning a large range of temperatures [@GMV+04], which one would expect for an ongoing merger [@Tuc98]. We do not know the true temperature structure, however, only the emission-weighted line-of-sight projected temperature distribution, which is also folded through the effective area and is thus dependent on the calibration and energy band. For the global spectrum, we are not particularly concerned with describing the true temperature structure, since that is not possible. Instead, we wish to accurately represent the part of the temperature distribution seen by , which is more heavily weighted toward the hotter regions and thus may not entirely agree with the projected temperature structure measured within a lower energy bandpass. Because thermal continua are fairly featureless, the 2T model will likely encompass the full range of significant gas temperatures. If the IC emission is sufficiently bright, however, then the higher temperature component of the 2T model will be skewed to an unphysically high value. In this case, the T+IC model should provide a better description of the overall spectrum. Although the thermal component would be imperfectly suited to the true thermal distribution, the harder non-thermal component would better capture the spectral shape at higher energies. Note that the statistical power resides at low energies where the majority of counts are, so the non-thermal excess at high energies must be sufficiently strong to overcome the worsening of the fit quality at the low end.
For the non-thermal component of the T+IC model, we fix the power law photon index to 1.86, the best-fit value found by @Aje+10. We also allowed the index to be a free parameter, but in nearly all cases the index became steeper ($\Gamma \sim 2.4$), where it was most likely mimicking the lower temperature component of the 2T model. This appropriating of the IC component directly results from the greater statistical power of the counts at the low end of the energy range driving the fit. Although the radio synchrotron spectrum basically agrees with this best-fit index, implying $\Gamma \sim 2.3$ for the IC index [@LHB+00], the electrons producing the radio emission are more energetic (for $B \sim 0.2$ $\mu$G, $\gamma \sim$ 23,000 where $\gamma$ is the “relativistic gamma” of the electron) than the ones producing the IC ($\gamma \sim 5000$ at 30 keV), so there is no guarantee the photon index would directly follow, and there is good reason to assume the index flattens at lower energies as is seen in, e.g., the Coma cluster [@TKW03]. For comparison purposes and for our primary result, we choose to fix $\Gamma = 1.86$, which is similar to the typically assumed value of $\sim 2$ in any case.
### Fitting the Three Models {#sec:analy:spec:fits}
Typical fits to the four spectra, using a background spectrum generated from the nominal background model, illustrate the subtle differences between the 1T, 2T, and T+IC descriptions of the Bullet cluster’s spectrum in Figures \[fig:specallmodels\] and \[fig:specratio\]. (Note: all uncertainties quoted in this subsection are purely statistical and are derived using the the nominal backgrounds displayed in the above figures.) In each of these figures, the data, backgrounds, and models for the four spectra have been grouped together for clarity, although the models are folded through each response separately during the fit. The background is also shown to highlight where the spectrum becomes dominated by the background. To zeroth order, the 1T fit is quite good, with a typical $\chi^2_{\rm red} \sim 1.01$. The global temperature of $kT \sim 14.2^{+0.3}_{-0.2}$ keV agrees quite well with the acceptable best-fit temperature range of 13.6 keV $\la kT \la 14.8$ keV [@MGD+02], which varies depending on the value of $N_H$ used. Although higher than the best-fit global temperature of $\sim 12 \pm 0.5$ keV [@PML06], we would expect the average temperature in the 3–30 keV band to be slightly higher than measured in the 1–10 keV band. The and temperature disagreement almost certainly comes down to their respective calibrations, e.g., @NDG10. Despite the somewhat coarse spectral resolution around the Fe lines, the large effective area and exposure time allows the abundance to be well constrained at $0.23 \pm 0.03$ of solar, consistent with those determined from previous observatories, such as [$0.24 \pm 0.04$, @PML06].
Although a 1T model can largely explain the detected emission, a very slight curvature in the residuals of the fit indicates that the spectrum is not of a truly isothermal plasma. Because our sensitivity extends up to higher energies, we can test whether that extra curvature is more likely to come from the true multi-temperature structure of the cluster or an IC component. The 2T model approximates what is actually a fairly smooth, somewhat bimodal, temperature distribution [e.g., @APM07], so the best fit thermal components in this model only roughly correspond to the actual temperatures. Even so, the temperatures we find for the two components are reasonable, with $kT_{\rm high} = 15.3^{+8.4}_{-3.6}$ keV and $kT_{\rm low} = 5.3^{+3.4}_{-3.0}$ keV. Figure \[fig:specallmodels\] shows the relative importance of the fainter component, with the lower temperature accounting for only $\sim 5$% of the 3–30 keV flux. The hard spectral tail up to 30 keV is fully consistent with a thermal spectrum of $\sim 15.3$ keV, only a little higher than the ambient, non-“bullet” ICM temperature of $\sim 14$ keV seen with [@GMV+04] and the 14.2 keV temperature found here with the 1T model. Given that at a minimum there is recently shocked gas at much higher temperatures [@Mar06], a rise in $kT_{\rm high}$ of this magnitude is not surprising. Also, the range in temperatures for $kT_{\rm low}$ agrees very well with temperatures common in both the “bullet” region and nearby [@APM07].
Lastly, we evaluate the likelihood of an IC excess at high energies with the T+IC model. The near success of the 1T model suggests that if a detectable IC component lies at harder energies, the thermal emission should be well accounted for by a single temperature component. We again find a very reasonable temperature of $kT = 13.8^{+0.5}_{-0.2}$ keV, and the resulting 50–100 keV IC flux is $(0.58 \pm 0.40) \times 10^{-12}$ ergs s$^{-1}$ cm$^2$. Again, the thermal component is entirely consistent with that found with previous observatories. The IC flux, on the other hand, falls nearly a factor of 3 below the expected value of $(1.58^{+0.43}_{-0.47}) \times 10^{-12}$ ergs s$^{-1}$ cm$^2$ [@Aje+10]. Unfortunately, the best-fit IC component only surpasses the thermal component at $\ga 40$ keV, where the cluster emission becomes so faint it is lost in statistical fluctuations of the background. To first order, the spectrum appears equally well-fit by the addition of a non-thermal model as by the addition of another thermal model, and the statistical significance of the IC component is possibly high enough to warrant a detection. The inclusion of systematic uncertainties and a detailed comparison of the 2T and T+IC fitting results outlined in the following subsection, however, preclude us from making such a claim.
### Relative Performance of Each Model {#sec:analy:spec:disc}
The mean parameter values and statistical errors were reported in Section \[sec:analy:spec:fits\]; however, true uncertainty ranges must include the impact of both statistical and systematic fluctuations in the background on the fits. The distribution of best-fit temperatures for the three models – found using 1000 realizations of the background for our 4 spectra – are shown in Figure \[fig:ktdist\], immediately illustrating the impact of both background uncertainties and the model we choose to use on our ability to evaluate the spectrum. When the shape of the model is determined by one parameter, as in the 1T case, background uncertainties have only a slight effect on the temperature, creating a spread of only 0.18 keV (compared to the statistical uncertainty of $\sim 0.25$ keV). Adding another parameter that more finely controls the broadband shape of the model (2T or T+IC cases) allows background fluctuations to play a more significant role. For the 2T model, the best-fit temperatures for each component – shown in green in the main panel and inset panel of Figure \[fig:ktdist\] – are much more sensitive to background fluctuations than in either the 1T or T+IC cases, mostly owing to the greater flexibility of the model to adjust to small changes in the shape of the spectrum. Background variations primarily affect the $kT_{\rm high}$ component, since a slightly higher/lower background will cause the spectrum to turn over at a lower/higher energy, thus pushing $kT_{\rm high}$ to lower/higher values. The $kT_{\rm low}$ component then adjusts to “correct” the low energy part of the spectrum; the two temperatures are strongly correlated for a given fit, such that a higher than typical $kT_{\rm high}$ will have a higher than typical $kT_{\rm low}$.
In the T+IC model, the temperature component dominates at all relevant energies and thus maintains the precision of the 1T model’s temperature (despite having a larger statistical error of 0.35 keV). The IC flux, in principle, should be much more sensitive to background systematics than to statistical uncertainties, for the simple reason that its shape more closely matches the background and any systematic shift up or down of the background will correspondingly shift the IC normalization. Tellingly, the uncertainty due to the background on the IC flux ($0.33 \times 10^{-12}$ ergs s$^{-1}$ cm$^2$, 50–100 keV) is slightly less than its statistical uncertainty ($0.4 \times 10^{-12}$ ergs s$^{-1}$ cm$^2$). A true non-thermal excess at high energies should be more affected by background fluctuations. Since the IC flux is not, it is likely driven more by “correcting” (as in the 2T case) the model shape at lower energies, where the signal-to-noise is higher. The IC component is not accounting for truly non-thermal flux in these fits; instead, it substitutes for additional thermal components missing from the single temperature model, likely at both the high [*and*]{} low energy ends of the spectrum. Allowing the IC photon index to be a free parameter further confirms this explanation. For the nominal background, the index steepens to $\sim 2.4$ and mirrors the contribution of the $kT_{\rm low}$ component of the 2T model at low energies, where its continuum shape is nearly identical to that of a $\sim 5$ keV plasma. As shown by the cyan histograms in Figure \[fig:ktdist\], the temperature of the thermal component in this T+IC model is actually hotter than for the 1T case. The hard emission is modeled entirely by the thermal component, while the IC appears to be mimicking the $kT_{\rm low}$ component. This argument alone does not invalidate the IC hypothesis, since in principle the gas could be sufficiently isothermal to allow IC emission to be contributing excess flux at lower energies where the component is mostly being constrained, and it is only coincidental that the spectrum runs out of counts just when the IC component begins to dominate the hard emission. Assuming this viewpoint, a combination of the statistical and systematic uncertainties gives a most likely IC flux of $0.58 \pm 0.52 \times 10^{-12}$ ergs s$^{-1}$ cm$^2$, or a positive fluctuation of less than 2$\sigma$. Considering we have strong reason to believe this component is thermal in nature, it is clear we do not detect IC emission in the global Bullet cluster spectrum with . The results for each model, including statistical and systematic uncertainties, are summarized in Table \[tab:fit\].
Even so, the T+IC model may fit the spectrum better than the 2T model, in which case we might still argue that the spectrum shows evidence of an IC component. The relative quality of the T+IC versus 2T fits depends on the background realization being used, and Figure \[fig:delcstat\] demonstrates that certain backgrounds do in fact favor the T+IC over the 2T model. In this figure, the fits with the 1000 background realizations have been binned according to the difference in C-statistic values between these two models, with values to the right of the vertical lines favoring the 2T model and values to their left favoring the T+IC model. The solid histogram/vertical line correspond to fits with $\Gamma$ fixed to a value of 1.86 and the dashed versions to fits with $\Gamma$ as a free parameter. In the majority of background realizations, the 2T model is preferred, and in only 1.2% of them can the same be said for the T+IC model where $\Gamma$ is fixed. The T+IC model is favored 7.6% of the time when $\Gamma$ is free, although in this case the IC component may simply be mimicking a second thermal component. So while it is most likely the case that the spectrum can best be characterized with a pure thermal model, we cannot rule out an IC flux within the range of fluxes in the right panel of Figure \[fig:ktdist\].
Based on this analysis, a fair 90% upper limit on the IC flux should correctly incorporate both the systematic and statistical uncertainties already discussed. To capture the fact that the T+IC fits prefer a non-zero IC flux, we sum the mean flux with the quadrature-summed uncertainties, yielding an upper limit of $1.1 \times 10^{-12}$ ergs s$^{-1}$ cm$^2$ in the 50–100 keV band.
[lcccccc]{} 1T & $14.2^{+0.3,+0.2}_{-0.2,-0.2}$ & $0.23 \pm 0.03,0.01$ & $1.61 \pm 0.02,0.01$ & - & - & $5717^{+138}_{-138}$\
2T & $15.3^{+8.4,+2.6}_{-3.6,-0.9}$ & $0.22 \pm 0.04,0.01$ & $1.45^{+0.03,+0.12}_{-1.05,-0.32}$ & $5.3^{+3.0,+2.4}_{-3.4,-1.8}$ & $0.22^{+1.12,+0.56}_{-0.26,-0.12}$ & $5708^{+137}_{-138}$\
T+IC & $13.8^{+0.5,+0.2}_{-0.2,-0.2}$ & $0.24 \pm 0.04,0.01$ & $1.51^{+0.10,+0.06}_{-0.03,-0.06}$ & 1.86(fixed) & $0.58^{+0.40,+0.35}_{-0.40,-0.32}$ & $5713^{+137}_{-141}$\
T+IC & $14.6^{+0.4,+0.3}_{-0.4,-0.5}$ & $0.26 \pm 0.05,0.02$ & $1.49^{+0.1,+0.06}_{-0.1,-0.07}$ & $2.4^{\tablenotemark{e},+0.4}_{-1.3,-0.4}$ & $0.12^{+0.06,+0.44}_{-0.06,-0.09}$ & $5710^{+136}_{-140}$
Summary and Discussion {#sec:disc}
======================
Brief Summary {#sec:disc:summary}
-------------
The Bullet cluster was observed by in two epochs for a cumulative 266 ks of conservatively-cleaned exposure time. The cluster is clearly detected below $\sim 30$ keV with an energy-dependent morphology consistent with the extrapolation of projected temperature maps obtained with and . Above $\sim 30$ keV, potential emission associated with the ICM consists of $< 10$% of the counts per channel. The average temperature of the global spectrum is $14.2 \pm 0.3$ keV, in good agreement with estimates from $+$[14.5 keV, @LHB+00] and [14.8 keV, @MGD+02], but somewhat higher than independent estimates from and [$\sim 12$ keV, @PML06]. Given the differences between instrument sensitivity and the accuracy of their respective calibrations, we do not suggest any significant discrepancy.
In order to search for a non-thermal excess above the thermal emission at hard energies, we invested a good deal of effort to understand the largest uncertain factor: the background. We constructed an empirical, spatial-spectral model of the background from blank sky data and applied it to our observations to derive a “most likely” model background spectrum for the region containing cluster emission. After evaluating the important systematic uncertainties in the model, 1000 realizations of the background are generated and each subtracted from the spectrum, which is fit with three spectral models representing a simple (1T) or more realistic (2T) thermal-only origin, or a significant IC component at the highest detectable energies (T+IC), for the emission. In over 98% of the fits, the 2T model was statistically favored over the T+IC model, and reasonable values are obtained for both temperatures in the former. We therefore conclude that no significant non-thermal emission has been detected in the observations of the Bullet cluster and place an upper limit on the IC flux of $1.1 \times 10^{-12}$ ergs s$^{-1}$ cm$^{-2}$ (50–100 keV). This flux falls below that reported by and .
Comparison to and Implications Regarding Previous Results {#sec:disc:prevresults}
---------------------------------------------------------
As mentioned in Section \[sec:intro\], @PML06 first suggested the existence of significant IC emission at hard energies in the Bullet cluster based on a joint analysis of and spectra. The uncertainty in the measurement of $(3.1 \pm 1.9) \times 10^{-12} $ ergs s$^{-1}$ cm$^{-2}$ (50–100 keV) is too large to justify a claim of detection. However, a more recent analysis [@Aje+10], using a BAT spectrum found a flux of $(1.6 \pm 0.5) \times 10^{-12} $ ergs s$^{-1}$ cm$^{-2}$ (50–100 keV), roughly consistent with that from @PML06. Both fluxes are only barely in conflict with our conservative upper limit, but our most likely IC flux of $(0.58 \pm 0.52) \times 10^{-12} $ ergs s$^{-1}$ cm$^{-2}$ (50–100 keV) is clearly inconsistent with these previous measurements.
The origin of the discrepancy has two potential explanations: either the spectra from the various instruments disagree; or the approach to modeling the spectra disagree. While even minor calibration differences between the characterization of the telescope responses and of the backgrounds can significantly affect results, a comparison of the , , and spectra fit to 1T or 2T models implies these are not responsible. None of the instruments on these satellites reliably detect emission above 30 keV from the Bullet, and below this energy there is no compelling excess above a reasonable thermal-only model in Figure 2 of @PML06, the lower left panel of Figure 5 of @Aje+10, or Figure \[fig:specallmodels\] of this paper. At higher energies, the background dominates the count rate and its treatment becomes crucial, where even small fluctuations can result in a false IC signal. It is beyond the scope of this paper to evaluate the backgrounds from the other two missions, but no causes for worry are evident in the analyses of the and data.
If the spectra are all consistent with each other, we must attribute the conflicting conclusions to differences in how the spectra are modeled. In principle there should be no difference, since 1T, 2T, and T+IC models are each tried in all three analyses. The crucial distinction between them is the minimum energy used in the fits: 1 keV [@PML06], 0.5 keV [@Aje+10], or 3 keV (this work). The lower end of the energy range matters because the thermal gas of the Bullet cluster is decidedly [*not*]{} isothermal [@MGD+02], and the fraction of the emission any temperature component contributes strongly varies with energy, with low temperature components dominating at soft energies but essentially disappearing from the hard band. Merging clusters, especially those like the Bullet where one subcluster hosts a cool core, may have components of roughly equal emission measure that span a factor of two in temperature. In particular, the emission coming from the cool core ranges from $kT \la 4$ keV up to 7 keV, has a higher abundance, and mostly contributes at the lowest energies. The gas associated with the main subcluster is hotter, with a central $kT \sim 12$ keV and shocked regions to the W and also to the slight SE with $kT \ga 16$ keV (M. Markevitch, priv. comm.). Given the extreme range in temperatures, even a 2T model may provide an insufficient description of the data over a broad energy range. Ironically, the T$+$IC model might better fit the [*purely thermal*]{} emission more successfully in this case, since a power law with free photon index is able to simultaneously account for emission from components at either extreme of the temperature distribution [e.g., A3112, @BNL07; @LNB+10].
By including data below 3 keV in order to better constrain the thermal component, in all likelihood the larger consequence is to bias the characterization of the thermal component, since only simple spectral models are considered. Because the response of ’s EPIC instruments peaks between 1–2 keV and shot noise, which has a fractional error decreasing with energy, sets the signal-to-noise ratio, fit minimization routines are overly biased to find good fits at these lower energies. The second model in the multi-component fits of @Aje+10, from this perspective, are focused on artificially “fixing” the residuals below 1 or 2 keV with either the second temperature or IC component, and the slope of the IC’s photon index is determined mostly by the data alone, given that the T$+$IC model over-predicts almost every BAT data point. This explanation is less compelling for the $+$analysis of @PML06. In this case, the fact that fits to both the (over 1–10 keV) and (over 3–30 keV) yield the same temperature despite the different energy bands is worrisome; given the multi-temperature structure, one would expect the 3–10 keV temperature from to be hotter than this average, and the 3–10 keV temperature from to be cooler or unchanged.
In contrast, the temperatures in our 2T model roughly agree with the approximately bimodal temperature distribution seen with , lending credence to the still imperfect thermal model approximated with only two components. The much improved spectral resolution of over that of and undoubtedly helps the fit find physical temperatures. For the T$+$IC model, when the photon index is left free, it tends toward a somewhat larger or steeper value where it only influences the lowest energy channels. The IC component, when exhibiting this behavior, mimics a lower temperature thermal component more than it tries to account for any excess emission at high energies, further refuting the existence of a significant non-thermal excess.
By combining the synchrotron spectrum at radio frequencies with an IC estimate or upper limit, we can directly constrain the volume averaged magnetic field strength. Following the arguments and expression for $B$ in Equation 14 of @WSF+09, we use the total radio halo flux of 78 mJy at 1300 MHz and a radio spectral index of 1.2–1.4 [@LHB+00]. The radio spectrum exhibits no flattening at lower frequencies as in @TKW03 for the Coma cluster, so we assume the spectrum continues as a power law to lower frequencies where the electron population producing the synchrotron is the same as those producing the IC. The upper limit on IC emission translates to a lower limit on the magnetic field strength of $B \ga 0.2$ $\mu$G, which is comparable to values found in other clusters using and data [e.g., @Ota12; @Wik+12]. Unlike estimates of $B \sim 0.1$–0.2 $\mu$G, such lower limits are more consistent with equipartition estimates [$\sim$ 1 $\mu$G for the Bullet cluster, @PML06] and Faraday rotation measure estimates in other clusters, which typically place the field strength at a few $\mu$G [e.g., @KKD+90; @CKB01; @BFM+10]. While it is possible to reconcile these estimates with a lower volume averaged value of $B$, our lower limit does not requires it.
Implications for Future IC Searches {#sec:disc:genconcl}
-----------------------------------
In order to detect diffuse, faint IC emission in galaxy clusters, the IC signal must be teased from both thermal and instrumental “backgrounds,” both of which are likely to be brighter than the IC emission itself. While going to harder energies reduces contaminating emission from the thermal gas, it requires a large effective area at high energies and/or low and well-characterized instrumental and/or cosmic backgrounds. Regarding the background, focusing optics like those onboard have clear advantages over non-focusing ones, such as collimators and coded-mask telescopes. The effective area or equivalent sensitivity, however, remains a greater challenge for reflective optics due to the large number – and thus weight – of mirror shells needed. IC photon intensity also declines rapidly with energy, making it exceedingly difficult to detect such emission at high energies given the statistical fluctuations of a realistic background level without a very large effective area. In the foreseeable future, IC emission in hot clusters will only be detectable as a subtle inflection of the thermal tail. Such non-thermal inflections, however, are complicated by having plausible alternative origins, such as background AGN, clumps of super hot gas, and slightly underestimated overall backgrounds. These difficulties, combined with magnetic field equipartition estimates nearly an order of magnitude larger than the field strengths inferred by IC measurements, emphasize the need for a conservative approach. The recent history of IC searches seems to justify this view. @Ota12 nicely summarizes some , , , and detections and upper limits in their Figure 10, which shows that clusters may exhibit an IC signal in the dataset of one observatory but not another – sometimes, but not often, contradictorily. The reasons behind these differences are not always clear, but likely include some combination of relative instrumental calibration, background treatment, and telescope capabilities. Detections are only mildly statistically significant and are in danger of being compromised by the complications mentioned above. The clusters expected to host IC-producing electrons are those undergoing mergers, which produce – possibly extreme – multi-temperature distributions. Such distributions should in principle be straightforward to separate from a non-thermal component, [*if*]{} the IC component begins to dominate the spectrum at an energy where the signal-to-noise is sufficiently high, including systematic uncertainties. For the Bullet cluster, we reach this point around 20–30 keV. The next mission capable of detecting IC emission associated with radio halos is [*Astro-H*]{}, which will include a Hard X-ray Telescope (HXT) and Imager (HXI), with a sensitivity similar to , as well as substantial soft X-ray capabilities with the Soft X-ray Imager (SXI) and X-ray Calorimeter Spectrometer (XCS). Although the HXI alone provides for no improvement over , the SXI and especially the XCS should allow for a more detailed and complete accounting of the thermal components of target clusters through emission line diagnostics. A better understanding of the thermal continuum will make marginal non-thermal-like excesses at hard energies more significant and upper limits more constraining.
If the average magnetic field strength in galaxy clusters hosting radio halos is typically closer to $\sim 1$ $\mu$G than the $\sim 0.2$ $\mu$G implied by past detections, even [*Astro-H*]{} is unlikely to be enough of a technical advance. Because the ratio of synchrotron to IC flux scales with the energy density of the of the magnetic field ($\propto B^2$), a $5\times$ stronger $B$ requires a $25\times$ more sensitive telescope than currently exists. IC emission at this level would only compete with the thermal emission of a Bullet-like cluster between 30–50 keV, and given how faint the cluster is at these energies relative to the background (e.g., Figures \[fig:specallmodels\] and \[fig:specsig\]), it is likely that most of the sensitivity gain will come from increasing the effective area. An increase in effective area over of not quite an order of magnitude would be achieved by the proposed probe class [*HEX-P*]{} mission[^1], so a substantial decrease in background and its systematic uncertainty would still be necessary.
In terms of past IC detections, it may be the case that what has been measured is not IC emission associated with large scale radio halos. Instead of being associated with the electrons producing radio halos and relics, the IC emission might originate from electrons accelerated by accretion shocks at the virial radius [e.g., @KW10; @KKL+12]. Non-imaging telescopes – unlike – would pick up this emission, which peaks in surface brightness $\ga$ Mpc from cluster centers. Given our restricted extraction region around the Bullet cluster, we are not sensitive to these electrons. However, the FOV does partially include the virial region, where we characterized the background, so in principle this IC emission could exist at very faint levels; a cursory check for a non-thermal component was made when the background was fit, but no such signal beyond the generic background model was apparent. Note that these observations are not ideally suited for searches of this emission, which would be better served by several offset pointings around the periphery of the cluster. Even so, the emission would be strongest at the low energy end, where we attribute extra flux detected in the background regions to scattered thermal photons. It should be feasible to constrain these models, but only after a more detailed accounting of the Bullet cluster’s thermal structure has been undertaken, in order to separate local emission from scattered photons from various regions in the cluster. We will address this issue in a future paper focussed on the hard X-ray weighted temperature structure, including extreme temperature shock regions.
This research was supported by an appointment to the NASA Postdoctoral Program at the Goddard Space Flight Center, administered by Oak Ridge Associated Universities through a contract with the National Aeronautics and Space Administration (NASA) and made use of data from the mission, a project led by the California Institute of Technology, managed by the Jet Propulsion Laboratory, and funded by NASA. We thank the Operations, Software and Calibration teams for support with the execution and analysis of these observations. This research has made use of the Data Analysis Software (NuSTARDAS) jointly developed by the ASI Science Data Center (ASDC, Italy) and the California Institute of Technology (USA). The authors wish to thank Maxim Markevitch for providing a 0.5 Ms image of the Bullet cluster to confirm the cluster region lacks bright point sources.
[46]{} natexlab\#1[\#1]{}
Ajello, M., Rebusco, P., Cappelluti, N., et al. 2009, , 690, 367 Ajello, M., Rebusco, P., Cappelluti, N., Reimer, O., Böhringer, H., La Parola, V., & Cusumano, G. 2010, , 725, 1688 Andersson, K., Peterson, J. R., & Madejski, G. 2007, , 670, 1010 Barcons, X., Mateos, S., & Ceballos, M. T. 2000, , 316, L13 Boldt, E. 1987, Observational Cosmology, 124, 611 Bonafede, A., Feretti, L., Murgia, M., et al. 2010, , 513, A30 Bonamente, M., Nevalainen, J., & Lieu, R. 2007, , 668, 796 Brunetti, G., & Blasi, P. 2005, , 363, 1173 Buote, D. A. 2001, , 553, L15 Clarke, T. E., Kronberg, P. P., B[ö]{}hringer, H. 2001, , 547, L111 Govoni, F., & Feretti, L. 2004, International Journal of Modern Physics D, 13, 1549 Govoni, F., Markevitch, M., Vikhlinin, A., et al. 2004, , 605, 695 Harrison, F. A., Craig, W. W., Christensen, F. E., et al. 2013, , 770, 103 Humphrey, P. J., Liu, W., & Buote, D. A. 2009, , 693, 822 Kaastra, J. S. 1992, An X-Ray Spectral Code for Optically Thin Plasmas (Internal SRON-Leiden Rep., updated version 2.0) Keshet, U., Kushnir, D., Loeb, A., & Waxman, E. 2012, arXiv:1210.1574 Kim, K.-T., Kronberg, P. P., Dewdney, P. E., & Landecker, T. L. 1990, , 355, 29 Kim, M., Wilkes, B. J., Kim, D.-W., et al. 2007, , 659, 29 Kushnir, D., & Waxman, E. 2010, Journal of Cosmology and Astroparticle Physics, 2, 25 Leccardi, A., & Molendi, S. 2007, , 472, 21 Lehto, T., Nevalainen, J., Bonamente, M., Ota, N., & Kaastra, J. 2010, , 524, A70 Liang, H., Hunstead, R. W., Birkinshaw, M., & Andreani, P. 2000, , 544, 686 Mantz, A., Allen, S. W., Ebeling, H., & Rapetti, D. 2008, , 387, 1179 Markevitch, M., Gonzalez, A. H., David, L., et al. 2002, , 567, L27 Markevitch, M. 2006, The X-ray Universe 2005, 604, 723 Markwardt, C. B. 2009, Astronomical Data Analysis Software and Systems XVIII, 411, 251 Mitchell, R. J., Culhane, J. L., Davison, P. J. N., & Ives, J. C. 1976, , 175, 29P Moretti, A., Gastaldello, F., Ettori, S., & Molendi, S. 2011, , 528, A102 Nevalainen, J., David, L., & Guainazzi, M. 2010, , 523, A22 Nynka, M., Hailey, C. J., Mori, K., et al. 2013, , 778, L31 Ota, N. 2012, Research in Astronomy and Astrophysics, 12, 973 Petrosian, V., Madejski, G., & Luli, K. 2006, , 652, 948 Rephaeli, Y. 1979, , 227, 364 Rephaeli, Y. 1987, , 225, 851 Rephaeli, Y., & Gruber, D. E. 1988, , 333, 133 Rephaeli, Y., Nevalainen, J., Ohashi, T., & Bykov, A. M. 2008, , 134, 71 Revnivtsev, M., Gilfanov, M., Sunyaev, R., Jahoda, K., & Markwardt, C. 2003, , 411, 329 Sarazin, C. L. 1999, , 520, 529 Schuecker, P., B[ö]{}hringer, H., Reiprich, T. H., & Feretti, L. 2001, , 378, 408 Shafer, R. A. 1983, PhD. Thesis, University of Maryland Snowden, S. L., Mushotzky, R. F., Kuntz, K. D., & Davis, D. S. 2008, , 478, 615 Solinger, A. B., & Tucker, W. H. 1972, , 175, L107 Thierbach, M., Klein, U., & Wielebinski, R. 2003, , 397, 53 Tucker, W., Blanco, P., Rappoport, S., et al. 1998, , 496, L5 T[ü]{}rler, M., Chernyakova, M., Courvoisier, T. J.-L., et al. 2010, , 512, A49 Wik, D. R., Sarazin, C. L., Finoguenov, A., Matsushita, K., Nakazawa, K., & Clarke, T. E. 2009, , 696, 1700 Wik, D. R., Sarazin, C. L., Zhang, Y.-Y., et al. 2012, , 748, 67 Vanderlinde, K., et al. 2010, arXiv:1003.0003 Vikhlinin, A., et al. 2009, , 692, 1060 ZuHone, J. A., Markevitch, M., & Lee, D. 2011, , 743, 16
Definition of the Background Model {#sec:appendixbgd}
==================================
Overview {#sec:appendixbgd:overview}
--------
The observatory design gives rise to various, independent background components that vary spatially across the FOV, complicating standard background estimation techniques. The purpose of this Appendix is to describe the behavior of the spatial variation of the various background components in the detector plane and how to translate to the sky frame used in source analysis. In this framework, we make assumptions about the spatial and spectral features of the several background components, that are physically motivated but empirically determined, to allow a model of the background for the entire FOV to be based on the characterization of only non-source regions. The model itself derives from fits to stacked “blank” field observations, taken from the deep (ECDFS) and medium (COSMOS) survey data, listed in Table \[tab:app:blankfields\].
[lcc]{} 60022001\_ECDFS\_MOS001 & 60022001002 & 41.2\
60022002\_ECDFS\_MOS002 & 60022002001 & 43.1\
60022003\_ECDFS\_MOS003 & 60022003001 & 43.2\
60022004\_ECDFS\_MOS004 & 60022004001 & 43.7\
60022005\_ECDFS\_MOS005 & 60022005001 & 42.3\
60022006\_ECDFS\_MOS006 & 60022006001 & 41.6\
60022007\_ECDFS\_MOS007 & 60022007002 & 44.3\
60022008\_ECDFS\_MOS008 & 60022008001 & 43.2\
60022009\_ECDFS\_MOS009 & 60022009001 & 41.7\
60022010\_ECDFS\_MOS010 & 60022010001 & 42.7\
60022011\_ECDFS\_MOS011 & 60022011001 & 43.4\
60022012\_ECDFS\_MOS012 & 60022012001 & 43.9\
60022013\_ECDFS\_MOS013 & 60022013001 & 44.6\
60022014\_ECDFS\_MOS014 & 60022014001 & 44.8\
60022015\_ECDFS\_MOS015 & 60022015001 & 45.1\
60022016\_ECDFS\_MOS016 & 60022016001 & 42.1\
60022001\_ECDFS\_MOS001 & 60022001003 & 40.9\
60022002\_ECDFS\_MOS002 & 60022002002 & 41.2\
60022003\_ECDFS\_MOS003 & 60022003002 & 41.0\
60022004\_ECDFS\_MOS004 & 60022004002 & 41.2\
60022005\_ECDFS\_MOS005 & 60022005002 & 41.2\
60022006\_ECDFS\_MOS006 & 60022006002 & 41.3\
60022007\_ECDFS\_MOS007 & 60022007003 & 41.7\
60022008\_ECDFS\_MOS008 & 60022008002 & 41.8\
60022009\_ECDFS\_MOS009 & 60022009003 & 41.6\
60022010\_ECDFS\_MOS010 & 60022010002 & 27.9\
60022010\_ECDFS\_MOS010 & 60022010004 & 13.2\
60022011\_ECDFS\_MOS011 & 60022011002 & 41.6\
60022012\_ECDFS\_MOS012 & 60022012002 & 41.9\
60022013\_ECDFS\_MOS013 & 60022013002 & 41.6\
60022014\_ECDFS\_MOS014 & 60022014002 & 44.1\
60022015\_ECDFS\_MOS015 & 60022015003 & 43.5\
60022016\_ECDFS\_MOS016 & 60022016003 & 43.9
[lcc]{} 60021001\_COSMOS\_MOS001 & 60021001002 & 18.8\
60021002\_COSMOS\_MOS002 & 60021002001 & 22.6\
60021003\_COSMOS\_MOS003 & 60021003001 & 20.5\
60021004\_COSMOS\_MOS004 & 60021004001 & 21.8\
60021005\_COSMOS\_MOS005 & 60021005001 & 21.5\
60021006\_COSMOS\_MOS006 & 60021006001 & 21.7\
60021007\_COSMOS\_MOS007 & 60021007001 & 22.9\
60021008\_COSMOS\_MOS008 & 60021008001 & 23.3\
60021009\_COSMOS\_MOS009 & 60021009002 & 22.8\
60021010\_COSMOS\_MOS010 & 60021010001 & 24.4\
60021011\_COSMOS\_MOS011 & 60021011001 & 25.9\
60021012\_COSMOS\_MOS012 & 60021012001 & 23.0\
60021013\_COSMOS\_MOS013 & 60021013001 & 25.3\
60021014\_COSMOS\_MOS014 & 60021014001 & 22.7\
60021015\_COSMOS\_MOS015 & 60021015001 & 23.4\
60021016\_COSMOS\_MOS016 & 60021016001 & 25.5\
60021017\_COSMOS\_MOS017 & 60021017001 & 22.9\
60021018\_COSMOS\_MOS018 & 60021018001 & 24.0\
60021019\_COSMOS\_MOS019 & 60021019001 & 28.9\
60021020\_COSMOS\_MOS020 & 60021020002 & 28.0\
60021021\_COSMOS\_MOS021 & 60021021001 & 27.4\
60021022\_COSMOS\_MOS022 & 60021022001 & 22.0\
60021023\_COSMOS\_MOS023 & 60021023001 & 24.6\
60021024\_COSMOS\_MOS024 & 60021024001 & 25.4\
60021025\_COSMOS\_MOS025 & 60021025001 & 22.1\
60021026\_COSMOS\_MOS026 & 60021026001 & 28.9\
60021027\_COSMOS\_MOS027 & 60021027002 & 23.9\
60021028\_COSMOS\_MOS028 & 60021028001 & 22.9\
60021029\_COSMOS\_MOS029 & 60021029001 & 22.9\
60021030\_COSMOS\_MOS030 & 60021030001 & 24.0\
60021031\_COSMOS\_MOS031 & 60021031001 & 22.2\
60021032\_COSMOS\_MOS032 & 60021032001 & 25.8\
60021033\_COSMOS\_MOS033 & 60021033001 & 21.7\
60021034\_COSMOS\_MOS034 & 60021034001 & 19.4\
60021034\_COSMOS\_MOS034 & 60021034003 & 9.4\
60021035\_COSMOS\_MOS035 & 60021035002 & 21.4\
60021036\_COSMOS\_MOS036 & 60021036002 & 22.7\
60021037\_COSMOS\_MOS037 & 60021037002 & 24.1\
60021038\_COSMOS\_MOS038 & 60021038001 & 23.1\
60021039\_COSMOS\_MOS039 & 60021039001 & 22.3\
60021040\_COSMOS\_MOS040 & 60021040001 & 23.8\
60021041\_COSMOS\_MOS041 & 60021041001 & 22.2\
60021042\_COSMOS\_MOS042 & 60021042002 & 22.4\
60021043\_COSMOS\_MOS043 & 60021043001 & 23.8\
60021044\_COSMOS\_MOS044 & 60021044002 & 21.7\
60021046\_COSMOS\_MOS046 & 60021046002 & 17.6\
60021046\_COSMOS\_MOS046 & 60021046004 & 13.0\
60021047\_COSMOS\_MOS047 & 60021047002 & 24.9\
60021048\_COSMOS\_MOS048 & 60021048002 & 24.9\
60021049\_COSMOS\_MOS049 & 60021049002 & 24.7\
60021053\_COSMOS\_MOS053 & 60021053002 & 11.4\
60021053\_COSMOS\_MOS053 & 60021053004 & 5.7\
60021053\_COSMOS\_MOS053 & 60021053006 & 6.2
consists of two separate telescopes (two sets of optics, housed in the optics module, focusing onto two focal planes, housed in the focal plane module) The telescopes, or associated data/response functions, are referred to as A and B. Each focal plane consists of a 2$\times$2 array of CdZnTe detectors with a 32$\times$32 array of pixels. In principle, each pixel has a unique background response, but in practice all the pixels on a single detector – excepting edge pixels – behave similarly. Due to differences in thickness and other properties of the detectors, the instrumental background for each detector is somewhat unique.
The benches containing the optics and detectors are separated on two ends of an unenclosed mast. Pointing variations throughout a given observation cause a given detector pixel to sample several times more sky than without this wobble. Because the light path is open to space, stray light from the cosmic X-ray background (CXB) is able to skirt between the optics bench and the aperture stops in front of the two focal planes (Figure \[fig:app:geometry\]); the geometry of this window produces highly non-uniform background gradients across the detectors at low energy ($E \la 15$ keV). At the lowest energies, scattered solar X-rays reflected from other parts of the observatory structure are visible to the detectors, due to its open design. The low altitude and inclination orbit of minimizes SAA activation and proton flares, so the instrumental background dominating at higher energies is low and stable.
![[*Left:*]{} A schematic of the observatory that illustrates how far off-axis sources can directly shine on the detectors through the aperture stop, producing the “Aperture” background. In this example, rays from the source are shielded from striking the left detector plane by the optics bench, but other rays from the same source have an unimpeded path through the aperture stop to shine on a corner of the right detector plane. [*Right:*]{} The location of sources on the sky, as visible from the detector plane, that produce the “Aperture” background for Telescopes A (top) and B (bottom). The images are weighted (darker) by the number of detector pixels a given source shines on. The crosses give the approximate position of the source shown in the left panel. \[fig:app:geometry\]](aperturebgd.eps "fig:"){height="12cm"} ![[*Left:*]{} A schematic of the observatory that illustrates how far off-axis sources can directly shine on the detectors through the aperture stop, producing the “Aperture” background. In this example, rays from the source are shielded from striking the left detector plane by the optics bench, but other rays from the same source have an unimpeded path through the aperture stop to shine on a corner of the right detector plane. [*Right:*]{} The location of sources on the sky, as visible from the detector plane, that produce the “Aperture” background for Telescopes A (top) and B (bottom). The images are weighted (darker) by the number of detector pixels a given source shines on. The crosses give the approximate position of the source shown in the left panel. \[fig:app:geometry\]](apsky.ps "fig:"){height="12cm"}
The background spectrum can generally be decomposed into four broadband components of fixed spectral shape. For certain observations near the Galactic plane, an additional component to account for diffuse Galactic Ridge emission (GRXE) may also be needed. The spectral components, fit to a stacked spectrum of the observations from Table \[tab:app:blankfields\] for the entire FOV, are shown in Figure \[fig:app:stack\].
![Initial independent fits to stacked blank sky data for telescopes A (black) and B (grey or red). The spectra for each of the four detectors in each focal plane are averaged together when combined, such that the rates shown are also per detector. The major contributions are labeled according to their source with the “Aperture,” “fCXB,” and “Solar” components having a cosmic origin and the “Int. Cont.” and “Int. Lines” having an “internal” or instrumental origin, due largely to the spacecraft environment. Because the spectral shapes are identical, the “Aperture" component consists of both emission directly from the CXB through the aperture stop and from CXB emission reflected/scattered off the backside of the aperture stop and/or other parts of the telescope. \[fig:app:stack\]](appspecAB.ps){width="11cm"}
Below $\sim 20$ keV, the background is dominated by stray light from unblocked sky leaking through the aperture stop; when the origin of this emission is the CXB, this component is referred to as the “Aperture” background. By their nature, the background produced by the CXB, GRXE, or bright sources with a line of sight through the stop is spatially non-uniform. The FOV samples a solid area of 37.2 deg$^2$ on the sky, with any individual pixel exposed to something in the range 0.3–10 deg$^2$. Below $\sim 5$ keV, there is a strong, soft additional component (“Solar”) that is most likely due to reflected solar X-rays as evidenced by its persistence in spectra from Earth observations when the satellite is illuminated by the Sun and its absence when not. This component can undergo significant fluctuations due to solar activity. Although thought to come primarily from reflections off the backside of the aperture stop, this conjecture has yet to be confirmed and thus we have no way to predict the spatial pattern it produces on the detectors, but it is likely non-uniform as well. The other low energy contributor to the background is from the CXB “focused” by the optics (“fCXB"). The “fCXB" includes both truly focused events (photons reflected off of both mirrors) and scattered events or ghost rays (photons reflecting off of only one mirror) from the many unresolved sources both within and outside the FOV. Its shape is roughly flat across the detector plane despite vignetting due to an increase in scattered light from sources outside the FOV at larger off-axis angles.
Above $\sim 15$–$20$ keV, the internal or instrumental background dominates. It is made up of gamma rays Compton scattered by the detector and shield, lines activated by interactions between the spacecraft/detectors and the radiation environment in orbit, and a few fluorescence lines. Most of the lines are driven by frequent – if glancing – passages through the SAA, when protons activate material in the focal plane module near or in the detectors. Unstable elements are created by proton spallation and secondary neutron capture by cadmium, which then radioactively decay with half-lives typically longer than ’s orbital period. The strongest of these activation lines appear in the complex from 22–25 keV. While the strengths of these lines depend on the spacecraft’s recent orbital history, there is as yet no evidence for spatial variations across individual detectors, and the relative strength of a given line between detectors – which depends on properties unique to each detector such as its thickness – does not vary. The strongest instrumental lines are due to K-shell fluorescence of Cesium and Iodine at 28 keV and 31 keV, respectively, residing in the anti-coincidence shield.
The continuum, meant to represent the Compton scattered component and any other featureless instrumental components, is modeled as a broken power law with a break at 124 keV. The lines and line complexes are modeled with 29 Lorentzian-profile lines, [*empirically*]{} added to the spectra in Figure \[fig:app:stack\] until the fit can no longer be reasonably improved. Initially, the line energies and widths, which are tied between the A and B spectra, are allowed some freedom during the fitting process – as is the temperature describing the “Solar” component and the indices of the internal continuum – but at some arbitrary point the model is deigned to be “good enough” and those parameters fixed thereafter. The internal (and perhaps “Solar”) components exhibit no detectable spatial variation within individual detectors, but they do between detectors.
Spatial Distribution of the “Aperture” Background {#sec:appendixbgd:aperbgd}
-------------------------------------------------
To first order, the CXB has a constant surface brightness across the sky. The intensity detected by a given pixel thus depends on the solid angle of visible sky, which is solely a function of the observatory’s geometry. Each pixel “sees” a solid angle of $\sim 12$ deg$^2$ defined by the circular aperture stop. The view is blocked, however, by the apparent position of the optics bench, which depends on the location of the pixel in the focal plane, so the level of CXB flux smoothly varies across the detectors. Despite understanding this geometry, the absolute position of the focal plane detectors in the bench is uncertain at the 1 mm level. Also, just as for the “Solar” component, CXB emission from the entire rear hemisphere of the sky – except that blocked by the Earth – can be scattered by the backside of the aperture stop and other parts of the observatory into the focal plane, thus modulating its spatial distribution. Using CXB focal plane maps generated by ray traces through the observatory’s geometry, we can adjust the precise position of the detectors within the focal plane and the proportion of unmodulated, scattered CXB flux until we obtain a good match to stacked images from the blank sky observations.
To isolate the “Aperture” component, we stack 7–15 keV images from all the observations in detector (DET1) coordinates, which has a finer spatial resolution than the native pixels (possible due to probability distribution functions relating to event grades obtained from pencil-beam ground calibrations); the stacked images are shown in the left panels of Figure \[fig:app:det1im\]. The 40$\times$40 mm detector plane is binned into cells with sides 2 to 4 mm long and fit to the ray trace model using the $\chi^2$ statistic and minimization package [MPFIT]{} [@Mar09]. In addition to fitting for the “Aperture” model $x$ and $y$ positions and normalizations, we also include simple spatial models for the internal, “Solar,” and “fCXB” components, the normalizations of which are allowed to vary. The relative flux assigned to each component is initially inconsistent with the expectation from Figure \[fig:app:stack\], when the “Aperture" spectral component only includes direct emission through the aperture stop. To reconcile the spatial and spectral “Aperture" models, the spatial model requires extra flat emission of uncertain origin. One possibility is that Earth albedo or CXB photons are scattered off of the backside of the aperture stop and elsewhere and into the detector housing. Spectra extracted during Earth-occulted periods, for example, exhibit a component below 15 keV with the same spectral shape as the “Aperture” component, even though the CXB is not directly visible (see Figure \[fig:app:solar\]). Alternatively, a contributor to the internal continuum component may rise with decreasing energy instead of following the simple power law spectrum we assume. The spectral shape of this component is hard to predict, and we make no attempt to do so, so such a rise is very plausible and would be consistent with it being spatially flat. For simplicity, the spatial model of the “Aperture” component is modified to include this extra emission, rather than adjusting the spectral shape of internal continuum. The amount of extra emission added is increased until we achieve self-consistency between spectral and spatial fits to the data without large shifts in the position of the detectors. We also performed simultaneous spectro-spatial fits of similarly binned regions, and while they are the most comprehensive, they are too computationally intensive and fickle to arrive at best-fit detector offsets and extra “Aperture” emission. However, these fits were useful to explore the parameter space, as were fits to the full FOV spectra, informing the level to which each component should contribute to the 7–15 keV images. This iterative procedure results in extra “Aperture" fractions of 13% for each focal plane and position offsets of ($-3.4$, $2.0$) and ($-3.5$, $1.6$) in ($x$, $y$) for A and B, respectively. These exact values are irrelevant as long as the “Aperture” shape is correct; they only matter if one wants to extract an absolute flux for the CXB using this data. The model created with these values is shown in the right panels of Figure \[fig:app:det1im\].
Determining the Complete Background Model {#sec:appendixbgd:modeldef}
-----------------------------------------
Because the origin of the “Aperture” component of the background is well understood, we were able to characterize its variation across the FOV with high confidence. The spatial distribution of the two other cosmic sources – “fCXB” and “Solar” – of the background have yet to be as well-constrained. Although the “fCXB” distribution can in principle be simulated, at the time of this writing the model of the mirror modules is still being refined to account for observed ghost ray patterns. (Ghost rays are photons typically scattered by the optics, usually once-reflected, from sources within $\sim 1\arcdeg$ of the optical axis.) The pre-launch model predicts a distribution somewhat following the vignetting function; however, the blank sky fields show no evidence of such a spatial modulation. Observed ghost ray patterns produced by bright sources near to but outside the FOV show an additional halo farthest from the source, likely due to reflections off the back sides of the mirrors. The extra contribution due to this halo from CXB sources outside the FOV may act to compensate for the drop in flux from higher off-axis sources within the FOV. Empirically, the spatial shape of the “fCXB” is consistent with a flat distribution, although due to its relative faintness it is difficult to discern otherwise. We assume a flat distribution hereafter for simplicity.
The “Solar” component has only recently been recognized as originating from the Sun through reflections off the observatory structure. No study of its likely spatial distribution has been undertaken, and in any case the distribution may vary with Sun angle. To allow for spatial variations, we treat the “Solar” continuum and associated 3.5 keV and 4.5 keV lines as if they had an instrumental origin and thus should only vary between detectors and not within them. Although this treatment amounts to a very coarsely defined spatial model, this component is typically only important below 5 keV where sources are brightest. We note, however, that X-ray emission from the Sun is highly variable and that during flares this component can dominate up to 10 keV; such periods are not currently handled by the background model described here, since the spectrum itself is likely to evolve from the quiescent one we include. When data are split between the periods that the spacecraft is and is not illuminated by the Sun, the correlation between the soft emission and a solar origin is clear, as shown in the examples in Figure \[fig:app:solar\], the four panels of which also demonstrate its variability.
In contrast, the components of the background with an instrumental origin should not depend on position within the FOV, as long as the detectors are all uniform and identical. Consisting of single CdZnTe crystals, each individual detector should be very close to uniform, which agrees with the lack of spatial fluctuations across any given detector in the stacked high energy images of the blank sky fields. The detectors are not identical, however, and the variation between them in thickness and charge transport properties lead to slight differences in overall background level and line strengths. For any given observation, the overall level and strengths depend on the orbital history through the South Atlantic Anomaly and other higher radiation zones. Since all the detectors share this history, the [*relative*]{} strengths of the internal components should always be the same. To complete our empirical spectro-spatial model of the background, we simply need to determine the ratios between these components for the detectors on each of the two focal planes. We separate each full FOV spectrum from Figure \[fig:app:stack\] into four spectra corresponding to each detector, which each share the same spectral model shapes. Having previously determined the “Aperture” and “fCXB” model spatial shapes, the relative proportion of their flux falling on each detector is fixed appropriately, but all other model normalizations are left free. The four spectra are then fit simultaneously, and independently for each telescope, and the resulting fits are shown in Figure \[fig:app4det\]. Many of the lines have similar strengths on each detector, but that is not universally true. Table \[tab:app:ratios\] gives the fraction of the model normalization of each component associated with each detector.
![The background model fit to the blank sky spectra from focal planes A (left panel) and B (right panel), separated by detector: Det0 (black), Det1 (red), Det2 (green), and Det3 (blue). The spectral shapes of each component are fixed, but the normalizations are free to vary to account for differences between detectors. \[fig:app4det\]](bgdcombA.ps "fig:"){width="9cm"} ![The background model fit to the blank sky spectra from focal planes A (left panel) and B (right panel), separated by detector: Det0 (black), Det1 (red), Det2 (green), and Det3 (blue). The spectral shapes of each component are fixed, but the normalizations are free to vary to account for differences between detectors. \[fig:app4det\]](bgdcombB.ps "fig:"){width="9cm"}
[rr|cccc|cccc]{} 3.54 & 0.43 & 0.217 & 0.228 & 0.326 & 0.228 & 0.235 & 0.231 & 0.209 & 0.324\
4.51 & 0.54 & 0.165 & 0.141 & 0.393 & 0.301 & 0.305 & 0.219 & 0.157 & 0.319\
10.20 & 0.64 & 0.035 & 0.000 & 0.425 & 0.540 & 0.308 & 0.194 & 0.148 & 0.350\
19.65 & 0.23 & 0.295 & 0.057 & 0.220 & 0.427 & 0.167 & 0.156 & 0.407 & 0.270\
21.89 & 0.57 & 0.182 & 0.140 & 0.248 & 0.430 & 0.179 & 0.259 & 0.330 & 0.232\
22.97 & 0.15 & 0.257 & 0.204 & 0.235 & 0.305 & 0.230 & 0.228 & 0.311 & 0.231\
24.75 & 1.96 & 0.204 & 0.166 & 0.273 & 0.357 & 0.225 & 0.248 & 0.290 & 0.237\
25.30 & 0.15 & 0.316 & 0.256 & 0.231 & 0.198 & 0.250 & 0.245 & 0.279 & 0.226\
27.75 & 1.71 & 0.543 & 0.457 & 0.000 & 0.000 & 0.232 & 0.252 & 0.265 & 0.251\
28.08 & 2.06 & 0.000 & 0.000 & 0.511 & 0.489 & 0.250 & 0.250 & 0.250 & 0.250\
28.55 & 0.27 & 0.285 & 0.310 & 0.195 & 0.210 & 0.266 & 0.251 & 0.244 & 0.239\
30.17 & 0.71 & 0.222 & 0.182 & 0.281 & 0.315 & 0.245 & 0.220 & 0.277 & 0.257\
30.86 & 0.45 & 0.273 & 0.272 & 0.235 & 0.220 & 0.248 & 0.259 & 0.255 & 0.238\
32.19 & 0.86 & 0.236 & 0.201 & 0.259 & 0.304 & 0.267 & 0.250 & 0.244 & 0.239\
35.03 & 0.82 & 0.249 & 0.277 & 0.277 & 0.198 & 0.265 & 0.255 & 0.227 & 0.254\
39.25 & 9.13 & 0.238 & 0.150 & 0.299 & 0.313 & 0.252 & 0.235 & 0.253 & 0.260\
39.40 & 0.52 & 0.231 & 0.215 & 0.234 & 0.320 & 0.225 & 0.239 & 0.269 & 0.267\
47.56 & 6.82 & 0.261 & 0.193 & 0.286 & 0.260 & 0.246 & 0.252 & 0.245 & 0.256\
52.50 & 1.60 & 0.241 & 0.193 & 0.260 & 0.306 & 0.233 & 0.245 & 0.266 & 0.256\
57.99 & 4.72 & 0.237 & 0.162 & 0.257 & 0.345 & 0.247 & 0.255 & 0.233 & 0.265\
65.01 & 5.24 & 0.230 & 0.162 & 0.270 & 0.338 & 0.228 & 0.218 & 0.271 & 0.283\
67.06 & 0.53 & 0.233 & 0.279 & 0.193 & 0.296 & 0.278 & 0.267 & 0.210 & 0.245\
75.18 & 5.59 & 0.204 & 0.300 & 0.191 & 0.305 & 0.242 & 0.205 & 0.259 & 0.294\
85.82 & 7.58 & 0.219 & 0.236 & 0.274 & 0.271 & 0.274 & 0.241 & 0.260 & 0.225\
87.90 & 0.58 & 0.299 & 0.279 & 0.199 & 0.223 & 0.241 & 0.260 & 0.238 & 0.260\
92.67 & 0.64 & 0.279 & 0.282 & 0.210 & 0.228 & 0.242 & 0.257 & 0.247 & 0.254\
105.36 & 0.46 & 0.272 & 0.335 & 0.172 & 0.221 & 0.237 & 0.246 & 0.238 & 0.279\
122.74 & 2.30 & 0.275 & 0.458 & 0.076 & 0.192 & 0.204 & 0.261 & 0.263 & 0.272\
144.56 & 0.74 & 0.300 & 0.294 & 0.193 & 0.213 & 0.253 & 0.260 & 0.229 & 0.258\
& 0.222 & 0.189 & 0.216 & 0.373 & 0.279 & 0.262 & 0.195 & 0.264\
& 0.239 & 0.252 & 0.243 & 0.266 & 0.254 & 0.244 & 0.252 & 0.250
Based on the above description, each identified component making up ’s background has been assigned a fixed spectral shape and spatial distribution across the FOV. Given these assumptions, one can directly measure a “local” background for any subset of the FOV and use that to accurately predict the background for anywhere in the entire FOV. The quality of the background is of course limited by the statistics available in the observation used to constrain the background model, but one advantage of separating out the different components is that separate systematic uncertainties associated with each component can be applied individually.
Application of the Background Model: [nuskybgd]{} {#sec:appendixsim}
=================================================
Determining the Background of the Bullet Cluster Observations
-------------------------------------------------------------
Now that we have a background model, we can use the events far from the cluster to determine the precise level of each background component, which are unique to the conditions of these observations. To apply the model defined in Appendix \[sec:appendixbgd\], we have developed a small suite of IDL routines called [nuskybgd]{}, whose purpose is to take regions defined in sky coordinates, compute the relative strengths of each background component based on their location in the detector plane, and create an [XSPEC]{}-readable script that sets up and fits for all observation-specific component normalizations, much in the spirit of the background treatment in the Extended Source Analysis Software package [as introduced in @SMK+08]. Those normalizations correspond to a complete spectro-spatial background model from which images in any energy band or spectra for any region can be produced.
In principle, we could extract a single spectrum of the non-cluster part of the FOV for each telescope and epoch and fit the model to that. The downside of this approach is that all spatial information is lost, which can cause the various components – especially the “Aperture” component – to obtain unphysical best-fit normalizations. To incorporate this information while also keeping the computational load to a minimum, we divide the non-cluster area into four rectangular regions for each focal plane and epoch, shown in Figure \[fig:bgdreg\]. We also try to minimize the “contamination” of these regions with cluster emission, mostly originating from the brightest parts of the cluster and carried far away from its true location by the wings of the PSF. The ellipse in Figure \[fig:bgdreg\] indicates the parts of the background regions excluded for this reason. Even so, residual cluster emission remains, which we must also model to avoid biasing the background model.
![The regions from which spectra are extracted to characterize the background; events inside the ellipse are excluded. Data from telescopes A and B are shown in the left and right panels, and the first and second epochs are shown in the top and bottom panels, respectively. The data are the same as in Figure \[fig:rawimgs\]: the images have been smoothed by Gaussian kernel of width 5 pixels and scaled between 0 (white) and 1 (black) counts to bring out structure in the background. \[fig:bgdreg\]](bgdregA0.ps "fig:"){width="9cm"} ![The regions from which spectra are extracted to characterize the background; events inside the ellipse are excluded. Data from telescopes A and B are shown in the left and right panels, and the first and second epochs are shown in the top and bottom panels, respectively. The data are the same as in Figure \[fig:rawimgs\]: the images have been smoothed by Gaussian kernel of width 5 pixels and scaled between 0 (white) and 1 (black) counts to bring out structure in the background. \[fig:bgdreg\]](bgdregB0.ps "fig:"){width="9cm"} ![The regions from which spectra are extracted to characterize the background; events inside the ellipse are excluded. Data from telescopes A and B are shown in the left and right panels, and the first and second epochs are shown in the top and bottom panels, respectively. The data are the same as in Figure \[fig:rawimgs\]: the images have been smoothed by Gaussian kernel of width 5 pixels and scaled between 0 (white) and 1 (black) counts to bring out structure in the background. \[fig:bgdreg\]](bgdregA1.ps "fig:"){width="9cm"} ![The regions from which spectra are extracted to characterize the background; events inside the ellipse are excluded. Data from telescopes A and B are shown in the left and right panels, and the first and second epochs are shown in the top and bottom panels, respectively. The data are the same as in Figure \[fig:rawimgs\]: the images have been smoothed by Gaussian kernel of width 5 pixels and scaled between 0 (white) and 1 (black) counts to bring out structure in the background. \[fig:bgdreg\]](bgdregB1.ps "fig:"){width="9cm"}
For the most part, the regions from each telescope and epoch are fit independently, but the CXB components between telescopes and epochs are correlated and can be tied together to improve their precision. The “fCXB” component, being the unresolved contribution of sources in that region of the sky, will be identical for telescopes A and B as long as the regions are roughly coincident. Similarly, because the roll angle is very similar between the two epochs, the part of the sky producing the “Aperture” component for each telescope is almost entirely identical. Fitting all 16 spectra simultaneously therefore permits the normalizations of these parameters to be appropriately tied together, reducing the number of free parameters and the chance that any component gets pushed to an unphysical value by preventing the fit from heading down a local minimum. We pursue this strategy because below $\sim 15$ keV all of the background components contribute at non-negligible levels, making it easier for the fit minimization procedure to be misled by mere statistical fluctuations.
Despite having a conservative exclusion region around the cluster, a small but noticeable number of cluster photons are scattered into the background regions, roughly at the level of the “fCXB” component. Because its contribution is fairly modest, the spectral model used to account for its emission does not have to be extremely accurate; we take a single temperature model at the global average temperature of 14.1 keV and abundance relative to solar of 0.15 convolved with the same ARF used by the “fCXB” component. The spectral shapes and normalizations of the scattered cluster emission and “fCXB” components turn out to be very similar, which means that if both were left free they may very well take on unphysical values. Even under these circumstances, however, the overall background model should not suffer, since the scattered component is not included in it and the “fCXB” normalization, while on average constant across the FOV, can significantly vary location-to-location due to cosmic variance. As that flux in the background regions cannot be directly applied at the cluster location, we simply fix the “fCXB" component to its average value and allow the scattered cluster emission component to be free, which may compensate for variations in the CXB flux in each region as well. The fit to all background region spectra is shown in Figure \[fig:bgdspec\].
{width="9cm"} {width="9cm"} {width="9cm"} {width="9cm"}
Applying the Background Model to the Bullet Cluster Spectra
-----------------------------------------------------------
The background model is defined both spatially and spectrally, and its parameters have now been determined for our specific observations, allowing a background spectrum to be generated from the model for any location in the FOV with [nuskybgd]{}. To realistically assess the impact of both statistical and systematic uncertainties associated with the background on fits to the cluster spectrum, we perform Monte Carlo simulations of the background including those uncertainties as fluctuations from the expected model. Because the background is broken up into separate components, each one can be varied based on its own systematic error, as specified in Section \[sec:cal:bgd:sys\]. Each of the simulated background spectra is generated in two steps from the predicted model for the source region. First, the normalizations of each component are randomly shifted, assuming a normal distribution about their systematic uncertainty. Then, a counts spectrum with Poisson fluctuations is created from the adjusted model for an exposure time equal to that of the observation using the [fakeit]{} command in [XSPEC]{}. While counting statistics should not bias the modeling of the Bullet cluster spectrum in principle, the true background can be thought of as one such realization; a conspiracy of high or low shot noise at just the right energies would act just like a systematic offset. Our procedure captures the likelihood of such occurrences and thus more realistic error ranges for the cluster model parameters.
We simulate 1000 background spectra, enough to characterize the standard deviation at each energy and confirm the naive expectation that the fluctuations are roughly Gaussian out to $\sim 3\sigma$. Considering the full gamut of likely background spectra, as opposed to the nominal model derived from local background regions, puts several intriguing or worrying features in the proper context. In Figure \[fig:specsig\], the hard band of the Bullet cluster spectrum is shown relative to the range bound by the background simulations (red/green or gray shaded regions) and relative to the average 1T thermal model (blue/dashed line). Above $\sim 50$ keV, the spectrum generally agrees well with the mean expectation of the background, and deviations from the mean fall appropriately distributed within the range. A few energy ranges, however, show more systematic deviations from the mean. From $\sim 20$–22 keV, the spectrum quickly rises above the 1T model, and from $\sim35$–50 keV the spectrum stays slightly, but consistently, below the mean background level. When considered relative to the allowed range of the background, it is clear that the deviations are not worryingly extreme. A common systematic fluctuation in the 35–50 keV background could cause the $\sim 1\sigma$ offset, and the blip at 22 keV is most likely an imperfectly calibrated background line or lines (see Table \[tab:app:ratios\]).
[^1]: http://pcos.gsfc.nasa.gov/studies/rfi/Harrison-Fiona-RFI.pdf
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Sequential dynamics are a key feature of many modern recommender systems, which seek to capture the ‘context’ of users’ activities on the basis of actions they have performed recently. To capture such patterns, two approaches have proliferated: Markov Chains (MCs) and Recurrent Neural Networks (RNNs). Markov Chains assume that a user’s next action can be predicted on the basis of just their last (or last few) actions, while RNNs in principle allow for longer-term semantics to be uncovered. Generally speaking, MC-based methods perform best in extremely sparse datasets, where model parsimony is critical, while RNNs perform better in denser datasets where higher model complexity is affordable. The goal of our work is to balance these two goals, by proposing a self-attention based sequential model (SASRec) that allows us to capture long-term semantics (like an RNN), but, using an attention mechanism, makes its predictions based on relatively few actions (like an MC). At each time step, SASRec seeks to identify which items are ‘relevant’ from a user’s action history, and use them to predict the next item. Extensive empirical studies show that our method outperforms various state-of-the-art sequential models (including MC/CNN/RNN-based approaches) on both sparse and dense datasets. Moreover, the model is an order of magnitude more efficient than comparable CNN/RNN-based models. Visualizations on attention weights also show how our model adaptively handles datasets with various density, and uncovers meaningful patterns in activity sequences.'
author:
- |
UC San Diego\
{wckang,jmcauley}@ucsd.edu
bibliography:
- 'sigproc.bib'
title: 'Self-Attentive Sequential Recommendation'
---
Introduction
============
The goal of sequential recommender systems is to combine personalized models of user behavior (based on historical activities) with some notion of ‘context’ on the basis of users’ recent actions. Capturing useful patterns from sequential dynamics is challenging, primarily because the dimension of the input space grows exponentially with the number of past actions used as context. Research in sequential recommendation is therefore largely concerned with how to capture these high-order dynamics succinctly.
Markov Chains (MCs) are a classic example, which assume that the next action is conditioned on only the previous action (or previous few), and have been successfully adopted to characterize short-range item transitions for recommendation [@rendle2010fpmc]. Another line of work uses Recurrent Neural Networks (RNNs) to summarize all previous actions via a hidden state, which is used to predict the next action [@DBLP:journals/corr/HidasiKBT15].
Both approaches, while strong in specific cases, are somewhat limited to certain types of data. MC-based methods, by making strong simplifying assumptions, perform well in high-sparsity settings, but may fail to capture the intricate dynamics of more complex scenarios. Conversely RNNs, while expressive, require large amounts of data (an in particular *dense* data) before they can outperform simpler baselines. Recently, a new sequential model *Transfomer* achieved state-of-the-art performance and efficiency for machine translation tasks [@transform]. Unlike existing sequential models that use convolutional or recurrent modules, Transformer is purely based on a proposed attention mechanism called ‘self-attention,’ which is highly efficient and capable of uncovering syntactic and semantic patterns between words in a sentence.
Inspired by this method, we seek to apply self-attention mechanisms to sequential recommendation problems. Our hope is that this idea can address both of the problems outlined above, being on the one hand able to draw context from all actions in the past (like RNNs) but on the other hand being able to frame predictions in terms of just a small number of actions (like MCs). Specifically, we build a Self-Attention based Sequential Recommendation model (*SASRec*), which adaptively assigns weights to previous items at each time step (Figure \[fig:diagram\]).
![A simplified diagram showing the training process of SASRec. At each time step, the model considers all previous items, and uses attention to ‘focus on’ items relevant to the next action.[]{data-label="fig:diagram"}](figs/fig1){width=".9\linewidth"}
The proposed model significantly outperforms state-of-the-art MC/CNN/RNN-based sequential recommendation methods on several benchmark datasets. In particular, we examine performance as a function of dataset sparsity, where model performance aligns closely with the patterns described above. Due to the self-attention mechanism, SASRec tends to consider long-range dependencies on dense datasets, while focusing on more recent activities on sparse datasets. This proves crucial for adaptively handling datasets with varying density.
Furthermore, the core component (i.e., the self-attention block) of SASRec is suitable for parallel acceleration, resulting in a model that is an order of magnitude faster than CNN/RNN-based alternatives. In addition, we analyze the complexity and scalability of SASRec, conduct a comprehensive ablation study to show the effect of key components, and visualize the attention weights to qualitatively reveal the model’s behavior.
Related Work
============
Several lines of work are closely related to ours. We first discuss general, followed by temporal, recommendation, before discussing sequential recommendation (in particular MCs and RNNs). Last we introduce the attention mechanism, especially the self-attention module which is at the core of our model.
General Recommendation
----------------------
Recommender systems focus on modeling the compatibility between users and items, based on historical feedback (e.g. clicks, purchases, likes). User feedback can be *explicit* (e.g. ratings) or *implicit* (e.g. clicks, purchases, comments) [@WRMF; @rendle2009bpr]. Modeling implicit feedback can be challenging due to the ambiguity of interpreting ‘non-observed’ (e.g. non-purchased) data. To address the problem, *point-wise* [@WRMF] and *pairwise* [@rendle2009bpr] methods are proposed to solve such challenges.
Matrix Factorization (MF) methods seek to uncover latent dimensions to represent users’ preferences and items’ properties, and estimate interactions through the inner product between the user and item embeddings [@Handbook; @korenSurvey]. In addition, another line of work is based on Item Similarity Models (ISM) and doesn’t explicitly model each user with latent factors (e.g. FISM [@kabbur2013fism]). They learn an item-to-item similarity matrix, and estimate a user’s preference toward an item via measuring its similarities with items that the user has interacted with before.
Recently, due to their success in related problems, various deep learning techniques have been introduced for recommendation [@DBLP:journals/corr/ZhangYS17aa]. One line of work seeks to use neural networks to extract item features (e.g. images [@wang2017your; @DBLP:conf/icdm/KangFWM17], text [@DBLP:conf/kdd/WangWY15; @DBLP:conf/recsys/KimPOLY16], etc.) for content-aware recommendation. Another line of work seeks to replace conventional MF. For example, NeuMF [@NeuMF] estimates user preferences via Multi-Layer Perceptions (MLP), and AutoRec [@sedhain2015autorec] predicts ratings using autoencoders.
Temporal Recommendation
-----------------------
Dating back to the *Netflix* Prize, temporal recommendation has shown strong performance on various tasks by explicitly modeling the timestamp of users’ activities. TimeSVD++ [@timeSVD] achieved strong results by splitting time into several segments and modeling users and items separately in each. Such models are essential to understand datasets that exhibit significant (short- or long-term) temporal ‘drift’ (e.g. ‘how have movie preferences changed in the last 10 years,’ or ‘what kind of businesses do users visit at 4pm?’, etc.) [@DBLP:conf/wsdm/WuABSJ17; @xiong2010temporal; @timeSVD]. Sequential recommendation (or next-item recommendation) differs slightly from this setting, as it only considers the order of actions, and models sequential patterns which are independent of time. Essentially, sequential models try to model the ‘context’ of users’ actions based on their recent activities, rather than considering temporal patterns *per se*.
Sequential Recommendation
-------------------------
Many sequential recommender systems seek to model item-item transition matrices as a means of capturing sequential patterns among successive items. For instance, FPMC fuses an MF term and an item-item transition term to capture long-term preferences and short-term transitions respectively [@rendle2010fpmc]. Essentially, the captured transition is a first-order Markov Chain (MC), whereas higher-order MCs assume the next action is related to several previous actions. Since the last visited item is often the key factor affecting the user’s next action (essentially providing ‘context’), first-order MC based methods show strong performance, especially on sparse datasets [@DBLP:conf/recsys/HeKM17]. There are also methods adopting high-order MCs that consider more previous items [@DBLP:conf/recsys/HeFWM16; @DBLP:conf/icdm/HeM16]. In particular, Convolutional Sequence Embedding (Caser), a CNN-based method, views the embedding matrix of $L$ previous items as an ‘image’ and applies convolutional operations to extract transitions [@DBLP:conf/wsdm/TangW18]. Other than MC-based methods, another line of work adopts RNNs to model user sequences [@DBLP:conf/wsdm/JingS17; @DBLP:journals/corr/HidasiKBT15; @DBLP:conf/icdm/LiuWWLW16; @DBLP:conf/wsdm/BeutelCJXLGC18]. For example, GRU4Rec uses Gated Recurrent Units (GRU) to model click sequences for session-based recommendation [@DBLP:journals/corr/HidasiKBT15], and an improved version further boosts its Top-N recommendation performance [@DBLP:journals/corr/HidasiK17]. In each time step, RNNs take the state from the last step and current action as its input. These dependencies make RNNs less efficient, though techniques like ‘session-parallelism’ have been proposed to improve efficiency [@DBLP:journals/corr/HidasiKBT15].
Attention Mechanisms
--------------------
Attention mechanisms have been shown to be effective in various tasks such as image captioning [@DBLP:conf/icml/XuBKCCSZB15] and machine translation [@DBLP:journals/corr/BahdanauCB14], among others. Essentially the idea behind such mechanisms is that sequential outputs (for example) each depend on ‘relevant’ parts of some input that the model should focus on successively. An additional benefit is that attention-based methods are often more interpretable. Recently, attention mechanisms have been incorporated into recommender systems [@DBLP:conf/sigir/ChenZ0NLC17; @DBLP:conf/ijcai/XiaoY0ZWC17; @DBLP:conf/aaai/WangHCHL018]. For example, Attentional Factorization Machines (AFM) [@DBLP:conf/ijcai/XiaoY0ZWC17] learn the importance of each feature interaction for content-aware recommendation.
However, the attention technique used in the above is essentially an *additional* component to the original model (e.g. attention+RNNs, attention+FMs, etc.). Recently, a purely attention-based sequence-to-sequence method, Transfomer [@transform], achieved state-of-the-art performance and efficiency on machine translation tasks which had previously been dominated by RNN/CNN-based approaches [@wu2016google; @DBLP:journals/tacl/ZhouCWLX16]. The Transformer model relies heavily on the proposed ‘self-attention’ modules to capture complex structures in sentences, and to retrieve relevant words (in the source language) for generating the next word (in the target language). Inspired by Transformer, we seek to build a new sequential recommendation model based upon the self-attention approach, though the problem of sequential recommendation is quite different from machine translation, and requires specially designed models.
Methodology
===========
[lX]{} Notation&Description\
$\mathcal{U},\mathcal{I}$ & user and item set\
$\mathcal{S}^u$ & historical interaction sequence for a user $u$: $(\mathcal{S}^u_1, \mathcal{S}^u_2, ... , \mathcal{S}^u_{|\mathcal{S}^u|})$\
$d\in \mathbb{N}$ & latent vector dimensionality\
$n\in \mathbb{N}$ & maximum sequence length\
$b\in \mathbb{N}$ & number of self-attention blocks\
$\M\in\mathbb{R}^{|\mathcal{I}|\times d}$ & item embedding matrix\
$\P\in\mathbb{R}^{n\times d}$ & positional embedding matrix\
$\widehat{\E}\in\mathbb{R}^{n\times d}$ & input embedding matrix\
$\S^{(b)}\in\mathbb{R}^{n\times d}$ & item embeddings after the $b$-th self-attention layer\
$\F^{(b)}\in\mathbb{R}^{n\times d}$ & item embeddings after the $b$-th feed-forward network\
In the setting of sequential recommendation, we are given a user’s action sequence $\mathcal{S}^u=(\mathcal{S}^u_1, \mathcal{S}^u_2, \ldots , \mathcal{S}^u_{|\mathcal{S}^u|})$, and seek to predict the next item. During the training process, at time step $t$, the model predicts the next item depending on the previous $t$ items. As shown in Figure \[fig:diagram\], it will be convenient to think of the model’s input as $(\mathcal{S}^u_1, \mathcal{S}^u_2, \ldots , \mathcal{S}^u_{|\mathcal{S}^u|-1})$ and its expected output as a ‘shifted’ version of the same sequence: $(\mathcal{S}^u_2, \mathcal{S}^u_3, \ldots , \mathcal{S}^u_{|\mathcal{S}^u|})$. In this section, we describe how we build a sequential recommendation model via an embedding layer, several self-attention blocks, and a prediction layer. We also analyze its complexity and further discuss how SASRec differs from related models. Our notation is summarized in Table \[tb:notation\].
Embedding Layer
---------------
We transform the training sequence $(\mathcal{S}^u_1, \mathcal{S}^u_2, ... , \mathcal{S}^u_{|\mathcal{S}^u|-1})$ into a fixed-length sequence $s=(s_1,s_2,\dots,s_n)$, where $n$ represents the maximum length that our model can handle. If the sequence length is greater than $n$, we consider the most recent $n$ actions. If the sequence length is less than $n$, we repeatedly add a ‘padding’ item to the left until the length is $n$. We create an item embedding matrix $\M\in\mathbb{R}^{|\mathcal{I}|\times d}$ where $d$ is the latent dimensionality, and retrieve the input embedding matrix $\E\in\mathbb{R}^{n\times d}$, where $\E_i=\M_{s_i}$. A constant zero vector $\0$ is used as the embedding for the padding item.
### **Positional Embedding** {#positional-embedding .unnumbered}
As we will see in the next section, since the self-attention model doesn’t include any recurrent or convolutional module, it is not aware of the positions of previous items. Hence we inject a learnable position embedding $\P\in\mathbb{R}^{n\times d}$ into the input embedding: $$\widehat{\E}=\left [
\begin{tabular}{c}
$\M_{s_1}+\P_{1}$ \\
$\M_{s_2}+\P_{2}$ \\
$\dots$ \\
$\M_{s_n}+\P_{n}$
\end{tabular}
\right ]$$ We also tried the fixed position embedding as used in [@transform], but found that this led to worse performance in our case. We analyze the effect of the position embedding quantitatively and qualitatively in our experiments.
Self-Attention Block
--------------------
The scaled dot-product attention [@transform] is defined as: $$\text{Attention}(\Q,\K,\V)=\text{softmax}\left (\frac{\Q\K^T}{\sqrt{d}}\right )\V,$$ where $\Q$ represents the queries, $\K$ the keys and $\V$ the values (each row represents an item). Intuitively, the attention layer calculates a weighted sum of all values, where the weight between query $i$ and value $j$ relates to the interaction between query $i$ and key $j$. The scale factor $\sqrt{d}$ is to avoid overly large values of the inner product, especially when the dimensionality is high.
### **Self-Attention layer** {#self-attention-layer .unnumbered}
In NLP tasks such as machine translation, attention mechanisms are typically used with $\K=\V$ (e.g. using an RNN encoder-decoder for translation: the encoder’s hidden states are keys and values, and the decoder’s hidden states are queries) [@DBLP:journals/corr/BahdanauCB14]. Recently, a self-attention method was proposed which uses the same objects as queries, keys, and values [@transform]. In our case, the self-attention operation takes the embedding $\widehat{\E}$ as input, converts it to three matrices through linear projections, and feeds them into an attention layer: $$\S=\text{SA}(\widehat{\E})=\text{Attention}(\widehat{\E}\W^Q,\widehat{\E}\W^K,\widehat{\E}\W^V),
\label{eq:sa}$$ where the projection matrices $\W^Q,\W^K,\W^V\in\mathbb{R}^{d\times d}$. The projections make the model more flexible. For example, the model can learn asymmetric interactions (i.e., <query $i$, key $j$> and <query $j$, key $i$> can have different interactions).
### **Causality** {#causality .unnumbered}
Due to the nature of sequences, the model should consider only the first $t$ items when predicting the $(t+1)$-st item. However, the $t$-th output of the self-attention layer ($\S_t$) contains embeddings of subsequent items, which makes the model ill-posed. Hence, we modify the attention by forbidding all links between $\Q_i$ and $\K_j$ ($j>i$).
### **Point-Wise Feed-Forward Network** {#point-wise-feed-forward-network .unnumbered}
Though the self-attention is able to aggregate all previous items’ embeddings with adaptive weights, ultimately it is still a linear model. To endow the model with nonlinearity and to consider interactions between different latent dimensions, we apply a point-wise two-layer feed-forward network to all $\S_i$ identically (sharing parameters): $$\F_i = \text{FFN}(\S_i)=\text{ReLU}(\S_i\W^{(1)}+\b^{(1)})\W^{(2)}+\b^{(2)},
\label{eq:ffn}$$ where $\W^{(1)},\W^{(2)}$ are $d\times d$ matrices and $\b^{(1)},\b^{(2)}$ are $d$-dimensional vectors. Note that there is no interaction between $\S_i$ and $\S_j$ ($i\neq j$), meaning that we still prevent information leaks (from back to front).
Stacking Self-Attention Blocks
------------------------------
After the first self-attention block, $\F_i$ essentially aggregates all previous items’ embeddings (i.e., $\widehat{\E}_j, j\leq i$). However, it might be useful to learn more complex item transitions via another self-attention block based on $\F$. Specifically, we stack the self-attention block (i.e., a self-attention layer and a feed-forward network), and the $b$-th ($b>1$) block is defined as: $$\begin{array}{c}
\S^{(b)}=\text{SA}(\F^{(b-1)}),\\
\F_i^{(b)}=\text{FFN}(\S_i^{(b)}),\ \ \forall i\in \{1,2,\dots,n\},
\end{array}
$$ and the $1$-st block is defined as $\S^{(1)}=\S$ and $\F^{(1)}=\F$.
However, when the network goes deeper, several problems become exacerbated: 1) the increased model capacity leads to overfitting; 2) the training process becomes unstable (due to vanishing gradients etc.); and 3) models with more parameters often require more training time. Inspired by [@transform], We perform the following operations to alleviate these problems: $$g(x)=x+\text{Dropout}(g(\text{LayerNorm}(x))),$$ where $g(x)$ represents the self attention layer or the feed-forward network. That is to say, for layer $g$ in each block, we apply layer normalization on the input $x$ before feeding into $g$, apply dropout on $g$’s output, and add the input $x$ to the final output. We introduce these operations below.
### **Residual Connections** {#residual-connections .unnumbered}
In some cases, multi-layer neural networks have demonstrated the ability to learn meaningful features hierarchically [@zeiler2014visualizing]. However, simply adding more layers did not easily correspond to better performance until residual networks were proposed [@he2016deep]. The core idea behind residual networks is to propagate low-layer features to higher layers by residual connection. Hence, if low-layer features are useful, the model can easily propagate them to the final layer. Similarly, we assume residual connections are also useful in our case. For example, existing sequential recommendation methods have shown that the last visited item plays a key role on predicting the next item [@rendle2010fpmc; @DBLP:conf/icdm/HeM16; @DBLP:conf/recsys/HeKM17]. However, after several self-attention blocks, the embedding of the last visited item is entangled with all previous items; adding residual connections to propagate the last visited item’s embedding to the final layer would make it much easier for the model to leverage low-layer information.
### **Layer Normalization** {#layer-normalization .unnumbered}
Layer normalization is used to normalize the inputs across features (i.e., zero-mean and unit-variance), which is beneficial for stabilizing and accelerating neural network training [@DBLP:journals/corr/BaKH16]. Unlike batch normalization [@DBLP:conf/icml/IoffeS15], the statistics used in layer normalization are independent of other samples in the same batch. Specifically, assuming the input is a vector $\x$ which contains all features of a sample, the operation is defined as: $$\text{LayerNorm}(\x)=\bm{\alpha}\odot\frac{\x-\mu}{\sqrt{\sigma^2+\epsilon}}+\bm{\beta},$$ where $\odot$ is an element-wise product (i.e., the Hadamard product), $\mu$ and $\sigma$ are the mean and variance of $\x$, $\bm{\alpha}$ and $\bm{\beta}$ are learned scaling factors and bias terms.
### **Dropout** {#dropout .unnumbered}
To alleviate overfitting problems in deep neural networks, ‘Dropout’ regularization techniques have been shown to be effective in various neural network architectures [@DBLP:journals/jmlr/SrivastavaHKSS14]. The idea of dropout is simple: randomly ‘turn off’ neurons with probability $p$ during training, and use all neurons when testing. Further analysis points out that dropout can be viewed as a form of ensemble learning which considers an enormous number of models (exponential in the number of neurons and input features) that share parameters [@DBLP:journals/corr/Warde-FarleyGCB13]. We also apply a dropout layer on the embedding $\widehat{\E}$.
Prediction Layer
----------------
After $b$ self-attention blocks that adaptively and hierarchically extract information of previously consumed items, we predict the next item (given the first $t$ items) based on $\F_t^{(b)}$. Specifically, we adopt an MF layer to predict the relevance of item $i$: $$r_{i,t}=\F_t^{(b)}\N_i^T,$$ where $r_{i,t}$ is the relevance of item $i$ being the next item given the first $t$ items (i.e., $s_1,s_2,\ldots,s_t$), and $\N\in\mathbb{R}^{|\mathcal{I}|\times d}$ is an item embedding matrix. Hence, a high interaction score $r_{i,t}$ means a high relevance, and we can generate recommendations by ranking the scores.
### **Shared Item Embedding** {#shared-item-embedding .unnumbered}
To reduce the model size and alleviate overfitting, we consider another scheme which only uses a single item embedding $\M$: $$r_{i,t}=\F_t^{(b)}\M_i^T.$$ Note that $\F_t^{(b)}$ can be represented as a function depending on the item embedding $\M$: $\F_t^{(b)}=f(\M_{s_1},\M_{s_2},\dots,\M_{s_t})$. A potential issue of using homogeneous item embeddings is that their inner products cannot represent asymmetric item transitions (e.g. item $i$ is frequently bought after $j$, but not vise versa), and thus existing methods like FPMC tend to use heterogeneous item embeddings. However, our model doesn’t have this issue since it learns a nonlinear transformation. For example, the feed forward network can easily achieve the asymmetry with the same item embedding: $\text{FFN}(\M_i)\M_j^T\neq \text{FFN}(\M_j)\M_i^T$. Empirically, using a shared item embedding significantly improves the performance of our model.
### **Explicit User Modeling** {#explicit-user-modeling .unnumbered}
To provide personalized recommendations, existing methods often take one of two approaches: 1) learn an *explicit* user embedding representing user preferences (e.g. MF [@MF], FPMC [@rendle2010fpmc] and Caser [@DBLP:conf/wsdm/TangW18]); 2) consider the user’s previous actions, and induce an *implicit* user embedding from embeddings of visited items (e.g. FSIM [@kabbur2013fism], Fossil [@DBLP:conf/icdm/HeM16], GRU4Rec [@DBLP:journals/corr/HidasiKBT15]). Our method falls into the latter category, as we generate an embedding $\F^{(b)}_n$ by considering all actions of a user. However, we can also insert an explicit user embedding at the last layer, for example via addition: $r_{u,i,t}=(\U_u+\F_t^{(b)})\M_i^T$ where $\U$ is a user embedding matrix. However, we empirically find that adding an explicit user embedding doesn’t improve performance (presumably because the model already considers all of the user’s actions).
Network Training
----------------
Recall that we convert each user sequence (excluding the last action) $(\mathcal{S}^u_1, \mathcal{S}^u_2, \ldots , \mathcal{S}^u_{|\mathcal{S}^u|-1})$ to a fixed length sequence $s=\{s_1,s_2,\dots,s_n\}$ via truncation or padding items. We define $o_t$ as the expected output at time step $t$: $$o_t=\begin{cases}
\texttt{<pad>} & \text{if } s_t \text{ is a padding item}\\
s_{t+1} & 1\leq t<n\\
\mathcal{S}^{u}_{|\mathcal{S}^u|} & t=n
\end{cases},$$ where $\texttt{<pad>}$ indicates a padding item. Our model takes a sequence $s$ as input, the corresponding sequence $o$ as expected output, and we adopt the binary cross entropy loss as the objective function: $$-\sum_{\mathcal{S}^{u}\in\mathcal{S}}\sum_{t\in[1,2,\dots,n]}\left [\log(\sigma(r_{o_t,t}))+\sum_{j\not\in \mathcal{S}^{u}}\log(1-\sigma(r_{j,t}))\right ].$$ Note that we ignore the terms where $o_t=\texttt{<pad>}$.
The network is optimized by the *Adam* optimizer [@DBLP:journals/corr/KingmaB14], which is a variant of Stochastic Gradient Descent (SGD) with adaptive moment estimation. In each epoch, we randomly generate one negative item $j$ for each time step in each sequence. More implementation details are described later.
Complexity Analysis {#sec:complexity}
-------------------
### **Space Complexity** {#space-complexity .unnumbered}
The learned parameters in our model are from the embeddings and parameters in the self-attention layers, feed-forward networks and layer normalization. The total number of parameters is $O(|\mathcal{I}|d+nd+d^2)$, which is moderate compared to other methods (e.g. $O(|\mathcal{U}|d+|\mathcal{I}|d)$ for FPMC) since it does not grow with the number of users, and $d$ is typically small in recommendation problems.
### **Time Complexity** {#time-complexity .unnumbered}
The computational complexity of our model is mainly due to the self-attention layer and the feed-forward network, which is $O(n^2d+nd^2)$. The dominant term is typically $O(n^2d)$ from the self-attention layer. However, a convenient property in our model is that the computation in each self-attention layer is fully parallelizable, which is amenable to GPU acceleration. In contrast, RNN-based methods (e.g. GRU4Rec [@DBLP:journals/corr/HidasiKBT15]) have a dependency on time steps (i.e., computation on time step $t$ must wait for results from time step $t\text{-}1$), which leads to an $O(n)$ time on sequential operations. We empirically find our method is over ten times faster than RNN and CNN-based methods with GPUs (the result is similar to that in [@transform] for machine translation tasks), and the maximum length $n$ can easily scale to a few hundred which is generally sufficient for existing benchmark datasets.
During testing, for each user, after calculating the embedding $\F_n^{(b)}$ , the process is the same as standard MF methods. ($O(d)$ for evaluating the preference toward an item).
### **Handing Long Sequences** {#handing-long-sequences .unnumbered}
Though our experiments empirically verify the efficiency of our method, ultimately it cannot scale to very long sequences. A few options are promising to investigate in the future: 1) using restricted self-attention [@poveytime] which only attends on recent actions rather than all actions, and distant actions can be considered in higher layers; 2) splitting long sequences into short segments as in [@DBLP:conf/wsdm/TangW18].
Discussion
----------
We find that SASRec can be viewed as a generalization of some classic CF models. We also discuss conceptually how our approach and existing methods handle sequence modeling.
### **Reduction to Existing Models** {#reduction-to-existing-models .unnumbered}
- *Factorized Markov Chains*: FMC factorizes a first-order item transition matrix, and predicts the next item $j$ depending on the last item $i$: $$P(j|i)\propto\M_i^T\N_j.$$ If we set the self-attention block to zero, use unshared item embeddings, and remove the position embedding, SASRec reduces to FMC. Furthermore, SASRec is also closely related to Factorized Personalized Markov Chains (FPMC) [@rendle2010fpmc], which fuse MF with FMC to capture user preferences and short-term dynamics respectively: $$P(j|u,i)\propto\left [\U_u,\M_i\right ]\N^T_j.$$ Following the reduction operations above for FMC, and adding an explicit user embedding (via concatenation), SASRec is equivalent to FPMC.
- *Factorized Item Similarity Models [@kabbur2013fism]*: FISM estimates a preference score toward item $i$ by considering the similarities between $i$ and items the user consumed before: $$P(j|u)\propto\left (\frac{1}{|\mathcal{S}^u|}\sum_{i\in\mathcal{S}^u}\M_i\right)\N^T_j.$$
If we use one self-attention layer (excluding the feed-forward network), set uniform attention weights (i.e., $\frac{1}{|\mathcal{S}_{u}|}$) on items, use unshared item embeddings, and remove the position embedding, SASRec is reduced to FISM. Thus our model can be viewed as an adaptive, hierarchical, sequential item similarity model for next item recommendation.
### **MC-based Recommendation** {#mc-based-recommendation .unnumbered}
Markov Chains (MC) can effectively capture local sequential patterns, assuming that the next item is only dependent on the previous $L$ items. Exiting MC-based sequential recommendation methods rely on either first-order MCs (e.g. FPMC [@rendle2010fpmc], HRM [@hrm], TransRec [@DBLP:conf/recsys/HeKM17]) or high-order MCs (e.g. Fossil [@DBLP:conf/icdm/HeM16], Vista [@DBLP:conf/recsys/HeFWM16], Caser [@DBLP:conf/wsdm/TangW18]). The first group of methods tend to perform best on sparse datasets. In contrast, higher-order MC based methods have two limitations: (1) The MC order $L$ needs to be specified before training, rather than being chosen adaptively; (2) The performance and efficiency doesn’t scale well with the order $L$, hence $L$ is typically small (e.g. less than 5). Our method resolves the first issue, since it can adaptively attend on related previous items (e.g. focusing on just the last item on sparse dataset, and more items on dense datasets). Moreover, our model is essentially conditioned on $n$ previous items, and can empirically scale to several hundred previous items, exhibiting performance gains with moderate increases in training time.
### **RNN-based Recommendation** {#rnn-based-recommendation .unnumbered}
Another line of work seeks to use RNNs to model user action sequences [@DBLP:journals/corr/HidasiKBT15; @DBLP:journals/corr/HidasiK17; @DBLP:conf/wsdm/WuABSJ17]. RNNs are generally suitable for modeling sequences, though recent studies show CNNs and self-attention can be stronger in some sequential settings [@transform; @DBLP:journals/corr/abs-1803-01271]. Our self-attention based model can be derived from item similarity models, which are a reasonable alternative for sequence modeling for recommendation. For RNNs, other than their inefficiency in parallel computation (Section \[sec:complexity\]), their maximum path length (from an input node to related output nodes) is $O(n)$. In contrast, our model has $O(1)$ maximum path length, which can be beneficial for learning long-range dependencies [@hochreiter2001gradient].
Experiments
===========
In this section, we present our experimental setup and empirical results. Our experiments are designed to answer the following research questions:
**RQ1:**
: Does SASRec outperform state-of-the-art models including CNN/RNN based methods?
**RQ2:**
: What is the influence of various components in the SASRec architecture?
**RQ3:**
: What is the training efficiency and scalability (regarding $n$) of SASRec?
**RQ4:**
: Are the attention weights able to learn meaningful patterns related to positions or items’ attributes?
Datasets
--------
We evaluate our methods on four datasets from three real world applications. The datasets vary significantly in domains, platforms, and sparsity:
- **Amazon:** A series of datasets introduced in [@VisualSIGIR], comprising large corpora of product reviews crawled from *Amazon.com*. Top-level product categories on *Amazon* are treated as separate datasets. We consider two categories, ‘Beauty,’ and ‘Games.’ This dataset is notable for its high sparsity and variability.
- **Steam:** We introduce a new dataset crawled from *Steam*, a large online video game distribution platform. The dataset contains 2,567,538 users, 15,474 games and 7,793,069 English reviews spanning October 2010 to January 2018. The dataset also includes rich information that might be useful in future work, like users’ play hours, pricing information, media score, category, developer (etc.).
- **MovieLens:** A widely used benchmark dataset for evaluating collaborative filtering algorithms. We use the version (MovieLens-1M) that includes 1 million user ratings.
We followed the same preprocessing procedure from [@DBLP:conf/recsys/HeKM17; @DBLP:conf/icdm/HeM16; @rendle2010fpmc]. For all datasets, we treat the presence of a review or rating as implicit feedback (i.e., the user interacted with the item) and use timestamps to determine the sequence order of actions. We discard users and items with fewer than 5 related actions. For partitioning, we split the historical sequence $\mathcal{S}^u$ for each user $u$ into three parts: (1) the most recent action $\mathcal{S}^u_{|\mathcal{S}^u|}$ for testing, (2) the second most recent action $\mathcal{S}^u_{|\mathcal{S}^u|-1}$ for validation, and (3) all remaining actions for training. Note that during testing, the input sequences contain training actions and the validation action.
Data statistics are shown in Table \[tb:data\]. We see that the two *Amazon* datasets have the fewest actions per user and per item (on average), *Steam* has a high average number of actions per item, and *MovieLens-1m* is the most dense dataset.
----------------- --------- -------- ------- ------- ------
*Amazon Beauty* 52,024 57,289 7.6 6.9 0.4M
*Amazon Games* 31,013 23,715 9.3 12.1 0.3M
*Steam* 334,730 13,047 11.0 282.5 3.7M
*MovieLens-1M* 6,040 3,416 163.5 289.1 1.0M
----------------- --------- -------- ------- ------- ------
: Dataset statistics (after preprocessing) \[exp:dataset\][]{data-label="tb:data"}
----------- --------- -------- -------- -------- -------- -------- -------- -------- -------- ------------ ------- -------
Hit@10 0.4003 0.3775 0.3771 0.4310 0.2125 0.3949 0.4264 **0.4854** 5.4% 13.8%
NDCG@10 0.2277 0.2183 0.2477 0.2891 0.1203 0.2556 0.2547 **0.3219** 6.6% 25.9%
\[1.5mm\] Hit@10 0.4724 0.4853 0.6358 0.6802 0.2938 0.6599 0.5282 **0.7410** 8.5% 12.3%
NDCG@10 0.2779 0.2875 0.4456 0.4680 0.4557 0.1837 0.3214 **0.5360** 14.5% 12.6%
\[1.5mm\] Hit@10 0.7172 0.7061 0.7731 0.7710 0.7624 0.4190 0.7874 **0.8729** 13.2% 8.9%
NDCG@10 0.4535 0.4436 0.5193 0.5011 0.4852 0.2691 0.5381 **0.6306** 21.4% 12.7%
\[1.5mm\] Hit@10 0.4329 0.5781 0.6986 0.7599 0.6413 0.5581 0.7501 **0.8245** 8.5% 4.6%
NDCG@10 0.2377 0.3287 0.4676 0.5176 0.3969 0.3381 0.5513 **0.5905** 14.1% 6.6%
----------- --------- -------- -------- -------- -------- -------- -------- -------- -------- ------------ ------- -------
Comparison Methods
------------------
To show the effectiveness of our method, we include three groups of recommendation baselines. The first group includes general recommendation methods which only consider user feedback without considering the sequence order of actions:
- **PopRec**: This is a simple baseline that ranks items according to their popularity (i.e., number of associated actions).
- **Bayesian Personalized Ranking (BPR) [@rendle2009bpr]**: BPR is a classic method for learning personalized rankings from implicit feedback. Biased matrix factorization is used as the underlying recommender.
The second group contains sequential recommendation methods based on first order Markov chains, which consider the last visited item:
- **Factorized Markov Chains (FMC):** A first-order Markov chain method. FMC factorizes an item transition matrix using two item embeddings, and generates recommendations depending only on the last visited item.
- **Factorized Personalized Markov Chains (FPMC) [@rendle2010fpmc]**: FPMC uses a combination of matrix factorization and factorized first-order Markov chains as its recommender, which captures users’ long-term preferences as well as item-to-item transitions.
- **Translation-based Recommendation (TransRec) [@DBLP:conf/recsys/HeKM17]**: A state-of-the-art first-order sequential recommendation method which models each user as a translation vector to capture the transition from the current item to the next.
The last group contains deep-learning based sequential recommender systems, which consider several (or all) previously visited items:
- **GRU4Rec [@DBLP:journals/corr/HidasiKBT15]**: A seminal method that uses RNNs to model user action sequences for session-based recommendation. We treat each user’s feedback sequence as a session.
- **GRU4Rec$^\text{+}$ [@DBLP:journals/corr/HidasiK17]**: An improved version of GRU4Rec, which adopts a different loss function and sampling strategy, and shows significant performance gains on Top-N recommendation.
- **Convolutional Sequence Embeddings (Caser) [@DBLP:conf/wsdm/TangW18]**: A recently proposed CNN-based method capturing high-order Markov chains by applying convolutional operations on the embedding matrix of the $L$ most recent items, and achieves state-of-the-art sequential recommendation performance.
Since other sequential recommendation methods (e.g. PRME [@feng2015prme], HRM [@hrm], Fossil [@DBLP:conf/icdm/HeM16]) have been outperformed on similar datasets by baselines among those above, we omit comparison against them. We also don’t include temporal recommendation methods like TimeSVD++ [@timeSVD] and RRN [@DBLP:conf/wsdm/WuABSJ17], which differ in setting from what we consider here.
For fair comparison, we implement BPR, FMC, FPMC, and TransRec using *TemsorFlow* with the Adam [@DBLP:journals/corr/KingmaB14] optimizer. For GRU4Rec, GRU4Rec$^\text{+}$, and Caser, we use code provided by the corresponding authors. For all methods except PopRec, we consider latent dimensions $d$ from $\{10,20,30,40,50\}$. For BPR, FMC, FPMC, and TransRec, the $\ell_2$ regularizer is chosen from $\{0.0001,0.001,0.01,0.1,1\}$. All other hyperparameters and initialization strategies are those suggested by the methods’ authors. We tune hyper-parameters using the validation set, and terminate training if validation performance doesn’t improve for 20 epochs.
implementation Details
----------------------
For the architecture in the default version of SASRec, we use two self-attention blocks ($b=2$), and use the learned positional embedding. Item embeddings in the embedding layer and prediction layer are shared. We implement SASRec with *TensorFlow*. The optimizer is the *Adam* optimizer [@DBLP:journals/corr/KingmaB14], the learning rate is set to $0.001$, and the batch size is $128$. The dropout rate of turning off neurons is 0.2 for *MovieLens-1m* and 0.5 for the other three datasets due to their sparsity. The maximum sequence length $n$ is set to 200 for *MovieLens-1m* and 50 for the other three datasets, i.e., roughly proportional to the mean number of actions per user. Performance of variants and different hyper-parameters is examined below.
*All code and data shall be released at publication time*.
Evaluation Metrics
------------------
We adopt two common Top-N metrics, Hit Rate@10 and NDCG@10, to evaluate recommendation performance [@NeuMF; @DBLP:conf/recsys/HeKM17]. Hit@10 counts the fraction of times that the ground-truth next item is among the top 10 items, while NDCG@10 is a position-aware metric which assigns larger weights on higher positions. Note that since we only have one test item for each user, Hit@10 is equivalent to Recall@10, and is proportional to Precision@10.
To avoid heavy computation on all user-item pairs, we followed the strategy in [@koren2008factorization; @NeuMF]. For each user $u$, we randomly sample 100 negative items, and rank these items with the ground-truth item. Based on the rankings of these 101 items, Hit@10 and NDCG@10 can be evaluated.
Recommendation Performance
--------------------------
Table \[tab:recommendation\] presents the recommendation performance of all methods on the four datasets (**RQ1**). By considering the second best methods across all datasets, a general pattern emerges with non-neural methods (i.e., (a)-(e)) performing better on sparse datasets and neural approaches (i.e., (f)-(h)) performing better on denser datasets. Presumably this owes to neural approaches having more parameters to capture high order transitions (i.e., they are expressive but easily overfit), whereas carefully designed but simpler models are more effective in high-sparsity settings. Our method SASRec outperforms all baselines on both sparse and dense datasets, and gains 6.9% Hit Rate and 9.6% NDCG improvements (on average) against the strongest baseline. One likely reason is that our model can adaptively attend items within different ranges on different datasets (e.g. only the previous item on sparse datasets and more on dense datasets). We further analyze this behavior in Section \[sec:vis\].
In Figure \[fig:K\] we also analyze the effect of a key hyper-parameter, the latent dimensionality $d$, by showing NDCG@10 of all methods with $d$ varying from 10 to 50. We see that our model typically benefits from larger numbers of latent dimensions. For all datasets, our model achieves satisfactory performance with $d\geq40$.
{width=".99\linewidth"}
[0.245]{} {width="\linewidth"}
[0.245]{} {width="\linewidth"}
[0.245]{} {width="\linewidth"}
[0.245]{} {width="\linewidth"}
Ablation Study
--------------
Since there are many components in our architecture, we analyze their impacts via an ablation study (**RQ2**). Table \[tb:ablation\] shows the performance of our default method and its **8** variants on all four dataset (with $d=50$). We introduce the variants and analyze their effect respectively:
Architecture *Beauty* *Games* *Steam* *ML-1M*
----------------------- -------------------- -------------------- -------------------- --------------------
\(0) Default 0.3142 0.5360 0.6306 0.5905
\(1) Remove PE **0.3183** 0.5301 0.6036 0.5772
\(2) Unshared IE 0.2437$\downarrow$ 0.4266$\downarrow$ 0.4472$\downarrow$ 0.4557$\downarrow$
\(3) Remove RC 0.2591$\downarrow$ 0.4303$\downarrow$ 0.5693 0.5535
\(4) Remove Dropout 0.2436$\downarrow$ 0.4375$\downarrow$ 0.5959 0.5801
\(5) 0 Block ($b$=0) 0.2620$\downarrow$ 0.4745$\downarrow$ 0.5588$\downarrow$ 0.4830$\downarrow$
\(6) 1 Block ($b$=1) 0.3066 **0.5408** 0.6202 0.5653
\(7) 3 Blocks ($b$=3) 0.3078 0.5312 0.6275 **0.5931**
\(8) Multi-Head 0.3080 0.5311 0.6272 0.5885
: Ablation analysis (NDCG@10) on four datasets. Performance better than the default version is boldfaced. ‘$\downarrow$’ indicates a severe performance drop (more than 10%).
\[tb:ablation\]
- \(1) *Remove PE (Positional Embedding)*: Without the positional embedding $\P$, the attention weight on each item depends only on item embeddings. That is to say, the model makes recommendations based on users’ past actions, but their order doesn’t matter. This variant might be suitable for sparse datasets, where user sequences are typically short. This variant performs better then the default model on the sparsest dataset (*Beauty*), but worse on other denser datasets.
- \(2) *Unshared IE (Item Embedding)*: We find that using two item embeddings consistently impairs the performance, presumably due to overfitting.
- \(3) *Remove RC (Residual Connections)*: Without residual connections, we find that performance is significantly worse. Presumably this is because information in lower layers (e.g. the last item’s embedding and the output of the first block) can not be easily propagated to the final layer, and this information is highly useful for making recommendations, especially on sparse datasets.
- \(4) *Remove Dropout*: Our results show that dropout can effectively regularize the model to achieve better test performance, especially on sparse datasets. The results also imply the overfitting problem is less severe on dense datasets.
- (5)-(7) *Number of blocks*: Not surprisingly, results are inferior with zero blocks, since the model would only depend on the last item. The variant with one block performs reasonably well, though using two blocks (the default model) still boosts performance especially on dense datasets, meaning that the hierarchical self-attention structure is helpful to learn more complex item transitions. Using three blocks achieves similar performance to the default model.
- \(8) *Multi-head*: The authors of Transformer [@transform] found that it is useful to use ‘multi-head’ attention, which applies attention in $h$ subspaces (each a $d/h$-dimensional space). However, performance with two heads is consistently and slightly worse than single-head attention in our case. This might owe to the small $d$ in our problem ($d=512$ in Transformer), which is not suitable for decomposition into smaller subspaces.
Training Efficiency & Scalability
---------------------------------
We evaluate two aspects of the training efficiency (**RQ3**) of our model: Training speed (time taken for one epoch of training) and convergence time (time taken to achieve satisfactory performance). We also examine the scalability of our model in terms of the maximum length $n$. All experiments are conducted with a single GTX-1080 Ti GPU.
### **Training Efficiency** {#training-efficiency .unnumbered}
Figure \[fig:speed\] demonstrates the efficiency of deep learning based methods with GPU acceleration. GRU4Rec is omitted due to its inferior performance. For fair comparison, there are two training options for Caser and GRU4Rec$^{\text{+}}$: using complete training data or just the most recent 200 actions (as in SASRec). For computing speed, SASRec only spends 1.7 seconds on model updates for one epoch, which is over 11 times faster than Caser (19.1s/epoch) and 18 times faster than GRU4Rec$^{\text{+}}$ (30.7s/epoch). We also see that SASRec converges to optimal performance within around 350 seconds on *ML-1M* while other models require much longer. We also find that using full data leads to better performance for Caser and GRU4Rec$^{\text{+}}$.
![Training efficiency on *ML-1M*. SASRec is an order of magnitude faster than CNN/RNN-based recommendation methods in terms of training time per epoch and in total.[]{data-label="fig:speed"}](figs/speed){width=".40\textwidth"}
[0.245]{} {width="\linewidth"}
[0.245]{} {width="\linewidth"}
[0.245]{} {width="\linewidth"}
[0.245]{} {width="\linewidth"}
### **Scalability** {#scalability .unnumbered}
As with standard MF methods, SASRec scales linearly with the total number of users, items and actions. A potential scalability concern is the maximum length $n$, however the computation can be effectively parallelized with GPUs. Here we measure the training time and performance of SASRec with different $n$, empirically study its scalability, and analyze whether it can handle sequential recommendation in most cases. Table \[tab:scale\] shows the performance and efficiency of SASRec with various sequence lengths. Performance is better with larger $n$, up to around $n=500$ at which point performance saturates (possibly because 99.8% of actions have been covered). However, even with $n=600$, the model can be trained in 2,000 seconds, which is still faster than Caser and GRU4Rec$^{\text{+}}$. Hence, our model can easily scale to user sequences up to a few hundred actions, which is suitable for typical review and purchase datasets. We plan to investigate approaches (discussed in Section \[sec:complexity\]) for handling very long sequences in the future.
$n$ 10 50 100 200 300 400 500 600
--------- ------- ------- ------- ------- ------- ------- ------- ------- --
Time(s) 75 101 157 341 613 965 1406 1895
NDCG@10 0.480 0.557 0.571 0.587 0.593 0.594 0.596 0.595
: Scalability: performance and training time with different maximum length $n$ on *ML-1M*.
\[tab:scale\]
Visualizing Attention Weights {#sec:vis}
-----------------------------
Recall that at time step $t$, the self-attention mechanism in our model adaptively assigns weights on the first $t$ items depending on their position embeddings and item embeddings. To answer **RQ4**, we examine all training sequences and seek to reveal meaningful patterns by showing the average attention weights on positions as well as items.
### **Attention on Positions** {#attention-on-positions .unnumbered}
Figure \[fig:vis\_pos\] shows four heatmaps of average attention weights on the last 15 positions at the last 15 time steps. Note that when we calculate the average weight, the denominator is the number of *valid* weights, so as to avoid the influence of padding items in short sequences.
We consider a few comparisons among the heatmaps:
- *(a) vs. (c)*: This comparison indicates that the model tends to attend on more recent items on the sparse dataset *Beauty*, and less recent items for the dense dataset *ML-1M*. This is the key factor that allows our model to adaptively handle both sparse and dense datasets, whereas existing methods tend to focus on one end of the spectrum.
- *(b) vs. (c)*: This comparison shows the effect of using positional embeddings (PE). Without them attention weights are essentially uniformly distributed over previous items, while the default model (c) is more sensitive in position as it is inclined to attend on recent items.
- *(c) vs. (d)*: Since our model is hierarchical, this shows how attention varies across different blocks. Apparently, attention in high layers tends to focus on more recent positions. Presumably this is because the first self-attention block already considers all previous items, and the second block does not need to consider far away positions.
Overall, the visualizations show that the behavior of our self-attention mechanism is *adaptive*, *position-aware*, and *hierarchical*.
### **Attention Between Items** {#attention-between-items .unnumbered}
Showing attention weights between a few cherry-picked items might not be statistically meaningful. To perform a broader comparison, using *MovieLens-1M*, where each movie has several categories, we randomly select two disjoint sets where each set contains 200 movies from 4 categories: *Science Fiction (Sci-Fi)*, *Romance*, *Animation*, and *Horror*. The first set is used for the query and the second set as the key. Figure \[fig:vis\_item\] shows a heatmap of average attention weights between the two sets. We can see the heatmap is approximately a block diagonal matrix, meaning that the attention mechanism can identify similar items (e.g. items sharing a common category) and tends to assign larger weights between them (without being aware of categories in advance).
![Visualization of average attention between movies from four categories. This shows our model can uncover items’ attributes, and assigns larger weights between similar items.[]{data-label="fig:vis_item"}](figs/cat){width=".45\textwidth"}
Conclusion
==========
In this work, we proposed a novel self-attention based sequential model *SASRec* for next item recommendation. SASRec models the entire user sequence (without any recurrent or convolutional operations), and adaptively considers consumed items for prediction. Extensive empirical results on both sparse and dense datasets show that our model outperforms state-of-the-art baselines, and is an order of magnitude faster than CNN/RNN based approaches. In the future, we plan to extend the model by incorporating rich context information (e.g. dwell time, action types, locations, devices, etc.), and to investigate approaches to handle very long sequences (e.g. clicks).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Over the last few decades, classical density-functional theory (DFT) and its dynamic extensions (DDFTs) have become powerful tools in the study of colloidal fluids. Recently, previous DDFTs for spherically-symmetric particles have been generalised to take into account both inertia and hydrodynamic interactions, two effects which strongly influence non-equilibrium properties. The present work further generalises this framework to systems of anisotropic particles. Starting from the Liouville equation and utilising Zwanzig’s projection-operator techniques, we derive the kinetic equation for the Brownian particle distribution function, and by averaging over all but one particle, a DDFT equation is obtained. Whilst this equation has some similarities with DDFTs for spherically-symmetric colloids, it involves a translational-rotational coupling which affects the diffusivity of the (asymmetric) particles. We further show that, in the overdamped (high friction) limit, the DDFT is considerably simplified and is in agreement with a previous DDFT for colloids with arbitrary shape particles.'
author:
- 'Miguel A.'
- 'Benjamin D.'
- Serafim Kalliadasis
title: Dynamical density functional theory for orientable colloids including inertia and hydrodynamic interactions
---
Introduction {#intro}
============
The study of colloidal fluids, which typically involve particles of micrometer size suspended in a simple fluid bath (where the particles are the atoms-molecules themselves, hence of nanometer size), goes back to the 19th Century with the work of @brown_brief_1828. Since then, numerous studies have been devoted to the description of the dynamics of such systems. A particular challenge is the multiscale nature of the dynamics due to the mass separation of the bath and colloidal particles which in turn implies that the velocities of the bath particles are much higher than those of the colloidal particles. While the time evolution of colloidal fluids can be formally studied by modelling the full systems, i.e. considering the Newtonian dynamics of both bath and colloidal particles, such an approach is computationally intractable, due to both the number of particles and the very different timescales one has to consider[@goddard_general_2012; @goddard_unification_2013], and thus there is a need for coarse graining leading to reduced models. One such reduced model consists of the stochastic Langevin Equations (LEs) for colloidal positions and momenta or equivalently the corresponding Fokker-Planck equation (FPE) for the $N$-body probability distribution function. The LEs were originally proposed heuristically, but can be formally justified as shown in Appendix \[app:projection-operator-tech\] (where for the first time a detailed microscopic derivation of the rotational part of the FPE is offered starting from the full system of bath and colloidal particles). A consequence of this coarse graining down to the dynamics of only the colloidal-particle degrees of freedom is that we no longer deal with deterministic but with random variables. The advantage is that it removes the enormous number of variables related to the dynamical variables of the bath, traditionally via projection-operator techniques which effectively allow to average out the bath[@kirkwood_statistical_1946; @murphy_brownian_1972; @deutch_molecular_1971; @wilemski_derivation_1976; @ermak_brownian_1978; @grabert_microdynamics_1980; @darve_computing_2009]. Despite such a simplification, the resultant equations are still computationally prohibitive because of the large number of particles that need to be taken into consideration to describe the behavior of colloidal fluids. A theoretical way out is to obtain mean field approximations which are independent of the number of particles. For an $N$-particle system, a standard procedure in the statistical mechanics of classical fluids is to average over $N-n$ (with $n < N$) colloidal particles, resulting in a time-evolution equation for the $n$-body reduced distribution function. However, to obtain a closed set of equations for a given $n$ requires knowledge of the relationships between the full and the reduced distributions. Dynamical density functional theories (DDFTs) perform this procedure for $n=1$, i.e. for the one-body density, and typically result in a continuity equation for the density and a time-evolution equation for the current, that is now a functional of the density. The functional that relates the current and the density is generally unknown, but, for a practical implementation of DDFTs it needs to be approximated. A standard approach, for the overdamped dynamics, is to take it as the free-energy functional of a system with the same density at equilibrium. This functional has been well-studied in the statistical mechanics of fluids[@evans_nature_1979; @lutsko_recent_2010]. Also, approximating the functional relating the current and the density with the free energy at equilibrium also ensures that the DDFTs reduce to the corresponding equilibrium DFTs[@goddard_general_2012; @goddard_unification_2013].
The first derivations of a DDFT go back to the early works of @evans_nature_1979 and @dieterich_nonlinear_1990. These initial formulations were mostly phenomenological and subsequent efforts focused on deriving rigorously the same results and also DDFTs that include additional effects such as inertia and hydrodynamic interactions (HIs)[@marconi_dynamic_2000; @archer_dynamical_2004; @archer_dynamical_2009; @espanol_derivation_2009; @goddard_overdamped_2012; @goddard_general_2012; @goddard_unification_2013; @donev_dynamic_2014].
There have also been several efforts to test the derived DDFTs and their predictive capabilities. In fact there have been successful applications to a wide spectrum of physical settings, from hard rod and hard sphere systems[@goddard_unification_2013], mixtures[@goddard_multi-species_2013], thermodynamic phase transitions such as nucleation and spinodal decomposition[@van_teeffelen_colloidal_2008; @lutsko_dynamical_2012] to the calculation of the van Hove distribution function for Brownian hard spheres[@hopkins_van_2010] and crystallization[@neuhaus_density_2014].
DDFTs have been derived for both the overdamped and inertial regimes and although they were initially applied to spherically-symmetric colloids, they have also been generalised to systems where both orientational and translational motions of particles are taken into account, either under weak[@rex_dynamical_2007] or strong[@wittkowski_dynamical_2011] coupling between the two motions. However, these generalisations are restricted to the overdamped regime and neglect HIs. The main goal of the present study is precisely to formulate a DDFT for orientable particles to include both inertia and HIs. Our methodology follows closely the momentum-moments approach adopted by @archer_dynamical_2009 to obtain a DDFT for spherically-symmetric colloidal particles with inertia and also by @goddard_general_2012[@goddard_general_2012; @goddard_unification_2013] to include both inertia and HIs in this DDFT.
In Section \[sec:evolution-equations\] we introduce the equations of motion for a system of $N$ interacting orientable colloidal particles immersed in a thermal bath. Analogous to the spherically-symmetric case, these equations are the LEs for the phase-space coordinates: positions, Eulerian angles, linear and angular (also referred to as rotational) momenta. These equations have already been derived and examined thoroughly before[@dickinson_brownian_1985; @wolynes_dynamical_1977; @hernandez-contreras_brownian_1996], but for the sake of clarity and completeness they are reviewed in Appendices \[app:projection-operator-tech\] and \[app:neglecting-bath-inertia\], in particular highlighting their range of applicability. The derived LEs naturally contain the full rotational-translational coupling which is crucial in the DDFT derivation that follows. In Section \[sec:continuity-current-equation\], we follow[@goddard_general_2012; @goddard_unification_2013] and integrate the Kramer’s equation over all but one particle, yielding a continuity equation for the density distribution and a time-evolution equation for the current. At this point we require closure approximations to obtain a final equation dependent on the density distribution only. These approximations, discussed in detail in Section \[sec:extended-unified-DDFT\], are: (i) the local-equilibrium approximation for the one-body distribution function; (ii) the adiabatic approximation involving the equilibrium free-energy functional; (iii) separation of the $N$-body distribution into local equilibrium and non-equilibrium contributions; and (iv) an approximation for the two-body interactions. We then arrive at the central result of this work, i.e. a generalised DDFT equation to describe systems of orientable particles taking account of both inertia and HIs. Finally, we compare our derived DDFT with previous DDFTs. As expected, if the rotational degrees of freedom are ignored, our DDFT reduces to the one obtained for point-like or spherical particles in the presence of both inertia and HIs[@goddard_general_2012; @goddard_unification_2013], and hence also to several other DDFTs[@marconi_dynamic_2000; @archer_dynamical_2009; @archer_dynamical_2004; @espanol_derivation_2009; @goddard_overdamped_2012] which have neglected inertia or HIs or both. For orientable particles, neglecting inertial effects and HIs allows us to connect our DDFT with previous ones, recovering in particular the DDFTs derived by @rex_dynamical_2007 and @wittkowski_dynamical_2011. Concluding remarks along with a discussion of open problems are offered in Section \[sec:conclusions\].
Evolution equations {#sec:evolution-equations}
===================
Consider a closed system of $N$ identical, asymmetric colloidal particles with mass $m$ immersed in a fluid of $n\gg N$ identical bath particles with mass $m_b \ll m$. This mass separation plays a vital role in deriving the Brownian equations of motion governing colloidal dynamics via projection-operator techniques[@kirkwood_statistical_1946; @murphy_brownian_1972; @deutch_molecular_1971; @wilemski_derivation_1976; @ermak_brownian_1978; @grabert_microdynamics_1980; @darve_computing_2009] (see also Appendix \[app:projection-operator-tech\]). We start by describing the configuration state of the colloidal particles by the position vectors $\mathbf{r}^N=\{\mathbf{r}_1,\dots,\mathbf{r}_N\}$ for their centres of mass and the Eulerian angles ${\boldsymbol{\alpha}}^N=\{{\boldsymbol{\alpha}}_1,\dots,{\boldsymbol{\alpha}}_N\}$, with ${\boldsymbol{\alpha}}_i=(\theta_i,\phi_i,\chi_i)$ defined in the same manner as @condiff_brownian_1966 [@goldstein_classical_2002] or @jose_classical_2013. These angles characterise the orientation of principal-axes frame of the particles, $\mathfrak{B}$, relative to the space-fixed frame, $\mathfrak{S}$. We also denote the translational and angular momenta as ${\mathbf{p}}^N=\{{\mathbf{p}}_1,\dots,{\mathbf{p}}_N\}$ and ${\mathbf{L}}^N=\{{\mathbf{L}}_1,\dots,{\mathbf{L}}_N\}$, respectively. The angular velocity of the $i$th colloidal particle, $\boldsymbol{\omega}_i$, is related to its angular momentum by the equation ${\mathbf{L}}_i=\mathbb{I}_i\boldsymbol{\omega}_i$ with $\mathbb{I}_i$ the inertia tensor. In the principal-axes frames $\mathfrak{B}$, such a tensor is diagonal and constant, $\mathbb{I}_i\equiv\mathbb{I}=\text{diag}(I_{1},I_{2},I_{3})$. The transformation between the frames $\mathfrak{B}$ and $\mathfrak{S}$ is given by the rotation operator $\mathcal{R}_i$,[@jose_classical_2013] such that $\mathbf{x}=\mathcal{R}_i\,\mathbf{x}'$ where $\mathbf{x}\in\mathfrak{B}$ and $\mathbf{x}'\in\mathfrak{S}$. Hence $\boldsymbol{\omega}_i=\mathcal{R}_i\boldsymbol{\omega}_i'$, with $\boldsymbol{\omega}_i'$ the angular velocity in $\mathfrak{S}$, and $\mathbb{I}_i'=\mathcal{R}_i^\top\,\mathbb{I}\,\mathcal{R}_i$, which is neither diagonal nor constant. The angular velocity ${\boldsymbol{\omega}}_i$ can be related to the Eulerian velocities, $\dot{{\boldsymbol{\alpha}}}_i$, via the transformation ${\boldsymbol{\omega}}_i({\boldsymbol{\alpha}}_i) \equiv
\dot{{\boldsymbol{\Phi}}_i}({\boldsymbol{\alpha}}_i) = {\boldsymbol{\Lambda}}_i^\top \dot{{\boldsymbol{\alpha}}}_i$, where the dot denotes time derivative and ${\boldsymbol{\Phi}}_i$ highlights the fact that ${\boldsymbol{\omega}}_i$ is the vector accounting for the rate of change of angular displacement over the Cartesian frame $\mathfrak{B}$. The exact definition of ${\boldsymbol{\Lambda}}^\top$ can be found in Appendix \[app:projection-operator-tech\].
For simplicity, we will consider angular quantities with respect to the body-fixed frames. As such, the dynamics of colloidal particles are described by the Langevin equations[@dickinson_brownian_1985; @dickinson_brownian_1985-1] (see detailed derivation in Appendix \[app:projection-operator-tech\]) $$\begin{aligned}
\dot{{\mathbf{r}}}_j(t)=&\,\frac{1}{m}{\mathbf{p}}_j,\quad
\dot{{\boldsymbol{\Phi}}}_j(t)=\,\mathbb{I}^{-1}{\mathbf{L}}_j,\label{eq:1}\\
\dot{{\mathbf{p}}}_j(t)=&\, -\frac{\partial}{\partial {\mathbf{r}}_j}V({\mathbf{r}}^N,{\boldsymbol{\alpha}}^N)-\sum_{k=1}^N\left({\boldsymbol{\Gamma}}_{jk}^{TT}{\mathbf{p}}_k+{\boldsymbol{\Gamma}}_{jk}^{TR}{\mathbf{L}}_k\right)+\sum_{k=1}^N \left( \mathbf{A}_{jk}^{TT}\mathbf{f}_k(t)+\mathbf{A}_{jk}^{TR}\mathbf{t}_k(t) \right),
\notag
\\
\dot{{\mathbf{L}}}_j(t)=&
-\frac{\partial}{\partial {\boldsymbol{\Phi}}_j}{V}({\mathbf{r}}^N,{\boldsymbol{\alpha}}^N)-(\mathbb{I}^{-1}{\mathbf{L}}_j)\times{\mathbf{L}}_j
\,
-\sum_{k=1}^N\left({\boldsymbol{\Gamma}}_{jk}^{RT}{\mathbf{p}}_k+{\boldsymbol{\Gamma}}_{jk}^{RR}{\mathbf{L}}_k\right)
+\sum_{k=1}^N \left( \mathbf{A}_{jk}^{RT}\mathbf{f}_k(t)+\mathbf{A}_{jk}^{RR}\mathbf{t}_k(t) \right) \notag,\end{aligned}$$ or, more compactly, $$\begin{aligned}
\dot{{\boldsymbol{\mathfrak{r}}}}_j(t)=&\ {\boldsymbol{\mathfrak{m}}}_j^{-1}{\boldsymbol{\mathfrak{p}}}_j(t)\label{eq:2},\\
\dot{{\boldsymbol{\mathfrak{p}}}}_j(t)=&\ {\boldsymbol{\mathfrak{F}}}_j-\sum_{k=1}^N{\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N){\boldsymbol{\mathfrak{p}}}_k + \sum_{k=1}^N\mathbf{A}_{jk}\boldsymbol{\xi}_k(t)\notag,\end{aligned}$$ with ${\boldsymbol{\mathfrak{m}}}_j = \text{diag}(m\mathbf{1},\mathbb{I})$, $\boldsymbol{\mathfrak{r}}_j=(\mathbf{r}_j,\boldsymbol{\Phi}_j)$, $\boldsymbol{\mathfrak{p}}_j=(\mathbf{p}_j,\mathbf{L}_j)$ and ${\boldsymbol{\mathfrak{F}}}_j=(-\frac{\partial V}{\partial {\mathbf{r}}_j},-\frac{\partial V}{\partial {\boldsymbol{\Phi}}_j}-(\mathbb{I}^{-1}{\mathbf{L}}_j)\times{\mathbf{L}}_j)$, where $V({\boldsymbol{\mathfrak{r}}}^N)$ is the solvent-averaged interaction-potential of mean force[@snook_langevin_2006]. The gradient operators are defined by $\boldsymbol{\nabla}_{\mathfrak{r}_j}=(\frac{\partial}{\partial
\mathbf{r}_j},\frac{\partial }{\partial \boldsymbol{\Phi}_j})^\top$, with $\frac{\partial}{\partial {\boldsymbol{\Phi}}}=\mathbf{e}_x\frac{\partial}{\partial
\Phi_x}+\mathbf{e}_y\frac{\partial}{\partial
\Phi_y}+\mathbf{e}_z\frac{\partial}{\partial \Phi_z}$ the angular gradient \[see Equations (\[eq:a-17\])–(\[eq:a-18\])\], and $\boldsymbol{\nabla}_{\boldsymbol{\mathfrak{p}}_j}= (\frac{\partial
}{\partial \mathbf{p}_j},\frac{\partial }{\partial \mathbf{L}_j})^\top$. Here ${\boldsymbol{\xi}}_j=(\mathbf{f}_j,\mathbf{t}_j)^\top$ is a 6-dimensional Gaussian white noise vector where $\mathbf{f}_j$ and $\mathbf{t}_j$ are random forces and torques acting upon the $j$-th particle, respectively, such that $\langle
\xi_j^a(t)\rangle=0$ and $\langle \xi_j^a(t)\xi_k^b(t')\rangle =
2\delta_{jk}\delta^{ab}\delta(t-t')$. As for the spherical case, the motion of the colloidal particles induces flows in the bath which in turn result in forces on the colloids, what we have already referred to in Section \[intro\] as HIs. These HIs are represented by the $6\times6$ friction tensors, $
{\boldsymbol{\Gamma}}_{jk}=({\boldsymbol{\Gamma}}_{jk}^{\mu\nu})
$ where the superscripts $\mu,\nu\in\{T,R\}$ denote the coupling between the translational and angular momenta. Furthermore, the interaction between bath and colloidal particles has a random component representing unpredictable collisions of the bath particles with the colloidal particles. The strength of these random forces is given by the $6\times 6$ tensors $\mathbf{A}_{jk}=(\mathbf{A}_{jk}^{\mu\nu})$ which are coupled to the friction tensors via the fluctuation-dissipation relation,[@wolynes_dynamical_1977; @dickinson_brownian_1985] $$\begin{aligned}
{\boldsymbol{\mathfrak{m}}}\,{\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N) = \beta\sum_{l=1}^N\mathbf{A}_{jl}({\boldsymbol{\mathfrak{r}}}^N)\mathbf{A}_{kl}({\boldsymbol{\mathfrak{r}}}^N)\label{eq:3}.\end{aligned}$$ Here $\beta=1/k_BT$ with $k_B$ the Boltzmann constant and $T$ the temperature, is assumed to be constant. We define the $6N\times6N$ total-mass tensor ${\boldsymbol{\mathfrak{m}}}_{_N}=\text{diag}({\boldsymbol{\mathfrak{m}}}_j)$, along with the $6N\times 6N$ tensors ${\boldsymbol{\Gamma}}=({\boldsymbol{\Gamma}}_{jk})$, and $\mathbf{A}=(\mathbf{A}_{jk})$, transforming Equation (\[eq:3\]) to ${\boldsymbol{\mathfrak{m}}}_{_N}{\boldsymbol{\Gamma}}= \beta\mathbf{A}\mathbf{A}^\top$. Due to the asymmetry of the particles, the friction and noise matrices are not generally symmetric. However, as discussed by @brenner_stokes_1964-2 and @condiff_brownian_1966, the asymmetric part of ${\boldsymbol{\Gamma}}_{ij}$ is exclusively due to nondissipative gyrostatic forces and torques acting on the particles. If such effects can be neglected[@condiff_brownian_1966], then one can assume ${\boldsymbol{\Gamma}}_{jk}^{TR}=({\boldsymbol{\Gamma}}_{jk}^{RT})^\top$. The fluctuation-dissipation relation is then $\mathbf{A} = \sqrt{k_BT {\boldsymbol{\mathfrak{m}}}_{_N}
{\boldsymbol{\Gamma}}}$, analogous to that for spherical particles.
Starting from the Smoluchowski equation, a fluctuating DDFT equation was derived recently[@donev_dynamic_2014]. The authors of this study argued that any equation of motion that accounts for inertial effects of the colloidal particles must include the inertia of the fluid bath. Moreover, they claim that there are only two consistent ways of dealing with this problem: either *fluctuating hydrodynamics*[@hauge_fluctuating_1973] or the overdamped LEs based on the work of @hinch_application_1975 and @roux_brownian_1992. They justify this claim using the estimate obtained by @roux_brownian_1992, who also refers to @hinch_application_1975, of the typical relaxation time of the bath, eventually invalidating the inertial LEs. Nevertheless, while the inertial equations (\[eq:1\]) are neither formally exact nor completely general, it can be shown that there exist very sensible regimes in which neglecting inertia in the bath is appropriate, namely when the colloidal-particle density, $\rho_B$, is much larger than that of the bath particles, $\rho_b$. This is done in Appendix \[app:neglecting-bath-inertia\] using dimensional analysis showing that there is indeed a mass-induced time-scale separation when $\rho_b/\rho_B\ll1$. Under such circumstances the LEs and the associated FPE must be applicable, as explicitly pointed out also by @roux_brownian_1992.
From the system of stochastic differential equations (SDEs) in Eq. (\[eq:2\]), one can obtain[@risken_fokker-planck_1996] (see Appendix \[app:projection-operator-tech\]) the time-evolution equation of the phase-space probability density function (PDF), ${f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t)$, to find each particle $j$ with positions ${\mathbf{r}}_j,\ {\boldsymbol{\alpha}}_j$ and momenta ${\mathbf{p}}_j,\ {\mathbf{L}}_j$ at time $t$, $$\begin{aligned}
\partial_t {f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t) +&
\sum_{j=1}^N{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_j}\cdot \big({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_j{f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t) \big)+\sum_{j=1}^N{\boldsymbol{\mathfrak{F}}}_j\cdot{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j}{f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t) \label{eq:4}\\
&= \sum_{j,k=1}^N \,{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j}\cdot{\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N)\left({\boldsymbol{\mathfrak{p}}}_k+k_BT{\boldsymbol{\mathfrak{m}}}\,{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_k}\right){f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t)\notag.\end{aligned}$$ It is well-known that Equations (\[eq:2\]) or (\[eq:4\]) are computationally tractable only for small systems. This problem can be circumvented to an extent by assuming a very simple structure for the friction tensor, namely ${\boldsymbol{\Gamma}}= \text{diag}(\gamma_{TT}\mathbf{1},\gamma_{RR}\mathbf{1})$. While such an approximation certainly reduces the numerical complexity \[since determining $\mathbf{A}$ is now $O(N)$ rather than $O(N^3)$\], it physically corresponds to ignoring HIs, thus neglecting crucial physical information. Another approach to tackle the problem, which is the one we adopt, is to derive a coarse-grained/mean-field model by averaging over the degrees of freedom of all but a few particles[@wu_density-functional_2007; @chan_time-dependent_2005; @goddard_unification_2013; @goddard_multi-species_2013]. This method yields a lower-dimensional system but requires knowledge of the functional relation between ${f^{(N)}}$ and the reduced distribution function. Therefore, convenient and accurate closures for the dynamics of the reduced distribution function must be found.
Continuity and flux equations \[sec:continuity-current-equation\]
=================================================================
Given (\[eq:4\]), we can reduce to a lower dimensional problem by averaging over all but a few particles. Our first assumption involves the solvent-averaged interaction potential $V({\boldsymbol{\mathfrak{r}}}^N)$. Inspired by the decomposition proposed for spherical particles[@goddard_unification_2013] we assume that this potential can be split into linear combinations of $n$-particle interactions, i.e. $$\begin{aligned}
V({\boldsymbol{\mathfrak{r}}}^N;t)= \sum_{j=1}^NV_1({\boldsymbol{\mathfrak{r}}}_j;t)+\sum_{n=2}^N\frac{1}{n!}\sum_{k_1\neq\dots\neq k_n=1}^n V_n({\boldsymbol{\mathfrak{r}}}_{{k}_1},\dots,{\boldsymbol{\mathfrak{r}}}_{{k}_n})\label{eq:5}.\end{aligned}$$ For ease of notation we also adopt the decomposition ${\boldsymbol{\Gamma}}_{ij}^{\mu\nu}=\gamma_{\mu\nu}\delta_{ij}\mathbf{1}+\gamma_{\mu\nu}\widetilde{{\boldsymbol{\Gamma}}}_{ij}^{\mu\nu}$, recalling that we are not making use of Einstein’s summation convention. Thus, $$\begin{aligned}
{\boldsymbol{\Gamma}}_{ij}=\delta_{ij}{\boldsymbol{\Upsilon}}+\widetilde{{\boldsymbol{\Upsilon}}}_{ij} ,\end{aligned}$$ with the $6\times 6$ tensor ${\boldsymbol{\Upsilon}}= (\gamma_{\mu\nu}\mathbf{1})$, where $\gamma_{\mu\nu}$ is the constant friction coefficient for an isolated particle and $\widetilde{{\boldsymbol{\Upsilon}}}_{ij}=(\gamma_{\mu\nu}\widetilde{{\boldsymbol{\Gamma}}}_{ij}^{\mu\nu})$, where $\widetilde{{\boldsymbol{\Gamma}}}_{ij}^{\mu\nu}$ are the translational-rotational HI tensors. In the special case when $\gamma_{\mu\nu}=0,\ \forall\mu,\nu$, the FPE (\[eq:4\]) becomes the Liouville equation initially derived by @condiff_brownian_1966[@condiff_brownian_1966; @condiff_transport_1969; @dahler_kinetic_1963] for molecular fluids consisting of non-spherical particles. In this case one usually invokes a Kirkwood-type time smoothing technique[@condiff_brownian_1966; @kirkwood_statistical_1946], rather than a local-equilibrium approximation involving a free-energy functional (see Section \[sec:extended-unified-DDFT\]).
We now define the reduced distributions obtained by integration of the phase-space PDF, $$\begin{aligned}
{f^{(n)}}({\boldsymbol{\mathfrak{r}}}^n,{\boldsymbol{\mathfrak{p}}}^n;t)=\frac{N!}{(N-n)!} \int d{\boldsymbol{\mathfrak{r}}}^{N-n}d{\boldsymbol{\mathfrak{p}}}^{N-n}\,{f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t)
\label{eq:7}\end{aligned}$$ where $d{\boldsymbol{\mathfrak{r}}}^{N-n}\equiv d{\boldsymbol{\mathfrak{r}}}_{n+1}\dots d{\boldsymbol{\mathfrak{r}}}_N$, with an analogous expression for $d{\boldsymbol{\mathfrak{p}}}$, and the corresponding configuration-space distribution, $$\begin{aligned}
{\rho^{(n)}}({\boldsymbol{\mathfrak{r}}}^n;t) = \int d{\boldsymbol{\mathfrak{p}}}^n\,{f^{(n)}}({\boldsymbol{\mathfrak{r}}}^n,{\boldsymbol{\mathfrak{p}}}^n;t).
\label{eq:8}\end{aligned}$$ We denote the domains of integration by $\mathfrak{V}=\mathcal{V}\times\mathcal{S}$, with $\mathcal{V}=\mathbb{R}^3$ and $\mathcal{S}=[0,2\pi)\times[0,\pi)\times[0,2\pi)$, and so $\int d{\boldsymbol{\mathfrak{r}}}\equiv \int_{\mathcal{V}}d\mathcal{V}\int_{\mathcal{S}}d\mathcal{S}$ and $\int d{\boldsymbol{\mathfrak{p}}}\equiv \int_{\mathcal{V}}d\mathcal{V}\int_{\mathcal{V}}d\mathcal{V}$, where $$\begin{aligned}
\int_{\mathcal{V}}d\mathcal{V}=& \int_{0}^\infty dx \int_{0}^\infty dy \int_{0}^\infty dz \label{eq:9}\\
\int_{\mathcal{S}}d\mathcal{S}=&\int_{0}^{2\pi}d\theta \int_{0}^{\pi}d\phi\, \sin(\phi)\int_{0}^{2\pi} d\chi.
\notag\end{aligned}$$ For ease of notation $\rho^{(1)}$ is replaced with $\rho$, also known as the one-body density. We now take successive momentum moments of Equation (\[eq:4\]) starting at zeroth order and integrating over all but one-particle position and momentum, i.e. $N \int \text{(\ref{eq:4}) } d{\boldsymbol{\mathfrak{r}}}^{N-1} d{\boldsymbol{\mathfrak{p}}}^{N}$, resulting in the continuity equation $$\begin{aligned}
\partial_t\rho({\boldsymbol{\mathfrak{r}}}_1;t) + {\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot{\boldsymbol{\mathfrak{j}}}({\boldsymbol{\mathfrak{r}}}_1;t) = 0,
\label{eq:10}\end{aligned}$$ with the flux defined as ${\boldsymbol{\mathfrak{j}}}({\boldsymbol{\mathfrak{r}}}_1;t)=\int d{\boldsymbol{\mathfrak{p}}}\,({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1)f^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1;t)$.
To obtain the first-momentum moment equation we use $N\int ({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1)\otimes \text{(\ref{eq:4}) }d{\boldsymbol{\mathfrak{r}}}^{N-1}d{\boldsymbol{\mathfrak{p}}}^N$ giving $$\begin{aligned}
\partial_t{\boldsymbol{\mathfrak{j}}}({\boldsymbol{\mathfrak{r}}}_1;t)&+{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot\int d{\boldsymbol{\mathfrak{p}}}_1({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1\otimes{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1)f^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1;t)\label{eq:11}\\
&+{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}V_1({\boldsymbol{\mathfrak{r}}}_1;t)\rho({\boldsymbol{\mathfrak{r}}}_1;t)
+{\boldsymbol{\mathfrak{m}}}^{-1}\sum_{n=2}^N\int d{\boldsymbol{\mathfrak{r}}}_2\dots d{\boldsymbol{\mathfrak{r}}}_n{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}V_n({\boldsymbol{\mathfrak{r}}}^n;t)\rho^{(n)}({\boldsymbol{\mathfrak{r}}}^n;t)\notag\\
&+{\boldsymbol{\Upsilon}}_{\parallel}{\boldsymbol{\mathfrak{j}}}({\boldsymbol{\mathfrak{r}}}_1;t)+{\boldsymbol{\Upsilon}}_{\bot}{\boldsymbol{\mathfrak{j}}}_{\bot}({\boldsymbol{\mathfrak{r}}}_1;t)+N{\boldsymbol{\mathfrak{m}}}^{-1}\sum_{j=1}^{N}\int d{\boldsymbol{\mathfrak{r}}}^{N-1}d{\boldsymbol{\mathfrak{p}}}^N\,\widetilde{{\boldsymbol{\Upsilon}}}_{1j}({\boldsymbol{\mathfrak{r}}}^N){\boldsymbol{\mathfrak{p}}}_j{f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t) = 0\notag\end{aligned}$$ where $\otimes$ is the dyadic product. Here we have defined ${\boldsymbol{\Upsilon}}_{\parallel}=\text{diag}(\gamma_{TT}\mathbf{1},\gamma_{RR}\mathbf{1})$, ${\boldsymbol{\Upsilon}}_{\bot}=\text{diag}(\gamma_{TR}\mathbf{1},\gamma_{RT}\mathbf{1})$ and ${\boldsymbol{\mathfrak{j}}}_{\bot}=\int
d{\boldsymbol{\mathfrak{p}}}_1({\boldsymbol{\mathfrak{m}}}^{-1}_{\bot}{\boldsymbol{\mathfrak{p}}}_1)f^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1;t)$, with ${\boldsymbol{\mathfrak{m}}}^{-1}_{\bot}=\text{adiag}(m^{-1}\mathbf{1},\mathbb{I}^{-1})$ the anti-diagonal block matrix with submatrices, $({\boldsymbol{\mathfrak{m}}}_{\bot}^{-1})_{1,2}=
m^{-1}\mathbf{1}$, $({\boldsymbol{\mathfrak{m}}}_\bot^{-1})_{2,1}=\mathbb{I}^{-1}$ and $({\boldsymbol{\mathfrak{m}}}_\bot^{-1})_{i,j}=\mathbf{0}$ otherwise. For particles where coupling cannot be neglected, such as screw-like colloids[@wittkowski_dynamical_2011], this first-momentum moment equation includes an additional term compared to that for spherical particles. This term, involving ${\boldsymbol{\mathfrak{j}}}_\bot({\boldsymbol{\mathfrak{r}}}_1;t)$, reflects the coupled translational and rotational forces. When such a coupling vanishes, as in the case of orthotropic-like particles[@wittkowski_dynamical_2011], the rotational and translational motions are completely decoupled and Equation (\[eq:11\]) can be split into two independent equations.
Equations (\[eq:10\])–(\[eq:11\]), are the basis of our DDFT. In the next section we introduce the closures which link the terms $f^{(1)}$, $\rho^{(n)}$ and ${f^{(N)}}$ of Eq (\[eq:11\]) to the one-body density, following the same reasoning as that applied for the spherical-particle case.
Extending the unified DDFT equation to non-spherical particles\[sec:extended-unified-DDFT\]
===========================================================================================
The closure procedure will be performed in four steps: (i) the term involving the dyadic product will be rewritten using the definition of the kinetic stress, ${\overline{\overline{\boldsymbol{\Pi}}}}({\boldsymbol{\mathfrak{r}}}_1;t)$, and the equilibrium distribution function, $f_{eq}^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1)$; (ii) an adiabatic approximation will be used to substitute the $n$-body density terms with an expression involving the free-energy functional, $\mathcal{F}[\rho]$; (iii) the $N$-body distribution function will be expressed as a sum of a local-equilibrium and non-equilibrium distributions, yielding an extension of the result obtained for spherical-colloidal fluids, and finally; (iv) we restrict our attention to the case of two-body HIs and use an approximation.
Equilibrium distribution, kinetic stress and adiabatic approximation \[subsec:equilibrium-kinetic-stress-adiabatic-approximation\]
----------------------------------------------------------------------------------------------------------------------------------
At equilibrium the one-body distribution function has a Maxwell-Boltzmann structure[@condiff_brownian_1966] $$\begin{aligned}
f_{eq}^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1)=\frac{\sqrt{|\mathbb{I}^{-1}|}}{m^{3/2}(2\pi k_BT)^3}\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)\exp\left(-\frac{({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1)\cdot{\boldsymbol{\mathfrak{p}}}_1}{2k_BT}\right)
\label{eq:12},\end{aligned}$$ with $|\mathbb{I}^{-1}|=\prod_{i}(I_{ii})^{-1}$. Hence, one can easily check that $
\int d{\boldsymbol{\mathfrak{p}}}_1 ({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1\otimes {\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1)f_{eq}^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1)= k_BT{\boldsymbol{\mathfrak{m}}}^{-1}\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)$. By adding and subtracting ${\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot \left(
k_BT{\boldsymbol{\mathfrak{m}}}^{-1}\rho({\boldsymbol{\mathfrak{r}}}_1)\right)$ on the left-hand side of Equation (\[eq:11\]) we can write, $$\begin{aligned}
k_BT{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\rho({\boldsymbol{\mathfrak{r}}}_1;t)+{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot{\overline{\overline{\boldsymbol{\Pi}}}}({\boldsymbol{\mathfrak{r}}}_1;t)&=\,{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot\int d{\boldsymbol{\mathfrak{p}}}_1({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1\otimes{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1)f^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1;t) \label{eq:13}\end{aligned}$$ with the kinetic-stress tensor defined as[@goddard_unification_2013] $$\begin{aligned}
{\overline{\overline{\boldsymbol{\Pi}}}}({\boldsymbol{\mathfrak{r}}}_1;t)&=\,\int d{\boldsymbol{\mathfrak{p}}}_1\left(({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1\otimes{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_1)-k_BT{\boldsymbol{\mathfrak{m}}}^{-1}\right)f^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1;t)\label{eq:14}.\end{aligned}$$ According to this definition, the term involving the one-body distribution function and the dyadic product in Equation (\[eq:11\]) can be identified as the remainder of the spatial and angular divergence of the kinetic stress. Although at this stage Equation (\[eq:13\]) only seems to be a more convenient way of expressing the second term on the left-hand side of Equation (\[eq:11\]), the advantage will become clear in Section \[subsec:local-equilibrium\] where we discuss how to deal with out-of-equilibrium distributions.
But before this, we will make use of the adiabatic approximation. This approximation identifies, at any time $t$, the one-body density, $\rho({\boldsymbol{\mathfrak{r}}}_1;t)$, with the equilibrium one, $\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)$. We now consider the Helmholtz free-energy functional[@evans_nature_1979; @lutsko_recent_2010] $$\begin{aligned}
\mathcal{F}[\rho_{eq}]={\textcolor{black}{k_BT}}\int_{\mathfrak{V}}d{\boldsymbol{\mathfrak{r}}}_1\,\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)\left(\ln\left(\Lambda^{3}\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)-1\right)\right)
+\mathcal{F}_{exc}[\rho_{eq}]+\int_{\mathfrak{V}}d{\boldsymbol{\mathfrak{r}}}_1\,\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)V_1({\boldsymbol{\mathfrak{r}}}_1)
\label{eq:15}\end{aligned}$$ where the first term on the right-hand side accounts for the ideal-rotator gas component, the last term contains the external potential contribution to the free energy, $\mathcal{F}_{exc}$ is an excess over ideal gas contribution and $\Lambda$ is the de Broglie wavelength (which will turn out to be irrelevant). From classical DFT[@lutsko_recent_2010] we know that there is a unique external potential, $\mathfrak{W}({\boldsymbol{\mathfrak{r}}}_1)=w({\boldsymbol{\mathfrak{r}}}_1)+V_1({\boldsymbol{\mathfrak{r}}}_1)$, for which $\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)$ is the equilibrium density. $${\textcolor{black}{
\frac{k_BT{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\rho({\boldsymbol{\mathfrak{r}}}_1)}{\rho({\boldsymbol{\mathfrak{r}}}_1)}+{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\mathfrak{W}({\boldsymbol{\mathfrak{r}}}_1) =
-\left.{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}_{exc}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}\right|_{\rho=\rho_{eq}}}}
\label{eq:16-a}$$ $$ {\textcolor{black}{{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}w({\boldsymbol{\mathfrak{r}}}_1)=-\left.{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}\right|_{\rho=\rho_{eq}}\equiv
-{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\mu({\boldsymbol{\mathfrak{r}}}_1)
}}
\label{eq:16}$$
Furthermore, under the same circumstances Equation (\[eq:11\]) becomes the first equation of Yvon-Born-Green (YBG) hierarchy[@barrat_basic_2003; @liboff_kinetic_2003] $$k_BT{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)+{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\mathfrak{W}({\boldsymbol{\mathfrak{r}}}_1;t)\rho_{eq}({\boldsymbol{\mathfrak{r}}}_1)=
{\boldsymbol{\mathfrak{m}}}^{-1}\sum_{n=2}^N\int d{\boldsymbol{\mathfrak{r}}}_2\dots d{\boldsymbol{\mathfrak{r}}}_n{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}V_n({\boldsymbol{\mathfrak{r}}}^n)\rho^{(n)}_{eq}({\boldsymbol{\mathfrak{r}}}^n)
\label{eq:17}$$ Making the approximation that Equations (\[eq:15\])–(\[eq:17\]) hold out of equilibrium, i.e. substitute $\rho_{eq}^{(n)}$ by $\rho^{(n)}(t)$ for all $n$, yields, $$\begin{aligned}
\rho({\boldsymbol{\mathfrak{r}}}_1;t){\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}=&\,
\rho({\boldsymbol{\mathfrak{r}}}_1;t){\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}V_1({\boldsymbol{\mathfrak{r}}}_1;t)+k_BT{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\rho({\boldsymbol{\mathfrak{r}}}_1;t)\label{eq:18}\\
&+\sum_{n=2}^N\int d{\boldsymbol{\mathfrak{r}}}_2\dots d{\boldsymbol{\mathfrak{r}}}_n{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}V_n({\boldsymbol{\mathfrak{r}}}^n)\rho^{(n)}({\boldsymbol{\mathfrak{r}}}^n;t). \notag\end{aligned}$$ Substituting Equations (\[eq:13\]) and (\[eq:18\]) into Equation (\[eq:11\]) produces, $$\begin{aligned}
\partial_t{\boldsymbol{\mathfrak{j}}}({\boldsymbol{\mathfrak{r}}}_1;t)+&{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot{\overline{\overline{\boldsymbol{\Pi}}}}({\boldsymbol{\mathfrak{r}}}_1;t)+{\boldsymbol{\mathfrak{m}}}^{-1}\rho({\boldsymbol{\mathfrak{r}}}_1;t){\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}
+{\boldsymbol{\Upsilon}}_{\parallel}{\boldsymbol{\mathfrak{j}}}({\boldsymbol{\mathfrak{r}}}_1;t)+{\boldsymbol{\Upsilon}}_{\bot}{\boldsymbol{\mathfrak{j}}}_{\bot}({\boldsymbol{\mathfrak{r}}}_1;t)\label{eq:19}\\
+&N{\boldsymbol{\mathfrak{m}}}^{-1}\sum_{j=1}^{N}\int d{\boldsymbol{\mathfrak{r}}}^{N-1}d{\boldsymbol{\mathfrak{p}}}^N\,\widetilde{{\boldsymbol{\Upsilon}}}_{1j}({\boldsymbol{\mathfrak{r}}}^N){\boldsymbol{\mathfrak{p}}}_j{f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t) = 0.\notag\end{aligned}$$
Local-equilibrium approximation and beyond\[subsec:local-equilibrium\]
----------------------------------------------------------------------
The next step of the derivation is the introduction of a widely used approximation in statistical physics[@pottier_nonequilibrium_2014; @liboff_kinetic_2003; @goddard_unification_2013], namely the local-equilibrium approximation. It allows rewriting the kinetic stress in terms of the one-body density. With the definitions of local density, $\rho$, velocity, ${\boldsymbol{\mathfrak{v}}}$, and temperature, $T$, given by $$\begin{aligned}
\int d{\boldsymbol{\mathfrak{p}}}\left\{\begin{array}{c}
1\\
{\boldsymbol{\mathfrak{p}}}\\
({\boldsymbol{\mathfrak{m}}}^{-1}\mathbf{c})\cdot\mathbf{c}
\end{array}\right\}
f^{(1)}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)=
\left\{\begin{array}{c}
1\\
{\boldsymbol{\mathfrak{m}}}{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t)\\
6k_BT
\end{array}\right\}\rho({\boldsymbol{\mathfrak{r}}};t)
\label{eq:20},\end{aligned}$$ with[@liboff_kinetic_2003] $\mathbf{c}\equiv\mathbf{c}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)={\boldsymbol{\mathfrak{p}}}-{\boldsymbol{\mathfrak{m}}}{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t)$ the deviation-from-the-mean microscopic momentum, the local-equilibrium approximation to the true probability distribution, $$\begin{aligned}
f^{(1)}_{leq}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)=&\, \frac{|\mathbb{I}^{-1}|^{1/2}\rho({\boldsymbol{\mathfrak{r}}};t)}{m^{3/2}(2\pi k_BT)^{3}}\label{eq:21}
\exp\left(-\frac{({\boldsymbol{\mathfrak{m}}}^{-1}\mathbf{c})\cdot\mathbf{c}}{2k_BT}\right)\end{aligned}$$ fulfils the normalisation conditions given in Equation (\[eq:20\]). While approximating $f^{(1)}$ with Equation (\[eq:21\]) can be reasonable as a first-order approach, there is in fact no actual reason to impose such a restriction. We instead introduce the expansion $f^{(1)}=f^{(1)}_{leq}\,\left[1+\varPsi \right]\equiv
f^{(1)}_{leq}+f^{(1)}_{neq}$, which takes into account both local-equilibrium and non-equilibrium effects. The non-equilibrium term must satisfy $$\begin{aligned}
\int d{\boldsymbol{\mathfrak{p}}}\left\{\begin{array}{c}
1\\
{\boldsymbol{\mathfrak{p}}}\\
({\boldsymbol{\mathfrak{m}}}^{-1}\mathbf{c})\cdot\mathbf{c}
\end{array}\right\}
f^{(1)}_{neq}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)=
\left\{\begin{array}{c}
0\\
\mathbf{0}\\
0
\end{array}\right\}.\label{eq:22}\end{aligned}$$ Hence, $f_{neq}^{(1)}$ only contributes to higher moments such as the stress, ${\overline{\overline{\boldsymbol{\Pi}}}}$, and the heat flux, $\boldsymbol{\mathfrak{Q}}$. Using Equations (\[eq:20\])–(\[eq:22\]) it is readily found that ${\boldsymbol{\mathfrak{j}}}({\boldsymbol{\mathfrak{r}}};t)={\boldsymbol{\mathfrak{j}}}_{leq}({\boldsymbol{\mathfrak{r}}};t)=\rho({\boldsymbol{\mathfrak{r}}};t){\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t)$. Thus, recalling the definition given in Equation (\[eq:11\]), the kinetic-stress tensor satisfies the decomposition, $$\begin{aligned}
{\overline{\overline{\boldsymbol{\Pi}}}}({\boldsymbol{\mathfrak{r}}};t)=& ({\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t)\otimes{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t))\rho({\boldsymbol{\mathfrak{r}}};t)+
\int d {\boldsymbol{\mathfrak{p}}}\left({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}\otimes {\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}\right)f_{neq}^{(1)}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)\label{eq:23} ={\overline{\overline{\boldsymbol{\Pi}}}}_{leq}({\boldsymbol{\mathfrak{r}}};t) + {\overline{\overline{\boldsymbol{\Pi}}}}_{neq}({\boldsymbol{\mathfrak{r}}};t),\end{aligned}$$ where the first term stands for the local-equilibrium stress, which embodies the kinetic contribution[@condiff_transport_1969] due to translational, rotational and coupled velocities. On the other hand, the heat flux only shows a non-equilibrium contribution and, given the symmetry of $f_{leq}^{(1)}$, $$\begin{aligned}
\boldsymbol{\mathfrak{Q}}({\boldsymbol{\mathfrak{r}}};t)=\frac{1}{2}\int d{\boldsymbol{\mathfrak{p}}}\,|\mathbf{c}|^2\,{\boldsymbol{\mathfrak{m}}}^{-1}\mathbf{c}\,f^{(1)}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)
=\frac{1}{2}\int d{\boldsymbol{\mathfrak{p}}}\,|\mathbf{c}|^2\,{\boldsymbol{\mathfrak{m}}}^{-1}\mathbf{c}\,f_{neq}^{(1)}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)\label{eq:24}.
\end{aligned}$$ Now, making use of the identity $$\begin{aligned}
{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}}\cdot\left[({\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t)\otimes{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t))\rho({\boldsymbol{\mathfrak{r}}};t)\right] =
{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t){\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}}\cdot\left({\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t)\rho({\boldsymbol{\mathfrak{r}}};t)\right)+\rho({\boldsymbol{\mathfrak{r}}};t)\left({\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t)\cdot{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}}\right){\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t),\label{eq:25}\end{aligned}$$ in Equations (\[eq:23\]) and (\[eq:19\]), we arrive at the time-evolution equation $$\begin{aligned}
\rho({\boldsymbol{\mathfrak{r}}}_1;t)\mathcal{D}_t{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)\label{eq:26}+&{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot {\overline{\overline{\boldsymbol{\Pi}}}}_{neq}({\boldsymbol{\mathfrak{r}}}_1;t)+{\boldsymbol{\mathfrak{m}}}^{-1}\rho({\boldsymbol{\mathfrak{r}}}_1;t){\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}
+({\boldsymbol{\Upsilon}}_{\parallel}+{\boldsymbol{\Upsilon}}_{\bot}{\boldsymbol{\epsilon}})\rho({\boldsymbol{\mathfrak{r}}}_1;t){\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)\\
+&N{\boldsymbol{\mathfrak{m}}}^{-1}\sum_{j=1}^{N}\int d{\boldsymbol{\mathfrak{r}}}^{N-1}d{\boldsymbol{\mathfrak{p}}}^N\,\widetilde{{\boldsymbol{\Upsilon}}}_{1j}({\boldsymbol{\mathfrak{r}}}^N){\boldsymbol{\mathfrak{p}}}_j{f^{(N)}}({\boldsymbol{\mathfrak{r}}}^N,{\boldsymbol{\mathfrak{p}}}^N;t) = 0\notag\end{aligned}$$ where we define the material derivative as $\mathcal{D}_t=(\partial_t+{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}};t)\cdot{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}})$ and ${\boldsymbol{\epsilon}}=\text{adiag}(m^{-1}\mathbb{I},m\mathbb{I}^{-1})$. At this stage, there are two terms which still have to be treated, namely the non-equilibrium and the friction tensor terms. The former will be analysed in what follows and the latter in Section \[subsec:enskog-closure\].
To treat the non-equilibrium term, we follow Grad’s method[@grad_kinetic_1949; @liboff_kinetic_2003]. Specifically, the one-body distribution is rewritten as $f^{(1)}=f^{(1)}_{leq}\,\left[1+\varPsi
\right]$ and $\Psi({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)$ is expanded in a basis of orthogonal polynomials of the $\mathbf{p}\oplus\mathbf{L}$ space. In particular, we choose a basis of the multiple generalised Hermite polynomials[@dunkl_classical_2001] on $\mathbb{R}^d$, $$\begin{aligned}
P^{[\alpha]}_{n,\boldsymbol{\ell}}(x_1,\dots,x_d)=H_{\ell_1}^{[\alpha]}(x_1)\dots H_{\ell_d}^{[\alpha]}(x_d)\label{eq:27},
\quad \sum\ell_k=n\end{aligned}$$ with $n\in\mathbb{N}_0=\mathbb{Z}^+\cup\{0\}$, $\boldsymbol{\ell}\in\mathbb{N}^d_0$ and $H_{k}^{[\alpha]}(x)\equiv
H_{k}(x/\sqrt{\alpha})/\sqrt{k!}$ the generalised Hermite polynomials, which are an orthogonal basis for $L^2\left(\mathbb{R}\right)$ with respect to the weight $(2\pi\alpha)^{-1/2} \exp( -x^2/(2\alpha) )$. Denoting $\mathbf{e}_j=(\delta_{jk})$, the first $P^{[\alpha]}_{n,\boldsymbol{\ell}}$ are given by $$\begin{aligned}
P^{[\alpha]}_{0,\boldsymbol{0}}(\mathbf{x})=\,1,\label{eq:28}\quad P^{[\alpha]}_{1,\mathbf{e}_{j}}(\mathbf{x})=\,\frac{x_j}{\sqrt{\alpha}},\quad
P^{[\alpha]}_{2,\mathbf{e}_{j}+\mathbf{e}_{k}}(\mathbf{x})=\,\frac{x_jx_k}{\alpha}-\delta_{jk}.\end{aligned}$$ Setting now $\alpha= \sqrt{k_BT}$, the one-body distribution function can be expanded as $$\begin{aligned}
f^{(1)}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)=f^{(1)}_{leq}({\boldsymbol{\mathfrak{r}}},{\boldsymbol{\mathfrak{p}}};t)\left(1+\sum_{n,\boldsymbol{\ell}}B_{n,\boldsymbol{\ell}}({\boldsymbol{\mathfrak{r}}};t)\,
P^{[\alpha]}_{n,\boldsymbol{\ell}}\left({\boldsymbol{\mathfrak{m}}}^{-1/2}\mathbf{c}\right)\right).\label{eq:29}\end{aligned}$$ with $B_{n,\boldsymbol{\ell}}({\boldsymbol{\mathfrak{r}}};t)$ the moments of the distribution $f_{neq}^{(1)}$. From Equation (\[eq:22\]), we then find that $B_{0,\mathbf{0}}=B_{1,\mathbf{e}_j}=0$ and $\sum_{j}{B}_{2,2\mathbf{e}_j}=0$. Finally, the non-equilibrium term in Equation (\[eq:26\]) can be expressed as $$\begin{aligned}
\boldsymbol{\Pi}_{jk}^{neq}({\boldsymbol{\mathfrak{r}}};t) = k_B T \rho({\boldsymbol{\mathfrak{r}}};t) B_{2,\mathbf{e}_j+\mathbf{e}_k}({\boldsymbol{\mathfrak{r}}};t),
\label{eq:30}\end{aligned}$$ which represents an extension of the analogous relationship found for the spherical case (Section 3.2 of Ref. ). This relationship does not necessarily imply the non-local kinetic pressure to be small, since the functions $B_{2,\mathbf{e}_j+\mathbf{e}_k}({\boldsymbol{\mathfrak{r}}};t)$ do not depend on $\rho,\, T$ or ${\boldsymbol{\mathfrak{v}}}$. However, at this point we make the assumption that $f_{neq}^{(1)}$ can be either approximated or neglected, inspired by previous rigorous analyses for spherical particles[@goddard_overdamped_2012; @goddard_unification_2013]. Therefore, to obtain a DDFT, we need to deal with the last term in Equation (\[eq:26\]).
Two-body interactions and closure equation\[subsec:enskog-closure\]
-------------------------------------------------------------------
With the aim of closing Equation (\[eq:26\]) in $\rho$, we restrict our attention to systems where the HIs are given by linear combinations of two-body interactions. In particular, we assume the general structure[@goddard_unification_2013; @goddard_overdamped_2012] $$\begin{aligned}
\widetilde{{\boldsymbol{\Upsilon}}}_{jk}({\boldsymbol{\mathfrak{r}}}^N)=\delta_{jk}\sum_{r\neq j}\mathbf{Z}_1({\boldsymbol{\mathfrak{r}}}_j,{\boldsymbol{\mathfrak{r}}}_r)+(1-\delta_{jk})\mathbf{Z}_2({\boldsymbol{\mathfrak{r}}}_j,{\boldsymbol{\mathfrak{r}}}_k)\label{eq:31}.\end{aligned}$$ The integral of Equation (\[eq:26\]) depends only on $f^{(2)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2,{\boldsymbol{\mathfrak{p}}}_1,{\boldsymbol{\mathfrak{p}}}_2;t)$ and we now need a closure relation between the one- and two-body distributions. From a previous rigorous derivation of DDFT[@chan_time-dependent_2005], it is known that the one-body density determines the $N$-body distribution and, therefore, all the time-dependent properties of the system can be expressed as functionals of $\rho({\boldsymbol{\mathfrak{r}}};t)$. Thus, we make the approximation $$\begin{aligned}
f^{(2)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2,{\boldsymbol{\mathfrak{p}}}_1,{\boldsymbol{\mathfrak{p}}}_2;t) = f^{(1)}({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{p}}}_1;t)f^{(1)}({\boldsymbol{\mathfrak{r}}}_2,{\boldsymbol{\mathfrak{p}}}_2;t)g({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2;[\rho]),
\label{eq:32}\end{aligned}$$ although the exact form of the functional $g({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2;[\rho])$ is unknown, accurate approximations are known at equilibrium. However, the problem of obtaining an expression for $g$ is beyond the scope of the present study. From now on, we assume there exists at least a good approximation.
Substituting Equations (\[eq:31\]) and (\[eq:32\]) into (\[eq:30\]) yields $$\begin{aligned}
\rho({\boldsymbol{\mathfrak{r}}}_1;t)\mathcal{D}_t{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)+&{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot {\overline{\overline{\boldsymbol{\Pi}}}}_{neq}({\boldsymbol{\mathfrak{r}}}_1;t)+{\boldsymbol{\mathfrak{m}}}^{-1}\rho({\boldsymbol{\mathfrak{r}}}_1;t){\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}
+({\boldsymbol{\Upsilon}}_{\parallel}+{\boldsymbol{\Upsilon}}_{\bot}{\boldsymbol{\epsilon}})\rho({\boldsymbol{\mathfrak{r}}}_1;t){\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)\label{eq:33}\\
+&{\boldsymbol{\mathfrak{m}}}^{-1}\rho({\boldsymbol{\mathfrak{r}}}_1;t)\int d{\boldsymbol{\mathfrak{r}}}_2\left[
\mathbf{Z}_1({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2){\boldsymbol{\mathfrak{m}}}{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)+\mathbf{Z}_2({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2){\boldsymbol{\mathfrak{m}}}{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_2;t)
\right]\,\rho({\boldsymbol{\mathfrak{r}}}_2;t)\,g({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2;[\rho]) = 0. \notag\end{aligned}$$ Neglecting the non-local equilibrium term as discussed above, and dividing by the one-body density, we finally obtain our equation for the evolution of the flux, $$\begin{aligned}
\mathcal{D}_t{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)+&{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}\label{eq:34}
+({\boldsymbol{\Upsilon}}_{\parallel}+{\boldsymbol{\Upsilon}}_{\bot}{\boldsymbol{\epsilon}}){\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)\\
+&{\boldsymbol{\mathfrak{m}}}^{-1}\int d{\boldsymbol{\mathfrak{r}}}_2\left[
\mathbf{Z}_1({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2){\boldsymbol{\mathfrak{m}}}{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)+\mathbf{Z}_2({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2){\boldsymbol{\mathfrak{m}}}{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_2;t)
\right]\,\rho({\boldsymbol{\mathfrak{r}}}_2;t)\,g({\boldsymbol{\mathfrak{r}}}_1,{\boldsymbol{\mathfrak{r}}}_2;[\rho]) = 0\notag.\end{aligned}$$ Combining this equation with the conservation law resulting from Equations (\[eq:20\]) and (\[eq:22\]) in (\[eq:10\]), namely $$\begin{aligned}
\partial_t\rho({\boldsymbol{\mathfrak{r}}}_1;t)+{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot(\rho({\boldsymbol{\mathfrak{r}}};t){\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t))=0\label{eq:35},\end{aligned}$$ we finally have the generalised DDFT for systems of orientable particles.
We can readily check that for the special cases of point-like and smooth spherical particles (with perfect slip on the surface[@dickinson_brownian_1985]), these equations reduce to the unified DDFT recently derived by @goddard_unification_2013. Section \[subsec:wl-eq\] will be devoted to showing how Equation (\[eq:34\]) turns into the equation derived by @rex_dynamical_2007 and @wittkowski_dynamical_2011 for systems of arbitrary-shape colloids by setting $\mathbf{Z}_1=\mathbf{Z}_2=0$, and showing that the generalised Einstein relationships[@dickinson_brownian_1985; @peters_smoluchowski_2000] can be recovered in such a case. Finally, we note that the overall effect of the last two terms in Equation (\[eq:34\]) will be a friction-like retardation of the translational and angular velocities ${\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1,t)$, akin to what happened for spherical particles[@goddard_unification_2013].
Connection with previous DDFTs\[subsec:wl-eq\]
----------------------------------------------
Here we connect our DDFT with the DDFTs obtained by @rex_dynamical_2007 and @wittkowski_dynamical_2011 for systems of arbitrary-shape colloids with overdamped dynamics, neglecting HIs. Within the context of our study, this is equivalent to setting $\mathbf{Z}_1=\mathbf{Z}_2=0$. We then have $$\begin{aligned}
({\boldsymbol{\Upsilon}}_{\parallel}+{\boldsymbol{\Upsilon}}_{\bot}{\boldsymbol{\epsilon}}){\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t) = -\mathcal{D}_t{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)-&{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}\label{eq:36}.\end{aligned}$$ Using this result in the continuity equation (\[eq:35\]) gives $$\begin{aligned}
\partial_t\rho({\boldsymbol{\mathfrak{r}}}_1;t)={\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot\left(\rho({\boldsymbol{\mathfrak{r}}}_1;t)\boldsymbol{\zeta}^{-1}\left({\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}+{\boldsymbol{\mathfrak{m}}}\mathcal{D}_t{\boldsymbol{\mathfrak{v}}}({\boldsymbol{\mathfrak{r}}}_1;t)\right)\right)\label{eq:37}\end{aligned}$$ where we defined $$\begin{aligned}
({\boldsymbol{\Upsilon}}_{\parallel}+{\boldsymbol{\Upsilon}}_{\bot}{\boldsymbol{\epsilon}}){\boldsymbol{\mathfrak{m}}}=\boldsymbol{\zeta}\equiv\,\begin{pmatrix}
\boldsymbol{\zeta}_{\text{TT}} & \boldsymbol{\zeta}_{\text{TR}}\\
\boldsymbol{\zeta}_{\text{RT}} & \boldsymbol{\zeta}_{\text{RR}}
\end{pmatrix}\label{eq:38}
\end{aligned}$$ along with $\boldsymbol{\zeta}_{\text{TT}} =\, m\gamma_{TT}\mathbf{1},\ \boldsymbol{\zeta}_{\text{TR}} =\, m^{-1}\gamma_{TR}\mathbb{I}^2,\ \boldsymbol{\zeta}_{\text{RT}} =\, m^2\gamma_{RT}\mathbb{I}^{-1},$ and $\boldsymbol{\zeta}_{\text{RR}} =\, \gamma_{RR}\mathbb{I}$. In the overdamped regime, inertial forces and unsteady acceleration are neglected, i.e. ${\boldsymbol{\mathfrak{m}}}\,\mathcal{D}_t{\boldsymbol{\mathfrak{v}}}\rightarrow0$. Therefore, Equation (\[eq:38\]) becomes $$\begin{aligned}
\partial_t\rho({\boldsymbol{\mathfrak{r}}}_1;t)={\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot\left(\rho({\boldsymbol{\mathfrak{r}}}_1;t)\beta\mathbf{D}({\boldsymbol{\mathfrak{r}}}_1;t)\left({\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}\right)\right)\label{eq:39}\end{aligned}$$ where $\beta\mathbf{D}=\boldsymbol{\zeta}^{-1}$ is the diffusion tensor. Making use of the block-matrix inversion formula \[e.g. Equation (2.8.18) of Ref. \] we can check that the diffusion tensor fulfils the generalised Einstein relationships[@dickinson_brownian_1985; @peters_smoluchowski_2000] $$\begin{aligned}
\beta\mathbf{D}_{\text{TT}} =& \left(\boldsymbol{\zeta}_{\text{TT}}-\boldsymbol{\zeta}_{\text{TR}}\boldsymbol{\zeta}_{\text{RR}}^{-1}\boldsymbol{\zeta}_{\text{RT}}\right)^{-1}\label{eq:40}\\
\beta\mathbf{D}_{\text{TR}} =& -\left(\boldsymbol{\zeta}_{\text{TT}}-\boldsymbol{\zeta}_{\text{TR}}\boldsymbol{\zeta}_{\text{RR}}^{-1}\boldsymbol{\zeta}_{\text{RT}}\right)^{-1}\boldsymbol{\zeta}_{\text{TR}}\boldsymbol{\zeta}_{\text{RR}}^{-1}\equiv
\left(\boldsymbol{\zeta}_{\text{RT}}-\boldsymbol{\zeta}_{\text{TT}}\boldsymbol{\zeta}_{\text{TR}}^{-1}\boldsymbol{\zeta}_{\text{RR}}\right)^{-1}\notag\\
\beta\mathbf{D}_{\text{RT}} =& -\left(\boldsymbol{\zeta}_{\text{TT}}-\boldsymbol{\zeta}_{\text{TR}}\boldsymbol{\zeta}_{\text{RR}}^{-1}\boldsymbol{\zeta}_{\text{RT}}\right)^{-1}\boldsymbol{\zeta}_{\text{RT}}\boldsymbol{\zeta}_{\text{TT}}^{-1}
\equiv\left(\boldsymbol{\zeta}_{\text{TR}}-\boldsymbol{\zeta}_{\text{RR}}\boldsymbol{\zeta}_{\text{RT}}^{-1}
\boldsymbol{\zeta}_{\text{TT}}\right)^{-1}\notag\\
\beta\mathbf{D}_{\text{RR}} =& \left(\boldsymbol{\zeta}_{\text{RR}}-\boldsymbol{\zeta}_{\text{RT}}\boldsymbol{\zeta}_{\text{TT}}^{-1}\boldsymbol{\zeta}_{\text{TR}}\right)^{-1}\notag
\end{aligned}$$ In the special case when the translational and rotational motions are fully decoupled, $\mathbf{D}_{\text{TR}}=\mathbf{D}_{\text{RT}}\equiv 0$, we recover the equation derived by @rex_dynamical_2007. This is indeed the case for spherically isotropic colloids \[e.g. Equation (37) of Ref. \] far from a wall. Furthermore, if the colloidal particles are smooth spheres we recover the DDFT derived by @marconi_dynamic_2000, as the stress tensor is then uniquely determined by the translational component. However, the strength of the hydrodynamic translation-rotation coupling increases rapidly as the distance of the particles from a wall decreases[@dickinson_brownian_1985; @beenakker_many-sphere_1984]. The coupling is then unavoidable near the walls even for the simple spherical case. In this regard, Equation (\[eq:39\]) can be seen as a generalisation of the uncoupled DDFT obtained by @rex_dynamical_2007.
Of course, our overdamped DDFT (\[eq:39\]) does not include the self-propulsion forces present in the DDFT of @wittkowski_dynamical_2011 as these forces were not present in our original equations to begin with (see Appendix \[app:projection-operator-tech\]). Nevertheless, they can readily be included as an additional term, ${\boldsymbol{\mathfrak{F}}}_{i}^{A}({\boldsymbol{\mathfrak{r}}}_i;t)$, inside the solvent-averaged interaction forces and torques ${\boldsymbol{\mathfrak{F}}}$. This results in Equation (\[eq:11\]) having the extra term, $-{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{F}}}_{1}^{A}({\boldsymbol{\mathfrak{r}}}_1;t)\rho({\boldsymbol{\mathfrak{r}}}_1;t)$. Following exactly the same steps as in (\[eq:12\])–(\[eq:16\]), the same first equation of the YBG hierarchy (\[eq:17\]) can be obtained, by setting ${\boldsymbol{\mathfrak{F}}}_1^A=0$ at equilibrium[@wittkowski_dynamical_2011]. Therefore, the additional term $-{\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{F}}}_{1}^{A}({\boldsymbol{\mathfrak{r}}}_1;t)\rho({\boldsymbol{\mathfrak{r}}}_1;t)$ would survive throughout the whole derivation so that the left-hand side of our DDFT equation (\[eq:34\]) would contain it. Eventually, this leads us to $$\begin{aligned}
\partial_t\rho({\boldsymbol{\mathfrak{r}}}_1;t)={\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\cdot\left(\rho({\boldsymbol{\mathfrak{r}}}_1;t)\beta\mathbf{D}({\boldsymbol{\mathfrak{r}}}_1;t)\left({\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_1}\frac{\delta \mathcal{F}[\rho]}{\delta \rho({\boldsymbol{\mathfrak{r}}}_1)}-{\boldsymbol{\mathfrak{F}}}_{1}^{A}({\boldsymbol{\mathfrak{r}}}_1;t)\right)\right)\label{eq:41}\end{aligned}$$ which is now in complete agreement with the equation previously derived within the overdamped regime[@wittkowski_dynamical_2011], with the diffusion tensor still satisfying the generalised Einstein relationships (\[eq:40\]).
Conclusions {#sec:conclusions}
===========
In this work we have formulated a DDFT for orientable particles with inertia and HIs. There are numerous examples where orientation of particles plays a key role. From the fundamental study of perfectly-rough spheres[@condiff_brownian_1966], loaded spherocylinders[@condiff_brownian_1966; @dahler_kinetic_1963], nematic solutions[@hartel_towing_2010; @zhang_isotropic-nematic_2006] or liquid-crystal nucleation[@schilling_self-poisoning_2004; @zhang_isotropic-nematic_2006; @miller_hierarchical_2009], to the study of biological processes such as[@dickinson_brownian_1985] protein adsorption and trapping, antibody-antigen interaction, biochemical assembly by monomer aggregation or polymerization, bone formation[@cantaert_nanoscale_2013; @gomez-morales_progress_2013] or *in vivo* protein crystallization[@bechtel_electron_1976; @koopmann_vivo_2012], to name but a few. We believe that all these problems could be tackled with the aid of the DDFT developed here.
The present study also addresses the problem posed by @wittkowski_dynamical_2011: generalising their work to include HIs. But our study goes a step further and also includes the inertia of the colloidal particles. We are currently in the process of implementing our DDFT equations in the pseudospectral scheme, based on Chebychev collocation[@boyd_chebyshev_2001], we have developed for the numerical treatment of DDFT for spherical particles and which has been successfully applied to a wide spectrum of physical settings; from fluids in a confining potential[@goddard_unification_2013] and mixtures in such a potential[@goddard_multi-species_2013] to adsorbed films on a substrate[@yatsyshin_spectral_2012], fluids in confined geometry[@yatsyshin_geometry-induced_2013; @yatsyshin_wetting_2015; @yatsyshin_density_2015] and contact lines[@nold_fluid_2014; @nold_nanoscale_2015]. While the theory presented here is valid for three-dimensional orientable particles, from a computational point of view systems of two-dimensional orientable particles would be more tractable. Finally, interesting extensions of the framework developed here would be to include torsional degrees of freedom to describe systems of flexible-chain molecules[@evans_momentum_1980], consider multiple-particle species, thus extending the previously developed DDFT for mixtures of spherical particles[@goddard_multi-species_2013], and confined geometry[@yatsyshin_geometry-induced_2013; @yatsyshin_wetting_2015; @yatsyshin_density_2015]. We shall examine these and related questions in future studies.
We are grateful to the anonymous referees for useful comments and suggestions and to Andreas Nold for stimulating discussions that led to the scaling arguments in Appendix B. We acknowledge financial support from the European Research Council via Advanced Grant No. 247031 and from EPSRC via grants No. EP/L020564 and EP/L025159.
Derivation of generalised Langevin equations {#app:projection-operator-tech}
============================================
The Fokker-Planck Equation
--------------------------
Here we outline the derivation of the time-evolution equation for the probability distribution function of a system of arbitrary-shape particles. This derivation can be thought as an application of the works of @kirkwood_statistical_1946, @murphy_brownian_1972, @deutch_molecular_1971 and @wilemski_derivation_1976 in conjunction with the results of @dahler_kinetic_1963, @condiff_brownian_1966, @evans_cumulant_1976 and @condiff_transport_1969, to consider arbitrary-shape particles. .
As in Section \[sec:evolution-equations\], consider $N$ identical, asymmetric colloidal particles with mass $m$ immersed in a fluid of $n\gg N$ identical bath particles with mass $m_b$. Throughout this section upper case letters will refer to colloidal particles while lower case ones indicate bath particles. The phase-space coordinates of the $i$-th bath particle are $\mathbf{x}_j\doteq(\mathbf{r}_j,\,\mathbf{p}_j)$, with “$\doteq$” denoting “by definition", $\mathbf{r}_j$ being the position vector and $\mathbf{p}_j=m_b\dot{\mathbf{r}}_j$ the canonical momentum. In addition, the dynamical state of the $j$-th colloidal particle is determined by $\mathbf{X}_j\doteq(\mathbf{R}_j,\mathbf{P}_j)$, where $\mathbf{R}_j$ is its centre-mass position vector and $\mathbf{P}_j=m\dot{\mathbf{R}}_j$ is its conjugate momentum, along with the pair $\boldsymbol{\Omega}_j\doteq(\boldsymbol{\alpha}_j,\boldsymbol{\pi}_j)$ which comprises the rotational degrees of freedom, with $\boldsymbol{\alpha}_j$ being the Eulerian angles as in Section \[sec:evolution-equations\] and $\boldsymbol{\pi}_j$ their conjugate momenta.[@dahler_kinetic_1963] The angular velocity of a particle, $\boldsymbol{\omega}_j$, is determined by the time derivative of the Euler angles. In the principal-axes frame $\mathfrak{B}$, we have the relation $\boldsymbol{\omega}_j=\boldsymbol{\Lambda}_j^\top\dot{\boldsymbol{\alpha}}_j$, where[@jose_classical_2013; @condiff_brownian_1966] $$\boldsymbol{\Lambda}_j^\top = \begin{pmatrix}
\cos\chi_j & \sin\theta_j\sin\chi_j & 0\\
-\sin\chi_j & \sin\theta_j\cos\chi_j & 0\\
0 & \cos\theta_j & 1
\end{pmatrix}\ \Leftrightarrow
\boldsymbol{\Lambda}_j^{-1} =
\begin{pmatrix}
\cos\chi_j & \csc\theta_j\cos\chi_j & -\cot\theta_j\sin\chi_j\\
-\sin\chi_j & \csc\theta_j\sin\chi_j & -\cot\theta_j\cos\chi_j\\
0 & 0 & 1
\end{pmatrix}.
\label{eq:a-1}$$ Accordingly, $\boldsymbol{\omega}_j'=\mathcal{R}_j^\top\boldsymbol{\omega}_j=
\boldsymbol{\Xi}_j\,\dot{\boldsymbol{\alpha}}_j$ under the space-fixed frame $\mathfrak{S}$, with $\boldsymbol{\Xi}_j\doteq\mathcal{R}_j^\top\boldsymbol{\Lambda}_j^\top$. These can be related with their corresponding angular momenta by, $\mathbf{L}_j=\mathbb{I}\,\boldsymbol{\omega}_j$ and $\mathbf{L}'_j=\mathbb{I}'_j\boldsymbol{\omega}'_j$ respectively.
Thus, the dynamical state of the system at any given instant represents a single point in a $6(N+n)-$dimensional space, $\Gamma$,[@liboff_kinetic_2003] $$\mathfrak{s}(t)\doteq(\mathbf{x}_1(t),\dots,\mathbf{x}_n(t),\mathbf{X}_1(t),\dots,\mathbf{X}_N(t),\boldsymbol{\Omega}_1(t),\dots,\boldsymbol{\Omega}_N(t))\equiv
(\mathbf{x}^n(t),\mathbf{X}^N(t),\boldsymbol{\Omega}^N(t))\in\Gamma
\label{eq:a-2}$$ where we made use of the notation $\mathbf{x}^n= \mathbf{x}_1\dots
\mathbf{x}_n$, $\mathbf{X}^N=\mathbf{X}_1\dots\mathbf{X}_N$ and $\boldsymbol{\Omega}^N=\boldsymbol{\Omega}_1,\dots,\boldsymbol{\Omega}_N$. From classical mechanics, the evolution of the system is completely determined by the initial conditions for positions and momenta of all particles. This time evolution, which prescribes a unique trajectory $\mathfrak{s}(t;t_0)$, is fully described by the Lagrangian and the Hamiltonian of the system. In the following we use the former to get the relation between $\boldsymbol{\pi}_j$ and $\dot{\boldsymbol{\alpha}}_j$ in order to construct the Hamiltonian. Then Hamilton’s equations along with Liouville’s theorem will be employed to get the time-evolution equation for the PDF of the system so that its phases lie in a differential region of $\Gamma$ with centre placed at $\mathfrak{s}(t)$. First, the Lagrangian of the system can be written as,[^1] $$\begin{aligned}
\mathcal{L}=&\mathcal{L}_b+\mathcal{L}_B,\label{eq:a-3}\\
\mathcal{L}_b=&\sum_{i=1}^n\frac{1}{2}m_b\dot{\mathbf{r}}_j\cdot\dot{\mathbf{r}}_j-\mathcal{U}(\mathbf{r}^N,\mathbf{R}^n,\boldsymbol{\alpha}^n),\notag\\
\mathcal{L}_B=&\sum_{i=1}^N\frac{1}{2}
\left(m\dot{\mathbf{R}}_j\cdot\dot{\mathbf{R}}_j+\dot{\boldsymbol{\alpha}}_j\cdot(\boldsymbol{\Xi}_j^\top\mathbb{I}'_j\boldsymbol{\Xi}_j) \dot{\boldsymbol{\alpha}}_j\right)-V(\mathbf{R}^N,\boldsymbol{\alpha}^N)\notag,\end{aligned}$$ with $V(\mathbf{R}^N,\boldsymbol{\alpha}^N)$ the potential energy due to short-range interactions exclusively between colloidal particles, and $$\mathcal{U}(\mathbf{r}^n,\mathbf{R}^N,\boldsymbol{\alpha}^N) = U(\mathbf{r}^n)+\sum_{\mu=1}^N\mathfrak{u}_\mu(\mathbf{r}^N,\mathbf{R}_\mu,\boldsymbol{\alpha}_\mu),\label{eq:a-4}$$ the short-range intermolecular potential energy coming from the interaction between bath particles, $U(\mathbf{r}^n)$, and the interaction of each colloidal particle with the whole bath, $\mathfrak{u}_\mu(\mathbf{r}^N,\mathbf{R}_\mu), \forall \mu=1,\dots,N$. From these equations we obtain $$\begin{aligned}
\mathbf{p}_j=\frac{\partial \mathcal{L}}{\partial \dot{\mathbf{r}}_j}=m_b\dot{\mathbf{r}_j},\quad
\mathbf{P}_j=\frac{\partial \mathcal{L}}{\partial \dot{\mathbf{R}}_j}=m\dot{\mathbf{R}_j},\quad
\boldsymbol{\pi}_j=\frac{\partial \mathcal{L}}{\partial \dot{\boldsymbol{\alpha}}_j}= (\boldsymbol{\Xi}_j^\top\mathbb{I}'_j\boldsymbol{\Xi}_j)\dot{\boldsymbol{\alpha}_j},
\label{eq:a-5}\end{aligned}$$ Therefore, $\boldsymbol{\pi}_j$ can be easily related with the angular momentum $\mathbf{L}_j$ via[@condiff_brownian_1966; @condiff_transport_1969] $$\boldsymbol{\pi}_j = (\boldsymbol{\Xi}_j^\top\mathbb{I}'_j\boldsymbol{\Xi}_j)\dot{\boldsymbol{\alpha}}_j
\equiv
\boldsymbol{\Lambda}_j{\mathbf{L}}_j,
\label{eq:a-6}$$ which will be used to perform the appropriate change of variables later on. We now write the Hamiltonian function of the system as $$\begin{aligned}
\mathcal{H} =& \mathcal{H}_b + \mathcal{H}_B \label{eq:a-7}\\
\mathcal{H}_b =& \sum_{i=1}^n\frac{\mathbf{p}_j\cdot\mathbf{p}_j}{2m_b}+\mathcal{U}(\mathbf{r}^N,\mathbf{R}^n,\boldsymbol{\alpha}^n)\notag\\
\mathcal{H}_B =&\sum_{i=1}^N\left(\frac{\mathbf{P}_j\cdot\mathbf{P}_j}{2m}+\frac{1}{2}\boldsymbol{\pi}_j\cdot(\boldsymbol{\Xi}_j^\top\mathbb{I}'_j\boldsymbol{\Xi}_j)^{-1}\boldsymbol{\pi}_j
\right)+V(\mathbf{R}^N,\boldsymbol{\alpha}^n).\notag\end{aligned}$$ According to Liouville’s theorem, the $n+N$ particle distribution function, $\varrho^{(n+N)}(\mathbf{x}^n,\mathbf{X}^N;t)$, will evolve according to $$\partial_t \varrho^{(n+N)}(\mathbf{x}^n,\mathbf{X}^N,\boldsymbol{\Omega}^N;t)+\left(i\mathfrak{L}_b+i\mathfrak{L}_B^T+i\mathfrak{L}_B^R\right)\,\varrho^{(n+N)}(\mathbf{x}^n,\mathbf{X}^N;t)=0
\label{eq:a-8}$$ with the Liouvillian, $\mathfrak{L}\doteq \mathfrak{L}_b+\mathfrak{L}_B^T+\mathfrak{L}_B^R$, $$\begin{aligned}
i\mathfrak{L}_b =& \sum_{j=1}^n\left(\frac{\mathbf{p}_j}{m_b}\cdot\frac{\partial }{\partial \mathbf{r}_j}+
\mathbf{f}_j\cdot\frac{\partial }{\partial \mathbf{p}_j}\right)\label{eq:a-9}\\
i\mathfrak{L}_B^T =& \sum_{j=1}^N\left(\frac{\mathbf{P}_j}{m}\cdot\frac{\partial }{\partial \mathbf{R}_j}+
\mathbf{F}_j\cdot\frac{\partial }{\partial \mathbf{P}_j}\right)\notag\\
i\mathfrak{L}_B^R =& \sum_{j=1}^N\left(\dot{\boldsymbol{\alpha}}_j\cdot\frac{\partial }{\partial \boldsymbol{\alpha}_j}+
\dot{\boldsymbol{\pi}}_j\cdot\frac{\partial }{\partial \boldsymbol{\pi}_j}\right)\notag\end{aligned}$$ where $\mathbf{f}_j=-\frac{\partial}{\partial\mathbf{r}_j}(U+\sum_{\mu=1}^N\mathfrak{u}_\mu)$ and $\mathbf{F}_j=-\frac{\partial}{\partial \mathbf{R}_j}(V+\mathfrak{u}_j)$ are the instantaneous forces acting on bath and colloidal particles, respectively. From a conceptual point of view,[@dahler_kinetic_1963] the quantities ${\boldsymbol{\alpha}}_j$ and ${\boldsymbol{\pi}}_j$ are less convenient than the angular velocities and momenta, $\boldsymbol{\omega}_j$ and $\mathbf{L}_j$. This can be easily checked by considering the rotational kinetic energy $$\begin{aligned}
K_{R}= \sum_{j=1}^N \frac{\dot{\boldsymbol{\alpha}}_j}{2}\cdot(\boldsymbol{\Xi}_j^\top\mathbb{I}'_j\boldsymbol{\Xi}_j) \dot{\boldsymbol{\alpha}}_j \equiv &\ \sum_{j=1}^N \frac{\boldsymbol{\omega}_j}{2}\cdot \mathbb{I}\boldsymbol{\omega}_j\label{a-10}\\
K_{R}= \sum_{j=1}^N \frac{\boldsymbol{\pi}_j}{2}\cdot(\boldsymbol{\Xi}_j^\top\mathbb{I}'_j\boldsymbol{\Xi}_j)^{-1}\boldsymbol{\pi}_j
\equiv&\ \sum_{i=1}^N \frac{\mathbf{L}_j}{2}\cdot(\mathbb{I}^{-1}\mathbf{L}_j).\notag\end{aligned}$$ Transforming now Equation (\[eq:a-8\]) into an equivalent equation for $$\begin{aligned}
F^{(n+N)}(\mathbf{x}^n,\mathbf{X}^N,\boldsymbol{\alpha}^N,\mathbf{L}^N;t)=&\left|
\frac{\partial(\mathbf{x}^n,\mathbf{X}^N,\boldsymbol{\alpha}^N,\boldsymbol{\pi}^N)}{\partial (\mathbf{x}^n,\mathbf{X}^N,\boldsymbol{\alpha}^N,\mathbf{L}^N)}
\right|\varrho^{(n+N)}(\mathbf{x}^n,\mathbf{X}^N,\boldsymbol{\Omega}^N;t)\label{eq:a-11}\\
=&\prod_{j}\sin\theta_j\,\varrho^{(n+N)}(\mathbf{x}^n,\mathbf{X}^N,\boldsymbol{\Omega}^N;t)\notag\end{aligned}$$ gives $$\begin{aligned}
\partial_t F^{(n+N)}(\mathbf{x}^n,\mathbf{X}^N,\boldsymbol{\Omega}^N;t)+\left(i\mathfrak{L}_b+i\mathfrak{L}_B^T+i\widetilde{\mathfrak{L}}_B^R\right)\,F^{(n+N)}(\mathbf{x}^n,\mathbf{X}^N,\boldsymbol{\Omega}^N;t)=0
\label{eq:a-12}\end{aligned}$$ with[@condiff_brownian_1966; @condiff_transport_1969; @dahler_kinetic_1963] $$i\widetilde{\mathfrak{L}}_B^R = \sum_{j=1}^N\dot{\boldsymbol{\alpha}}_j \cdot \hat{\partial}_{\boldsymbol{\alpha}_j}+\mathbf{T}_j\cdot\frac{\partial }{\partial \mathbf{L}_j}
\label{eq:a-13}$$ where the operator[@condiff_brownian_1966] $\hat{\partial}_{\boldsymbol{\alpha_j}}\doteq
(\csc\theta_j\frac{\partial}{\partial \theta_j}\sin\theta_j,
\frac{\partial}{\partial \phi_j}, \frac{\partial}{\partial\chi_j})^\top$, deduced by using the chain rule[@risken_fokker-planck_1996], has been introduced. In the latter equation, $\mathbf{T}_j = \mathbf{N}_j -
\boldsymbol{\omega}_j\times\mathbf{L}_j$ denotes the net torque acting on the $j$-th colloidal particle, with[@condiff_brownian_1966] $$\begin{aligned}
\mathbf{N}_j &= - \boldsymbol{\Lambda}_j^{-1}\frac{\partial}{\partial \boldsymbol{\alpha}_j}\left[V(\mathbf{R}^N,\boldsymbol{\alpha}^n)+\mathfrak{u}_j(\mathbf{r}^N,\mathbf{R}_j,\boldsymbol{\alpha}_j)\right]
\label{eq:a-14}\end{aligned}$$ the torque due to intermolecular interactions along the principal axes of inertia. At this point, it is necessary to introduce the angular-gradient operator (also known as orientational gradient[@peters_smoluchowski_2000]) $$\begin{aligned}
\frac{\partial }{\partial \boldsymbol{\Phi}_j}&\doteq
\boldsymbol{\Lambda}_j^{-1}\frac{\partial }{\partial \boldsymbol{\alpha}_j}\equiv \hat{\partial}_{\boldsymbol{\alpha}_j}\cdot(\boldsymbol{\Lambda}^\top)^{-1}\label{eq:a-15}\end{aligned}$$ such that $\frac{\partial }{\partial {\boldsymbol{\Phi}}}= \mathbf{e}_x\frac{\partial
}{\partial \Phi_{x}}+\mathbf{e}_y\frac{\partial }{\partial
\Phi_{y}}+\mathbf{e}_z\frac{\partial }{\partial \Phi_{z}}$, where $\mathbf{e}_i$ is the unitary vector along axis $i\in\{x,y,z\}$ of the Cartesian frame $\mathfrak{B}$, and $$\begin{aligned}
\frac{\partial }{\partial \Phi_{x}^j}&= \cos\chi_j\frac{\partial}{\partial \theta_j}
+\csc\theta_j\cos\chi_j\frac{\partial }{\partial \phi_j}
-\cot\theta_j\sin\chi_j\frac{\partial }{\partial \chi_j}\label{eq:a-16}\\
\frac{\partial }{\partial \Phi_y^j}&=
-\sin\chi_j\frac{\partial}{\partial \theta_j}
+\csc\theta_j\sin\chi_j\frac{\partial }{\partial \phi_j}
-\cot\theta_j\cos\chi_j\frac{\partial }{\partial \chi_j}
\notag\\
\frac{\partial }{\partial \Phi_z^j}&=\frac{\partial }{\partial\chi_j}.\notag\end{aligned}$$ It is worth mentioning here that the derivative operators $(\frac{\partial
}{\partial \Phi_{x}},\frac{\partial }{\partial \Phi_{y}},\frac{\partial
}{\partial \Phi_{z}})$ are the generators of rotations of a rigid body about the body-fixed Cartesian frame[@gray_theory_1984]. This results in $$\begin{aligned}
i\widetilde{\mathfrak{L}}_B^R=\sum_{j=1}^N \frac{\partial }{\partial \boldsymbol{\Phi}_j}\cdot\boldsymbol{\omega}_j + \left(\mathbf{N}_j+\mathbf{L}_j\times\boldsymbol{\omega}_j\right)\cdot\frac{\partial }{\partial\mathbf{L}_j}
\label{eq:a-17}.\end{aligned}$$ Now we can define ${\boldsymbol{\mathfrak{m}}}\doteq \text{diag}(m\mathbf{1},\mathbb{I})$ along with the vectors $\boldsymbol{\mathfrak{r}}_j\doteq
(\mathbf{R}_j,\boldsymbol{\Phi}_j)$, $\boldsymbol{\mathfrak{p}}_j\doteq(\mathbf{P}_j,\mathbf{L}_j)$ and $\boldsymbol{\mathfrak{f}}_j\doteq(\mathbf{F}_j,\mathbf{T}_j)$, and the operators $\boldsymbol{\nabla}_{\mathfrak{r}_j}\doteq(\frac{\partial}{\partial
\mathbf{R}_j},\frac{\partial }{\partial \boldsymbol{\Phi}_j})^\top$ and $\boldsymbol{\nabla}_{\boldsymbol{\mathfrak{p}}_j}\doteq (\frac{\partial
}{\partial \mathbf{P}_j},\frac{\partial }{\partial \mathbf{L}_j})^\top$, enabling us to rewrite Equation (\[eq:a-12\]) in a more compact and convenient way, $$\begin{aligned}
\partial_t F^{(n+N)}(t)+
\sum_{j=1}^n\left(\frac{\mathbf{p}_j}{m_b}\cdot\frac{\partial }{\partial \mathbf{r}_j}+
\mathbf{f}_j\cdot\frac{\partial }{\partial \mathbf{p}_j}\right){F^{(n+N)}}(t)
+ \sum_{j=1}^N\left({\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_j}\cdot {\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_j + {\boldsymbol{\mathfrak{f}}}_j\cdot{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j}
\right){F^{(n+N)}}(t)=0\label{eq:a-18},\end{aligned}$$ where explicit dependence on the phase-space coordinates was omitted, but recalled through the superscript $(n+N)$. Following @murphy_brownian_1972, we introduce the scaling quantity, ${\boldsymbol{\lambda}}\doteq {\boldsymbol{\mathfrak{m}}}^{-1/2}$, so that $\widetilde{{\boldsymbol{\mathfrak{p}}}}_j={\boldsymbol{\lambda}}{\boldsymbol{\mathfrak{p}}}_j$ and hence, the last term of the previous equation becomes $$\begin{aligned}
\sum_{j=1}^N{\boldsymbol{\lambda}}\left({\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_j}\cdot {\widetilde{\boldsymbol{\mathfrak{p}}}}_j + {\boldsymbol{\mathfrak{f}}}_j\cdot{\boldsymbol{\nabla}}_{{\widetilde{\boldsymbol{\mathfrak{p}}}}_j}
\right){F^{(n+N)}}(t). \label{eq:a-19}\end{aligned}$$ Substitution of Equation (\[eq:a-19\]) into Equation (\[eq:a-18\]) the results in an equation resembling Liouville’s equation for spherical colloidal particles. Such a result provides a description of the time evolution of the full system. However, our interest rests exclusively on colloidal particles. Thus, our aim is to get the time-evolution equation for the $N$-particle distribution, $$\begin{aligned}
{f^{(N)}}(t) \doteq \int d{\mathbf{x}}^n {F^{(n+N)}}(t).\label{eq:a-20}\end{aligned}$$ For this purpose, Zwanzig’s projection technique can be applied as in the case of spherical colloidal particles[@deutch_molecular_1971; @murphy_brownian_1972; @wilemski_derivation_1976; @ermak_brownian_1978]. Following the work of @murphy_brownian_1972, for the arbitrary initial state at $t=-t_I$ we choose one where the bath particles are in equilibrium with the instantaneous positions of the colloidal particles. This means that $${F^{(n+N)}}({\mathbf{x}}^n,{\mathbf{X}}^N,{\boldsymbol{\Omega}}^N;-t_I)=\rho_n^\dag({\mathbf{x}}^n)\,{f^{(N)}}({\mathbf{X}}^N,{\boldsymbol{\Omega}}^N;-t_I)\label{eq:a-21},$$ with $\rho_n^\dag$ the canonical distribution of the $n$ bath particles in the instantaneous potential created by the colloidal particles. The last step before integrating Equation (\[eq:a-18\]) to remove the dependence on fast variables involves the definition of the projection operator, $$\hat{\mathcal{P}}\doteq \rho_n^\dag({\mathbf{x}}^n)\int d{\mathbf{x}}^n\,,$$ and its complementary operator $\hat{\mathcal{Q}}=1-\hat{\mathcal{P}}$. Thus, the integration of Equation (\[eq:a-18\]) over the fast variables ${\mathbf{x}}^n$ is equivalent to applying $(\rho_n^\dag)^{-1}\hat{\mathcal{P}}$ on both sides of such an equation. Although considerable algebraic manipulations are required, we can follow the work of @murphy_brownian_1972 and @lebowitz_microscopic_1965 step by step to finally reach the desired time-evolution equation for the projected distribution function ${f^{(N)}}$, $$\begin{aligned}
\partial_t{f^{(N)}}(t)+\sum_{j=1}^N&{\boldsymbol{\lambda}}\left({\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_j}\cdot {\widetilde{\boldsymbol{\mathfrak{p}}}}_j+{\boldsymbol{\mathfrak{F}}}_j\cdot{\boldsymbol{\nabla}}_{{\widetilde{\boldsymbol{\mathfrak{p}}}}_j}\right){f^{(N)}}(t) \label{eq:a-23}\\
=& \sum_{j,k=1}^N {\boldsymbol{\lambda}}^2\,{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j}\cdot\int_{-t_I}^{t}dt'{\boldsymbol{\gamma}}_{jk}({\boldsymbol{\lambda}};t,t')\left({\boldsymbol{\nabla}}_{{\widetilde{\boldsymbol{\mathfrak{p}}}}_k}+\beta\,{\widetilde{\boldsymbol{\mathfrak{p}}}}_k\right){f^{(N)}}(t')\notag\end{aligned}$$ where $\beta=1/k_BT$, where $k_B$ is the Boltzmann constant, $T$ is the temperature imposed by the bath, and ${\boldsymbol{\mathfrak{F}}}$ is the equilibrium average force and torque, i.e. $$\begin{aligned}
{\boldsymbol{\mathfrak{F}}}_j &\equiv
\left\langle
\begin{array}{l}
{\mathbf{F}}_j\\
{\mathbf{T}}_j
\end{array}
\right\rangle^\dag
=\int d{\mathbf{x}}^n \rho_n^\dag({\mathbf{x}}^n)\,{\boldsymbol{\mathfrak{f}}}_j({\mathbf{r}}^n,{\mathbf{R}}^N,{\boldsymbol{\alpha}}^N)
\label{eq:a-24}\\
&=-\left[
\begin{array}{l}
-\frac{\partial }{\partial {\mathbf{R}}_j} V({\mathbf{R}}^N,{\boldsymbol{\alpha}}^N)\\
-\frac{\partial }{\partial {\boldsymbol{\Phi}}_j}V({\mathbf{R}}^N,{\boldsymbol{\alpha}}^N)- {\boldsymbol{\omega}}_j\times{\mathbf{L}}_j
\end{array}
\right]+\left\langle{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_j}\mathfrak{u}({\mathbf{r}}^n,{\mathbf{R}}_j,{\boldsymbol{\alpha}}_j)\right\rangle^\dag =
\left[
\begin{array}{l}
-\frac{\partial }{\partial {\mathbf{R}}_j} (V+\psi)\\
-\frac{\partial }{\partial {\boldsymbol{\Phi}}_j}(V+\psi)- {\boldsymbol{\omega}}_j\times{\mathbf{L}}_j
\end{array} \right],
\notag\end{aligned}$$ with $\langle.\rangle^\dag$ the equilibrium average over the fast variables. Equation (\[eq:a-24\]) also includes the definition of the potential of mean force $\psi\doteq
\langle\mathfrak{u}\rangle^\dag$, i.e. the potential which gives rise to the average (over all configurations of the $n$ bath molecules) force and torque acting on the $j$th colloidal particle at any given configuration keeping all the colloidal particles frozen. Thus, the fluid-equilibrium average force and torque includes the contribution of a postulated vacuum colloid-colloid interaction potential, $V$, and the solvent contribution to the total force and torque, $\psi$. This combination in turn results in a solvent-averaged potential of mean force, $\widetilde{V}\doteq V+\psi$, which can be obtained from a given physical model, e.g. the DLVO theory for the interaction of charged colloidal particles[@snook_langevin_2006] or the ten Wolde-Frenkel potential[@wolde_enhancement_1997]. Finally, the tensor ${\boldsymbol{\gamma}}_{jk}$ is given by $$\begin{aligned}
{\boldsymbol{\gamma}}_{jk}({\boldsymbol{\lambda}};t,t')&=\left\langle {\boldsymbol{\mathfrak{f}}}_j(t)\otimes e^{i(t'-t)\hat{\mathcal{Q}}\mathfrak{L}}({\boldsymbol{\mathfrak{f}}}_k(t)-{\boldsymbol{\mathfrak{F}}}_k)\right\rangle^\dag
\label{eq:a-25}\\
&\equiv\left\langle
\begin{array}{ccc}
{\mathbf{F}}_j(t)\otimes e^{i(t'-t)\hat{\mathcal{Q}}\mathfrak{L}}({\mathbf{F}}_k(t)-\boldsymbol{\mathcal{F}}_k)
&& {\mathbf{F}}_j(t)\otimes e^{i(t'-t)\hat{\mathcal{Q}}\mathfrak{L}}({\mathbf{T}}_k(t)-\boldsymbol{\mathcal{T}}_k)\\
{\mathbf{T}}_j(t)\otimes e^{i(t'-t)\hat{\mathcal{Q}}\mathfrak{L}}({\mathbf{F}}_k(t)-\boldsymbol{\mathcal{F}}_k)
&& {\mathbf{T}}_j(t)\otimes e^{i(t'-t)\hat{\mathcal{Q}}\mathfrak{L}}({\mathbf{T}}_k(t)-\boldsymbol{\mathcal{T}}_k)
\end{array}
\right\rangle^\dag\notag.\end{aligned}$$ where $\boldsymbol{\mathcal{F}}\doteq\langle{\mathbf{F}}_k\rangle^\dag$ and $\boldsymbol{\mathcal{T}}\doteq\langle{\mathbf{T}}\rangle^\dag$. The behaviour of Equation (\[eq:a-23\]) as the product $m_b\,{\boldsymbol{\mathfrak{m}}}^{-1}$ vanishes can be obtained by simply letting ${\boldsymbol{\lambda}}\rightarrow \mathbf{0}$. In such a limit, the friction tensor ${\boldsymbol{\gamma}}$ can be approximated by the first term of a multi-power series in ${\boldsymbol{\lambda}}$,[@murphy_brownian_1972] $$\begin{aligned}
{\boldsymbol{\gamma}}_{jk}({\boldsymbol{\lambda}};t,t')\sim \left\langle
{\boldsymbol{\mathfrak{f}}}_j(t)\otimes ({\boldsymbol{\mathfrak{f}}}_k(t')-{\boldsymbol{\mathfrak{F}}}_k)
\right\rangle\label{eq:a-26}\end{aligned}$$ which yields $$\begin{aligned}
\partial_t{f^{(N)}}(t)+{\boldsymbol{\lambda}}\sum_{j=1}^N&\left({\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_j}\cdot {\widetilde{\boldsymbol{\mathfrak{p}}}}_j+{\boldsymbol{\mathfrak{F}}}_k\cdot{\boldsymbol{\nabla}}_{{\widetilde{\boldsymbol{\mathfrak{p}}}}_j}\right){f^{(N)}}(t) \label{eq:a-27}\\
=& \sum_{j,k=1}^N {\boldsymbol{\lambda}}^2\,{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j}\cdot\int_{-t_I}^{t}dt'\left\langle
{\boldsymbol{\mathfrak{f}}}_j(t)\otimes ({\boldsymbol{\mathfrak{f}}}_k(t')-{\boldsymbol{\mathfrak{F}}}_k)
\right\rangle\left({\boldsymbol{\nabla}}_{{\widetilde{\boldsymbol{\mathfrak{p}}}}_k}+\beta\,{\widetilde{\boldsymbol{\mathfrak{p}}}}_k\right){f^{(N)}}(t'). \notag\end{aligned}$$ The last step, the “Markovianization” of this time-evolution equation with memory, is the most controversial one[@snook_langevin_2006; @mazo_theory_1969; @mazur_molecular_1970; @bocquet_microscopic_1997] as it requires the assumption that ${f^{(N)}}(t')$ is very slowly varying compared to the correlation $\langle {\boldsymbol{\mathfrak{f}}}_j(t)\otimes{\boldsymbol{\mathfrak{f}}}_k(t) \rangle^\dag$, so that[@murphy_brownian_1972] $$\int_{-t_I}^{t}dt'\left\langle
{\boldsymbol{\mathfrak{f}}}_j(t)\otimes ({\boldsymbol{\mathfrak{f}}}_k(t')-{\boldsymbol{\mathfrak{F}}}_k)
\right\rangle\left({\boldsymbol{\nabla}}_{{\widetilde{\boldsymbol{\mathfrak{p}}}}_k}+\beta\,{\widetilde{\boldsymbol{\mathfrak{p}}}}_k\right){f^{(N)}}(t')\sim \int_{-\infty}^{t}dt'\left\langle
{\boldsymbol{\mathfrak{f}}}_j(t)\otimes ({\boldsymbol{\mathfrak{f}}}_k(t')-{\boldsymbol{\mathfrak{F}}}_k)
\right\rangle\left({\boldsymbol{\nabla}}_{{\widetilde{\boldsymbol{\mathfrak{p}}}}_k}+\beta\,{\widetilde{\boldsymbol{\mathfrak{p}}}}_k\right){f^{(N)}}(t). \label{eq:a-28}$$ Such an approximation is widely known to produce an incorrect description of the velocity correlation function if the colloidal particles have a similar density to that of the bath particles[@snook_langevin_2006; @mazo_theory_1969; @bocquet_microscopic_1997; @mazur_molecular_1970]. Equation (\[eq:a-28\]) gives an exponential decay, $\sim e^{-t}$, for the velocity correlation at large times, while with memory, such a decay is algebraic[@snook_langevin_2006], $\sim t^{-3/2}$. Nevertheless, it has also been pointed out that such long-time tails are very small compared to the exponential component predicted under the Markovianized theory[@snook_langevin_2006]. Although these could be reasons to avoid this critical step, the advantages of getting an FPE are significant. In contrast, the approximation is completely valid when both $m_b/m$ and $N_{\text{Kn}}\doteq r_0/R_0$ (the Knudsen number, with $r_0$ being a characteristic length scale for fluid intermolecular interactions and $R_0$ is a characteristic colloidal particle length scale) are considered very small[@peters_fokker-planck_1999]. Nevertheless, it was argued by @bocquet_microscopic_1997 that under such circumstances sedimentation of colloidal particles could occur. This seems to restrict the applicability of the resultant theory to microgravity scenarios. However, it is not clear at all what the actual significance of these algebraic long-time tails is and each case should be judged on its own merits[@mazo_theory_1969]. Thus Equation (\[eq:a-28\]) comprises, on the one hand, an uncontrolled approximation. On the other hand, it has been extensively used in statistical mechanics[@curtiss_kinetic_1957; @wolynes_dynamical_1977; @peters_fokker-planck_1999; @hernandez-contreras_brownian_1996]. For instance, the same hypothesis underlies a recent derivation of a unified DDFT to include inertia and HIs[@goddard_unification_2013], and is also involved in modern theories describing nucleation of colloidal systems and macromolecules[@lutsko_dynamical_2012; @lutsko_classical_2013; @duran-olivencia_mesoscopic_2015; @lutsko_two-parameter_2015]. Moreoever, the results obtained are in perfect agreement with experiments and simulations, corroborating the smallness of the error related to Equation (\[eq:a-28\]) when it comes to describing systems of interacting colloidal particles. More akin to the problem at hand, the hypothesis is tacitly assumed within the seminal work of @dickinson_brownian_1985, where a generalised algorithm to simulate protein diffusional problems is proposed. Therefore, while we cannot really justify such an assumption in a rigorous manner it does, nevertheless, represents the state-of-the-art in modelling colloidal systems. With this in mind, we can finally obtain the FPE related to Equation (\[eq:a-27\]) when Equation (\[eq:a-28\]) is taken into consideration, $$\begin{aligned}
\partial_t{f^{(N)}}(t)+\sum_{j=1}^N\left({\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_j}\cdot {\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_j+{\boldsymbol{\mathfrak{F}}}_j\cdot{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j}\right){f^{(N)}}(t) = \sum_{j,k=1}^N \,{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j}\cdot{\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N)\left({\boldsymbol{\mathfrak{p}}}_k+k_BT{\boldsymbol{\mathfrak{m}}}\,{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_k}\right){f^{(N)}}(t),
\label{eq:a-29}\end{aligned}$$ with the translational, rotational and coupled translational-rotational friction tensors, $${\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N)
\equiv
\left(
\begin{array}{cc}
{\boldsymbol{\Gamma}}_{jk}^{TT} & {\boldsymbol{\Gamma}}_{jk}^{TR}\\
{\boldsymbol{\Gamma}}_{jk}^{RT} & {\boldsymbol{\Gamma}}_{jk}^{RR}\\
\end{array}
\right)
\doteq \beta {\boldsymbol{\mathfrak{m}}}^{-1}\int_{0}^{\infty}ds\,\left\langle
{\boldsymbol{\mathfrak{f}}}_j(t)\otimes ({\boldsymbol{\mathfrak{f}}}_k(t-s)-{\boldsymbol{\mathfrak{F}}}_k)
\right\rangle. \label{eq:a-30}$$
Equations of motion
-------------------
The FPE previously derived can be rewritten in the less compact but more explicit form, $$\begin{aligned}
\partial_t {f^{(N)}}(t)+&\sum_{j=1}^N
{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{r}}}_j}\cdot
\left({\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_j{f^{(N)}}(t)\right)
-
{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j}
\cdot \left[\left({\boldsymbol{\mathfrak{F}}}_j-\sum_{k=1}^N{\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N){\boldsymbol{\mathfrak{p}}}_k
\right){f^{(N)}}(t)\right]\label{eq:a-31}\\
&= \sum_{j,k=1}^N
\left[
{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_j} \otimes
{\boldsymbol{\nabla}}_{{\boldsymbol{\mathfrak{p}}}_k}
\right]:\left(k_BT{\boldsymbol{\mathfrak{m}}}\,{\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N){f^{(N)}}(t)\right)\notag\end{aligned}$$ which is equivalent to the system of SDEs[@risken_fokker-planck_1996; @kampen_stochastic_2011; @berendsen_simulating_2007] $$\begin{aligned}
\dot{{\boldsymbol{\mathfrak{r}}}}_j(t)=&\ {\boldsymbol{\mathfrak{m}}}^{-1}{\boldsymbol{\mathfrak{p}}}_j(t)\label{eq:a-32}\\
\dot{{\boldsymbol{\mathfrak{p}}}}_j(t)=&\ {\boldsymbol{\mathfrak{F}}}_j-\sum_{k=1}^N{\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N){\boldsymbol{\mathfrak{p}}}_k + \sum_{k=1}^N\mathbf{A}_{jk}\boldsymbol{\xi}_k(t)\notag\end{aligned}$$ where ${\boldsymbol{\xi}}_j=(\mathbf{f}_j,\mathbf{t}_j)^\top$ is a 6-dimensional Gaussian white noise representing the random forces, $\mathbf{f}_j$, and torques, $\mathbf{t}_j$, acting upon the $j$-th particle, such that $\langle
\xi_j^a(t)\rangle=0$ and $\langle \xi_j^a(t)\xi_k^b(t')\rangle =
2\delta_{jk}\delta^{ab}\delta(t-t')$, where $\langle.\rangle$ refers to the average over an ensemble of the white-noise realisations. The strength of these random forces and torques is given by the tensor $\mathbf{A}_{jk}$ which obeys the fluctuation-dissipation relation, $$k_BT {\boldsymbol{\mathfrak{m}}}\,{\boldsymbol{\Gamma}}_{jk}({\boldsymbol{\mathfrak{r}}}^N) = \sum_{l=1}^N\mathbf{A}_{jl}({\boldsymbol{\mathfrak{r}}}^N)\mathbf{A}_{kl}({\boldsymbol{\mathfrak{r}}}^N)
\label{eq:a-33},$$ and $$\begin{aligned}
\mathbf{A}_{jk}{\boldsymbol{\xi}}_k\equiv
\begin{pmatrix}
\mathbf{A}_{jk}^{TT}&\mathbf{A}_{jk}^{TR}\\
\mathbf{A}_{jk}^{RT}&\mathbf{A}_{jk}^{RR}
\end{pmatrix}
\begin{pmatrix}
\mathbf{f}_{k}\\
\mathbf{t}_{k}
\end{pmatrix}.
\label{eq:a-34}\end{aligned}$$ Coming back to the expanded notation, the system of equations (\[eq:a-32\]) becomes $$\begin{aligned}
\frac{d{\mathbf{R}}_j}{dt}=&\,\frac{1}{m}{\mathbf{P}}_j,\label{eq:a-35}\\
\frac{d{\boldsymbol{\Phi}}_j}{dt}=&\,\mathbb{I}^{-1}{\mathbf{L}}_j\label{eq:a-36},\\
\frac{d{\mathbf{P}}_j}{dt}=&\, -\frac{\partial}{\partial {\mathbf{R}}_j}\widetilde{V}({\mathbf{R}}^N,{\boldsymbol{\alpha}}^N)-\sum_{k=1}^N\left({\boldsymbol{\Gamma}}_{jk}^{TT}{\mathbf{P}}_k+{\boldsymbol{\Gamma}}_{jk}^{TR}{\mathbf{L}}_k\right)+\sum_{k=1}^N\mathbf{A}_{jk}^{TT}\mathbf{f}_k(t)+\mathbf{A}_{jk}^{TR}\mathbf{t}_k(t)
\label{eq:a-37}\\
\frac{d{\mathbf{L}}_j}{dt}\equiv\mathbb{I}\frac{d{\boldsymbol{\omega}}_j}{dt}=&
-\frac{\partial}{\partial {\boldsymbol{\Phi}}_j}\widetilde{V}({\mathbf{R}}^N,{\boldsymbol{\alpha}}^N)-{\boldsymbol{\omega}}_j\times{\mathbf{L}}_j
\,
-\sum_{k=1}^N\left({\boldsymbol{\Gamma}}_{jk}^{RT}{\mathbf{P}}_k+{\boldsymbol{\Gamma}}_{jk}^{RR}{\mathbf{L}}_k\right)+\sum_{k=1}^N\mathbf{A}_{jk}^{RT}\mathbf{f}_k(t)+\mathbf{A}_{jk}^{RR}\mathbf{t}_k(t)
\label{eq:a-38}\end{aligned}$$ with Equation (\[eq:a-36\]) equivalent to the relation ${\boldsymbol{\omega}}_j={\boldsymbol{\Lambda}}_j^\top\dot{{\boldsymbol{\alpha}}_j}$, as can be verified by using Equation (\[eq:a-15\]). Equations (\[eq:a-35\]-\[eq:a-38\]), a much less convenient representation of the rotational-translational Langevin equation (\[eq:a-32\]), have been introduced in the studies by @wolynes_dynamical_1977, @dickinson_brownian_1985[@dickinson_brownian_1985; @dickinson_brownian_1985-1] and @hernandez-contreras_brownian_1996 but these authors started from postulated equations instead of the detailed microscopic derivation from the full system of bath and colloidal particles offered here.
On the rapid relaxation of the fluid: neglecting inertia in the bath {#app:neglecting-bath-inertia}
====================================================================
Here we address the question of when the inertia of the fluid bath can be neglected while having finite viscous forces. With this aim we will make use of some of the results derived by @peters_fokker-planck_1999 in his study on the FPE for coupled rotational and translational motions of structured Brownian particles. The main conclusion, the rapid relaxation of the fluid bath, is ultimately connected with the assumption of negligible inertial effects in the fluid bath whilst considering inertial effects of the colloids.
In a very detailed study, @peters_fokker-planck_1999 applied the multiple time-scale expansion to the derivation of the FPE for arbitrary-shape colloids. With this method @peters_fokker-planck_1999 showed that Equation (\[eq:a-29\]) is the formal time-evolution equation for the distribution function, up to $(m/M)^3$, when both $m/M$ and $N_{\text{Kn}}$ are considered small. It was also argued that the rapid relaxation of the fluid depends upon $N_{\text{Kn}}$ for a system with $m/M$ small. If we now make use of the fact that $m \sim r_0^3\rho_b$ and $M\sim R_0^3\rho_B$, the conditions for the FPE would to be a good description of the colloidal system can be reduced to requiring that $\rho_b/\rho_B$ must be small, which is indeed the condition pointed out by @bocquet_microscopic_1997 and many others[@roux_brownian_1992; @hauge_fluctuating_1973; @hinch_application_1975; @masters_time-scale_1986; @michaels_long-time_1975]. In the following we analyse whether or not this is possible while the ratio between inertia in the fluid bath and viscous forces is small. That is, whether is possible to have a low Reynolds number for the fluid bath along with the condition on the ratio between bath and colloidal densities. In such a case, it would be justified to neglect inertial effects in the bath.
To this end, we consider the case of solid and spherical colloids (far from walls) so that the coupling components of the friction and diffusion tensors vanish, and neglect HIs for the moment. If the radius of the colloidal particles is denoted by $R_0$, the friction tensor takes the simple form[@dickinson_brownian_1985], $$\begin{aligned}
{\boldsymbol{\Gamma}}_{jk}^{TT} = \gamma_{T}\delta_{jk}\,\mathbf{1},\quad {\boldsymbol{\Gamma}}_{jk}^{RR}=\gamma_{R}\delta_{jk}\,\mathbf{1},\quad {\boldsymbol{\Gamma}}_{jk}^{TR}={\boldsymbol{\Gamma}}_{jk}^{RT}=0,\label{eq:b-1}\end{aligned}$$ with $\gamma_{T}=6\pi\eta\, R_0/M$ and $\gamma_{R}=8\pi\eta\,R_0^3/I=20\pi\eta\,R_0/M$, and $\eta$ the dynamic viscosity, satisfying the Stokes-Einstein formula[@dickinson_brownian_1985]. For the sake of generality, the friction components will be considered equally important, i.e. $\gamma_{T}/\gamma_{R}\sim\mathcal{O}(1)$. For this reason, we can define the following two time scales $$\begin{aligned}
t_0^T = \left(\frac{MR_0^2\,\gamma_{TT}}{k_BT}\right),\quad t_0^R=\left(\frac{MR_0^2\,\gamma_{RR}}{k_BT}\right)
\label{eq:b-2},\end{aligned}$$ which are indeed of the same order-of-magnitude. Considering the natural physical scales of the system, the following dimensionless translational and rotational variables (denoted by an asterisk) are pertinent $$\begin{aligned}
\begin{array}{ccccc}
t = t_0^T\,t*,\ &\ {\mathbf{R}}_j = R_0\,{\mathbf{R}}^*_j,\ &\ {\mathbf{P}}_j = \frac{k_BT}{\gamma_TR_0}\,{\mathbf{P}}_j^*, \ &\ \mathbf{F}_j = \frac{k_BT}{R_0}\,\mathbf{F}_j^*,\ &\ \mathbf{F}_{j,\text{noise}}= \frac{k_BT}{R_0}\,\mathbf{F}_{j,\text{noise}}^*,\ \\
t = t_0^R\,t*,\ &\ {\boldsymbol{\omega}}_j = \frac{k_BT}{MR_0^2\gamma_R}{\boldsymbol{\omega}}^*_j,\ &\ {\mathbf{L}}_j = \frac{k_BT}{\gamma_R}\,{\mathbf{L}}_j^*, \ &\ \mathbf{T}_j = k_BT\,\mathbf{T}_j^*,\ &\ \mathbf{T}_{j,\text{noise}}= k_BT\,\mathbf{T}_{j,\text{noise}}^*,\ \\
\end{array}\label{eq:b-3}\end{aligned}$$ which, when applied in Equations (\[eq:a-35\])-(\[eq:a-38\]), yield the dimensionless equations of motion, $$\begin{aligned}
\frac{d{\mathbf{R}}_j^*}{dt^*}=&\,{\mathbf{P}}_j^*,\label{eq:b-4}\quad \frac{d{\boldsymbol{\Phi}}_j^*}{dt^*}=\,{\boldsymbol{\omega}}_j^*\\
\frac{d{\mathbf{P}}_j^*}{dt^*}=&\,\gamma_T^{*^2}\left(\mathbf{F}_j^*-{\mathbf{P}}_j^*+\mathbf{F}_{j,\text{noise}}^*(t)\right)\notag\\
\frac{d{\mathbf{L}}_j^*}{dt^*}=&\,\gamma_R^{*^2}\left(\mathbf{T}_j^*-{\mathbf{L}}_j^*+\mathbf{T}_{j,\text{noise}}^*(t)\right)\notag\end{aligned}$$ along with the definitions $$\begin{aligned}
\gamma_T^{*}=\gamma_T R_0\sqrt{\frac{M}{k_BT}},\quad \gamma_R^*=\gamma_{R} R_0\sqrt{\frac{M}{k_BT}}.
\label{eq:b-5}\end{aligned}$$ What we wish to test is whether or not it is possible to have finite viscous forces, i.e. $\gamma_T^{*}\sim\mathcal{O}(1)$, such that at the same time inertia forces in the fluid bath are negligible. For this purpose both translational and rotational Reynolds numbers must be small. They can be defined as[@happel_low_1981], $$\begin{aligned}
\text{Re}^T=\frac{U\,R_0\,\rho_b}{\eta},\quad \text{Re}^R=\frac{\Omega\,R_0^2\,\rho_b}{\eta},
\label{eq:b-6}\end{aligned}$$ respectively. The quantities $U$ and $\Omega$ represent typical linear and angular velocity scales, which can be obtained from the nondimensionalised momenta, $$U=\frac{k_BT}{MR_0\gamma_{T}},\quad \Omega=\frac{k_BT}{MR_0^2\gamma_{R}}.\label{eq:b-7}$$ Making use of (\[eq:b-2\]), (\[eq:b-3\]), (\[eq:b-5\]) and (\[eq:b-7\]) into (\[eq:b-6\]), we finally reach $$\text{Re}^T=\frac{9}{2\gamma_{T}^{*^2}}\frac{\rho_b}{\rho_B},\quad \text{Re}^R=\frac{15}{\gamma_{R}^{*^2}}\frac{\rho_b}{\rho_B}\label{eq:b-8}.$$ As already noted, the regime we consider involves, $\gamma_T^{*}\sim\gamma_{R}^*\sim\mathcal{O}(1)$ and $\text{Re}^T\sim\text{Re}^R\ll1$, which is undoubtedly satisfied when $\rho_b/\rho_B\ll 1$. Then, neglecting inertia forces in the fluid bath is consistent with the separation of time scales we already assumed to obtain the FPE. Further details of the physical interpretation and consequences of this limiting condition were given in Section \[app:projection-operator-tech\].
References {#references .unnumbered}
==========
[86]{} \[1\][\#1]{} \[1\][\#1]{} urlstyle \[1\][DOI \#1]{}
Archer, A.J.: Dynamical density functional theory for molecular and colloidal fluids: [A]{} microscopic approach to fluid mechanics. J. Chem. Phys. **130**(1), 014509 (2009)
Archer, A.J., Evans, R.: Dynamical density functional theory and its application to spinodal decomposition. J. Chem. Phys. **121**(9), 4246–4254 (2004)
Barrat, J., Hansen, J.: Basic [Concepts]{} for [Simple]{} and [Complex]{} [Liquids]{}. Cambridge University Press (2003)
Bechtel, D.B., Bulla, L.A.: Electron [Microscope]{} [Study]{} of [Sporulation]{} and [Parasporal]{} [Crystal]{} [Formation]{} in [Bacillus]{} thuringiensis. J. Bacteriol. **127**(3), 1472–1481 (1976)
Beenakker, C.W.J., Saarloos, W.v., Mazur, P.: Many-sphere hydrodynamic interactions. Physica A **127**(3), 451 – 472 (1984)
Berendsen, H.J.C.: Simulating the [Physical]{} [World]{}: [Hierarchical]{} [Modeling]{} from [Quantum]{} [Mechanics]{} to [Fluid]{} [Dynamics]{}. Cambridge University Press (2007)
Bernstein, D.S.: Matrix [Mathematics]{}: [Theory]{}, [Facts]{}, and [Formulas]{} with [Application]{} to [Linear]{} [Systems]{} [Theory]{}. Princeton University Press (2005)
Bocquet, L., Piasecki, J.: Microscopic derivation of non-[Markovian]{} thermalization of a [Brownian]{} particle. J. Stat. Phys. **87**(5-6), 1005–1035 (1997)
Boyd, J.P.: Chebyshev and [Fourier]{} [Spectral]{} [Methods]{}: [Second]{} [Revised]{} [Edition]{}. Courier Corporation (2001)
Brenner, H.: The [Stokes]{} resistance of an arbitrary particle[II]{}. Chem. Eng. Sci. **19**(9), 599 – 629 (1964)
Brown, R.: A brief account of microscopical observations made in the months of [June]{}, [July]{} and [August]{}, 1827, on the particles contained in the pollen of plants; and on the general existence of active molecules in organic and inorganic bodies. Phil. Mag. **4**, 161–173 (1828)
Cantaert, B., Beniash, E., Meldrum, F.C.: Nanoscale [Confinement]{} [Controls]{} the [Crystallization]{} of [Calcium]{} [Phosphate]{}: [Relevance]{} to [Bone]{} [Formation]{}. Chem.-Eur. J. **19**(44), 14918–14924 (2013)
Chan, G.K.L., Finken, R.: Time-[Dependent]{} [Density]{} [Functional]{} [Theory]{} of [Classical]{} [Fluids]{}. Phys. Rev. Lett. **94**(18), 183001 (2005)
Condiff, D.W., Brenner, H.: Transport [Mechanics]{} in [Systems]{} of [Orientable]{} [Particles]{}. Phys. Fluids **12**(3), 539–551 (1969)
Condiff, D.W., Dahler, J.S.: Brownian [Motion]{} of [Polyatomic]{} [Molecules]{}: [The]{} [Coupling]{} of [Rotational]{} and [Translational]{} [Motions]{}. J. Chem. Phys. **44**(10), 3988–4004 (1966)
Curtiss, C.F., Muckenfuss, C.: Kinetic [Theory]{} of [Nonspherical]{} [Molecules]{}. [II]{}. J. Chem. Phys. **26**(6), 1619–1636 (1957)
Dahler, J.S., Sather, N.F.: Kinetic [Theory]{} of [Loaded]{} [Spheres]{}. [I]{}. J. Chem. Phys. **38**(10), 2363–2382 (1963)
Darve, E., Solomon, J., Kia, A.: Computing generalized [Langevin]{} equations and generalized [Fokker]{}[Planck]{} equations. Proceedings of the National Academy of Sciences **106**(27), 10,884–10,889 (2009)
Deutch, J.M., Oppenheim, I.: Molecular [Theory]{} of [Brownian]{} [Motion]{} for [Several]{} [Particles]{}. J. Chem. Phys. **54**(8), 3547–3555 (1971)
Dickinson, E.: Brownian dynamic with hydrodynamic interactions: the application to protein diffusional problems. Chem. Soc. Rev. **14**(4), 421–455 (1985)
Dickinson, E., Allison, S.A., McCammon, J.A.: Brownian dynamics with rotationtranslation coupling. J. Chem. Soc., Faraday Trans. 2 **81**(4), 591–601 (1985)
Dieterich, W., Frisch, H.L., Majhofer, A.: Nonlinear diffusion and density functional theory. Z. Phys. B Cond. Mat. **78**(2), 317–323 (1990)
Donev, A., Vanden-Eijnden, E.: Dynamic density functional theory with hydrodynamic interactions and fluctuations. J. Chem. Phys. **140**(23), 234115 (2014)
Dunkl, C.F., Xu, Y.: Classical and [Generalized]{} [Classical]{} [Orthogonal]{} [Polynomials]{}. In: Orthogonal [Polynomials]{} of [Several]{} [Variables]{}, Encyclopedia of [Mathematics]{} and its [Applications]{}. Cambridge University Press (2001)
Dur[á]{}n-Olivencia, M.A., Lutsko, J.F.: Mesoscopic nucleation theory for confined systems: [A]{} one-parameter model. Phys. Rev. E **91**(2), 022402 (2015)
Ermak, D.L., McCammon, J.A.: Brownian dynamics with hydrodynamic interactions. J. Chem. Phys. **69**(4), 1352–1360 (1978)
Espa[ñ]{}ol, P., L[ö]{}wen, H.: Derivation of dynamical density functional theory using the projection operator technique. J. Chem. Phys. **131**(24), 244101 (2009)
Evans, G.T.: Cumulant expansion of a [Fokker]{}[Planck]{} equation: [Rotational]{} and translational motion in dense fluids. J. Chem. Phys. **65**(8), 3030–3039 (1976)
Evans, G.T.: Momentum space diffusion equations for chain molecules. J. Chem. Phys. **72**(7), 3849–3858 (1980)
Evans, R.: The nature of the liquid-vapour interface and other topics in the statistical mechanics of non-uniform, classical fluids. Adv. Phys. **28**(2), 143–200 (1979)
Goddard, B.D., Pavliotis, G.A., Kalliadasis, S.: The [Overdamped]{} [Limit]{} of [Dynamic]{} [Density]{} [Functional]{} [Theory]{}: [Rigorous]{} [Results]{}. Multiscale Model. Simul. **10**(2), 633–663 (2012)
Goddard, B.D., Nold, A., Kalliadasis, S.: Multi-species dynamical density functional theory. J. Chem. Phys. **138**(14), 144904 (2013)
Goddard, B.D., Nold, A., Savva, N., Pavliotis, G.A., Kalliadasis, S.: General [Dynamical]{} [Density]{} [Functional]{} [Theory]{} for [Classical]{} [Fluids]{}. Phys. Rev. Lett. **109**(12), 120603 (2012)
Goddard, B.D., Nold, A., Savva, N., Yatsyshin, P., Kalliadasis, S.: Unification of dynamic density functional theory for colloidal fluids to include inertia and hydrodynamic interactions: derivation and numerical experiments. J. Phys.: Condens. Matter **25**(3), 035101 (2013)
Goldstein, H., Poole, C.P., Safko, J.L.: Classical [Mechanics]{}. Addison Wesley (2002)
G[ó]{}mez-Morales, J., Iafisco, M., Delgado-L[ó]{}pez, J.M., Sarda, S., Drouet, C.: Progress on the preparation of nanocrystalline apatites and surface characterization: [Overview]{} of fundamental and applied aspects. Prog. Cryst. Growth Charact. Mater. **59**(1), 1–46 (2013)
Grabert, H., H[ä]{}nggi, P., Talkner, P.: Microdynamics and nonlinear stochastic processes of gross variables. J. Stat. Phys. **22**(5), 537–552 (1980)
Grad, H.: On the kinetic theory of rarefied gases. Commun. Pure Appl. Math. **2**(4), 331–407 (1949)
Gray, C.G., Gubbins, K.E.: Theory of [Molecular]{} [Fluids]{}: [I]{}: [Fundamentals]{}. OUP Oxford (1984)
Happel, J., Brenner, H.: Low [Reynolds]{} number hydrodynamics, *Mechanics of fluids and transport processes*, vol. 1. Springer Netherlands, Dordrecht (1981)
H[ä]{}rtel, A., Blaak, R., L[ö]{}wen, H.: Towing, breathing, splitting, and overtaking in driven colloidal liquid crystals. Phys. Rev. E **81**(5), 051703 (2010)
Hauge, E.H., Martin-L[ö]{}f, A.: Fluctuating hydrodynamics and [Brownian]{} motion. J. Stat. Phys. **7**(3), 259–281 (1973)
Hern[á]{}ndez-Contreras, M., Medina-Noyola, M.: Brownian motion of interacting nonspherical tracer particles: [General]{} theory. Phys. Rev. E **54**(6), 6573–6585 (1996)
Hinch, E.J.: Application of the [Langevin]{} equation to fluid suspensions. J. Fluid Mech. **72**(03), 499–511 (1975)
Hopkins, P., Fortini, A., Archer, A.J., Schmidt, M.: The van [Hove]{} distribution function for [Brownian]{} hard spheres: [Dynamical]{} test particle theory and computer simulations for bulk dynamics. J. Chem. Phys. **133**(22), 224505 (2010)
Jos[é]{}, J.V., Saletan, E.J.: Classical [Dynamics]{}: [A]{} [Contemporary]{} [Approach]{}. Cambridge University Press Textbooks (2013)
Kampen, N.G.V.: Stochastic [Processes]{} in [Physics]{} and [Chemistry]{}. Elsevier (2011)
Kirkwood, J.G.: The [Statistical]{} [Mechanical]{} [Theory]{} of [Transport]{} [Processes]{} [I]{}. [General]{} [Theory]{}. J. Chem. Phys. **14**(3), 180–201 (1946)
Koopmann, R., Cupelli, K., Redecke, L., Nass, K., DePonte, D.P., White, T.A., Stellato, F., Rehders, D., Liang, M., Andreasson, J., Aquila, A., Bajt, S., Barthelmess, M., Barty, A., Bogan, M.J., Bostedt, C., Boutet, S., Bozek, J.D., Caleman, C., Coppola, N., Davidsson, J., Doak, R.B., Ekeberg, T., Epp, S.W., Erk, B., Fleckenstein, H., Foucar, L., Graafsma, H., Gumprecht, L., Hajdu, J., Hampton, C.Y., Hartmann, A., Hartmann, R., Hauser, G., Hirsemann, H., Holl, P., Hunter, M.S., Kassemeyer, S., Kirian, R.A., Lomb, L., Maia, F.R.N.C., Kimmel, N., Martin, A.V., Messerschmidt, M., Reich, C., Rolles, D., Rudek, B., Rudenko, A., Schlichting, I., Schulz, J., Seibert, M.M., Shoeman, R.L., Sierra, R.G., Soltau, H., Stern, S., Str[ü]{}der, L., Timneanu, N., Ullrich, J., Wang, X., Weidenspointner, G., Weierstall, U., Williams, G.J., Wunderer, C.B., Fromme, P., Spence, J.C.H., Stehle, T., Chapman, H.N., Betzel, C., Duszenko, M.: In vivo protein crystallization opens new routes in structural biology. Nat. Methods **9**(3), 259–262 (2012)
Lebowitz, J.L., R[é]{}sibois, P.: Microscopic [Theory]{} of [Brownian]{} [Motion]{} in an [Oscillating]{} [Field]{}; [Connection]{} with [Macroscopic]{} [Theory]{}. Phys. Rev. **139**(4A), A1101–A1111 (1965)
Liboff, R.: Kinetic [Theory]{} - [Classical]{}, [Quantum]{}, and [Relativistic]{} [Descriptions]{}. Graduate [Texts]{} in [Contemporary]{} [Physics]{}, third edn. Springer-Verlag (2003)
Lutsko, J.F.: Recent [Developments]{} in [Classical]{} [Density]{} [Functional]{} [Theory]{}. In: Advances in [Chemical]{} [Physics]{}, pp. 1–92. John Wiley & Sons, Inc. (2010)
Lutsko, J.F.: A dynamical theory of nucleation for colloids and macromolecules. J. Chem. Phys. **136**(3), 034509 (2012)
Lutsko, J.F., Dur[á]{}n-Olivencia, M.A.: Classical nucleation theory from a dynamical approach to nucleation. J. Chem. Phys. **138**(24), 244908 (2013)
Lutsko, J.F., Dur[á]{}n-Olivencia, M.A.: A two-parameter extension of classical nucleation theory. J. Phys.: Condens. Matter **27**(23), 235101 (2015)
Marconi, U.M.B., Tarazona, P.: Dynamic density functional theory of fluids. J. Phys.: Condens. Matter **12**(8A), A413 (2000)
Masters, A.J.: Time-scale separations and the validity of the [Smoluchowski]{}, [Fokker]{}-[Planck]{} and [Langevin]{} equations as applied to concentrated particle suspensions. Mol. Phys. **57**(2), 303–317 (1986)
Mazo, R.M.: On the theory of [Brownian]{} motion. [I]{}. [Interaction]{} between [Brownian]{} particles. J. Stat. Phys. **1**(1), 89–99 (1969)
Mazur, P., Oppenheim, I.: Molecular theory of [Brownian]{} motion. Physica **50**(2), 241–258 (1970)
Michaels, I.A., Oppenheim, I.: Long-time tails and brownian motion. Physica A **81**(2), 221 – 240 (1975)
Miller, W.L., Cacciuto, A.: Hierarchical self-assembly of asymmetric amphiphatic spherical colloidal particles. Phys. Rev. E **80**(2), 021404 (2009)
Murphy, T.J., Aguirre, J.L.: Brownian [Motion]{} of [N]{} [Interacting]{} [Particles]{}. [I]{}. [Extension]{} of the [Einstein]{} [Diffusion]{} [Relation]{} to the [N]{}-[Particle]{} [Case]{}. J. Chem. Phys. **57**(5), 2098–2104 (1972)
Neuhaus, T., H[ä]{}rtel, A., Marechal, M., Schmiedeberg, M., L[ö]{}wen, H.: Density functional theory of heterogeneous crystallization. Eur. Phys. J. Spec. Top. **223**(3), 373–387 (2014)
Nold, A., Sibley, D.N., Goddard, B.D., Kalliadasis, S.: Fluid structure in the immediate vicinity of an equilibrium three-phase contact line and assessment of disjoining pressure models using density functional theory. Phys. Fluids **26**(7), 072001 (2014)
Nold, A., Sibley, D.N., Goddard, B.D., Kalliadasis, S.: Nanoscale [Fluid]{} [Structure]{} of [Liquid]{}-solid-vapour [Contact]{} [Lines]{} for a [Wide]{} [Range]{} of [Contact]{} [Angles]{}. Math. Model. Nat. Phenom. **10**(4), 111–125 (2015)
Peters, M.H.: Fokker-[Planck]{} equation and the grand molecular friction tensor for coupled rotational and translational motions of structured [Brownian]{} particles near structured surfaces. J. Chem. Phys. **110**(1), 528–538 (1999)
Peters, M.H.: The [Smoluchowski]{} diffusion equation for structured macromolecules near structured surfaces. J. Chem. Phys. **112**(12), 5488–5498 (2000)
Pottier, N.: Nonequilibrium [Statistical]{} [Physics]{}: [Linear]{} [Irreversible]{} [Processes]{}. Oxford University Press (2014)
Rex, M., Wensink, H.H., L[ö]{}wen, H.: Dynamical density functional theory for anisotropic colloidal particles. Phys. Rev. E **76**(2), 021403 (2007)
Risken, H.: The [Fokker]{}-[Planck]{} [Equation]{}: [Methods]{} of [Solutions]{} and [Applications]{}. 2nd ed. edn. Springer (1996)
Roux, J.N.: Brownian particles at different times scales: a new derivation of the [Smoluchowski]{} equation. Physica A **188**(4), 526–552 (1992)
Schilling, T., Frenkel, D.: Self-poisoning of crystal nuclei in hard-rod liquids. J. Phys.: Condens. Matter **16**(19), S2029 (2004)
Snook, I.: The [Langevin]{} and [Generalised]{} [Langevin]{} [Approach]{} to the [Dynamics]{} of [Atomic]{}, [Polymeric]{} and [Colloidal]{} [Systems]{}. Elsevier (2006)
van Teeffelen, S., Likos, C.N., L[ö]{}wen, H.: Colloidal [Crystal]{} [Growth]{} at [Externally]{} [Imposed]{} [Nucleation]{} [Clusters]{}. Phys. Rev. Lett. **100**(10), 108302 (2008)
Wilemski, G.: On the derivation of [Smoluchowski]{} equations with corrections in the classical theory of [Brownian]{} motion. J. Stat. Phys. **14**(2), 153–169 (1976)
Wittkowski, R., L[ö]{}wen, H.: Dynamical density functional theory for colloidal particles with arbitrary shape. Mol. Phys. **109**(23-24), 2935–2943 (2011)
Wolde, P.R.t., Frenkel, D.: Enhancement of [Protein]{} [Crystal]{} [Nucleation]{} by [Critical]{} [Density]{} [Fluctuations]{}. Science **277**(5334), 1975–1978 (1997)
Wolynes, P.G., Deutch, J.M.: Dynamical orientation correlations in solution. J. Chem. Phys. **67**(2), 733–741 (1977)
Wu, J., Li, Z.: Density-[Functional]{} [Theory]{} for [Complex]{} [Fluids]{}. Annu. Rev. Phys. Chem. **58**(1), 85–112 (2007)
Yatsyshin, P., Savva, N., Kalliadasis, S.: Spectral methods for the equations of classical density-functional theory: [Relaxation]{} dynamics of microscopic films. J. Chem. Phys. **136**(12), 124113 (2012)
Yatsyshin, P., Savva, N., Kalliadasis, S.: Geometry-induced phase transition in fluids: [Capillary]{} prewetting. Phys. Rev. E **87**(2), 020402(R) (2013)
Yatsyshin, P., Savva, N., Kalliadasis, S.: Density functional study of condensation in capped capillaries. J. Phys.: Condens. Matter **27**(27), 275104 (2015)
Yatsyshin, P., Savva, N., Kalliadasis, S.: Wetting of prototypical one- and two-dimensional systems: [Thermodynamics]{} and density functional theory. J. Chem. Phys. **142**(3), 034708 (2015)
Zhang, Z.X., Duijneveldt, J.S.v.: Isotropic-nematic phase transition of nonaqueous suspensions of natural clay rods. J. Chem. Phys. **124**(15), 154910 (2006)
[^1]: With $b$ and $B$ subscripts referring to the *bath* and *Brownian* (colloidal) particles respectively.
| {
"pile_set_name": "ArXiv"
} |
---
address: |
$^a$ Physik-Institut, Universität Zürich, Winterthurerstrasse 190, CH-8057 Zürich, Switzerland\
$^b$ Institute for Particle Physics Phenomenology, Durham University, Durham, DH1 3LE, UK\
$^c$ Theoretical Physics Department, CERN, CH-1211 Geneva 23, Switzerland
author:
- 'X. CHEN,$^a$ T. GEHRMANN,$^a$ N. GLOVER,$^b$ M. HÖFER,$^a$ A. HUSS$^c$'
title: 'ISOLATED PHOTON AND PHOTON+JET PRODUCTION AT NNLO QCD ACCURACY AND THE RATIO $R_{13/8}^\gamma$'
---
Introduction
============
Both inclusive isolated photon ($\gamma+X$) and photon-plus-jet ($\gamma+j$) production in $pp$ collisions present a means to test QCD dynamics using a colourless probe. Because their Born-level processes are $\bar{q}q\rightarrow g\gamma$ and $qg\rightarrow q\gamma$, related observables are sensitive to the gluon-distribution in the proton already at leading order (LO).
Recent experimental analyses by ATLAS [@ATLAS; @ATLAS2] and CMS [@CMS] pushed the experimental uncertainties down to a few percent. To match this accuracy also in theory calculations, the inclusion of NNLO QCD corrections is crucial. They have been calculated for $\gamma+X$ and $\gamma+j$ at $\sqrt{s}=8~\mathrm{TeV}$ by the MCFM [@MCFM] collaboration. In our recent paper [@paper] we present an independent calculation of the NNLO corrections, using the framework. is a parton-level event generator which uses the antenna subtraction method [@Antenna] to subtract the infrared (IR) QCD divergencies. The matrix elements for $\gamma+X$ and $\gamma+j$ are implemented up to NNLO in analytic form.
In the experimental environment it is necessary to separate any photon produced in the hard partonic scattering process from photons of other origin, for example radiation occurring during the hadronization process. One therefore measures the hadronic energy in the vicinity of the photon and defines conditions for its shape and amount. If these are met, the photon is said to be isolated.
When reconstructing the experimental isolation procedure in fixed-order theory calculations, one has to deal with hadronic radiation arbitrarily collinear to the photon. This must be taken care of by either including photon fragmentation functions to the perturbative order under consideration, which to NNLO has not been done so far, or by modifying the isolation prescription to eliminate the collinear configurations. In the latter approach a systematic difference between isolation procedures used in experiment and theory emerges.
Photon Isolation
================
There are several prescriptions for the photon isolation. They mainly differ in how exactly the “vicinity” of the photon is defined and how the hadronic energy therein may be distributed. The two most common ones are the fixed (hard) cone isolation, and the dynamical cone (Frixione [@Frixione]) isolation.
#### Fixed cone isolation -
A cone around the photon axis is defined by the distance $R=\sqrt{\Delta\eta^2+\Delta\phi^2}$, called the radius of the cone. The integrated hadronic transverse energy within the cone has to be smaller as a certain $E_T^\mathrm{max}$ for the photon to be considered as isolated. Often $E_T^\mathrm{max}$ is given as a simple linear function of the photon transverse momentum/ energy: $$E_T^\mathrm{max} = \varepsilon E_T^\gamma + E_T^\mathrm{thres}\,.$$ This isolation criterion is used in all experimental analyses so far. It allows, however, for hadronic radiation arbitrarily collinear to the photon, as long as its energy is not to large. This introduces a sensitivity to the photon fragmentation, which is difficult to describe from the theoretical viewpoint. On the other hand it is not possible to simply set $E_T^\mathrm{max}=0$, because this would, while indeed eliminating the fragmentation sensitivity, cut out part of the soft phase space, rendering observables IR unsafe.
#### Dynamical cone isolation -
Instead of a fixed $E_T^\mathrm{max}$ one defines a profile $E_T^\mathrm{max}(r_d)$ with $E_T^\mathrm{max}(r_d)\rightarrow0$ as $r_d\rightarrow 0$. $r_d$ is again the distance from the photon. For any sub-cone with $r_d$ smaller than some maximal radius $R_d$ the integrated energy within this sub-cone must not exceed $E_T^\mathrm{max}(r_d)$. The functional form of the profile conventionally used is $$E_T^\mathrm{max} = \varepsilon_dE_T^\gamma\left(\frac{1-\cos r_d}{1-\cos R_d}\right)^n \qquad \mathrm{for\ all\ \ } r_d<R_d\,.$$ This prescription both eliminates the fragmentation sensitivity and ensures IR safety. It can, however, only be approximated in experiments and so one has to tune the parameters of the dynamical isolation to fit the experimental setup as closely as possible.\
This difference in the isolation procedures used in experiment and theory is unsatisfactory, as it is a source of uncertainty, which is difficult to quantify. Only the inclusion of the photon fragmentation functions to the same order as the partonic calculation can solve this issue. To NNLO this has not been done so far. But an improvement over the current situation can already be achieved by combining both fixed and dynamical cone in a hybrid approach [@Hybrid], as used by ATLAS in their $\gamma+j$ study [@ATLAS2].
#### Hybrid cone isolation -
A dynamical cone with comparatively small $R_d$ is used to eliminate the fragmentation contribution. In a second stage of the isolation a fixed cone with $R^2\gg R_d^2$ is applied, the parameters of which are chosen according to any experimental analysis under consideration. In this way observables should retain the correct dependence on the parameters of the outer “physical” isolation cone. A residual dependence on the inner dynamical cone remains, but can in principle be made small for a suitable choice of parameters. In our paper [@paper] we present some technical studies on the choice of the inner cone parameters. We calculated the total cross section for $\gamma+X$ at $13~\mathrm{TeV}$ as a function of the outer isolation cone radius $R$, too. It would be interesting to see this analysis performed also in experiment.
\[fig:Ratio\] ![$R^\gamma_{13/8}$ as a function of the transverse energy/momentum of the isolated photon at LO, NLO and NNLO in four different rapidity bins, from central (top left) to most forward (bottom right). The theoretical uncertainty bands are derived by means of an independent variation of factorization and renormalization scales, both in the numerator and the denominator (see text for details). The results are compared to ATLAS data [@ATLAS2019].](iG_Et_gam_ATLAS_Ratio_138_31pt.pdf "fig:"){width="\textwidth"}
The ratio $R_{13/8}^\gamma$
===========================
In our paper [@paper] we calculated, using the hybrid isolation procedure, several differential distributions for $\gamma+X$ and $\gamma+j$ at $8~\mathrm{TeV}$ and $13~\mathrm{TeV}$, based on studies by ATLAS [@ATLAS; @ATLAS2] and CMS [@CMS]. We found that the inclusion of NNLO corrections leads to an significant improvement in both the accuracy of the predictions and the description of the data. Amounting to no more than a few percent, the theory uncertainties are now competitive with experimental errors.
Here we present a NNLO calculation of the ratio $R_{13/8}^\gamma$ of the $\gamma+X$ cross section at $8~\mathrm{TeV}$ and $13~\mathrm{TeV}$, differential in $p_T^\gamma$ and presented in four rapidity bins. It is based on a recent measurement of this quantity by ATLAS [@ATLAS2019]. Measuring ratios is a means to reduce the experimental systematic uncertainties.
Both the $8~\mathrm{TeV}$ and the $13~\mathrm{TeV}$ measurements of the $p_T^\gamma$-distribution in isolated photon production by ATLAS were performed in four different regions in rapidity $$|y^\gamma|<0.6\,,\qquad 0.6<|y^\gamma|<1.37\,,\qquad 1.56<|y^\gamma|<1.81\,,\qquad 1.81<|y^\gamma|<2.37\,,$$ which excludes the region $[1.37,1.56]$. The ratio is measured in the same bins, using the overlap of the phase-space regions of both measurements, with $p_T^\gamma>125~\mathrm{GeV}$.
For the prediction we use the NNPDF3.1 PDF set and a hybrid photon isolation with parameters $$\begin{array}{lll}
R_d = 0.1\,,& \qquad\varepsilon_d = 0.1\,,& \qquad n = 2\,,\\[5pt]
R = 0.4\,,& \qquad E_T^\mathrm{thres} = 4.8~\mathrm{GeV}\,,& \qquad \varepsilon = 0.0042\,,
\end{array}$$ where the fixed-cone parameters ($R,E_T^\mathrm{thres},\varepsilon$) correspond to the isolation set-up used by ATLAS.
The theory prediction for $R_{13/8}^\gamma$ has not been performed as an independent calculation, but rather has been derived using the two calculations for $8~\mathrm{TeV}$ and $13~\mathrm{TeV}$. For both the theoretical uncertainty is estimated by means of a seven-point scale variation, $\mu_F=a\,p_T^\gamma$, $\mu_R=b\,p_T^\gamma$ with $a,b\in\{1/2,1,2\}$, where we exclude the configurations with $a/b\in\{1/4,4\}$.
The uncertainty of $R_{13/8}^\gamma$ has now been estimated by forming the ratio for all possible combinations of the seven scale configurations for numerator and denominator, excluding again the combinations where the ratio of any two scales equals $1/4$ or $4$. This effectively corresponds to a generalisation of the seven-point scale variation for two scales to a 31-point variation for four scales.
In figure \[fig:Ratio\] we show the result in the four rapidity bins mentioned above and compare to ATLAS data [@ATLAS2019]. Except for the highest bins in $p_T^\gamma$ the description of the data is excellent. Like for the calculations [@paper] for individual $\sqrt{s}$ we see a significant reduction in the uncertainty when going from NLO to NNLO: While at NLO the uncertainty lies between $(+10,-9)\%$ and $(+17,-14)\%$, only slightly growing with $p_T^\gamma$ and $|y^\gamma|$, at NNLO it lies between $(+3.4,-2.8)\%$ and $(+6.5,-4.0)\%$.
References {#references .unnumbered}
==========
[99]{}
The ATLAS Collaboration, ,
The ATLAS Collaboration,
The CMS Collaboration, ,
J.M. Campbell [*et al*]{}, ,
X. Chen [*et al*]{}, arXiv:1904.01044
A. Gehrmann-De Ridder [*et al*]{}, , A. Daleo [*et al*]{} , J. Currie [*et al*]{}
S. Frixione,
F. Siegert,
The ATLAS Collaboration,
The NNPDF Collaboration,
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Topological spin configurations in proximity to a superconductor have recently attracted great interest due to the potential application of the former in spintronics and also as another platform for realizing non-trivial topological superconductors. Their application in these areas requires precise knowledge of the existing exchange fields and/or the stray-fields which are therefore essential for the study of these systems. Here, we determine the effective stray-field $\vb{H}_{str}$ and the Meissner currents $\vb{j}_{\text{S}}$ in a Superconductor/Ferromagnet/Superconductor (S/F/S) junction produced by various nonhomogenous magnetic textures $\vb{M}(\vb*{r})$ in the F. The inhomogeneity arises either due to a periodic structure with flat domain walls (DW) or is caused by an isolated chiral magnetic skyrmion (Sk). We consider both Bloch– and Sk and also analyze in detail the periodic structures of different types of DW’s— that is DW (BDW) and DW (NDW) of finite width with in- and out-of-plane magnetization vector $\vb{M}(x)$. The spatial dependence of the fields $\vb{H}_{str}(\vb*{r})$ and Meissner currents $\vb{j}_{\text{S}}(\vb*{r})$ are shown to be qualitatively different for the case of Bloch– and magnetic textures. While the spatial distributions in the upper and lower S are identical for Sk and DW’s they are asymmetric for the case of magnetic textures. The depairing factor, which determines the critical temperature $T_{c}$ and which is related to vector potential of the stray-field, can have its maximum at the center of a magnetic domain but also, as we show, above the DW. For Sk’s the maximum is located at a finite distance within the Sk radius $r_{\text{Sk}}$. Based on this, we study the nucleation of superconductivity in the presence of DW’s. Because of the asymmetry for structures, the critical temperature $T_{c}$ in the upper and lower S is expected to be different. The obtained results can also be applied to S/F bilayers.'
author:
- 'Samme M. Dahir, Anatoly F. Volkov, and Ilya M. Eremin'
bibliography:
- 'literature.bib'
date:
-
-
title: 'Meissner Currents Induced by Topological Magnetic Textures in Hybrid Superconductor/Ferromagnet Structures'
---
Over the past decades, continuous efforts have been made to study superconductor-ferromagnet heterostructures due to a variety of interesting features caused by the proximity effect, *i.e.*, the penetration of Cooper pairs from the superconductor (S) into the ferromagnet (F). The most interesting and well established effects are the sign reversal of the Josephson current in S/F/S junctions and the appearance of a long-ranged triplet component (see review articles [@golubov_current-phase_2004; @buzdin_proximity_2005; @bergeret_odd_2005; @eschrig_spin-polarized_2015; @linder_superconducting_2015; @balatsky_odd_2017; @ohnishi2020spintransport] and references therein).
Other interesting features involve the interplay of various types of topological defects that, under certain conditions, can be present in the superconductor and/or ferromagnet. One of these topological defects are the Abrikosov vortices which occur in superconductors[@Abrikosov:1956sx] in the magnetic field interval, . There are also several different topological structures that can be found in ferromagnets. The most prominent ones are magnetic domain walls (DW), where the magnetization vector $\vb{M}$ rotates by an angle $\pi$ across the DW. Another example of a topological defect that has received much attention recently due to its potential application in spintronics are the so-called magnetic Skyrmions (Sk)[@bogdanov1989thermodynamically; @Leonov2016; @fert_magnetic_2017; @Everschor2018]. These local whirl-like structures are topologically equivalent to two DW’s as one can map the inner part of the Sk on the stripes between two domains via conformal transformation. Similar to flat DW’s, where the magnetization vector $\vb{M}$ changes its direction by rotating either in the $(x,z)$-plane () or in the $(y,z)$-plane (), the winding of chiral Sk can either have a Bloch– or a Néel–like structure. Which type of chiral Sk is realized depends on the underlying chiral interaction. Note that there is already some work on the mutual interaction between topological defects occurring in ferromagnets and superconductors, see review[@lyuksyutov_ferromagnetsuperconductor_2005] and references therein. In the absence of the direct proximity effect (no direct contact between S and F), this interaction is realized through the magnetic stray-field $\vb{H}_{\text{str}}$ generated by the non-uniform magnetic textures in the F and the magnetic field associated with the superconducting vortices. The creation of Pearl and Abrikosov vortices in S/F structures with and without DW’s has been analyzed in Refs.[@lyuksyutov_magnetization_1998; @erdin_topological_2001; @lyuksyutov_ferromagnetsuperconductor_2005; @milosevic_interaction_2003; @laiho_penetration_2003]. More recently, the spontaneous creation of vortices in S/F structures with Sk’s with and without direct proximity effect was also studied theoretically[@Hals2016; @baumard_generation_2019; @dahir_interaction_2019; @Menezes2019; @Rex2019; @Palermo2020]
As it is well known, there is no stray-field $\vb{H}_{\text{str}}$ outside of uniformly magnetized infinite film [@landau_electrodynamics_nodate]. However, within the ferromagnet the magnetic induction $\vb{B}_{\text{F}}$ or the magnetic field $\vb{H}_{\text{F}}$ can still acquire finite values, [*i.e.*]{}, and for the in-plane magnetization and and for the out-of-plane magnetization. Therefore for a uniform magnetization $\vb{M}_{0}$ in the F of a S/F/S structure, both the $\vb{B}_{\text{S}}$, the $\vb{H}_{\text{S}}$ and the Meissner current $\vb{j}_{\text{S}}$ are equal zero in the superconductors where . Thus, non-zero stray-fields and Meissner currents can only occur if the magnetization of an infinite F is non-homogeneous. This was studied in S/F structures with DW’s of zero width in Refs.[@laiho_penetration_2003; @stankiewicz_magnetic_1997; @bulaevskii_ferromagnetic_2000]and for DW’s of finite width in Ref.[@Burmistrov2005]. In the presence of a proximity effect and spin-orbit coupling, the Meissner current was recently calculated in a bilayer S/F structure with a particular Sk in the F and a vortex in the S [@baumard_generation_2019].
Despite of existing literature, there are still no systematic studies of Meissner currents in S/F and S/F/S structures with different topological magnetic textures (Sk’s or flat DW’s) with different orientations of the magnetization vector $\vb{M}$. This is particular interesting due to potential realization of Majorana fermions in such heterostructures[@Yang2016; @Garnier2019; @Rex2019]. In the present paper we address this topic, by analyzing S/F/S systems with an isolated magnetic Sk (Bloch– and Sk) or with a periodic flat DW structure (out-of-plane and in-plane magnetization, Bloch and Néel DW’s) in the ferromagnetic material.
Assuming that there is no proximity effect present and that there are no Abrikosov vortices in the S/F/S structure , *i.e.*, magnetic stray-fields are supposed to be less than the critical field $H_{c1}$, we find the effective magnetic stray-field from which we deduce the induced screening currents $\vb{j}_{\text{S}}$. Note that the obtained spatial distribution of the current density $\vb{j}_{\text{S}}$ in the S in a Josephson system S/F/S is qualitatively similar to that in S/F bilayer. The knowledge of the effective stray-field $\vb{H}_{\text{str}}$ and the current density $\vb{j}_{\text{S}}$ allows to estimate the region where superconductivity nucleates upon decreasing the temperature $T$ below $T_{c}$: Either at the DW’s or in the center of the domains. Available experimental data point out that the nucleation of superconductivity preferably occurs at the DW’s[@Iavarone2014]. However it will be shown that the exact location depends on the considered type of DW’s. In addition, we show that new interesting and non-trivial features arise in the system under consideration. For example, we find an pronounced asymmetry in the $z$-dependence of $\vb{j}_{\text{S}}(z)$, which occurs for both Néel DW’s and magnetic Sk’s. This asymmetry is characterized by the in-plane dependence $\vb*{r}_{\perp}$ of the Meissner current $\vb{j}_{\text{S}}(\vb*{r}_{\perp})$ and stray-field $\vb{H}_{\text{str}}(\vb*{r}_{\perp})$, which differs greatly above and below the ferromagnet and can even result in a local sign change of the Meissner current. In the absence of a superconductor, the asymmetry of the stray-field for DW’s is already known[@mallinson_one-sided_1973]. For instance, it was recently demonstrated in experiments on artificial magnetic structures[@Marioni_Halbach_2018]. In the case of Bloch DW’s such an asymmetry does not arise. This difference between Bloch– or DW’s and Sk’s follows from the different orientation of the vector $\vb{\hat{e}}_{\text{rot}}$ which describes the rotation axis of the magnetization $\vb{M}$ along the domain wall. For instance, in the case of Bloch– and DW one can define a vector , where $\vb{\hat{e}}_{x}$ is a unit vector normal to the plane of the DW. For DW’s this vector is non-zero, while for DW’s the vector product is zero because the rotation vector $\vb{\hat{e}}_{\text{rot}}$ is collinear to the vector $\vb{\hat{e}}_{x}$. In the language of magnetic monopoles, which can be used for magnetic stray-fields, the presence of $\vb{\hat{e}}_{N}$ translates into the existence of magnetic bulk charges. In combination with the magnetic surface charges, the stray-field components of the bulk charges results in the aforementioned asymmetry. In the case of a S/F bilayer it generally makes no sense to speak about an asymmetry, but the spatial distribution of the Meissner $\vb{j}_{\text{S}}$ in the S still depends on the direction of the vector $\vb{\hat{e}}_{N}$ with respect to the F film (upward or downwards).
We will begin this paper by calculating a general expression for the effective magnetic stray-field $\vb{H}(\vb*{r})$ in an S/F/S structure generated by a nonhomogeneous two-component magnetization $\vb{M}$, see . From the stray-field, we extract an expression for the Meissner current in the two superconducting region, which is then applied to describe induced currents in the presence of isolated Néel– and Sk, see , as well as for various magnetic DW configurations, see . In , we use a Ginzburg-Landau model to estimate the nucleation of superconductivity in the presence of the DW structures, we considered earlier. Note that the obtained results in this sections are independent of the type of S, as we are working with unscreened magnetic stray-fields. The universal expressions for these unscreened fields can be easily extracted from our results in the previous section. We end this work with a conclusion in .
Stray field and Meissner current \[Section1\]
=============================================
We consider an S/F/S structure, that is, a ferromagnetic film of thickness $2d_{\text{F}}$ interfaced by two superconductors at . The magnetization $\vb{M}(\vb*{r})$ inside the ferromagnet can be written in the form $$\vb{M}(\vb*{r})=M_{0}\vb{n}(\vb*{r})$$ where the unit vector $\vb{n}(\vb*{r})$ is a function of the position vector with $\vb*{r}_{\perp}$ lying in the $(x,y)$-plane. In the following the magnetization is assumed to be independent of the $z$-coordinate .
We will now begin with determining the spatial distribution of the screened stray-field in the superconducting regions. The superconducting order parameters (OP) are assumed to be homogeneous. Any magnetic field inside the S must then satisfy the London equation which we write for the Fourier component $$\partial_{zz}^{2}\vb{H}_{\text{S}}(\vb*{k},z)-\kappa^{2}\vb{H}_{\text{S}}(\vb*{k},z)=0, \quad \text{S regions}\label{Eq: London Fourier}$$ where and $\lambda_{\text{L}}$ is the London penetration depth. In the general case, the two superconductors may have different London penetration depth $\lambda_{\text{L}}^{+}$ and $\lambda_{\text{L}}^{-}$. The solution of Eq.(\[Eq: London Fourier\]) is given by $$\begin{aligned}
\vb{H}_{\text{S}}(\vb*{k},z)=\vb{C}_{\pm}(\vb*{k})e^{-\kappa_{\pm}\abs{z}}, \quad \text{S regions}\end{aligned}$$ where the index $\pm$ of the constant and $\kappa^{\pm}$ indicates their values in the upper/lower superconducting regions, respectively.\
The stray-field generated by the magnetization $\vb{M}$ inside the F has to fulfill the magnetostatic condition which allows us to define a magnetic scalar potential $U$ with $$\vb{H}_{\text{F}}(\vb*{k},z)=-\qty(i\vb*{k}U(\vb*{k},z),\partial_{z}U(\vb*{k},z))\label{Eq: Relation Potential and H Field}\quad$$ In the absence of the proximity effect (PE) the potential $U$ is related to the magnetization $\vb{M}$ via so that we can write $$\partial_{zz}^{2}U(\vb*{k},z)-\abs{k}^{2}U(\vb*{k},z)=4\pi M_{0} i\vb{k}\vdot\vb{n}_{\perp}(\vb*{k}), \quad \text{F film} \label{Eq: DGL Potential}\quad$$ Solving Eq.(\[Eq: DGL Potential\]) for $U(\vb*{k},z)$ we obtain $$\begin{aligned}
U(\vb*{k},z)=&4\pi M_{0}\Bigg\{A(\vb*{k})\sinh(\abs{k}z)+B(\vb*{k})\cosh(\abs{k}z)\notag\\
&-\frac{i\vb*{k}\vdot\vb{n}_{\perp}(\vb*{k})}{\abs{k}^2}+C_{0}\delta(\vb*{k})+\frac{i\vb*{k}\vdot \vb{C}_{\vb*{r}_{\perp}}}{\abs{k}^{2}}\delta(\vb*{k})\Bigg\}
\end{aligned}$$ with the Dirac $\delta$-function $\delta(\vb*{k})$. The last two terms are contributions to the homogeneous solutions of Eq.(\[Eq: DGL Potential\]). In the coordinate representation it has the form: Note, the constant $C_{0}$ does not affect any physical quantity, so that we can set . The constant $\vb{C}_{\perp}$ on the other hand, is related to a non-compensated magnetic moment $\vb{M}_{un}$ in the F which turns to zero for . Using Eq.(\[Eq: Relation Potential and H Field\]) we can determine the stray-field $\vb{H}_{\text{F}}$ in the F film $$\begin{aligned}
\vb{H}_{\text{F}}^{\perp}(\vb*{k},z)=&-4\pi M_{0} i\vb*{k} \Bigg\{A(\vb*{k})\sinh(\abs{k}z)+B(\vb*{k})\cosh(\abs{k}z)\notag\\
&-\frac{i\vb*{k}\vdot\vb{\bar{n}}_{\perp}(\vb*{k})}{\abs{k}^2}\Bigg\}\\
\mathrm{H}_{\text{F}}^{z}(\vb*{k},z)=&-4\pi M_{0} \abs{k}\qty{A(\vb*{k})\cosh(\abs{k}z)+B(\vb*{k})\sinh(\abs{k}z)} \end{aligned}$$ where we defined $$\vb{\bar{n}}_{\perp}(\vb*{k}):=\qty(\vb{n}_{\perp}(\vb*{k})-\vb{C}_{\vb*{r}_{\perp}}\delta(\vb{k}))$$ The constants $A(\vb*{k})$, $B(\vb*{k})$ and $\vb{C}_{\pm}(\vb*{k})$ can be found using the matching conditions for the magnetic field and the magnetic induction at the S/F interfaces. They are reduced to the continuity of the tangential components of the in-plane field $\vb{H}^{\perp}(\vb*{k},z)$ and the normal component of the magnetic induction , *i.e.*, $$\begin{aligned}
\vb{H}_{\text{S}}^{\perp}(\pm d_{\text{F}})=&\vb{H}_{\text{F}}^{\perp}(\pm d_{\text{F}})\label{Eq: Interface Cond1}\\
\mathrm{H}_{\text{S}}^{z}(\pm d_{\text{F}})=&\mathrm{H}_{\text{F}}^{z}(\pm d_{\text{F}})+4\pi M_{0}\mathrm{n}_{z}(\vb*{k})\label{Eq: Interface Cond2}
\end{aligned}$$ In addition, the in-plane component of $\vb{H}_{\text{S}}$ is coupled to the normal component via the equation so that $$\vb{C}_{\pm}^{\perp}(\vb*{k})=\mp\frac{i\vb*{k}}{\abs{k}}\frac{\kappa_{\pm}}{\abs{k}}\mathrm{C}_{\pm}^{z}(\vb*{k}) \label{Eq: Divergenzless HS}$$
Using Eqs.(\[Eq: Interface Cond1\]-\[Eq: Divergenzless HS\]) we can determine the coefficients $A(\vb*{k})$ and $B(\vb*{k})$, which are given by $$\begin{aligned}
A(\vb*{k})=&\mathrm{n}_{z}(\vb*{k})\frac{\kappa_{+}D_{2}^{-}(k)+\kappa_{-}D_{2}^{+}(k)}{\abs{k}D(k)}+\vb{\bar{n}}_{\perp}(\vb*{k})\frac{i\vb*{k}}{\abs{k}}\frac{D_{2}^{-}(k)-D_{2}^{+}(k)}{D(k)}\\
B(\vb*{k})=&\mathrm{n}_{z}(\vb*{k})\frac{\kappa_{+}D_{1}^{-}(k)-\kappa_{-}D_{1}^{+}(k)}{\abs{k}D(k)}+\vb{\bar{n}}_{\perp}(\vb*{k})\frac{i\vb*{k}}{\abs{k}}\frac{D_{1}^{-}(k)+D_{1}^{+}(k)}{D(k)}
\end{aligned}$$ with $D(k)=D_{1}^{-}(k)D_{2}^{+}(k)+D_{1}^{+}(k)D_{2}^{-}(k)$, where $$\begin{aligned}
D_{1}^{\pm}(k)=&\abs{k}\sinh(\abs{k}d_{\text{F}})+\kappa_{\pm}\cosh(\abs{k}d_{\text{F}})\\
D_{2}^{\pm}(k)=&\abs{k}\cosh(\abs{k}d_{\text{F}})+\kappa_{\pm}\sinh(\abs{k}d_{\text{F}})\end{aligned}$$ and also the coefficient $\vb{C}_{\pm}(\vb*{k})$ given by $$\begin{aligned}
C_{\pm}^{z}(k)=-4\pi M_{0}\abs{k}\sinh(\abs{k}d_{\text{F}})e^{\kappa_{\pm} d_{\text{F}}}&\left[\pm\frac{i \vb*{k}}{\abs{k}}\frac{\vb{\bar{n}}_{\perp}(\vb*{k})}{D(k)}\qty(\qty{D_{1}^{-}(k)+D_{1}^{+}(k)}\pm\qty{\kappa_{-}-\kappa_{+}}\cosh(\abs{k}d_{\text{F}}))\right.\notag\\
&\left.-\frac{\mathrm{n}_{z}(\vb*{k})}{D(k)}\qty(\qty{D_{2}^{-}(k)+D_{2}^{+}(k)}\pm\qty{\kappa_{-}-\kappa_{+}}\sinh(\abs{k}d_{\text{F}}))\right]
\end{aligned}$$ For the sake of simplicity, we will from now on consider two identical superconducting materials, *i.e.*, $\lambda_{\text{L}}^{+}=\lambda_{\text{L}}^{-}$. In this case the expression for the coefficients can be reduced to $$\begin{aligned}
A(\vb*{k})=\frac{\kappa}{\abs{k}}\frac{\mathrm{n}_{z}(\vb*{k})}{D_{1}(k)},
&&B(\vb*{k})=\frac{i\vb*{k}}{\abs{k}}\frac{\vb{\bar{n}}_{\perp}(\vb*{k})}{D_{2}(k)}
\end{aligned}$$ and $$\mathrm{C}_{\pm}^{z}(\vb*{k})=-4\pi M_{0}\abs{k}\sinh(\abs{k}d_{\text{F}})e^{\kappa d_{\text{F}}}\qty[\pm\frac{i\vb*{k}}{\abs{k}}\frac{\vb{\bar{n}}_{\perp}(\vb*{k})}{D_{2}(k)}-\frac{\mathrm{n}_{z}(\vb*{k})}{D_{1}(k)}]$$ With this we obtain the $k$-space representation of the screened stray-field in an S/F/S junction for two identical superconductors.
In the S region $\abs{z}>d_{\text{F}}$: $$\begin{aligned}
\vb{\tilde{H}}_{\perp}^{(\text{S})}(\vb*{k},z)=&\kappa\sinh(\abs{k}d_{\text{F}})\frac{i\vb*{k}}{\abs{k}}\qty[\frac{i\vb*{k}}{\abs{k}}\frac{\vb{\bar{n}}_{\perp}(\vb*{k})}{D_{2}(k)}\mp\frac{\mathrm{n}_{z}(\vb*{k})}{D_{1}(k)}]e^{-\kappa\abs{z\mp d_{\text{F}}}}\label{Eq: HSperp kspace}\\
\mathrm{\tilde{H}}_{z}^{(\text{S})}(\vb*{k},z)=&-\abs{k}\sinh(\abs{k}d_{\text{F}})\qty[\pm\frac{i\vb*{k}}{\abs{k}}\frac{\vb{\bar{n}}_{\perp}(\vb*{k})}{D_{2}(k)}-\frac{\mathrm{n}_{z}(\vb*{k})}{D_{1}(k)}]e^{-\kappa\abs{z\mp d_{\text{F}}}}\label{Eq: Hz kspace}
\end{aligned}$$ In the F film $\abs{z}<d_{\text{F}}$: $$\begin{aligned}
\vb{\tilde{H}}_{\perp}^{(\text{F})}(\vb*{k},z)=&-\frac{i\vb*{k}}{\abs{k}}\qty[\frac{i\vb*{k}}{\abs{k}}\vb{\bar{n}}_{\perp}(\vb*{k})\qty(\frac{\abs{k}\cosh(\abs{k}z)}{D_{2}(k)}-1)+\frac{\mathrm{n}_{z}(\vb*{k})\kappa\sinh(\abs{k}z)}{D_{1}(k)}]\label{Eq: HFperp kspace}\\
\mathrm{\tilde{H}}_{z}^{(\text{F})}(\vb*{k},z)=&-
\qty[\frac{i\vb*{k}}{\abs{k}}\frac{\vb{\bar{n}}_{\perp}(\vb*{k})\abs{k}\sinh(\abs{k}z)}{D_{2}(k)}+\frac{\mathrm{n}_{z}(\vb*{k})\kappa\cosh(\abs{k}z)}{D_{1}(k)}] \label{Eq: HFz kspace}
\end{aligned}$$ where we expressed the results in terms of a dimensionless field $\tilde{\vb{H}}=\vb{H}/4\pi M_{0}$.
The obtained expressions describe the screened stray-field in an S/F/S structure. By taking , *i.e.*, , we can also extract the unscreened stray-field which would be present in the absence of superconductors. In this limit, the result describes the general distribution of the stray-field created by a nonhomogeneous magnetization in a F in vacuum. The associated vector potential is later used to estimate the nucleation of superconductivity. The origin of the screening field that leads to the effective stray-field in Eq.(\[Eq: HFperp kspace\],\[Eq: HFz kspace\]), are the induced supercurrents inside the S. The supercurrent (Meissner current) $\vb{j}_{\text{S}}$ can be determined using Ampère’s law . $$\vb{j}_{\text{S}}(\vb*{r}_{\perp},z)=\frac{c}{4\pi}\qty[\qty(\nabla_{\vb*{r}_{\perp}}+\hat{\vb{e}}_{z}\partial_{z})\times\qty(\vb{H}_{\perp}(\vb*{r}_{\perp},z)+\hat{\vb{e}}_{z}\mathrm{H}_{z}(\vb*{r}_{\perp},z))]$$ In the Fourier representation we further obtain $$\vb{j}_{\text{S}}(\vb*{k},z)=\frac{c}{4\pi}\qty[\qty(i\vb*{k}+\hat{\vb{e}}_{z}\partial_{z})\times\qty(\vb{H}_{\perp}(\vb*{k},z)+\hat{\vb{e}}_{z}\mathrm{H}_{z}(\vb*{k},z))]$$ It can easily be shown, that the supercurrent disappears within the F, which is the expected result when the PE is absent. Outside the ferromagnet , we obtain the following expression $$\vb{j}_{\text{S}}(\vb*{k},z)=\frac{cM_{0}\lambda_{\text{L}}^{-2}}{\abs{k}}\qty(\hat{\vb{e}}_{z}\times\frac{i\vb*{k}}{\abs{k}})\mathrm{\tilde{H}}_{z}^{(\text{S})}(\vb*{k},z)\label{Eq:MeissnerCurrent}$$ from which we can directly derive the vector potential $\vb{A}$ in the superconductor using $$\vb{A}^{(\text{S})}(\vb*{k},z)=-\frac{4\pi M_{0}}{\abs{k}}\qty(\hat{\vb{e}}_{z}\times\frac{i\vb*{k}}{\abs{k}})\mathrm{\tilde{H}}_{z}^{(\text{S})}(\vb*{k},z) \label{Eq: Vectorpotential S}$$
{width="\linewidth"}
Isolated Skyrmion \[Section2\]
==============================
In this section, we will set the magnetization profile $\vb{M}$ to describe an isolated magnetic skyrmion (Sk) in a ferromagnetic background. It is assumed that the Sk’s are stabilized by an underlying chiral interaction resulting in either Bloch– or Sk’s. The magnetization profile has a cylindrical symmetry and varies along the radial direction $\vb*{\rho}$ so that . The unit vector $\vb{n}$ of a chiral Bloch or Néel Sk can then be written as $$\begin{aligned}
\vb{n}_{\perp}(\vb*{\rho})=&\frac{\vb*{\rho}}{\rho}\sin(\theta(\rho))\Theta(r_{\text{Sk}}-\rho), \quad &&\text{N\'{e}el Sk} \label{Eq: M inplane Neel}\\
\vb{n}_{\perp}(\vb*{\rho})=&\frac{\vb{\hat{e}}_{z}\cross\vb*{\rho}}{\rho}\sin(\theta(\rho))\Theta(r_{\text{Sk}}-\rho), \quad &&\text{Bloch Sk} \label{Eq: M inplane Bloch}
\end{aligned}$$ for the in-plane component and $$\mathrm{n}_{z}(\vb*{\rho})=\qty[1+\cos(\theta(\rho))]\Theta(r_{\text{Sk}}-\rho)-1 , \quad \text{Bloch \& N\'{eel} Sk} \label{Eq: M out-ofplane}$$ for the out-of-plane component. Here, $\theta(\rho)$ describes the angular variation of the magnetization w.r.t. the $z$-axis and $\Theta(r_{\text{Sk}}-\rho)$ is a Heaviside step function with $r_{\text{Sk}}$ being the skyrmion radius. The Fourier components of $\vb{n}(\vb*{\rho})$ are equal to $$\begin{aligned}
\vb{n}_{\perp}(\vb*{k})=&-2\pi\frac{i\vb*{k}}{k} m_{\perp}(k), \quad &&\text{N\'{e}el Sk}\\
\vb{n}_{\perp}(\vb*{k})=&-2\pi\frac{\vb{\hat{e}}_{z}\cross\vb*{k}}{k} m_{\perp}(k), \quad &&\text{Bloch Sk}\end{aligned}$$ and $$\mathrm{n}_{z}(\vb*{k})=2\pi\qty[m_{z}(k)-2\pi\delta(\vb*{k})], \quad \text{Bloch \& N\'{eel} Sk}$$ The functions $m_{\perp}(k)$ and $m_{z}(k)$ are defined as $$\begin{aligned}
m_{\perp}(k)=&\int_{0}^{r_{\text{Sk}}}\dd{\rho}\rho J_{1}(k\rho)\sin(\theta(\rho))\\
m_{z}(k)=&\int_{0}^{r_{\text{Sk}}}\dd{\rho}\rho J_{0}(k\rho)\qty[1+\cos(\theta(\rho))]
\end{aligned}$$ where $J_{n}(x)$ is the Bessel-function of the first kind of order $n$. The angular dependence of the Sk profile $\theta(\rho)$, is described using a circular $360^{\circ}$–domain wall Ansatz[@Romming_field_2015]. $$\theta(\rho)=\sum_{\pm}\arcsin(\tanh(-\frac{\rho\mp c}{w/2}))\label{Eq: Magnetizationprofile}$$ with $c$ being the size of the domain core and $w$ is the domain wall width. For the remainder of this work, we set $c=0\,\mathrm{nm}$. Using Eq.(\[Eq: Magnetizationprofile\]), one can estimate the radius $r_{\text{Sk}}$ of the Sk. It should be noted that the expressions in this section can be used for any radially symmetric magnetization profile.
Using the obtained result from the previous section, we will begin analyzing the effective stray-field generated by a Sk in our S/F/S structure in the case of two identical superconductors. Afterwards we will determine the corresponding induced Meissner currents. Taking into account that for a Bloch Sk (see Eq.(\[Eq: M inplane Bloch\])), we see that the first term in Eqs.(\[Eq: HSperp kspace\]-\[Eq: HFz kspace\]) vanishes. This means that the individual components of the stray-field are either symmetric or anti-symmetric functions of $z$. On the other hand, the in-plane magnetization of a Néel Sk . Hence, in this case $\vb{H}(z)\neq\vb{H}(-z)$ which describes an asymmetry of the magnetic stray-field. This asymmetry is a typical feature of stray-fields generated by magnetic textures with Néel–like magnetization[@Mallinson_Onesided_1973; @Marioni_Halbach_2018].
In order to fully determine the magnetic stray-field and the Meissner current, we first need to specify the value of the constant $\vb{C}_{\vb*{r}_{\perp}}$. Using the condition that the spatial average of the in-plane component of the stray-field vanishes, *i.e.*, , we get an additional equation for $\vb{C}_{\vb*{r}_{\perp}}$. For the case of an isolated Sk this constant is equal to zero . The real-space representation of the screened stray-field in Eqs.(\[Eq: HSperp kspace\]-\[Eq: HFz kspace\]) can now be easily expressed as
{width="\linewidth"}
In the S region $\abs{z}>d_{\text{F}}$: $$\begin{aligned}
\vb{\tilde{H}}_{\perp}^{(\text{S})}(\vb*{\rho},z)=&-\frac{\vb*{\rho}}{\rho}\int_{0}^{\infty}\dd{k}kJ_{1}(k\rho)\kappa\sinh(kd_{\text{F}})\qty[\frac{\bar{m}_{\perp}(k)}{D_{2}(k)}\mp\frac{m_{z}(k)}{D_{1}(k)}]e^{-\kappa\abs{z\mp d_{\text{F}}}}\\
\mathrm{\tilde{H}}_{z}^{(\text{S})}(\vb*{\rho},z)=&-\int_{0}^{\infty}\dd{k}kJ_{0}(k\rho)k\sinh(kd_{\text{F}})\qty[\pm\frac{\bar{m}_{\perp}(k)}{D_{2}(k)}-\frac{m_{z}(k)}{D_{1}(k)}]e^{-\kappa\abs{z\mp d_{\text{F}}}}
\end{aligned}$$ In the F film : $$\begin{aligned}
\vb{\tilde{H}}_{\perp}^{(\text{F})}(\vb*{\rho},z)=&\frac{\vb*{\rho}}{\rho}\int_{0}^{\infty}\dd{k}kJ_{1}(k\rho)\qty[\bar{m}_{\perp}(k)\qty(\frac{k\cosh(kz)}{D_{2}(k)}-1)+m_{z}(k)\frac{\kappa\sinh(kz)}{D_{1}(k)}]\\
\mathrm{\tilde{H}}_{z}^{(\text{F})}(\vb*{\rho},z)=&-\int_{0}^{\infty}\dd{k}kJ_{0}(k\rho)\qty[\bar{m}_{\perp}(k)\frac{k\sinh(kz)}{D_{2}(k)}+m_{z}(k)\frac{\kappa\cosh(kz)}{D_{1}(k)}]+\frac{1}{2}\end{aligned}$$
where we inserted the magnetization profile Eq.(\[Eq: M inplane Neel\]-\[Eq: M out-ofplane\]) and defined $$\bar{m}_{\perp}(k)=\begin{cases}
m_{\perp}(k) , &\quad \text{N\'{e}el Sk}\\
0, &\quad \text{Bloch Sk}
\end{cases}$$ Analogously, the Meissner current can be found using Eq.(\[Eq:MeissnerCurrent\]), which has the following form in real-space $$\begin{aligned}
\vb{j}_{\text{S}}(\rho,z)=&cM_{0}\lambda_{\text{L}}^{-2}\int_{0}^{\infty}\dd{k}kJ_{1}(k\rho)\sinh(kd_{\text{F}})\notag\\
&\times\qty[\pm\frac{\bar{m}_{\perp}(k)}{D_{2}(k)}-\frac{m_{z}(k)}{D_{1}(k)}]e^{-\kappa\abs{z\mp d_{\text{F}}}}\vb{\hat{e}}_{\varphi}\label{Eq: SK Meissner current}
\end{aligned}$$ The stray-field induces circulating supercurrents pointing in $\vb{\hat{e}}_{\varphi}$-direction. Since the supercurrent is linked to the stray-field, the Meissner current also features the asymmetry which is related to the magnetization profile of the Néel Sk. Using Eq.(\[Eq: SK Meissner current\]), this asymmetry can be identified by the changing sign in the term associated with the in-plane contribution of the magnetization. In we show the spatial dependence of the Meissner current $\vb{j}_{\text{S}}(\rho,\pm d_{\text{F}})$ in the upper (a) and c)) and the lower (b) and d)) superconductors in the presence of a Sk in the ferromagnetic material. The curves are displayed for the parameters and . As expected, we observe a strong asymmetry in the dependence $\vb{j}_{\text{S}}(\rho, d_{\text{F}})$ and $\vb{j}_{\text{S}}(\rho,- d_{\text{F}})$ in the upper and lower superconductors. The current $\vb{j}_{\text{S}}(\rho,d_{\text{F}})$ in the upper S changes sign at some finite distance within the Sk region whereas the current $\vb{j}_{\text{S}}(\rho,- d_{\text{F}})$ remains negative for all $\rho$. Note that the sign reversal of the Meissner current in S/F systems has been found earlier [@bergeret_spin_2004; @volkov_spin_2019; @mironov_electromagnetic_2018], but its underlying mechanism was different as it was related to the proximity effect. In the case of Bloch Sk, all the mentioned features are missing and the Meissner current in both superconducting regions is the same, see .
{width="\linewidth"}
Flat Domain Walls\[Section3\]
=============================
In this section we consider the magnetization profiles of several different periodic flat DW’s. The alignment of magnetization changes across the DWs as a function of the $x$-coordinate, *i.e.*, $\vb*{r}_{\perp}=x\vb{\hat{e}}_{x}$ with $\vb{\hat{e}}_{x}$ being the corresponding unit-vector. The period of the structures is $2L_{0}$. This enables us to expand all function as a Fourier series: For example, the vector $\vb{n}(x,z)$ is represented as $$\vb{n}(x,z)=\sum_{n=-\infty}^{\infty}\vb{n}(k_{n},z)\exp(ik_{n}x)$$ with $$\vb{n}(k_{n},z)=\frac{1}{2L_{0}}\int_{-L_{0}}^{L_{0}}\dd{x}\vb{n}(x,z)\exp(-ik_{n}x)\label{Eq: FourierSeries}$$ where . Below we drop the subindex $n$ for brevity.
Now suppose that the vector $\vb{n}(x)$ depends only on the $x$-coordinate, [*i.e.*]{} it is completely described by its $x$-component . In this case, the expression for the normalized magnetic stray-field and the Meissner current $j_{\text{S}}$ can be obtained in the same manner as in . For instance, for two identical S, we obtain the magnetic stray-field $\vb{H}^{(\text{S})}(k,z)$ by substituting and in Eq.(\[Eq: HSperp kspace\]-\[Eq: Hz kspace\]). For periodic DW’s, one further needs to replace , which follows from the finite range of integration in Eq.(\[Eq: FourierSeries\]). Finally, the normalized field components in the superconducting regions are: $$\begin{aligned}
\tilde{\mathrm{H}}_{x}^{(\text{S})}(k,z)=&-\kappa\sinh(\abs{k}d_{\text{F}})\qty[\frac{\mathrm{\bar{n}}_{x}(k)}{D_{2}(k)}\pm\frac{ik}{\abs{k}}\frac{\mathrm{n}_{z}(k)}{D_{1}(k)}]e^{-\kappa\abs{z\mp d_{\text{F}}}}\label{Eq: DW HSx}\\
\tilde{\mathrm{H}}_{z}^{(\text{S})}(k,z)=&-\abs{k}\sinh(\abs{k}d_{\text{F}})\qty[\pm\frac{ik}{\abs{k}}\frac{\mathrm{\bar{n}}_{x}(k)}{D_{2}(k)}-\frac{\mathrm{n}_{z}(k)}{D_{1}(k)}]e^{-\kappa\abs{z\mp d_{\text{F}}}}\end{aligned}$$ with $$\mathrm{\bar{n}}_{x}(k):=\mathrm{n}_{x}(k)-C_{x}\frac{\sin(kL_{0})}{kL_{0}}\label{Eq: barnx}$$ and analogously within the ferromagnet $\abs{z}<d_{\text{F}}$: $$\begin{aligned}
\tilde{\mathrm{H}}_{x}^{(\text{F})}(k,z)=&\qty[\mathrm{\bar{n}}_{x}(k)\qty(\frac{\abs{k}\cosh(\abs{k}z)}{D_{2}(k)}-1)-\frac{ik}{\abs{k}}\frac{\mathrm{n}_{z}(k)\kappa\sinh(\abs{k}z)}{D_{1}(k)}]\\
\tilde{\mathrm{H}}_{z}^{(\text{F})}(k,z)=&-\qty[\frac{ik}{\abs{k}}\frac{\mathrm{\bar{n}}_{x}(k)\abs{k}\sinh(\abs{k}z)}{D_{2}(k)}+\frac{\mathrm{n}_{z}(k)\kappa\cosh(\abs{k}z)}{D_{1}(k)}]\end{aligned}$$ The Meissner current can be extracted from Eq.(\[Eq:MeissnerCurrent\]). The supercurrent flows in $y$-direction and has the magnitude $$j_{\text{S}}(k,z)=cM_{0}\lambda_{\text{L}}^{-2}\sinh(\abs{k}d_{\text{F}})\qty[\pm\frac{\mathrm{\bar{n}}_{x}(k)}{D_{2}(k)}+\frac{ik}{\abs{k}}\frac{\mathrm{n}_{z}(k)}{D_{1}(k)}]e^{-\kappa\abs{z\mp d_{\text{F}}}}\label{Eq: DW jS}$$ where once again the $\pm$ indicates the solution in the upper or lower S region, respectively. The current in coordinate representation $\vb{j}(x,z)$ can be calculated using $$\vb{j}(x,z)=\sum_{k}\vb{j}(k,z)\exp(ikx)$$ Having determined the expressions for $\vb{H}_{str}$ and $\vb{j}_{\text{S}}$ for an arbitrary type of DW, we need to specify the precise magnetic texture. Its components can be expressed in terms of the function $\mathrm{n}_{even}$ and $\mathrm{n}_{odd}$ which are characterized by an even or odd dependency on $x$ or $k$. Since we are interested in a qualitative spatial dependence of all quantities (the fields and the Meisner currents), we approximate $\mathrm{n}_{even,odd}$ by
$$\begin{aligned}
n_{odd}(x)=&\cos(\frac{\pi}{2}\frac{(x-L)}{w})\theta(w-\abs{x-L})-\cos(\frac{\pi}{2}\frac{(x+L)}{w})\theta(w-\abs{x+L})\label{Eq:noddx}\\
n_{even}(x)=&\qty[1-\sin(\frac{\pi}{2}\frac{(x-L)}{w})]\theta(w-\abs{x-L})+\qty[1+\sin(\frac{\pi}{2}\frac{(x+L)}{w})]\theta(w-\abs{x+L})+2\theta(L-w-\abs{x})-1
\end{aligned}$$
with the domain wall width $2w$ and the size of the domain . That is, we assume that the rotation angle of the vector $\vb{n}$ outside the DW’s remains constant whereas it changes linearly inside the DW’s . This approximation allows us to present results in a simple analytical form. Outside the interval , $n(x)$ is a periodic function of $x$: . The Fourier components of $n_odd(x)$ and $n_{even}(x)$ are equal to $$\begin{aligned}
n_{odd}(k)=&\frac{2\pi iw}{2L_{0}}\frac{\cos(kw)\sin(kL)}{k^2w^2-\qty(\pi/2)^{2}}=\frac{2\pi iw}{2L_{0}} f(k)\label{Eq:noddk}\\
n_{even}(k)=&\frac{\pi^{2}}{2L_{0}k}\frac{\cos(kw)\sin(kL)}{k^2w^2-\qty(\pi/2)^{2}}-\frac{2\sin(kL_{0})}{2L_{0}k}=\frac{2\pi w}{2L_{0}} F(k)\label{Eq:nevenk}
\end{aligned}$$ with $$\begin{aligned}
f(k)=&\frac{\cos(kw)\sin(kL)}{k^2w^2-\qty(\pi/2)^{2}}\label{Eq:foddk}\\
F(k)=&\frac{\pi}{2kw}f(k)-\frac{\sin(kL_{0})}{\pi kw}\label{Eq:fevenk}
\end{aligned}$$ Obviously, $f(k)$ is also an odd function, whereas $F(k)$ is an even function of $k$. It should be noted that the limiting case of the DW width $w=0$ was analyzed in Ref.[@Laiho2003; @stankiewicz_magnetic_1997; @bulaevskii_ferromagnetic_2000; @Sonin_ferromagnetic_2002].
In our model, the vector $\vb{n}$ has two non-zero components that allow the construction of six different magnetic textures (see ). They are characterized by vectors $\vb{n}$ with the following components: $(0,n_{even},n_{odd})$, $(0,n_{odd},n_{even})$, $(n_{even},0,n_{odd})$, $(n_{odd},0,n_{even})$ and $(n_{even},n_{odd},0)$, $(n_{odd},n_{even},0)$. Note that we are working with the underlying assumption of fixed chirality, *i.e*, the vector $\vb{n}$ rotates in the same direction within the DW’s, which is either clock-wise or counter-clockwise. Another chirality may be obtained if the rotation of the vector $\vb{n}$ in adjacent DW’s occurs in different directions; then the function $n_{odd}$ should be replaced by , where $\tilde{n}_{even}$ is equal to $$\begin{aligned}
\tilde{n}_{even}(x)=&\cos(\frac{\pi}{2}\frac{(x-L)}{w})\theta(w-\abs{x-L})+\cos(\frac{\pi}{2}\frac{(x+L)}{w})\theta(w-\abs{x+L})\\
\tilde{n}_{even}(q)=&-\frac{2\pi w}{2L_{0}}\frac{\cos(kw)\cos(kL)}{k^2w^2-\qty(\pi/2)^{2}}
\end{aligned}$$
{width="\linewidth"}
In order to obtain our final result for the magnetic stray-field and the Meissner current from Eq.(\[Eq: DW HSx\]-\[Eq: DW jS\]), we need to determine the constant $C_{x}$. As mentioned in , the average over the in-plane component $H_{x}^{(\text{F})}$ has to vanish, *i.e*, . From this follows $$\begin{aligned}
\expval{H_{x}^{(\text{F})}(x,z)}=&\frac{1}{2L_{0}}\int_{-L_{0}}^{L_{0}}\dd{x}\frac{1}{2d_{\text{F}}}\int_{-d_{\text{F}}}^{d_{\text{F}}}\dd{z}H_{x}^{(\text{F})}(x,z)\\
=&\frac{4\pi M_{0}\lambda_{\text{L}}^{-1}d_{\text{F}}}{1+\lambda_{\text{L}}^{-1}d_{\text{F}}}\qty[\mathrm{n}_{x}(k=0)-C_{x}]\overset{!}{=}0
\end{aligned}$$ where we used . Hence, the constant $C_{x}$ is given by $$C_{x}=\mathrm{n}_{x}(k=0,z)$$ The quantity can be either or . In the latter case, $\mathrm{n}_{x}(k=0,z)=0$ (see Eq.(\[Eq:noddk\],\[Eq:foddk\])) and therefore $C_{x}=0$. The other case is only realized for certain DW’s and results in $$C_{x}=\frac{2L}{L_{0}}-1$$ *i.e.*, the constant $C_{x}$ vanishes for $L_{0}=2L$. Otherwise, if , the domains with positive and negative magnetization differ in size, which leads to an uncompensated total magnetization $\vb{M}_{un}$ and . We will now examine the various possible magnetic textures that exhibit a chirality as defined in Eq.(\[Eq:noddx\]). Note that, the type of DW in a ferromagnetic sample is determined by the existing magnetic interaction and material specific parameters (temperature, thickness of the F film etc). Accordingly, the actual magnetic texture in the F corresponds to the configuration associated with the minimum of the thermodynamic potential. Nevertheless, we will find the spatial distribution of the Meissner currents for all possible configurations, bearing in mind that some of these textures might not be energetically favorable, but could be achieved in artificial magnetic structures [@Marioni_Halbach_2018].
Out-of-plane $\vb{n}$ (Néel and Bloch DW’s)
-------------------------------------------
For an out-of-plane magnetization, both Néel and/or Bloch DW’s (see can exist within the F. The Néel DW (NDW$^{(z)}$) is described by the following configuration $\vb{n}(x)$ $$\vb{n}(x)=(n_{odd},0,n_{even}), \qquad \text{NDW}^{(z)}$$
The superscript $(z)$ indicates the alignment of the vector $\vb{n}(x)$ across the domains, which is oriented along the $z$-axis. The Meissner current at the interfaces is obtain by inserting the corresponding Fourier components in Eq.(\[Eq: DW jS\]). $$j_{\text{S}}(k,\pm d_{\text{F}})=cM_{0}\lambda_{\text{L}}^{-2}\sinh(\abs{k}d_{\text{F}})\frac{2\pi iw}{2L_{0}}\qty[\pm\frac{f(k)}{D_{2}(k)}+\frac{\text{sgn}(k)F(k)}{D_{1}(k)}]$$ where $f(k)$ and $F(k)$ are given in Eq.(\[Eq:foddk\],\[Eq:fevenk\]). One can easily see that the current is an odd function of $k$. In the coordinate representation, we obtain the following result $$j_{\text{S}}(x,\pm d_{\text{F}})=-\frac{4\pi M_{0}cw\lambda_{\text{L}}^{-2}}{2L_{0}}\sum_{k=0}^{\infty}\sin(kx)\sinh(k d_{\text{F}})\qty[\pm\frac{f(k)}{D_{2}(k)}+\frac{F(k)}{D_{1}(k)}]$$
{width="\linewidth"}
In Fig. we plot the dependence of the normalized current for , , and . This plot shows a strong asymmetry between the upper and lower superconductors with currents flowing above/below the DW regions. The direction of the supercurrent depends on the direction of rotation of the DW. Varying the value for the London penetration depth reveals a sign change for the supercurrent within DW’s in the lower supercurrent (see . One can see that the Meissner currents at different DW’s flow in opposite directions. This means that the currents flow along closed loops. Unlike the case of Abrikosov vortices, there is no phase change along these loops. This sign change is similar to the behavior described for the Néel Sk’s.
In the case of a Bloch DW (BDW$^{(z)}$) the vector $\vb{n}(x)$ has the components $$\vb{n}(x)=(0,n_{odd},n_{even}), \qquad \text{BDW}^{(z)}$$ where the magnetization in the domain is once again oriented along the $z$-direction. The Meissner current in Fourier representation is given by $$j_{\text{S}}(k,\pm d_{\text{F}})=cM_{0}\lambda_{\text{L}}^{-2}\sinh(\abs{k}d_{\text{F}})\frac{2\pi iw}{2L_{0}}\frac{\text{sgn}(k)F(k)}{D_{1}(k)}$$ and in the coordinate representation by $$j_{\text{S}}(x,\pm d_{\text{F}})=-\frac{4\pi M_{0}cw\lambda_{\text{L}}^{-2}}{2L_{0}}\sum_{k=0}^{\infty}\sin(kx)\sinh(k d_{\text{F}})\frac{F(k)}{D_{1}(k)}$$ One can directly deduce that the resulting Meissner currents are identical in the upper and lower superconductors, which is due to the missing $x$-component of the magnetization. This is once again, similar to the Sk case, as there was also no asymmetry present for Bloch Sk’s. For both NDW$^{(z)}$ and BDW$^{(z)}$ follows that $j_{\text{S}}(x,z)$ is an odd function of $x$ so that the total current vanishes. In , we plot the dependence of the Meissner current for the considered case of a BDW$^{(z)}$. The parameter are the same as in .
In-plane $\vb{n}$ (Néel and Bloch DW’s)
---------------------------------------
Let us first consider a DW where the magnetization vector $\vb{n}(x)$ at the domains is oriented along the (see ), then $$\vb{n}(x)=(n_{odd},n_{even},0), \qquad \text{NDW}^{(y)}$$ The Fourier component of the Meissner current is equal to $$j_{\text{S}}(k,\pm d_{\text{F}})=\pm cM_{0}\lambda_{\text{L}}^{-2}\sinh(\abs{k}d_{\text{F}})\frac{2\pi iw}{2L_{0}}\frac{f(k)}{D_{2}(k)}$$ and in the coordinate representation $$j_{\text{S}}(x,\pm d_{\text{F}})=\mp\frac{4\pi M_{0}cw\lambda_{\text{L}}^{-2}}{2L_{0}}\sum_{k=0}^{\infty}\sin(kx)\sinh(k d_{\text{F}})\frac{f(k)}{D_{2}(k)}$$ The functions $j_{\text{S}}(x,d_{\text{F}})$ and $j_{\text{S}}(x,-d_{\text{F}})$ are shown in . Once again the currents differ in the two superconducting regions. The magnitude of the currents is the same, but the currents flow in opposite direction, resulting in an antisymmetric behavior.
The Bloch type DW (see ) is described by $$\vb{n}(x)=(0,n_{even},n_{odd}), \qquad \text{BDW}^{(y)}$$ The occurring currents $j_{\text{S}}(x,z)$ for BDW$^{(y)}$ are even function of $x$ given by $$j_{\text{S}}(x,\pm d_{\text{F}})=-\frac{4\pi M_{0}cw\lambda_{\text{L}}^{-2}}{2L_{0}}\sum_{k=0}^{\infty}\cos(kx)\sinh(k d_{\text{F}})\frac{f(k)}{D_{1}(k)}$$ For both the currents are equal. In , we plot the $x$-dependence of the functions $j_{\text{S}}(x,d_{\text{F}})$ and $j_{\text{S}}(x,-d_{\text{F}})$. The Meissner currents in the upper and lower S near the BDW flow in the same direction. The total current $j_{av}$ is zero. The results for $j_{\text{S}}(x,z)$ for the BDW$^{(y)}$ are similiar to those obtained by Burmistrov and Chtchelkatchev[@Burmistrov2005].
Other types of NDW
------------------
Other types of NDW’s correspond to a magnetization profile $\vb{n}(x)$ in which the alignment in the domain is along the $x$-direction (see ). The rotation of the vector $\vb{n}(x)$ occurs either in the $(x,z)$-plane or in the $(x,y)$-plane. Thus, the vector $\vb{n}(x)$ has the components $$\begin{aligned}
\vb{n}(x)=&(n_{even},n_{odd},0), \qquad \text{NDW}_{y}^{(x)}\\
\vb{n}(x)=&(n_{even},0,n_{odd}), \qquad \text{NDW}_{z}^{(x)}
\end{aligned}$$ Remember, that for the case the constant $C_{x}$ has a finite value given by . It follows that $$\mathrm{\bar{n}}_{x}(k)=n_{even}(x)-C_{x}\frac{\sin(kL_{0})}{kL_{0}}=\frac{2\pi w}{2L_{0}}\bar{F}(k)$$ with $$\bar{F}(k)=\frac{\pi}{2kw}f(k)-\frac{2L}{L_{0}}\frac{\sin(kL_{0})}{\pi kw}$$
With this expression, we obtain the Meissner current $$\begin{aligned}
j_{\text{S}}^{1}(x,\pm d_{\text{F}})=&\pm\frac{4\pi M_{0}cw\lambda_{\text{L}}^{-2}}{2L_{0}}\sum_{k=0}^{\infty}\cos(kx)\sinh(k d_{\text{F}})\frac{\bar{F}(k)}{D_{2}(k)}\\
j_{\text{S}}^{2}(x,\pm d_{\text{F}})=&\frac{4\pi M_{0}cw\lambda_{\text{L}}^{-2}}{2L_{0}}\sum_{k=0}^{\infty}\cos(kx)\sinh(k d_{\text{F}})\qty[\pm\frac{\bar{F}(k)}{D_{2}(k)}-\frac{f(k)}{D_{1}(k)}]
\end{aligned}$$
We continue to consider a compensated magnetization where so that .\
Once again, we find the typical asymmetry associated with magnetic textures. The plots for the two magnetization NDW$_{y}^{(x)}$ and NDW$_{z}^{(x)}$ can be found in and , respectively.
Nucleation of superconductivity\[Section4\]
===========================================
In this section, we analyze the nucleation of superconductivity in the S when the temperature drops below the critical value in a bulk superconductor $T_{cB}$. Naturally, the value of the critical current $T_{c}$ in the presence of a local depairing factor $V(\vb{r}_{\perp})$ varies from its bulk value. This was already analyzed in the case of ferromagnetic superconductors with DW’s in Ref.[@Yu_2003] and in Ref.[@Aladyshkin_2006] for the case of S/F bilayer. In the latter case the authors studied the dependence of $T_{c}$ on an external field $H_{ext}$ in the zero DW width approximation.
Near the critical temperature $T_{c}$ the superconducting order parameter $f$ is small and obeys the Ginzburg-Landau equations (see, for example, Ref. [@Aladyshkin_2006]) $$-d_{\text{S}}\nabla_{\vb*{r}_{\perp}}^{2}f+V(\vb*{r}_{\perp})f-Ef=-Ef^{3}\label{Eq: GL}$$ where . The “energy” $E$ is related to the coherence length , . The vector potential $A_{0}$ defines the stray-field in absence of superconductivity, which can be extracted from Eq.(\[Eq: Vectorpotential S\]) by taking the limit $\lambda_{\text{L}}\rightarrow\infty$. $$\vb{A}_{\text{0}}(\vb*{k},z)=\frac{4\pi M_{0}}{\abs{k}}\sinh(\abs{k}d_{\text{F}})\qty(\hat{\vb{e}}_{z}\times\frac{i\vb*{k}}{\abs{k}})\qty[\pm\frac{i\vb*{k}}{\abs{k}}\vb{n}_{\perp}(\vb*{k})-\mathrm{n}_{z}(\vb*{k})]e^{-\abs{k}\abs{z}}$$ For simplicity, we assume that the thickness of the S films $d_{\text{S}}$ is smaller than $\xi_{\text{S}}$, so that the order parameter (OP) $f$ depends only on the in-plane coordinates. At larger $d_{\text{S}}$, the factor $V(\vb*{r}_{\perp},z)$ depends on the coordinate $z$ and the effect of this depairing factor on the nucleation of superconductivity becomes weaker. Eq.(\[Eq: GL\]) is called the time-independent Gross-Pitaevskii [@Gross_1961; @pitaevskii_1961] equation or nonlinear Schrödinger equation. This equation is also used to analyze the nucleation of superconductivity near the critical magnetic field $H_{c2}$ (see Abrikosov’s book [@abrikosov_fundamentals_1988] and [@abrikosov_magnetic_1957]) and also near DW in a S/F system [@Aladyshkin_2006]. Note also Ref.[@Moor_2014], where this equation is applied for studying the appearance of an OP in a system with two competing OPs.
In the following, we will focus on DW structures where in Eq.(\[Eq: GL\]) . In this case the real space expression for the vector potential at the interface is given by $$\vb{A}_{0}(x,\pm d_{\text{F}})=-4\pi M_{0}\sum_{k}\frac{\sinh(\abs{k}d_{\text{F}})}{\abs{k}}e^{ikx}\qty[\pm \mathrm{n}_{x}(k)+\frac{ik}{\abs{k}}\mathrm{n}_{z}(k)]\vb{\hat{e}}_{y}$$
The associated depairing potential is shown in for the different magnetization configurations. The potential $V(x)$ has minima located either at the DW’s or in the center of the domains . The critical temperature $T_{c}$ is determined by the condition , where $E_{min}$ is the minimal “energy” at which Eq.(\[Eq: GL\]) has a non-trivial solution. We assume that the domain size $2L$ is much larger than the width of the DW $2w$. In the following we consider two possible cases.
{width="\linewidth"}
### $V(x)$ has a minimum at the DW
Consider first the case when the potential $V(x)$ has a sharp minimum at the DW, for example, at where . Since we are interested in a qualitative picture, we approximate the dependence $V(x)$ near the DW with a rectangular potential well: . Then, neglecting the r.h.s. in Eq.(\[Eq: GL\]) and using the matching conditions at ($f(x)$ and $\partial_{x}f(x)$ are continuous) we can write a solution in the form $$f(x)=C_{in}\cos(K_{\text{S}}\tilde{x}),\qquad \abs{\tilde{x}}<w$$ $$f(x)=C_{out}\begin{cases}
\exp(-K_{out}(\tilde{x}-w)), \qquad &\tilde{x}> w\\
\exp(-K_{out}(\tilde{x}+w)), \qquad &\tilde{x}< -w
\end{cases}$$ where , and . The integration constants $C_{in}$ and $C_{out}$ are related to each other. In the limiting cases of small and large , we have for $T_{c}$ and $C_{in}$, $C_{out}$
a\) ;\
b) .\
Thus, if the depairing potential $V(x)$ has a dip at the DW’s, superconductivity is nucleated at the DW’s. This happens in the following DW configurations: BDW$^{(y)}$, NDW$_{y}^{(x)}$, NDW$_{z}^{(x)}$ (the potentials are shown in ). The opposite case is realized for the magnetization profiles: NDW$^{(z)}$, BDW$^{(z)}$, NDW$^{(y)}$ (see also for the respecting potentials) and is considered in the next section.
For simplicity, we neglect the width $w$ in comparison with $L$. The solution outside the DW has the form $$f(x)=C_{out}\cos(K_{\text{S}x}), \qquad \abs{x}<L$$ The critical temperature $T_{c}$ is found from the matching condition . The constant is not zero provided that the condition $$\theta_{c}\tan(\theta_{c})=\lambda$$ is fulfilled, where . Eq.(\[Eq: GL\]) yields for the critical temperature $T_{c}$:\
a)\
b)\
The constant $C$ is found analogously to the case considered in Ref.[@Moor_2014] $$C^{2}=\frac{T-T_{c}}{T_{cB}-T_{c}}r(\theta_{L})$$ with $$r(\theta_{L})=\frac{\expval{\cos^{2}(K_{\text{S}}x)}}{\expval{\cos^{4}(K_{\text{S}}x)}}=\frac{2\theta_{L}+\sin(2\theta_{L})}{(3/2)\theta_{L}+\sin(2\theta_{L})+(1/8)\sin(4\theta_{L})}$$ where and . For the coefficient $r$ is equal to: .
### $V(x)$ has a mimimum at the center of the domain
We assume that the potential $V(x)$ in Eq.(\[Eq: GL\]) has the form $$V(x)=V_{0}\sum_{n}\delta(x-2nL)$$ First, we linearize Eq.(\[Eq: GL\]) and find the minimum “energy” $E$ of the Schrödinger like equation in the interval $$-\partial_{\tilde{x}\tilde{x}}^{2}f+V_{0}\delta(\tilde{x})f(0)=\tilde{E}f$$ where $\tilde{x}=x/\xi_{\text{S}0}$, $\tilde{E}=E(\xi_{\text{S}0}/d_{\text{S}})=1-T/T_{cB}$. A periodic solution ($f(\tilde{x}=f(\tilde{x}+2L/\xi_{\text{S}0}))$) can be represented in the form $$f(\tilde{x})=\begin{cases}
a\cos(q\tilde{x})+b\sin(q\tilde{x}),\qquad &0<\tilde{x}<L/\xi_{\text{S}0}\\
\bar{a}\cos(q\tilde{x})+\bar{b}\sin(q\tilde{x}),\qquad &-L/\xi_{\text{S}0}<\tilde{x}<0 \label{Eq: f center domain}
\end{cases}$$ where . The function $f(\tilde{x})$ has to fulfill the matching conditions $$\begin{aligned}
[f]= 0&& [\partial_{\tilde{x}}f]=V(\tilde{x})f(0)\notag\\
f(L)=f(-L) && \partial_{\tilde{x}}f(\tilde{x})|x=L= \partial_{\tilde{x}}f(\tilde{x})|x=-L
\end{aligned}$$ The solution (\[Eq: f center domain\]) exists if the condition $$\theta\tan(\theta)=v \equiv VL/2 \label{Cond: Center domain}$$ is satisfied where . The coefficients $a$ and b are coupled by the relations: , . From Eq.(\[Cond: Center domain\]) we find the critical temperature $$T_{c}/T_{cB}=\begin{cases}
1-(v\xi_{\text{S}0}/L)^{2},\qquad &v\ll 1\\
1-(\pi\xi_{\text{S}0}/2L)^{2},\qquad &v\gg 1
\end{cases}$$ If , superconductivity is suppressed completely.
Conclusion\[Section5\]
======================
To conclude, in this manuscript we calculated the magnetic stray-field $\vb{H}_{str}$ and Meissner current $\vb{j}_{\text{S}}$ in a superconductors S created by various non-homogeneous magnetic texture in a F film incorporated in an S/F/S system. Two types of topological structures were considered: isolated chiral magnetic skyrmions and periodic flat domain walls of Bloch (BDW) or (NDW). Considering a two-dimensional two-component magnetization $\vb{M}(\vb*{r}_{\perp})$, we investigated six different magnetic DW textures as well as magnetic Sk of Bloch and . Each of these different magnetic textures possesses a particular spatial dependence of the stray-field $\vb{H}_{\text{str}}(\vb*{r}_{\perp},z)$ and the induced Meissner current $\vb{j}_{\text{S}}(\vb*{r}_{\perp},z)$. The most apparent difference appears between the Bloch– and the magnetic structures. While the Neel-type structure yields a strong asymmetry $j_{\text{S}}(\vb*{r}_{\perp},z)\neq j_{\text{S}}(\vb*{r}_{\perp},-z)$, the remains always symmetric w.r.t $z$-component. For certain parameter, this asymmetry can be strong enough to cause a sign change of the Meissner current for $\vb*{r}_{\perp}$ within the DW region or within the Sk radius $\vb*{r}_{\perp}$. Note that a similar sign change can be obtained in S/F or S/F/S systems that feature a proximity effect[@bergeret_spin_2004; @volkov_spin_2019; @mironov_electromagnetic_2018].
The Meissner current $j_{\text{S}}$ is connected to the vector potential $A$ via which enters the Ginzburg-Landau equation and acts as a depairing factor where $A_{0}$ is the vector potential in absence of superconductivity. This factor determines the critical temperature of the superconducting transition in bulk superconductors[@abrikosov_fundamentals_1988] and in S/F heterostructures[@Yu_2003; @Burmistrov2005; @Aladyshkin_2006] and the superconductivity emerges first at places where $A_{0}$ has a minimum. As can be seen in the locations of the minima or maxima of depends on the type of DW’s. For magnetic skyrmions the depairing potential $V(\vb*{r}_{\perp})$ has its minimum in the center of the Sk. However, it can exhibit an additional local minimum for finite $\vb*{r}_{\perp}$ within the radius of a Néel Sk which is not present for skyrmions. Thus, by measuring the location of the superconducting nucleation like it was done previously [@Iavarone2014], one can determine the type of the DW or distinguish between Bloch– and skyrmions.
Acknowledgement
===============
The authors acknowledge support from the Deutsche Forschungsgemeinschaft Priority Program SPP2137, Skyrmionics, under Grant No. ER 463/10.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Increasingly large populations of disk galaxies are now being observed at increasingly high redshifts, providing new constraints on our knowledge of how such galaxies evolve. Are these observations consistent with a cosmology in which structures form hierarchically? To probe this question, we employ SPH/N-body galaxy scale simulations of late-type galaxies. We examine the evolution of these simulated disk galaxies from redshift 1 to 0, looking at the mass-size and luminosity-size relations, and the thickness parameter, defined as the ratio of scale height to scale length. The structural parameters of our simulated disks settle down quickly, and after redshift $z=1$ the galaxies evolve to become only slightly flatter. Our present day simulated galaxies are larger, more massive, less bright, and redder than at $z=1$. The inside-out nature of the growth of our simulated galaxies reduces, and perhaps eliminates, expectations of evolution in the size-mass relation.'
author:
- 'Chris B. Brook, Daisuke Kawata, Hugo Martel, Brad K. Gibson , Jeremy Bailin'
title: 'Disk evolution since $z\sim 1$ in a CDM Universe.'
---
INTRODUCTION
============
Galaxies are major structural components of our Universe, yet their formation and evolution remain an outstanding mystery of contemporary astrophysics. The latest generation of telescopes, including [*Keck*]{}, [*Subaru*]{}, the [*Very Large Telescope*]{}, [*Gemini*]{}, and the [*Hubble Space Telescope*]{}, provide snapshots of galaxies at various redshifts. The properties of disk galaxies out to $z=1$ have been the subject of several recent studies. These properties include the Tully-Fisher relation (e.g. @vogt96 [@bm02; @b04]), the luminosity-size, magnitude-size, and mass-size relations (e.g. @lilly98 [@simard99; @bs02; @b05]), the size distribution (e.g. @shen03 [@r04; @F04]) and the disk “thickness" (@rdc03, hereafter RDC, @elmegreen).
The evolution of the luminosity-size and mass-size relations of disk galaxies remains a controversial issue. In particular, the interpretation of such evolution depends acutely on the selection biases of the surveys. Several studies have found an increase in the $B$-band surface brightness of $\sim 1$ mag arcsec$^{-2}$, out to redshift 1 (e.g. @schade [@roche98; @lilly98]). Taking into account selection effects [@simard99] found minimal evolution with redshift, work that was supported by [@r04], who found that the luminosity-size relation evolves by less than 0.4 mag arcsec$^{-1}$ between $z=1.25$ and $z=0$. Yet these groups invoke a population of high redshift, high surface brightness galaxies in their interpretation. Different interpretations of completeness have resulted in [@trujillo04] and [@b05] finding an evolution in the rest frame $V$-band surface brightness of $\sim 0.8$ and $\sim 1.0$ mag arcsec$^{-2}$ out to $z\sim 0.7$ and $z\sim 1.0$, respectively. [@b05] also find that disks at a given absolute magnitude are bluer, with lower stellar mass to light ratios at $z\sim1$, resulting in weak evolution in the relation between stellar mass and disk size.
RDC selected 34 edge-on disk galaxies in the H[*ubble Deep Field*]{} with apparent diameters larger than 1.3 and unperturbed morphology. They found an evolution between distant and local disk galaxies in their relative thickness or flatness, as characterized by the ratio of scale height $h_z$ to scale length $h_l$. Their results indicate that disks at redshift $z\sim 1$ are smaller, in absolute value, than present day disk galaxies, and have a thickness ratio, $h_{\rm z}/h_{\rm l}$ larger by a factor $\sim 1.5$. support for such a thickening of disk galaxies at high redshifts also comes from [@elmegreen], who inferred from their study of galaxies in the Hubble Ultra-Deep-Field that high redshift disk galaxies are thicker than local spirals by a factor of $\sim 2$.
Studies such as these are clearly increasing our knowledge of the evolutionary stages of populations of disk galaxies, yet the path taken for galaxies to evolve between these stages remains unclear, and reliant on modeling. The general scenario of galaxy formation within the cosmological context, as envisioned by [@wr78], involves extended dark halos of galaxies forming hierarchically by gravitational clustering of non-dissipative dark matter. The luminous components form by a combination of the gravitational clustering and dissipative collapse. A general framework for the formation of disk galaxies was outlined in [@fe80]. In their model, a uniformly rotating, homogeneous protogalactic cloud of gas begins to collapse as it decouples from the Hubble flow. This protogalaxy is endowed with angular momentum from tidal torques driven by surrounding structure (@p69). It is assumed that the baryons destined to form the disk receive the same tidal torques as the dark matter before much dissipation occurs. If the collapse is smooth, then specific angular momentum is conserved and the gas forms a thin, rapidly rotating disk (e.g. @dss97 [@mmw98]).
Detailed direct modeling has highlighted that the above disk formation scenario is far from simple within a Cold Dark Matter (CDM) Universe . Complex baryonic processes involving multiple gas phases and energy feedback, for example from quasars, supernovae, and stellar winds, have proven difficult to model, yet these processes appear to be crucial in the formation and evolution of disk galaxies within the context of hierarchical structure assembly. Early simulations employed thermal feedback, following [@kg91], augmented somewhat in [@nw93], who added a variable fraction of supernova energy as kinetic energy. This type of feedback has been popular up until recent times (e.g. @sn02 [@abadi03]), even though it was known to be highly inefficient, with thermal energy radiating away over timescales shorter than the dynamical timescales [@katz92]. The effects of this inefficient feedback, in light of the the overcooling problem (@wr78), was the rapid formation of stars in the early collapsing halos. Subsequent mergers of these halos resulted in loss of angular momentum to the dark halo, resulting in disk galaxies being dominated by stellar halo components, and significantly deficient in angular momentum. [@weil] showed that the suppression of early radiative cooling in numerical simulations can result in more realistically sized disks. The problem was tackled in [@tc00] by turning off cooling for a fixed time in gas within the SPH smoothing kernel of a supernova event. Although lacking detailed insights into the multiphase gas processes involved in galaxy formation, these models succeeded in regulating star formation in small halos, and resulted in increased angular momentum in the simulated disk galaxy. In [@brooketal04a], we simulated galaxies [*[using identical initial conditions]{}*]{} with these two different feedback methods. The thermal feedback following [@kg91] resulted in galaxies which are dominated by a high mass, high metallicity spheroidal stellar component, with a low mass disk component. By regulating star formation in low mass building blocks, the adiabatic feedback model, which follows [@tc00], formed galaxies with a dominant disk component, and a low mass, low metallicity stellar halo. Sommer-Larsen and collaborators have examined other implementations of feedback, and also shown the importance of regulating cooling at early epochs (@sl99 [@sl01; @sl03]). Meanwhile, [@governato04] highlights the importance of resolution in simulating disk galaxies. Implementing these findings has allowed models such as [@robertsonetal04], [@okamoto05] and [@governato05] to make progress in reproducing disk galaxies with properties approaching those of observed disks.
Our chemo-dynamical models for simulating disk galaxies (which we will refer to as sGALS, to avoid confusion with real galaxies) have been successful in forming disk galaxies with a dominant young, metal-rich disk component, with a metal poor, old stellar halo, and an old, intermediate metallicity thick disk component. The sGALS have many features observed in the Milky Way and local disk galaxies, including metallicities, abundances, ages, and color gradients of the stellar populations of the thin and thick disks, metallicities and abundances of the stellar halo, and the relation between luminosity and metallicity of the stellar halo (@brooketal04a; @brooketal04b \[hereafter BKGF\]; [@brooketal05] \[hereafter BGMK\]; @rendaetal05). Here we report on the evolution of the disk component of such sGALS, specifically after redshift $\sim 1$. We deliberately follow the evolution of our sGALS at times after which they have settled into a disk morphology, and trace their luminosity, scale length, scale height, and central surface brightness. We hope to provide insight into how the disk component of late-type galaxies evolve in their relatively “quiescent phase,” during which mergers are only minor and star formation occurs almost exclusively in the disk regions.
In this paper, we concentrate on the evolution of the size-mass and size-central surface brightness relations and the “thickness” of four simulated disk galaxies (sGALS), and what these tell us about interpreting observations. We also examine the evolution of the magnitude and color of our simulated galaxies. Our study will not allow direct comparison with the observational studies which compare distant galaxy populations to “like” local galaxy populations; it is not clear which galaxies are likely progenitors of today’s galaxies. Yet, by providing evolutionary paths that disk galaxies can be expected to follow, our study will provide insights into the interpretation of such observations, in particular in how to interpolate between the two populations.
THE CODE AND MODELS
===================
The four sGALS studied here are the same as were studied in BGMK. They are formed using our galactic chemo-dynamical evolution code, [GCD+]{}, which self-consistently models the effects of gravity, gas dynamics, radiative cooling, and star formation. Full details of [GCD+]{} can be found in Kawata & Gibson (2003). All galaxies are formed with the [*Adiabatic Feedback Model*]{} from [@brooketal04a]. In this model, gas within the SPH smoothing kernel of SNe II explosions is prevented from cooling. This adiabatic phase is assumed to last for the lifetime of the lowest mass star which ends as a SNe II, i.e. the lifetime of an 8-M star (100 Myr). In the AFM, the energy released by SNe Ia, which do not trace starburst regions, is fed back as thermal energy.
[cccc]{} sGAL1 & 1.8 & 5$\times$10$^{11}$ & 0.0675\
sGAL2 & 1.9 & 5$\times$10$^{11}$ & 0.0675\
sGAL3 & 2.0 & 1$\times$10$^{12}$ & 0.0600\
sGAL4 & 2.2 & 5$\times$10$^{11}$ & 0.0675\
We employ a semi-cosmological version of [GCD+]{}, where the initial conditions consist of an isolated sphere of dark matter and gas [@kg91]. This top-hat overdensity has an amplitude, $\delta_i$, at initial redshift, $z_i$, which is approximately related to the collapse redshift, $z_c$, by $z_c=0.36\delta_i(1+z_i)-1$ [e.g. @pad93]. Small scale density fluctuations are superimposed on each sphere, parameterized by $\sigma_8$, seeding local collapse and subsequent star formation. Longer wavelength fluctuations are incorporated by imparting a solid-body rotation corresponding to a spin parameter, $\lambda=J|E|^{1/2}/GM_{\rm tot}^{5/2}$, to the initial sphere, where $J$ is the total angular momentum, $E$ the total energy, and $M_{\rm tot}$ is the total mass of the sphere. This simplified model allows us to run a suite of simulations at high resolution, whilst retaining the most important features of full cosmological simulations. The choice of large values of $\lambda$, and initial conditions in which no major merger occurs in late epochs ($z<1$), ensures that disk sGALS are formed. We use an Einstein-deSitter CDM model; parameters relevant to all sGALS include $\Omega_0=1$, baryon fraction, $\Omega_b=0.1$, $H_0=50\rm km\,s^{-1}\,Mpc^{-1}$, star formation efficiency, $c_*=0.05$, and $\sigma_8=0.5$. This choice was made merely for convenience. As we showed in BGMK, the consequences of using this cosmology with the initial conditions of this study, rather than the now standard $\Lambda$CDM cosmology, are negligible. Essentially, prior to the redshift of collapse $z_c\sim2$, a $\Lambda$CDM universe is almost identical to a Einstein-de Sitter universe, while after that redshift, the system is so dense that the background cosmology becomes irrelevant. For a detailed proof, we refer the reader to BGMK. Other parameters varied slightly between the four models, but all were chosen to result in Milky-Way-type sGALS. Table \[initial\], reproduced from BGMK, shows the values of total mass, $\lambda$, and $z_c$ for the four simulations. Along with these variations in parameters, different random seeds are incorporated in the initial conditions of the sGALS, creating evolutionary diversity in our sample. We employed 38911 dark matter and 38911 gas/star particles for each model.
RESULTS
=======
Our model results in the formation of late-type sGALS which have a dominant disk component, with a low mass, metal poor halo component (BKGF). Figure \[SGAL1\] shows the rest frame $B$-band surface brightness profile of sGAL1, both edge-on (upper panels) and face-on (lower panels), at three epochs: $z=0.7$ (corresponding to a lookback time of 7.3 Gyrs) $z=0.5$ (5.7 Gyrs), and $z=0$. Figures \[SGAL2\]-\[SGAL4\] show corresponding plots for sGALS 2-4, shown at $z=1$ (corresponding to lookback time of 8.5 Gyrs) $z=0.5$ ($z=0.4$ for sGAL2, lookback time 4.8 Gyrs), and $z=0$. These epochs are chosen to be: (a) after the thick disk formation epoch, when the old thin disk has begun to form; (b) at an intermediate time; and (c) the present, respectively. The different time chosen for the early epoch of sGAL1 ($z=0.7$) was taken as that simulated galaxy is undergoing a significant merger at $z=1$, and we need our sGALS to have a settled disk structure for a fair comparison with distant galaxies which have been selected for their disk features. Similarly, a satellite is being accreted to sGAL2 at the intermediate epoch, $z\sim$0.5, so we chose a later intermediate time, $z=0.4$ for this sGAL. The surface brightness is calculated by assuming that each star particle is a single stellar population, and using the SSPs of [@ka97], with a simple Salpeter initial mass function (IMF) parameterized by a power-law slope of 1.35 and lower and upper mass of 0.1 and $60{\rm M}_\odot$, respectively. The photometric properties of sGALS discussed here are all intrinsic values, where we avoid discussing the effect of the dust absorption.
Total $B$-band magnitudes of the sGALS are found by integrating the luminosity densities displayed in these images. The results are shown in column 3 of Table \[results\]. We see that the galaxies have an average $M_B\sim -21.1$ at the earliest epoch (which we will refer to as $z\sim 1$), an average of $-20.6$ at the intermediate epoch ($z\sim0.5$) and $-19.7$ at $z=0$. The higher luminosity at earlier times is largely due to the enhanced star formation at the epoch during which the thick disk forms (see BKGF), the epoch which precedes the “early epoch” ($z\sim1$) analyzed in this study. The rate of star formation is lower during the more quiescent later epochs, resulting in evolution toward fainter magnitudes at the present.
We repeat the above process in both the rest frame $V$ and $I$ bands, and derive $B-I$ and $V-I$ colors. The $B-I$ and $V-I$ colors are shown in columns 5 and 6 of Table \[results\], respectively. The galaxies become redder during the relatively quiescent late epochs studied, from an average $B-I$ ($V-I$) of 1.1 (0.7) at $z$$\sim$1, to 1.5 (0.9) at $z\sim0.5$, and 1.7 (1.0) at $z=0$. These redder colors are due to the effects of a decreasing star formation rate, aging stellar populations, and more metal rich stellar populations being born in the disk region. The redder colors are not sufficient to overcome the decrease in $B$-band luminosity; the sGALS also become fainter in the redder bands.
The $B$-band surface brightness profiles for the four sGALS are plotted against radius in Figure \[scalelengths\], at the three epochs shown in Figures \[SGAL1\]-\[SGAL4\]. We use open stars for the early phase of evolution ($z\sim 1$), open squares for the intermediate phase ($z\sim0.5$), and solid triangles for $z=0$ in this and all ensuing plots in order to allow evolution of the properties of the disk sGALS to be easily followed. Outside the central bulge regions, the disk components can be approximated reasonably by exponential fits. This allows us to derive disk scale lengths for each sGAL at each epoch. This is done in two stages. An approximate scale length $h_l^*$ is derived by fitting between 4 and 20 kpc. This allows us to then make fits between $1.5h_l^*$ (to exclude the bulge) and $4h_l^*$. Numerous tests verified that this technique gave robust results, which are shown column 9 in Table \[results\]. Scale lengths increase from an average of 2.9 at $z\sim1$ to an average of 3.5 at $z\sim0.5$, and to 4.1 by $z=0$. The implication is that the galaxies grow from the “inside-out.”
We plot the surface brightness profiles of the four sGALS perpendicular to the disk in Figure \[scaleheights\]. We exclude the bulge region when calculating how the surface brightness varies with distance from the disk plane. Again, the profiles are well fitted with exponentials, and we fit between 0 and 2.5 scale heights (essentially out to 1.5 kpc in Figure \[scaleheights\]) using similar techniques as employed to derive scale lengths. The derived scale heights are shown in column 8 of Table \[results\]. Scale heights remain close to constant, with an average of $0.63$ at $z\sim 1$ to an average of $0.62$ at $z\sim 0.5$, and an average of $0.65$ at $z=0$. We derive the thickness of the disks, defined as the ratio of scale height to scale length, $h_z/h_l$, and show the results column 10 in Table \[results\]. On average, the galaxies are “thicker” at the earlier epochs, by a factor of $1.4$ at $z\sim 1$. This is due to the no evolution of the scale hight and the increasing scale length with decreasing redshift. Our simulation does show little evolution of the scale height, because the thick disk forms first, followed by the thin disk formation, as discussed in BKGF. If the thick disk comes from heating of the thin disk, we should expect an increasing scale height with decreasing redshift, and a slower evolution of the disk thickness. Further observations of the thickness evolution will help constrain thick disk formation scenarios. The evolution of the thickness in our sGALS (in the sense that the galaxies become thinner) is driven by the inside-out thin disk formation. At the intermediate epoch, the sGALS show a marginal increase in thickness, by a factor of $1.1$ at $z\sim 0.5$; galaxies settle rapidly to disk morphologies after their last “significant” merger (i.e. a merger likely to alter the morphology of the galaxy).
Numerical heating is an issue when considering vertical structure in numerical simulations, but we have reason to believe that our results are not affected significantly. Previous studies at similar resolution to ours have shown numerical vertical heating to be relatively low (@font [@mayer]). Figure 2 of BGMK shows the vertical velocity dispersion versus age for four sGALS in this study, which demonstrates relatively little vertical heating during the disks’ quiescent period; in the order of 20% over the last $\sim 8$ Gyrs. Even if this were totally attributed to numerical heating (unlikely as processes such as spiral structure and substructure heating will contribute), our results will not be greatly affected, because the vertical scale height is proportional to the square root of the velocity dispersion, i.e $h_z\propto \sqrt{\sigma_z}$ (e.g. @bottema). Our measured scale height at $z=0$ will be, in the worst case, overestimated by $\sim$10%. As far as relevance to our study, the important point that we note is that any artificial heating will not affect our results [*[qualitatively]{}*]{}. In the absence of such effects, our conclusions that our simulated disk galaxies become [*[flatter]{}*]{} ($h_z/h_l$ decreases) will hold. Our [[*quantitative*]{}]{} estimates would then be underestimating this effect. Having said this, higher resolution studies will increase the reliability of the [*[quantitative]{}*]{} estimates
The way that the structural parameters, scale length and scale height, evolve relative to one another can be ascertained from Figure \[hzhl\]. Also plotted as crosses are observations of local disk galaxies from Schwarzkopf & Dettmar (2000). Plotted as open circles are observations of 34 edge-on disk galaxies at $z$$\sim$1, taken from RDC. The $z\sim1$ disk galaxies of RDC are found to be smaller and thicker than local galaxies, as determined by previous studies, such as @sd00, @dG98, @K02 and @bm02. Our sGALS evolve in this plot primarily by increasing in scale length, with less evolution in scale height, resulting in galaxies evolving across this diagram, from left to right.
Central surface brightness, again in the rest frame in the $B$-band, is derived by extrapolating the exponential fits from Figure \[scalelengths\] to $r=0$. Results are shown in column 7 of Table \[results\]. We find that our sGALS are brighter at earlier epochs, with an average $\mu_B=21.0$ at $z\sim 1$, compared with $\mu_B=21.8$ at $z\sim 0.5$, and $\mu_B=22.7$ at $z=0$. This is related to the higher star formation rates at the earlier epochs in our sGALS, as shown in Figure 5 of BGMK. This is consistent with observations, which point to star formation rates being higher at earlier epochs (e.g. @madau [@hammer; @flores; @juneau]). We plot the $B$-band central surface brightness against the flatness parameter, $h_l/h_z$ (the inverse of the thickness parameter), in Figure \[muBhlhz\]. The sGALS have higher surface brightness and are thicker at earlier times, closer to the merging epochs when the thick disks formed. As the thin disk grows larger in the more quiescent later stages of their evolution, the sGALS become less bright, and also thinner. At each epoch, it is also apparent that brighter sGALS are thicker, although this trend is not as strong as the trend with time, because the sGALS we have studied are all relatively quiescent throughout the epoch during which evolution is followed in this study. This supports the conclusions of RDC, who plotted the relation between $h_l/h_z$ and central surface brightness of $z\sim 1$ galaxies (their Fig .7), and found that the brightest objects have the thickest disk. They speculated that enhanced star formation and enhanced thickness are the results of ongoing interactions and mergings.
We derive the stellar mass $M_*$ of each sGAL at each epoch using the friends-of-friends algorithm, and confirm that sensitivity to our choice of linking length is insignificant. Results are displayed in column 4 of Table \[results\]. The inside-out nature of the growth of the thin disk in our sGALS is apparent in the relationship between the stellar mass and scale length, shown in Figure \[Masshl\]. The mass and size of the sGALS grow together. Although we do not want to overemphasize the quantitative nature of our results, using just four relatively homogeneous L$_*$ sGALS, we note that the ratio $M_*/h_l$ does not evolve, with the average at each epoch staying remarkably close to $10^{10}{\rm M}_{\Sun}\,\mathrm{kpc}^{-1}$. Qualitatively, it is clear that the thin disks of our sGALS increase in size as they grow in mass, and the disk evolves along the mass scale length relation.
[cccccccccc]{} $z=0$ & sGAL1 & -19.5 & 2.9$\times$10$^{10}$ &1.7 & 1.0 & 23.3 & 0.70 & 4.5 & 0.16\
$z=0$ & sGAL2 & -19.6 & 3.0$\times$10$^{10}$ &1.7 & 1.0 & 22.1 & 0.59 & 3.7 & 0.16\
$z=0$ & sGAL3 & -20.2 & 6.6$\times$10$^{10}$&1.8 & 1.1 & 21.4 & 0.60 & 3.5 & 0.17\
$z=0$ & sGAL4 & -19.6 & 3.6$\times$10$^{10}$&1.7 & 1.0 & 23.9 & 0.70 & 4.6 & 0.15\
$z=0$ & mean & -19.7 & 4.0 $\times$10$^{10}$&1.7 & 1.0 & 22.7 & 0.65 & 4.1 & 0.16\
$z=0.5$ & sGAL1 & -20.3 & 2.4$\times$10$^{10}$&1.5 & 0.9 & 22.2 & 0.64 & 3.8 & 0.17\
$z=0.4$ & sGAL2 & -20.6 &2.4 $\times$10$^{10}$&1.3 & 0.8 & 21.0 & 0.60 & 2.7 & 0.22\
$z=0.5$ & sGAL3 & -21.0 & 5.9$\times$10$^{10}$&1.5 & 0.9 & 20.6 & 0.56 & 3.2 & 0.18\
$z=0.5$ & sGAL4 & -20.3 & 3.2$\times$10$^{10}$&1.5 & 0.9 & 23.2 & 0.67 & 4.2 & 0.16\
$z\sim 0.5$ & mean & -20.6 & 3.5$\times$10$^{10}$&1.5 & 0.9 & 21.8 & 0.62 & 3.5 & 0.18\
$z=0.7$ & sGAL1 & -20.9 &2.1$\times$10$^{10}$& 1.1 & 0.7 & 21.8 & 0.69 & 3.2 & 0.22\
$z=1$ & sGAL2 & -20.9 & 1.4$\times$10$^{10}$&0.9 & 0.6 & 20.7 & 0.51 & 2.2 & 0.23\
$z=1$ & sGAL3 & -21.6 & 5.0$\times$10$^{10}$&1.3 & 0.8 & 20.2 & 0.75 & 3.0 & 0.25\
$z=1$ & sGAL4 & -21.1 & 2.7$\times$10$^{10}$&1.2 & 0.8 & 21.6 & 0.58 & 3.1 & 0.22\
$z\sim0.9$ & mean & -21.1 &2.8$\times$10$^{10}$ &1.1 & 0.7 & 21.0 & 0.63 & 2.9 & 0.23\
\[results\]
SUMMARY AND CONCLUSION
======================
The evolution of individual galaxies cannot be observed directly, but lookback studies are providing increasing information on properties of populations of “like” galaxies, out to redshift 1 and beyond. These studies remain difficult, in large part because it is unclear how to select which galaxies should be compared at each epoch. In addition, there are large uncertainties involved in correcting for dust and cosmological dimming, especially in the case of edge-on disk galaxies.
Several recent studies have compared the observed evolution of disk galaxy properties with semi-analytic and infall type models. For example, [@b05] and [@trujillo05] show that one of the most basic assumptions of the semi-analytic approach, where disk size scales in proportion to the virial radius of the dark halo [@mmw98], is at odds with observations; the observed size evolution is weaker than the predicted $R\propto H(z)^{-2/3}$, where $H(z)$ is the Hubble parameter as a function of redshift. The infall model of [@bs02] assumes a star formation rate at each radius and time, according to the local density (@Schmidt), and the infall of metal-free gas. This model is able to match observations of the size evolution, at given luminosity, out to $z\sim1.7$ (@trujillo05). In order to explain the evolution of disk galaxies, the necessity of accounting for effects such as mass build up, star formation, energy feedback, and gas infall self-consistently within a CDM hierarchical framework, motivates us to confront observations with direct numerical simulations.
We examine four simulated disk galaxies (sGALS), and do so in an idealized manner, free from the difficulties posed in observation. Our sGALS form thick disks at high redshift ($z\sim 1.5-2$) during a period of mergers of gas rich “building blocks” (BKGF), and rapidly settle into a disk galaxy morphology by $z\sim1$. We examine their subsequent evolution to the present, comparing various properties of the galaxies at high ($\sim1$), intermediate ($\sim0.5$), and zero redshift, where we find clear trends. Our main results are:
- Scale heights of disk galaxies show no evolution after $z\sim 1$. Scale lengths increase slowly, resulting in flatter disks at later times. This is consistent with the observations of RDC, who find that disk galaxies at $z\sim 1$ are smaller than present-day galaxies, and have a thickness ratio, $h_z/h_l$ larger by a factor of $\sim 1.5$.
- The “inside-out” growth of disks is reflected in the scale length increasing roughly proportional to the stellar mass, resulting in little to no evolution of the mass-size relation. This is consistent with interpretations of [@trujillo04; @b05; @trujillo05]. Previously, few direct numerical studies have probed these issues. In a study of sGALS of various morphologies in a fully self-consistent $\Lambda$CDM simulation, [@sl03] examined two disk galaxies, and found that while one clearly evolved inside-out, the other evolves outside-in. [@Samland] examine the build up of disk galaxies through a spherical gas infall model, perhaps an idealization of the quiescent period of disk galaxy formation, and finds that disks form inside-out.
- Disk sGALS become more massive, less bright, and redder between redshifts $z\sim1$ and 0. The evolution of the luminosity of disk galaxies remains observationally controversial, with selection effects largely determining interpretations (for a discussion see @b05). The mass build up in this epoch is relatively quiescent, with aging stellar populations and high metallicity stars born in the disk accounting for the galaxy reddening.
- Rapid periods of star formation due to mergers and interactions can result in both thickening of the disk, and galaxy brightening (see Figure \[muBhlhz\]). This is consistent with the results of RDC, as well as local observations which show that interacting galaxies are thicker (@rc97), and that a correlation exists between central surface brightness and “thickness,” in the sense that higher surface brightness galaxies are thicker (@b00 [@bm02; @bk04]).
The quantitative results of our study are summarized in Table \[results\]. These results are subject to uncertainties inherent in our models, such as star formation and feedback prescriptions. Also, our simulations do not cover the range of galaxy masses, sizes, morphologies, and histories of the large observed populations. We therefore feel that more modeling of a greater population of high- and low-redshift sGALS within a fully cosmological simulation is required in order to make more meaningful quantitative predictions. Having said this, clear trends have emerged in this study of the evolution of the properties of disk sGALS formed in a hierarchical, Cold Dark Matter Universe.
We thank Agostino Renda for his help. Simulations were performed on Swinburne University Supercomputer, those of the Australian Partnership for Advanced Computing, and on “Purplehaze,” the Supercomputer facility at Université Laval. CB and HM thank the Canada Research Chair program and NSERC for support. BKG, DK & JB acknowledge the financial support of the Australian Research Council through its Discovery Project program. DK acknowledges the financial support of the Japan Society for the Promotion of Science, through a Postdoctoral Fellowship for research abroad.
Abadi, M. G., Navarro, J. F., Steinmetz, M., Eke, V. R. 2003, ApJ, 591, 499
Barden M. et al. 2005, preprint (astro-ph/0502416)
Bizyaev, D. 2000, Sov. Astron. Lett, 26, 219
Bizyaev, D., & Kajsin, S. 2004, ApJ, 613, 886
Bizyaev, D., & Mitronova, S. 2002, A&A, 389, 795
Böhm, A. et al. 2004, A&A, 420, 97
Bottema, R. 1993 A&A, 275, 16
Bouwens, R., & Silk, J. 2002, ApJ, 568, 522
Brook, C. B., Gibson, B. K., Martel, H., & Kawata D. 2005, ApJ, in press (BGMK)
Brook, C. B., Kawata, D., Gibson, B. K., & Flynn C. 2004a, MNRAS, 349, 52
Brook, C. B., Kawata, D., Gibson, B. K., & Freeman K. 2004b, ApJ, 612, 894 (BKGF)
Dalcanton, J. J., Spergel, D. N., & Summers, F. J. 1997, ApJ, 482, 659
de Grijs, R. 1998, MNRAS, 299, 595
Elmegreen, D. M., Elmegreen, B. G., Rubin, D. S., & Schaffer, M. A. 2005, ApJ, 631, 85
Fall, S. M., & Efstathiou, G. 1980, MNRAS, 193, 189
Ferguson, H. C. et al. 2004, ApJ, 600, L107
Flores, H. et al. 1999, ApJ, 595, L9
Font, A. S., Navarro, J. F., Stadel, J., & Quinn, T. R. 2001, ApJ, 563L, 1
Governato, F., Stitson, G, Wadsley, J., Quinn, T. 2005, Proceedings for the Vth Marseille International Cosmology Conference: The Fabulous Destiny of Galaxies
Governato, F. et al. 2004, ApJ, 607, 688
Hammer, F. et al. 1997, ApJ, 481, 49
Juneau, S. et al. 2005, ApJ, 619, L135
Kawata, D., & Gibson, B. K. 2003, MNRAS, 340, 908
Katz, N., & Gunn, J. E. 1991, ApJ, 377, 565
Katz, N. 1992, ApJ, 391, 502
Kodama T., & Arimoto N. 1997, A&A, 320, 41
Kregel, M., van den Kruit, P. C., & de Grijs, R. 2002, MNRAS, 311, 565
Lilly, S. et al. 1998, ApJ, 500, 75
Madau, P., Ferguson, H. C., Dickinson, M. E., Giavalisco, M., Steidel, C. C., & Fletcher, A. 1996, MNRAS, 283, 1388
Mayer, L. 2004, in [*[Baryons in Dark Matter Haloes]{}*]{}, Novigrad, Croatia, 5-9 October 2004, Proc.Sci. BDMH2004 037
Mo, H. J., Mao, S., & White, S. D. M. 1998, MNRAS, 295, 319
Navarro, J. F., & White, S. D. M. 1993 MNRAS, 265, 271
Okamoto, T., Eke, V. R., Frenk, C. S., & Jenkins, A. 2005, preprint (astro-ph/0503676)
Padmanabhan, T. 1993, Structure Formation in the Universe (Cambridge: Cambridge University Press)
Peebles, P. J. E. 1969, ApJ, 155, 393
Ravindranath, S. et al 2004, ApJ, 604, L9
Renda, A., Gibson, B. K., Mouhcine, M., Ibata, R. A., Kawata, D., Flynn, C., & Brook, C. B. 2005, submitted to MNRAS
Reshetnikov, V. P., & Combes, F. 1997, A&A, 324, 80
Reshetnikov, V. P., Dettmar, R.-J. & Combes, F. 2003, A&A, 399, 879 (RDC)
Robertson, B., Yoshida, N., Springel, V., & Hernquist, L. 2004, ApJ, 606, 32
Roche, N., Ranatunga, K., Griffiths, R. E., Im, M., & Naim, A. 1998, MNRAS, 293, 157
Samland, M., & Gerhard, O. E. 2003, A&A, 399, 961
Schade, D., Lilly, S. J., Le Fèvre, O., Hammer, F., & Crampton, D. 1996, ApJ, 464, 79
Schmidt, M. 1959, ApJ, 129, 243
Schwarzkopf, U., & Dettmar, R.-J. 2000, A&AS, 144, 85
Shen, S., Mo, H. J., White, S. D. M., Blanton, M. R., Kauffmann, G., Voges, W., Brinkmann, J., & Csabai, I. 2003, MNRAS, 343, 978
Simard, L. et al. 1999, ApJ, 519, 563
Sommer-Larsen, J., & Dolgov, A. 2001, ApJ, 551, 608
Sommer-Larsen, J., Götz, M., & Portinari, L. 2003, ApJ, 596, 47
Sommer-Larsen, J., Gelato, S., & Vedel. H. 1999, ApJ, 519, 501
Steinmetz, M., & Navarro, J. F. 2002, NewA., 7, 155
Thacker, R. J., & Couchman, H. M. P. 2000, ApJ, 545, 728
Trujillo, I., & Aguerri, J. A. L. 2004, MNRAS, 355, 82
Trujillo, I. et al. 2005, submitted to ApJ (astro-ph/0504225)
Vogt, N. P., Forbes, D. A., Phillips, A. C., Gronwall, C., Faber, S. M., Illingworth, G. D., & Koo, D. C. 1996, ApJ, 604, 521
Weil, M. L., Eke, V. R., & Efstathiou, G. 1998, MNRAS, 300, 773
White, S. D. M., & Rees, M. J. 1978, MNRAS, 183, 341
=50.mm
=50.mm
=50.mm
=50.mm
=50.mm
=50.mm
=50.mm
=50.mm
=50.mm
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In contrast to discrete-variable teleportation, a quantum state is imperfectly transferred from a sender to a remote receiver in a continuous-variable setting. We recall the ingenious scheme proposed by Braunstein and Kimble for teleporting a one-mode state of the quantum radiation field. By analyzing this protocol, we have previously proven the factorization of the characteristic function of the output state. This indicates that teleportation is a noisy process that alters, to some extent, the input state. Teleportation with a two-mode Gaussian EPR state can be described in terms of the superposition of a distorting field with the input one. Here we analyze the one-mode Gaussian distorting-field state. Some of its most important properties are determined by the statistics of a positive EPR operator in the two-mode Gaussian resource state. We finally examine the fidelity of teleportation of a coherent state when using an arbitrary resource state.'
author:
- |
PAULINA MARIAN and TUDOR A. MARIAN\
Centre for Advanced Quantum Physics,\
University of Bucharest, P.O.Box MG-11,\
R-077125 Bucharest-Măgurele, Romania\
E-mail: [email protected]
title: |
CONTINUOUS-VARIABLE TELEPORTATION:\
A NEW LOOK
---
INTRODUCTION \[int\]
====================
Quantum teleportation within continuous-variable (CV) settings is based on the same ideas as in the discrete case: these were put forward in the seminal work of Bennett [*et al.*]{} [@Ben], who discovered the teleportation of qubits. The proposal of Braunstein and Kimble [@BK] was the first CV-teleportation scheme implemented experimentally. We find it useful to give here a succinct account of their protocol for teleporting a single-mode state of the quantum radiation field.
Two distant operators, Alice, at a sending station, and Bob, at a receiving terminal, share an entangled two-mode state $\rho_{AB}$. Mode $A$ is operated by Alice and mode $B$ is controlled by Bob. When the cross-correlations between modes are strong enough, Alice and Bob can exploit the non-local character of the bipartite state $\rho_{AB}$ as a quantum resource for teleporting an unknown one-mode state $\rho_{in}$. The inseparable state $\rho_{AB}$ that connects the two parties is usually called an Einstein-Podolsky-Rosen (EPR) state. Without going into details, we briefly recall the successive steps of the Braunstein-Kimble (BK) teleportation protocol.
Alice performs a von Neumann measurement of a pair of commuting continuous variables. She combines two optical operations on her modes: mode mixing and homodyne detection. More specifically, Alice mixes the input mode whose state $\rho_{in}$ is to be teleported with her $A$-mode of the EPR state $\rho_{AB}$ by employing a balanced lossless beam-splitter. As a result, she is ready to detect quadratures of the output modes. By applying convenient projectors, Alice chooses to measure simultaneously the pair of commuting quadratures $$\hat q_{A}:= \frac{1}{\sqrt2}(\hat q_{in}- \hat q_1),\;\;
\hat p_{A}:= \frac{1}{\sqrt2}(\hat p_{in} +\hat p_1). \label{1.1}$$ She conveys to Bob, through a [*classical channel*]{}, the result $\{q,p\}$ of her homodyne measurement as a complex amplitude, $\mu:=q+i p$. Any individual CV measurement performed by Alice is accompanied by a collapse of the initial tripartite state $\rho_{in} \otimes \rho_{AB}$. This results into a modified reduced $B$-mode state. Bob employs the value $\mu$ transmitted by Alice via the classical channel to perform a suitable displacement of the new reduced one-mode state at his side, $\rho_B(\mu)\rightarrow \rho^{\prime}_B(\mu)$.
The outcome $\mu$ is a continuous random variable. Therefore, Alice has to repeat her measurement under identical conditions in order to obtain a significant ensemble of results. She sends to Bob all these results, one by one, by successive classical communications. Every time, Bob operates the suitable displacement on the mode $B$ at his hand. He can thereby infer the distribution function ${\cal P}(\mu):={\cal P}(q,p)$ of the random variable $\mu.$ Bob is eventually able to build an imperfect replica of the initial state $\rho_{in}$ by averaging on the above-mentioned ensemble with the corresponding distribution function: $$\rho_{out}=\int {\rm d}^2 \mu {\cal P}(\mu)\rho^{\prime}_B(\mu).
\label{1.4}$$ Here we have denoted ${\rm d}^2 \mu:={\rm d}q \;{\rm d}p.$
Had we summarized the key steps of the BK protocol, this is explained in the framework of quantum mechanics, in Section 2, in terms of measurements and operations. Our main tool is the Weyl expansion of the density operators of the states involved. For subsequent use, we introduce a non-local positive operator that we call EPR operator. Section 3 deals with two-mode Gaussian EPR states. We first prove the existence of a one-mode [*distorting-field state*]{} that is entirely determined by the EPR state. Its properties are then carefully analyzed and we show that a [*Gaussian teleportation channel*]{} does exist. The accuracy of the CV teleportation, measured either by the amount of added noise or by the fidelity of teleporting a coherent state, is investigated in Section 4.
QUANTUM-MECHANICAL DESCRIPTION \[QM\]
======================================
We present the BK protocol in the Schrödinger picture. The initial three-mode state is the product $\rho_{in}\otimes \rho_{AB}$ of the one-mode state to be teleported, $$\begin{aligned}
\rho_{in}=\frac{1}{\pi}\int {\rm d}^2 \lambda \;\;\chi_{in} (\lambda)
D(-\lambda), \label{cf1}\end{aligned}$$ and of the two-mode EPR state, $$\begin{aligned}
\rho_{AB}=\frac{1}{\pi^2}\int {\rm d}^2 \lambda_1 {\rm d}^2
\lambda_2\;\;
\chi_{AB}(\lambda_1,\lambda_2)D_1(-\lambda_1)D_2(-\lambda_2).
\label{cf2}\end{aligned}$$ Equations (\[cf1\]) and (\[cf2\]) exhibit the Weyl expansions of the corresponding density operators. We have denoted here by $D(\alpha):=\exp{(\alpha {\hat a}^{\dag}-\alpha^*{\hat a})}$ a Weyl displacement operator on a single-mode Hilbert space: $\hat a$ is the mode annihilation operator. The states (\[cf1\]) and (\[cf2\]) are described in terms of their characteristic functions (CFs) $\chi_{in}(\lambda)$ and $\chi_{AB}(\lambda_1,\lambda_2)$, which are particular cases of the multimode definition $$\chi(\lambda_1,\lambda_2,\cdots ,\lambda_n):={\rm Tr}
[\rho D(\lambda_1)D(\lambda_2)\cdots D(\lambda_n)]. \label{cfs}$$ The compatible observables (\[1.1\]) have continuous spectra. Their common eigenfunction for an outcome $\{q,p\}$ of the homodyne measurement, $$\begin{aligned}
|\Phi(q,p)\rangle=\frac{1}{\sqrt{\pi}}
\int\limits_{-\infty}^{\infty}{\rm d}\eta{\rm e}^{i\sqrt{2} p\eta}
|\sqrt{2}q+\eta\rangle_{in}\otimes|\eta\rangle_A \label{eigen},\end{aligned}$$ satisfies the orthonormality condition $$\langle \Phi(q^{\prime},p^{\prime})|\Phi(q,p)\rangle=
\delta (q^{\prime}-q)\delta (p^{\prime}-p).$$ The distribution function of the continuous random variable $\mu$ is $${\cal P}(q,p)={\rm Tr}_{in,AB}\left[M(\mu)\right], \label{df}$$ where $M(\mu)$ is an operator on the Hilbert space ${\cal H}_{in} \otimes {\cal H}_A \otimes {\cal H}_B$: $$M(\mu):=\left[|\Phi(q,p)\rangle
\langle\Phi(q,p)|\otimes I_B\right](\rho_{in}\otimes\rho_{AB})
\label{M(mu)}.$$ As a result of the projective measurement performed by Alice, the initial product state $\rho_{in}\otimes \rho_{AB}$ collapses, so that the after-collapse $B$-mode reduction $\rho_B(\mu)$ can be written by tracing out the three-mode operator (\[M(mu)\]) on the Hilbert space ${\cal H}_{in} \otimes {\cal H}_A:$ $$\begin{aligned}
\rho_B(\mu)=\frac{1}{{\cal P}(\mu)}
{\rm Tr}_{in,A}\left[M(\mu)\right]. \label{rhoB}\end{aligned}$$ The information provided by Alice allows Bob to perform a suitable displacement, $\rho_B^{\prime}(\mu)=D_2(\mu)\rho_B(\mu)
D^{\dag}_2(\mu)$. Then the ensemble averaging (\[1.4\]) yields the state emerging by CV teleportation from the unknown input state $\rho_{in}$: $$\rho_{out}=\int {\rm d}^2 \mu \;{\cal P}(\mu)\;D_2(\mu)\rho_B(\mu)
D^{\dag}_2(\mu). \label{rhoout}$$ Note that, owing to the CV-teleportation protocol itself, the output single-mode state $\rho_{out}$, eq. (\[rhoout\]), is always a mixed one.
Our main previous result is a very simple formula connecting the one-mode states $\rho_{in}$ and $\rho_{out}$. It is expressed in terms of their normally-ordered CFs and a remnant of the CF $\chi_{AB}(\lambda_1,\lambda_2)$ of the EPR state [@PT03; @PT06; @PT08]: $$\chi_{out}^{(N)}(\lambda)= \chi_{in}^{(N)}(\lambda)
\chi_{AB}(\lambda^*,\lambda). \label{chinchi}$$ The factorization formula (\[chinchi\]) shows that CV teleportation is a noisy process, which always alters the input state $\rho_{in}.$ It is equivalent to the identity $$\chi_{out}(\lambda)= \chi_{in}(\lambda)
\chi_{AB}(\lambda^*,\lambda). \label{chichi}$$ The function $$\chi_{D}^{(N)}(\lambda):=\chi_{AB}(\lambda^*,\lambda)
\label{chind}$$ leads us to introduce the related one, $$\chi_{D}(\lambda):=\exp{\left (-\frac{1}{2}|\lambda|^2 \right )}
\chi_{D}^{(N)}(\lambda), \label{chid}$$ which enters the one-mode Weyl expansion $$\rho_D:=\frac{1}{\pi}\int {\rm d}^2 \lambda \;\;\chi_{D} (\lambda)
D_2(-\lambda). \label{rod}$$ Here $\rho_D$ is a self-adjoint Hilbert-Schmidt operator of unit trace, on the single-mode Hilbert space ${\cal H}_B$ at Bob’s side.
Unless the two-mode EPR state $\rho_{AB}$ is Gaussian, one could only conjecture the positivity of the operator $\rho_D$. However, in the Gaussian case, we will prove this property in the next section. Accordingly, for the special class of the two-mode Gaussian EPR states, the function $\chi_{D}(\lambda)$, eq. (\[chid\]), is the CF of a $B$-mode state $\rho_D$ that we have termed [*distorting-field state*]{} [@PT08]. The multiplication rule (\[chinchi\]) displays therefore the fact that $\rho_{out}$ is the $B$-mode state of a superposition of two single-mode fields: the input one and a remote distorting field in the state $\rho_D$, due to the imperfect character of CV teleportation.
We conclude this section by introducing some operators on the Hilbert space ${\cal H}_A \otimes {\cal H}_B$ that prove to be useful when analyzing the non-locality features of the two-mode resource state $\rho_{AB}$. The usual EPR observables are two commuting linear combinations of the single-mode canonical operators: the [*relative coordinate*]{} $\hat Q:= \hat q_1-\hat q_2$ and the [*total momentum*]{} $\hat P:=\hat p_1+\hat p_2$. In terms of them we define the EPR operator $$\hat \Delta:=\frac{1}{2}\left({\hat Q}^2+{\hat P}^2\right).
\label{d}$$ For later convenience, let us introduce a non-local normal operator, $$\begin{aligned}
\hat A:=\hat a_1-(\hat a_2)^{\dag}=\frac{1}{\sqrt{2}}
(\hat Q+i \hat P), \label{A}\end{aligned}$$ as a suitable amplitude of the positive EPR operator: $$\hat \Delta={\hat A}^{\dag} \hat A.
\label{a}$$
GAUSSIAN DISTORTING-FIELD STATE
===============================
In what follows we adopt a shorthand notation concerning the operator $\rho_D$, eq. (\[rod\]): we denote by ${\cal H}$ its one-mode Hilbert space and by $\hat a$ the corresponding annihilation operator. Moreover, in order to simplify the subsequent discussion, we assume for the moment that a distorting-field state $\rho_D$ does exist whatever EPR state $\rho_{AB}$. We eventually prove that this assumption is true for any two-mode Gaussian EPR state.
To start on our analysis, substitution into eq. (\[chind\]) of the Taylor expansions $$\begin{aligned}
\chi^{(N)}_D(\lambda)=\sum_{l, m=0}^{\infty}
\frac{1}{l!m!}
\lambda^{l}(-\lambda^{*})^{m}\langle({\hat a}^{\dag})^{l}
{\hat a}^{m}\rangle_D
\label{chin}\end{aligned}$$ and $$\begin{aligned}
\chi_{AB}(\lambda^*,\lambda)=\sum_{l, m=0}^{\infty}
\frac{1}{l!m!}
\lambda^{l}(-\lambda^{*})^{m}\langle({\hat A}^{\dag})^{l}
{\hat A}^{m}\rangle_{AB}
\label{ab}\end{aligned}$$ yields the correlation functions in the distorting-field state: $$\begin{aligned}
\langle({\hat a}^{\dag})^{l}{\hat a}^{m}\rangle_D\
=\langle({\hat A}^{\dag})^{l}{\hat A}^{m}\rangle_{AB}.
\label{cfs1}\end{aligned}$$ We will omit subsequently the pair of indices $AB$ when writing expectation values in the EPR state $\rho_{AB}$: $$\langle\ldots\rangle:=\langle\ldots\rangle_{AB}.$$ In particular, the $l$th-order correlation function is non-negative for any $l$: $$\langle({\hat a}^{\dag})^{l}{\hat a}^{l}\rangle_D\
=\langle{\hat \Delta}^{l}\rangle \geq 0.
\label{dcf}$$ The identity (\[cfs1\]) can be employed to evaluate the $2\times 2$ covariance matrix (CM) of the distorting-field state $\rho_D$ [@PTH01]: $$\begin{aligned}
{\cal V}_D=\left(\begin{array}{cc}\sigma_D(q, q)&\sigma_D(q, p)\\
\sigma_D(p, q)&\sigma_D(p, p)\end{array}\right)
\label{vm}\end{aligned}$$ Explicitly, the CM ${\cal V}_D$ has the following entries [@PT08]: $$\begin{aligned}
\sigma_D(q, q)=\frac{1}{2}+\langle{\hat Q}^2\rangle,\;\;\;
\sigma_D(q, p)=\langle\hat Q \hat P\rangle,\;\;\;
\sigma_D(p, p)=\frac{1}{2}+\langle{\hat P}^2\rangle. \label{ent}\end{aligned}$$ Owing to the Schwarz inequality for a quasi-inner product, $${\langle\hat Q \hat P\rangle}^2 \leq
\langle \hat Q^2\rangle \langle \hat P^2\rangle,
\label{Schwarz}$$ the Robertson-Schrödinger uncertainty relation holds: $${\cal V}_D+\frac{i}{2}J\geq 0, \;\;\;(J:=i\sigma_2),
\label{RS}$$ with $\sigma_2$ denoting the complex Pauli matrix. Condition (\[RS\]) is necessary for all single-mode states, but it is sufficient only for the Gaussian ones. By virtue of definition (\[chind\]), the function $\chi_{D}(\lambda)$, eq. (\[chid\]), is Gaussian if and only if the two-mode EPR state $\rho_{AB}$ is Gaussian too. In this case, $\chi_{D}(\lambda)$ is indeed the CF of a one-mode Gaussian state $\rho_D$: our assertion is therefore proven.
From now on, in this section, we deal only with Gaussian CV teleportation. The existence of a Gaussian distorting-field state $\rho_D$ allows us to read eq. (\[chinchi\]) as a multiplication rule of normally-ordered CFs: $$\chi_{out}^{(N)}(\lambda)= \chi_{in}^{(N)}(\lambda)
\chi_{D}^{(N)}(\lambda). \label{chinchin}$$ Note that if the Gaussian resource state $\rho_{AB}$ is undisplaced, so is the distorting-field state $\rho_D$. Further, according to eqs. (\[ent\]), $\rho_D$ is a mixed state, since $\det{\cal V}_D >1/4$, unless the random EPR variables $\hat Q$ and $\hat P$ are constants, meeting thus the ideal EPR demand.
Moreover, eqs. (\[ent\]) show that a matrix inequality holds, ${\cal V}_D \geq \frac{1}{2}I_2,$ pointing out that $\rho_D$ is a [*classical*]{} state. Therefore, the Gaussian distorting-field state $\rho_D$ has a regular Glauber-Sudarshan $P$ representation, $$P_D(\alpha)=\frac{1}{\pi}\int{\rm d}^{2}\lambda
\exp{(\alpha \lambda^*-\alpha^* \lambda)}
\chi_D^{(N)}(\lambda), \label{p}$$ which is a Gaussian distribution function. The multiplication rule (\[chinchin\]) is equivalent to the existence of a mapping $$\rho_{out}=\int{\rm d}^{2}\beta P_{D}(\beta)D(\beta)
\rho_{in}D^{\dag}(\beta) \label{channel}$$ between one-mode states on the Hilbert space ${\cal H}$. We call such a mapping a [*teleportation channel*]{}. In sum, Gaussian CV teleportation is described by a Gaussian channel (\[channel\]).
It is instructive to evaluate the Glauber $R$ function of the state $\rho_D$ [@Glauber], $$R_D(\beta^*, \beta^{\prime}):=\exp{\left(\frac{1}{2}(|\beta|^2
+|\beta^{\prime}|^2)\right)}\langle\beta |\rho_D|
\beta^{\prime}\rangle, \label{r}$$ as an integral [@PT93]: $$R_D(\beta^*, \beta^{\prime})=\exp(\beta^*\beta^{\prime})
\frac{1}{\pi}\int{\rm d}^{2}\lambda \chi_D^{(N)}(\lambda)
\exp{(-|\lambda|^2-\beta^* \lambda+\beta^{\prime}
\lambda^*)}. \label{rint}$$ Making use of the eqs. (\[chin\]) and (\[cfs\]), we find: $$R_D(\beta^*, \beta^{\prime})=\langle\exp{(-{\hat \Delta}
+\beta^{\prime}{\hat A}^{\dag}+\beta^*{\hat A})}\rangle.
\label{rd}$$ It follows that the Husimi function ($\equiv$ the Glauber $Q$ function) reads: $$Q_D(\beta):=\frac{1}{\pi}\langle\beta |\rho_D|\beta\rangle
=\frac{1}{\pi}\exp(-|\beta|^2)\langle\exp{(-{\hat \Delta}
+\beta{\hat A}^{\dag}+\beta^*{\hat A}})\rangle. \label{qd}$$ We take advantage of the Taylor expansion of the $R$ function [@Glauber], $$R_D(\beta^*, \beta^{\prime})=\sum_{l, m=0}^{\infty}
(\rho_D)_{lm}\frac{1}{\sqrt{l!m!}}(\beta^*)^{l}
(\beta^{\prime})^{m}, \label{rser}$$ to write down the density matrix $$(\rho_D)_{lm}=\frac{1}{\sqrt{l!m!}}\langle\exp{(-{\hat \Delta)}}
{\hat A}^{l}({\hat A}^{\dag})^{m}\rangle. \label{dm}$$ The corresponding photon-number distribution, $$(\rho_D)_{ll}=\frac{1}{l!}\langle{\hat \Delta}^{l}
\exp(-{\hat \Delta)}\rangle, \label{pnd}$$ has the generating function $G_D(s):=\sum_{l=0}^{\infty}s^l(\rho_D)_{ll},\;(|s|\leq 1)$: $$G_D(s)=\langle\exp((s-1){\hat \Delta})\rangle. \label{gen}$$ This distribution is entirely determined by the statistics of the EPR operator ${\hat \Delta}$, eq. (\[d\]), in the two-mode Gaussian EPR state $\rho_{AB}$. It is worth mentioning that eqs. (\[chinchin\])– (\[gen\]) hold also for any classical [*non-Gaussian*]{} distorting-field state $\rho_D$: in fact, classicality is required only to ensure the validity of eqs. (\[p\]) and (\[channel\]), playing no role for the other ones.
Let us specialize the above discussion to a zero-mean Gaussian resource state frequently used, namely, a two-mode squeezed vacuum state (SVS): $\rho_{AB}=|\Psi_{AB}\rangle\langle\Psi_{AB}|.$ The Schmidt decomposition of such a pure state in the standard Fock basis, $$|\Psi_{AB}\rangle=\frac{1}{\cosh r}\sum_{n=0}^{\infty}
(\tanh r)^n |n\rangle_A \otimes |n\rangle_B,$$ is parametrized with the squeezing factor $r>0$. The corresponding distorting-field state is thermal, $$\chi_{D}^{(N)}(\lambda)=\exp{(-{\rm e}^{-2 r}|\lambda|^2)},
\label{chiSVS}$$ and has the $P$ representation $${P}_D(\alpha)=\frac{{\rm e}^{2 r}}{\pi}
\exp{(-{\rm e}^{2 r}|\alpha|^2)}.
\label{PSVS}$$ With eqs. (\[chin\]) and (\[cfs1\]), we get the correlation functions $$\begin{aligned}
\langle({\hat a}^{\dag})^{l}{\hat a}^{m}\rangle_D\
={\delta}_{lm}l!({\rm e}^{-2 r})^l
=\langle({\hat A}^{\dag})^{l}{\hat A}^{m}\rangle.
\label{cfsSVS}\end{aligned}$$ Note the mean photon number, $\langle {\hat a}^{\dag} {\hat a} \rangle_D=\exp{(-2 r)}
=\langle{\hat \Delta}\rangle$, and the $l$th-order correlation function, $\langle ({\hat a}^{\dag})^l {\hat a}^l \rangle_D
=l!{\langle{\hat \Delta}\rangle}^l
=\langle{\hat \Delta}^l\rangle$.
For the sake of completeness, we write down further the $R$ function, $$R_D(\beta^*, \beta^{\prime})=\frac{1}{1+{\rm e}^{-2 r}}
\exp{\left(\frac{{\rm e}^{-2 r}}{1+{\rm e}^{-2 r}}
\beta^* \beta^{\prime}\right)}, \label{R}$$ the Husimi function, $$Q_D(\beta)=\frac{1}{\pi}\frac{1}{1+{\rm e}^{-2 r}}
\exp{\left(-\frac{|\beta|^2}{1+{\rm e}^{-2 r}}\right)},
\label{QSVS}$$ the density matrix, $$(\rho_D)_{lm}={\delta}_{lm}\frac{1}{1+{\rm e}^{-2 r}}
\left(\frac{{\rm e}^{-2 r}}{1+{\rm e}^{-2 r}}\right)^l,
\label{dmSVS}$$ and the generating function of the photon-number distribution, $$G_D(s)=\left[1+(1-s){\rm e}^{-2 r}\right]^{-1}. \label{genf}$$ All the above formulae are specific for a single-mode thermal state. Therefore, when using a SVS as the resource state, teleportation is described by a thermalization channel.
ACCURACY OF TELEPORTATION
=========================
Originally, the quality of the teleportation protocol was quantified by the input-output overlap for pure states [@BK], or by use of the Uhlmann fidelity for mixed Gaussian states [@PT03; @Ban2]. For a clear survey on the progress in CV teleportation we refer the reader to Ref. [@PM]. More recently [@PT06; @PT08], in analyzing CV teleportation, the present authors have introduced the distorting-field state $\rho_D$ and focused on its properties. We point out here the conspicuous role of the EPR operator ${\hat \Delta}$, eq. (\[d\]). Its expectation value in the resource state $\rho_{AB}$, called EPR uncertainty [@Duan], quantifies the non-locality of this state. As the EPR uncertainty $\langle {\hat \Delta} \rangle$ decreases, the non-local character of the two-mode state $\rho_{AB}$ becomes stronger. In particular, the inequality $\langle {\hat \Delta} \rangle < 1$ is a criterion of inseparability of the bipartite state $\rho_{AB}$.
We start by assuming first the existence of a one-mode remote-field state $\rho_D$: we have shown that this effectively happens [*at least*]{} for the class of the two-mode Gaussian EPR states. The quality of teleportation can be evaluated in terms of the mean photon number $\langle {\hat a}^{\dag}{\hat a} \rangle_D$ in the one-mode state $\rho_D$. For any undisplaced Gaussian EPR state, this can be seen as the amount of noise added by teleportation: it distorts the features of the input field state $\rho_{in}$. The smaller this noise, the higher the quality of the CV teleportation. According to eq. (\[cfs1\]), the added noise is equal to the EPR uncertainty: $$\langle {\hat a}^{\dag} {\hat a} \rangle_D
=\langle {\hat \Delta} \rangle. \label{noise}$$
Second, we make a conjecture that extends this result to an arbitrary undisplaced two-mode EPR state: [*The amount of noise distorting the properties of the input field state is equal to the EPR uncertainty $\langle {\hat \Delta} \rangle$*]{}.
A remarkable theorem proven by Giedke [*et al.*]{} [@G] states that among all equally entangled pure two-mode states, the SVS has the minimal EPR uncertainty $\langle {\hat \Delta} \rangle$, [*i.e.*]{}, the strongest non-local character. This theorem regarding the ranking of pure-state entanglement at a given EPR uncertainty enables us to notice an interesting property of CV teleportation: [*The SVS adds the minimal noise in teleportation with pure two-mode resource states having the same entanglement.*]{}
We finally give a new expression of another quantity that is widely employed to measure the teleportation accuracy: the fidelity of teleporting a coherent state, hereafter denoted by ${\cal F}_{coh}$. Recall that a coherent state $\rho_{in}=|\alpha\rangle \langle \alpha|$ has a Gaussian CF: $$\chi_{in}(\lambda)=\exp{(\alpha^* \lambda-\alpha \lambda^*)}
\exp{\left (-\frac{1}{2}|\lambda|^2 \right )}.$$ The fidelity of teleporting a coherent state is the probability of the transition $\rho_{in} \rightarrow \rho_{out}:$ $${\cal F}_{coh}=\langle \alpha|\rho_{out}|\alpha \rangle.$$ When writing this quantity in terms of the CFs of the states involved, eq. (\[chichi\]) provides an expression that is independent of the input coherent state: $${\cal F}_{coh}=\frac{1}{\pi} \int{\rm d}^{2}\lambda
\exp{(-|\lambda|^2)}\chi_{AB}(\lambda^*,\lambda).$$ Making use of the Taylor series (\[ab\]), we find: $${\cal F}_{coh}=\sum_{l=0}^{\infty}\frac{(-1)^l}{l!}
\langle{\hat \Delta}^l\rangle=\langle
\exp{(-\hat \Delta)}\rangle. \label{sum}$$ Equation (\[sum\]) is valid for any two-mode EPR state $\rho_{AB}$. It displays to what extent the EPR operator $\hat \Delta$ is involved in the structure of the fidelity of teleporting a coherent state. If a distorting-field state $\rho_D$ exists, then an inspection of eqs. (\[qd\]) and (\[sum\]) gives the identity $${\cal F}_{coh}=\pi Q_D(0). \label{fcoh}$$ For instance, let us consider again the case of a SVS chosen as a two-mode resource state. Then, by use of eqs. (\[QSVS\]) and (\[fcoh\]), we recover the formula $${\cal F}_{coh}=\frac{1}{1+\exp{(-2 r)}}, \label{svs}$$ in agreement with previous results [@BK; @Ban2].
To sum up, in this paper we have examined further the CF description of the BK teleportation protocol. For the class of the two-mode Gaussian EPR states, we have been able to identify a remote distorting field mode superposed on the input one. This originates in the noisy character of the CV teleportation. We have pointed out the main properties of a one-mode Gaussian distorting-field state. They are connected with the non-local features of the two-mode EPR state. The accuracy of the CV teleportation is measured either by the amount of added noise or by the fidelity of teleporting a coherent state. Both quantities depend on the degree of non-locality of the bipartite resource state expressed in terms of the EPR uncertainty.
### Acknowledgements {#acknowledgements .unnumbered}
This work was supported by the Romanian Ministry of Education and Research through Grant IDEI-995/2007 for the University of Bucharest.
[22]{} C.H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W.K. Wootters, Phys. Rev. Lett. [**70**]{}, 1895 (1993). S. L. Braunstein and H. J. Kimble, Phys. Rev. Lett. [**80**]{}, 869 (1998); A. Furusawa, J. L. Sorensen, S. L. Braunstein, C. A. Fuchs, H. J. Kimble, and E. S. Polzik, Science [**282**]{}, 706 (1998). Paulina Marian, T. A. Marian, and H. Scutaru, Rom. J. Phys. [**48**]{}, 727 (2003). Paulina Marian and T. A. Marian, Phys. Rev. A [**74**]{}, 042306 (2006). Paulina Marian and T. A. Marian, Int. J. of Quant. Inf. [**6**]{}, 721 (2008). Paulina Marian, T. A. Marian, and H. Scutaru, J. Phys A [**34**]{}, 6969 (2001). R. J. Glauber, Phys. Rev. [**131**]{}, 2766 (1963). Paulina Marian and T. A. Marian, Phys. Rev. A [**47**]{}, 4474 (1993). M. Ban, Phys. Rev. A [**69**]{}, 054304 (2004). S. Pirandola and S. Mancini, Laser Physics [**16**]{}, 1418 (2006). Lu-Ming Duan, G. Giedke, J. I. Cirac, and P. Zoller, Phys. Rev. Lett. [**84**]{}, 2722 (2000). G. Giedke, M. M. Wolf, O. Krüger, R. F. Werner, and J. I. Cirac, Phys. Rev. Lett. [**91**]{}, 107901 (2003).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We formulate three current models of discrete-time quantum walks in a combinatorial way. These walks are shown to be closely related to rotation systems and 1-factorizations of graphs. For two of the models, we compute the traces and total entropies of the average mixing matrices for some cubic graphs. The trace captures how likely a quantum walk is to revisit the state it started with, and the total entropy measures how close the limiting distribution is to uniform. Our numerical results indicate three relations between quantum walks and graph structures: for the first model, rotation systems with higher genera give lower traces and higher entropies, and for the second model, the symmetric 1-factorizations always give the highest trace.'
author:
- 'Chris Godsil[^1] [^2] Hanmeng Zhan'
bibliography:
- 'dqw.bib'
title: 'Discrete-Time Quantum Walks and Graph Structures'
---
Introduction
============
There are at least three different models of discrete-time quantum walks in current research. Some of them are equivalent to quantum search algorithms [@Shenvi2003], and some are shown to exhibit desired properties such as periodicity and perfect state transfer [@Barr2014; @Dukes2014]. Most of the existing studies focus on a specific model or a restricted family of graphs. In this paper, we analyze the dependence of these models on different combinatorial structures. To be more specific, we reformulate all three models in graph-theoretic language, and experiment on different classes of graphs to see how the properties of a graph affect the properties of a walk.
Unlike the continuous-time quantum walks, the discrete ones require more than just a graph to build. The extra inputs might include a selection of coins, a set of linear orders of the neighbors of vertices, a 1-factorization of the bipartite double cover, or a Markov chain on the graph. According to how the walks are constructed from these building blocks, we will name them the *arc-reversal model*, the *shunt-decomposition model* and the *two-reflection model*.
In the arc-reversal model, the walk alternately flips a unitary coin and reverses the arcs of the graph. We show that if the coins are circulants with simple eigenvalues, then for each rotation system of the graph, we can construct a unique arc-reversal walk. In the shunt-decomposition model, the walk alternately flips a unitary coin and hops between outgoing arcs in the same class. Each such class, or “shunt", determines a perfect matching of the bipartite double cover of the graph. Thus every shunt-decomposition walk inherits the properties of some 1-factorization. The two-reflection model is coinless, and the walk alternately applies two involutions with respect to two weighted partitions of the arcs. These partitions usually arise from a classical Markov chain.
After introducing different models, we review some asymptotic properties of a general quantum walk. It is well-known that the probability distribution of a quantum walk does not converge. However, the average probabilities over time converge to a limit, which we call the average mixing matrix. In Section \[avgmix\], we develop some theory of the average mixing matrix. Two parameters of the average mixing matrix are of our interest—the trace, which captures how likely a walk is to stay at home in the limit, and the total entropy, which captures how far the limiting distributions are from being uniform.
We then present some computational results indicating sensitivity of walk properties to details of the model. For each cubic graph on up to 12 vertices, we enumerate all its rotation systems and shunt-decompositions, compute the associated average mixing matrices for the arc-reversal model and the shunt-decomposition model, and compared their traces and total entropies. Our numerical results on the traces indicate two relations. For the first model, as the genus of the embedding increases, the trace of the average mixing matrix decreases, with only few exceptions for some cubic graphs on 12 vertices. For the second model, the trace distinguishes non-isomorphic shunt-decompositions, and the symmetric shunts, if any, always give the highest trace. Meanwhile, for the arc-reversal model, the total entropy distinguishes different graph embeddings, and increases as the genus increases for most of the time.
Besides the natural question arising from the aforementioned experiments, we list some other open problems in the end. It is not our current goal to give a charaterization of the most general form of a model. Rather, we aim to find some specific choice of the building blocks, such as the coins, which induce interesting walks that will reveal clear relation between the properties of the graphs and the properties of the walks.
Models
======
A discrete-time quantum walk, in its most general form, is a power of a unitary matrix, which updates the amplitudes on arcs at each step. While it seems natural to construct a unitary matrix acting on the vertices of a graph, not every graph admits a unitary weighted adjacency matrix. The path on three vertices is an example. In [@Severini2003], Severini discovered some properties of digraphs that admit unitary representations. Alternatively, for an arbitrary graph we can always construct a unitary matrix acting on a larger space. In each of the models we describe below, the transition matrix is acting on the arcs, or the ordered pairs of vertices, and is a product of two sparse unitary matrices. The first two types of walks are driven by a set of coins, while the last one is coinless.
Arc-Reversal Model
------------------
The first model we investigate dates back to 2001, when Watrous [@Watrous2001] presented a method to simulate random walks on regular graphs using quantum computations. In [@Kendon2003] Kendon extended his idea to a class of coined quantum walks on general graphs. Based on this model, Emms et al [@Emms2005; @Emms2009] proposed an algorithm for the graph isomorphism problem. They tested a large number of strongly regular graphs, and the spectrum of their proposed matrix successfully distinguished all of their non-isomorphic strongly regular graphs with the same parameters.
We will view an undirected graph $X$ as a directed graph, with each edge replaced by a pair of opposite arcs. Suppose $X$ has $n$ vertices and $m$ arcs. Our quantum walk takes place in the complex inner product space ${{\mathbb C}}^{m}$, spanned by the characteristic vectors $e_{u,v}$ of the arcs $(u,v)$. Let $R$ be the permutation matrix that reverses each arc, that is, $$R e_{u,v} = e_{v,u}.$$ This gives us the first sparse unitary matrix. To construct a second unitary, for each vertex $u$ we specify a linear order on its neighbors: $$f_u:\{1,2,\cdots,\deg(u)\} \to \{v: u\sim v\}.$$ The vertex $f_u(j)$ will be referred to as the $j$-th neighbor of $u$, and the arc $(u, f_u(j))$ $j$-th arc of $u$. Next, define a $\deg(u) \times \deg(u)$ unitary matrix $C_u$ that acts on the outgoing arcs of $u$. This matrix serves as a quantum coin: it sends the $j$-th arc of $u$ to a superposition of all outgoing arcs of $u$, in which the amplitudes come from the $j$-th column of $C_u$: $$C_u e_j = \sum_{k=1}^{\deg(u)} (e_k^T C_u e_j) e_k.$$ Denote by $C$ the block diagonal matrix $$C = {\begin{pmatrix} C_1 &&&\\ & C_2 &&\\ &&\ddots &\\ &&& C_n \end{pmatrix}}.$$ The transition matrix of an arc-reversal quantum walk is given by $$U = RC.$$ Thus, each iteration consists of two steps—a coin flip and an arc-reversal.
By way of example, consider the complete graph $K_3$. Pick the linear orders $f_1$, $f_2$ and $f_3$ such that $$\begin{gathered}
f_1(1)=2, \quad f_2(2)=3;\\
f_2(1)=1, \quad f_2(2)=3;\\
f_3(1)=1, \quad f_3(2)=2.\end{gathered}$$ Then the arc-reversal matrix can be written as $$R=
\begin{blockarray}{ccccccc}
& (1,2) & (1,3) & (2,1) & (2,3) & (3,1) & (3,2) \\
\begin{block}{c(cccccc)}
(1,2) & 0 & 0 & 1 & 0 & 0 & 0\\
(1,3) & 0 & 0 & 0 & 0 & 1 & 0\\
(2,1) & 1 & 0 & 0 & 0 & 0 & 0\\
(2,3) & 0 & 0 & 0 & 0 & 0 & 1\\
(3,1) & 0 & 1 & 0 & 0 & 0 & 0\\
(3,2) & 0 & 0 & 0 & 1 & 0 & 0\\
\end{block}
\end{blockarray}$$ Now let us give the same coin to all the vertices, so that the $k$-th arc of $u$ hops to the $j$-th arc of $u$ with amplitude $1/\sqrt{2} (-1)^{jk}$, for any $u$. The coin operator is thus $$C=
\begin{blockarray}{ccccccc}
& (1,2) & (1,3) & (2,1) & (2,3) & (3,1) & (3,2) \\
\begin{block}{c(cccccc)}
(1,2) & 1/\sqrt{2} & 1/\sqrt{2} & 0 & 0 & 0 & 0\\
(1,3) & 1/\sqrt{2} & -1/\sqrt{2} & 0 & 0 & 0 & 0\\
(2,1) & 0 & 0 & 1/\sqrt{2} & 1/\sqrt{2} & 0 & 0\\
(2,3) & 0 & 0 & 1/\sqrt{2} & -1/\sqrt{2} & 0 & 0\\
(3,1) & 0 & 0 & 0 & 0 & 1/\sqrt{2} & 1/\sqrt{2}\\
(3,2) & 0 & 0 & 0 & 0 & 1/\sqrt{2} & -1/\sqrt{2}\\
\end{block}
\end{blockarray}$$ With initial state $e_{1,3}$, after one iteration, the system will be in the state $$\frac{1}{\sqrt{2}} e_{2,1} - \frac{1}{\sqrt{2}} e_{3,1}.$$
The above coin is called the *Fourier coin*: $$F= \left(\frac{1}{\sqrt{d}}e^{2jk\pi i/d}\right)_{jk}$$ Another common choice would be the *Grover coin*: $$G = \frac{2}{d}J - I.$$ This coin treats the neighbors in a simple way – an arc $(u,v)$ stays at current position with amplitude $2/d$, and moves to any other outgoing arc of $u$ with amplitude $2/d-1$. Hence it is indifferent to the linear order $f_u$. The famous algorithm, Grover’s search, can be formulated as an arc-reversal walk on the looped $K_n$, with coin $-G$ assigned to the marked vertex, and coin $G$ to the unmarked ones.
Shunt-Decomposition Model
-------------------------
The next model has been mostly applied to graphs with some symmetry, such as the infinite paths, infinite grids, cycles, and cubes. It is formally defined by Aharonov et al [@Aharonov2000]. Similar to an arc-reversal model, each iteration is a coin flip followed by another unitary operation. Let $\{f_u: u\in V(X)\}$ be a set of linear orders as described in the previous section. For reasons that will become clear later, assume our graph $X$ is $d$-regular. The state space ${{\mathbb C}}^{m}$ is then isomorphic to the tensor product ${{\mathbb C}}^n \otimes {{\mathbb C}}^d$. With this decomposition, the arc $(u, f_u(j))$ can be represented by the vector $$e_u \otimes e_j.$$ The transition matrix is given by $$U=SC,$$ where $C$ is the coin operator, and $S$ shifts the arcs in the following way: for each vertex $u$, it maps its $j$-th arc to the $j$-th arc of $f_u(j)$: $$S ( e_u \otimes e_j) = e_{f_u(j)} \otimes e_j.$$ That is, $S$ moves every arc one step forward to the arc with the same label. By definition, $S$ is a $01$-matrix where each column has exactly one $1$. Since $S$ is unitary, it has to be a permutation matrix. Hence the linear orders $f_u$ cannot be arbitrary; for any vertex $u$, the values $$\{f_w^{-1}(v): w \sim v\}$$ must be distinct. Under this assumption, each label $j$ induces a permutation on $V(X)$ that maps one vertex to its $j$-th neighbor, denoted by the permutation matrix $P_j$. This explains why $X$ has to be regular. Now, $S$ is equivalent to the block diagonal matrix $${\begin{pmatrix} P_1 & & &\\
& P_2 & & \\
& & \ddots &\\
& & & P_d \end{pmatrix}}.$$ By the isomorphism between ${{\mathbb C}}^n\otimes {{\mathbb C}}^d$ and ${{\mathbb C}}^d \otimes {{\mathbb C}}^n$, $$U = (P_1\otimes E_{11}+\cdots+P_d\otimes E_{dd}) (E_{11}\otimes C_1+\cdots+E_{nn}\otimes C_n).$$
We consider an example. The following is a collection of linear orders for $K_3$: $$\begin{gathered}
f_1(1)=2, \quad f_1(2)=3;\\
f_2(1)=3, \quad f_2(2)=1;\\
f_3(1)=1, \quad f_3(2)=2.\end{gathered}$$ They give rise to an ordered coloring of the arcs of $K_3$, as shown in Figure \[arc\_coloring\].
\[every node/.style=[circle, draw]{}\]
\(1) at (0,1.6); (2) at (-1,0); (3) at (1,0);
/in [1/2, 2/3, 3/1]{} () to ();
/in [1/3, 3/2, 2/1]{} () to ();
\[arc\_coloring\]
\[every node/.style=[circle, draw]{}\]
\(1) at (0,1); (2) at (0,0); (3) at (0,-1);
(1’) at (2,1); (2’) at (2,0); (3’) at (2,-1);
/in [1/2’, 2/3’, 3/1’]{} () – ();
/in [1/3’, 2/1’, 3/2’]{} () – ();
\[1-factn\]
Each color determines a permutation on the vertices: $$P_1 =\begin{blockarray}{cccc}
& 1 & 2 & 3 \\
\begin{block}{c(ccc)}
1 & 0 & 0 & 1\\
2 & 1 & 0 & 0\\
3 & 0 & 1 & 0\\
\end{block}
\end{blockarray}
\qquad P_2 =\begin{blockarray}{cccc}
& 1 & 2 & 3 \\
\begin{block}{c(ccc)}
1 & 0 & 1 & 0\\
2 & 0 & 0 & 1\\
3 & 1 & 0 & 0\\
\end{block}
\end{blockarray}$$ which sum up to the adjacency matrix of $K_3$. Equivalently, we can realize the linear orders as an ordered edge coloring of the bipartite graph $K_2 \times K_3$, as shown in Figure \[1-factn\]. The edge $(j,k')$ is colored red if $f_j(1)=k$, and blue if $f_j(2)=k$.
In general, for a $d$-regular graph $X$, there is a one-to-one correspondence between a decomposition of its adjacency matrix into $d$ permutation matrices $$A(X) = P_1 + \cdots + P_d$$ and a 1-factorization of $K_2\times X$. Each permutation $P_j$ maps vertices to adjacent vertices. Such permutations are called *shunts*. We refer to the above decomposition as a *shunt-decomposition* of $X$. A set of coins together with an ordered shunt-decomposition determines a quantum walk.
Two-Reflections Model
---------------------
The third model is a direct quantization of the classical random walks, without introducing a coin register. It was extracted from Ambainis’s paper [@Ambainis2003] and formalized by Szegedy [@Szegedy2004]. Let $M$ be a Markov chain on $X$. We will construct two reflections $R_1$ and $R_2$ on the $n^2$ ordered pairs of vertices based on $M$, and set $$U = R_2R_1.$$ First, let $N$ denote the matrix obtained from $M$ by taking the square root of its entries. Next, define two partitions of the ordered pairs: $$\begin{aligned}
Q_1 &= {\begin{pmatrix} e_1\otimes (Ne_1) & \cdots & e_n\otimes (Ne_n) \end{pmatrix}}\\
Q_2 &= {\begin{pmatrix} (N^Te_1) \otimes e_1 & \cdots & (N^Te_n) \otimes e_n \end{pmatrix}}.\end{aligned}$$ Since $M$ is doubly stochastic, these are $n^2\times n$ matrices such that $$Q_1^TQ_1 = Q_2^TQ_2 = I.$$ Then for each $j$, the matrix $Q_jQ_j^T$ is a projection onto the column space of $Q_j$, and $$R_j = 2Q_jQ_j^T - I$$ is a reflection about the column space of $Q_j$.
For instance, take the simple random walk on $K_3$ with $$M = \begin{blockarray}{cccc}
& 1 & 2 & 3 \\
\begin{block}{c(ccc)}
1 & 0 & 1/2 & 1/2\\
2 & 1/2 & 0 & 1/2\\
3 & 1/2 & 1/2 & 0\\
\end{block}
\end{blockarray}.$$ Then we have $$Q_1 = \begin{blockarray}{cccc}
& 1 & 2 & 3 \\
\begin{block}{c(ccc)}
11 & 0 & 0 & 0\\
12 & 1/\sqrt{2} & 0 & 0\\
13 & 1/\sqrt{2} & 0 & 0\\
21 & 0 & 1/\sqrt{2} & 0\\
22 & 0 & 0 & 0\\
23 & 0 & 1/\sqrt{2} & 0\\
31 & 0 & 0 & 1/\sqrt{2}\\
32 & 0 & 0 & 1/\sqrt{2}\\
33 & 0 & 0 & 0\\
\end{block}
\end{blockarray}
\qquad
Q_2 = \begin{blockarray}{cccc}
& 1 & 2 & 3 \\
\begin{block}{c(ccc)}
11 & 0 & 0 & 0\\
12 & 0 & 1/\sqrt{2} & 0\\
13 & 0 & 0 & 1/\sqrt{2}\\
21 & 1/\sqrt{2} & 0 & 0\\
22 & 0 & 0 & 0\\
23 & 0 & 0 & 1/\sqrt{2}\\
31 & 1/\sqrt{2} & 0 & 0\\
32 & 0 & 1/\sqrt{2} & 0\\
33 & 0 & 0 & 0\\
\end{block}
\end{blockarray}$$ Effectively, $Q_1$ partitions the arcs according to their tails, and $Q_2$ partitions the arcs according to their heads. This is true for any simple random walk, as for any non-adjacent vertices $u$ and $v$, the row indexed by $(u,v)$ is zero in both $Q_1$ and $Q_2$. Thus, a quantization of a simple random walk can be viewed as a walk on the arcs only. In the above example, the initial state $e_{12}$ is mapped to $e_{13}$ by $R_1$, and then mapped to $e_{12}$ by $R_2$.
More generally, we can define a two-reflection walk by setting $$U = (2Q_1Q_1^T - I) (2Q_2 Q_2^T - I),$$ where $Q_1$ and $Q_2$ are normalized character matrices that represent two partitions of the arcs. These partitions may arise from other graph structures, such as orientable embeddings [@Zhan2017a]. The spectral decomposition of a two-reflection walk has been studied [@Szegedy2004], and used to find perfect state transfer in circulant graphs [@Zhan2017].
Similarities and Differences
----------------------------
In some cases the aforementioned models are equivalent, but there are also quantum walks that belong to only one of the frameworks. The shift matrix $S$ in the shunt-decomposition is symmetric if and only if every shunt $P_j$ is symmetric, and so the arc coloring arising from the shunt-decomposition is in fact an edge coloring. Hence, the arc-reversal model and the shunt-decomposition model overlap when the graph admits a $d$-edge-coloring, and the same coins and same linear orders are applied. The arc-reversal model can also be equivalent to the two-reflections model, if the first walk uses the Grover coins and lexicographical linear orders, and the second walk quantizes the simple random walk. They are equivalent in the sense that two steps of the first walk has the same effect as one step of the second walk. For more details, see the recent paper by Wong [@Wong2016]. Another correlation between these two models has been found by Portugal and Segawa [@Portugal2016], who showed that an arc-reversal walk with Grover coins can be converted into a two-reflection walk on the same graph with each edge subdivided once.
Shunt-Decompositions
====================
We have seen that a shunt-decomposition of a graph $X$ is equivalent to a 1-factorization of $K_2 \times X$. Every regular bipartite graph has at least one 1-factorization, but there might be more, in which case we would like to compare the properties of walks induced by different shunt-decompositions. To enumerate all 1-factorizations of a graph $Y$, notice that a matching of $Y$ is a coclique of the line graph $L(Y)$ of $Y$. Thus, every 1-factorization of $Y$ gives rise to a vertex partition of $L(Y)$ into maximum cocliques, and vise versa. We enumerated all shunt-decompositions of each cubic graph on up to 8 vertices, and compared the limiting average distributions of the corresponding quantum walks. More details will be discussed in Section \[data\_shunt\].
Some graphs come with natural shunt-decompositions. For example, the arcs in a Cayley graph $X(G, \operatorname{\mathcal{C}})$ are partitioned according to the elements in the connection set $\operatorname{\mathcal{C}}$. Hence each element $c_j \in \operatorname{\mathcal{C}}$ determines a shunt that sends vertex $u$ to vertex $c_i^{-1} u$. In the next section, we will consider another family of graphs with obvious shunt-decompositions.
Maps
====
Let $X$ be a graph embedded on some surface. For simplicity, we assume that every face is bounded by a cycle. Let $u$ be a vertex, $e$ an edge incident to $u$, and $f$ a face containing the edge $e$. The triple $(u,e,f)$ is called a flag of the embedding. Pictorially, a flag is a triangle in the barycentric subdivision of a face. In the following example, each black dot represents a flag in the planner embedding of $C_3$.
\[every node/.style=[circle,draw]{}\]
\(1) at (0,3.2); (2) at (-2,0); (3) at (2,0);
\(0) at (0,1); (1’) at (0,0); (2’) at (1,1.6); (3’) at (-1,1.6); (a) at (0,5.2); (a’) at (0,-1.6); (b) at (-4.8,-1.6); (b’) at (3,2.7); (c) at (4.8,-1.6); (c’) at (-3,2.7);
\(A) at (-0.3,2); (B) at (0.3,2); (C) at (1.1,0.8); (D) at (0.8,0.3); (E) at (-0.8,0.3); (F) at (-1.1,0.8);
\(G) at (-0.3,4); (H) at (0.3,4); (I) at (3,-0.2); (J) at (2.7,-0.6); (K) at (-2.7,-0.6); (L) at (-3,-0.2);
/in [1/2, 2/3, 3/1]{} () to ();
/in [0/1, 0/1’, 0/2, 0/2’, 0/3, 0/3’, 1/a,1’/a’,2’/b’, 3’/c’,2/b,3/c]{} () to ();
For each flag $(u,e,f)$, let $u'$ be the other endpoint of $e$, let $e'$ be the other edge in $f$ that is incident to $u$, and let $f'$ be the other face that contains $e$. Define three functions $$\begin{aligned}
\tau_0: &(u, e, f) \mapsto (u', e, f),\\
\tau_1: &(u, e, f) \mapsto (u, e', f),\\
\tau_2: &(u, e, f) \mapsto (u, e, f').\end{aligned}$$ We have the following observations.
(i) $\tau_0, \tau_1, \tau_2$ are fixed-point-free involutions.
(ii) $\tau_0\tau_2 = \tau_2\tau_0$, and $\tau_0\tau_2 $ is fixed-point-free.
(iii) The group ${\langle \tau_0, \tau_1, \tau_2\rangle}$ acts transitively on the flags.
Combinatorially, any tuple $(\tau_0, \tau_1, \tau_2)$ satisfying the above is called a map. More details on maps can be found in [@Archdeacon1994].
Coming back to our map from an embedding of $X$, the *graph-encoded-map*, or *gem*, is the graph whose vertices are the flags, and two flags are adjacent if they are swapped by one of $\tau_0, \tau_1, \tau_2$. We will denote this graph by $\Phi(X)$. Note that $\Phi(X)$ is cubic and $3$-edge-colorable. In Figure \[flag\_C3\], we draw the gem of the planar $C_3$ with red, blue and green edges. The gem of the planar $C_n$ is the prism graph $K_2\square C_{2n}$.
\[every node/.style=[circle,draw]{}\]
\(1) at (0,3.2); (2) at (-2,0); (3) at (2,0);
\(0) at (0,1); (1’) at (0,0); (2’) at (1,1.6); (3’) at (-1,1.6); (a) at (0,5.2); (a’) at (0,-1.6); (b) at (-4.8,-1.6); (b’) at (3,2.7); (c) at (4.8,-1.6); (c’) at (-3,2.7);
\(A) at (-0.3,2); (B) at (0.3,2); (C) at (1.1,0.8); (D) at (0.8,0.3); (E) at (-0.8,0.3); (F) at (-1.1,0.8);
\(G) at (-0.3,4); (H) at (0.3,4); (I) at (3,-0.2); (J) at (2.7,-0.6); (K) at (-2.7,-0.6); (L) at (-3,-0.2);
/in [1/2, 2/3, 3/1]{} () to ();
/in [0/1, 0/1’, 0/2, 0/2’, 0/3, 0/3’, 1/a,1’/a’,2’/b’, 3’/c’,2/b,3/c]{} () to ();
/in[B/C, D/E, F/A, H/I, J/K, L/G]{} () to ();
/in [A/B, C/D, E/F, G/H, I/J, K/L]{} () to ();
/in [A/G, B/H, C/I, D/J, E/K, F/L]{} () to ();
The map $(\tau_0, \tau_1, \tau_2)$ is *orientable* if the underlying surface is orientable. For a formal definition of orientable manifolds, see Lee [@Lee2000 Ch 15]. Here we only provide some intuition. Given a face on a surface, we can orient its boundary in two directions. If there is some orientation of the faces such that whenever an edge $e$ is contained in two faces $f$ and $f'$, the orientation $e$ receives in $f$ is opposite to the orientation it receives in $f'$, then the orietation is called *coherent*. Figure \[coherent\] is an example of a coherent orietation. A surface is called *orientable* if it admits a coherent orientation, and *non-orientable* otherwise.
\[every node/.style=[circle]{}\]
\(0) at (0,0) ; (1) at (2,0) ; (2) at (2,2) ; (3) at (0,2) ; (4) at (4,0) ; (5) at (4,2) ;
(0rt) at (0.2,0.2) ; (0lt) at (-0.2,0.2) ; (0rb) at (0.2,-0.2) ; (1rt) at (2.2,0.2) ; (1lt) at (1.8,0.2) ; (1rb) at (2.2,-0.2) ; (1lb) at (1.8,-0.2) ; (2rt) at (2.2,2.2) ; (2lt) at (1.8,2.2) ; (2rb) at (2.2,1.8) ; (2lb) at (1.8,1.8) ; (3rt) at (0.2,2.2) ; (3rb) at (0.2,1.8) ; (3lb) at (-0.2,1.8) ; (4rt) at (4.2,0.2) ; (4lt) at (3.8, 0.2) ; (4lb) at (3.8,-0.2) ; (5lt) at (3.8,2.2) ; (5rb) at (4.2,1.8) ; (5lb) at (3.8,1.8) ;
/in [0/1, 1/2, 2/3, 3/0, 1/4, 4/5, 5/2]{} () to ();
/in [0rt/3rb, 3rb/2lb,2lb/1lt, 1lt/0rt, 2rb/5lb, 5lb/4lt, 4lt/1rt, 1rt/2rb, 3lb/0lt, 0rb/1lb, 1rb/4lb, 4rt/5rb, 5lt/2rt, 2lt/3rt]{} () to ();
Alternatively, we can characterize orientability using the gem. The following result is due to Vince [@Vince1983].
The embedding of $X$ is orientable if and only if the gem $\Phi(X)$ is bipartite. [ ]{} \[orient\_bip\]
With the $3$-edge-coloring induced by $\tau_0, \tau_1, \tau_2$, we obtain a symmetric shunt-decomposition of $\Phi(X)$. This defines an arc-reversal walk on $\Phi(X)$. In the case where $X$ is embedded on an orientable surface, the gem $\Phi(X)$ is bipartite and isomorphic to $K_2\times Y$ for some cubic graph $Y$. Then the edge-coloring of $\Phi(X)$ also gives rise to a shunt-decomposition, possibly asymmetric, on the smaller graph $Y$. To illustrate this, recall that the gem of the planer $C_3$ is $K_2\square C_6$. It is isomorphic to $K_2\times (K_2\square K_3)$, as drawn in Figure \[flagrf\]. Figure \[K2\_K\_3\] shows the corresponding shunt-decomposition of the graph $K_2\square K_3$.
\[every node/.style=[circle, draw]{}\] (1) at (-0.9,1.9); (3’) at (0.9,1.9); (2) at (2.1,0); (2’) at (-2.1,0); (3) at (-0.9,-1.9); (1’) at (0.9,-1.9); (6’) at (-0.5,0.9); (5) at (0.5,0.9); (4) at (-1.1,0); (4’) at (1.1,0); (5’) at (-0.5,-0.9); (6) at (0.5,-0.9);
/in [1/2’, 2/3’,3/1’, 4/6’,6/5’,5/4’]{} () to (); /in [2/1’,3/2’,1/3’,6/4’,5/6’,4/5’]{} () to (); /in [1/6’, 6/1’, 2/4’, 4/2’, 3/5’, 5/3’]{} () to ();
\[flagrf\]
\[every node/.style=[circle, draw]{}\]
\(1) at (0,1.9); (2) at (-1,0.3); (3) at (1,0.3); (4) at (-1,-1.9); (5) at (1,-1.9); (6) at (0,-0.3);
/in [1/2, 2/3, 3/1, 4/6, 6/5, 5/4]{} () to (); /in [2/1, 3/2, 1/3, 6/4, 4/5, 5/6]{} () to (); /in [1/6, 2/4, 3/5, 6/1, 4/2, 5/3]{} () to ();
\[K2\_K\_3\]
In general, the correspondence between $\Phi(X)$ and $Y$ is not unique. However, for each $Y$ such that $\Phi(X)$ is isomorphic to $K_2\times Y$, the adjacency matrix of $\Phi(X)$ can be written as $$A(\Phi(X))= {\begin{pmatrix} 0 & A(Y) \\ A(Y) & 0 \end{pmatrix}}.$$ Given the shunt decomposition of $A$ $$A(\Phi(X))=P_1+P_2+\cdots+P_d$$ where each $P_j$ is of the form $$P_j = {\begin{pmatrix} 0 & Q_j^T \\ Q_j & 0 \end{pmatrix}},$$ we have the shunt-decomposition of $Y$: $$A(Y) = Q_1+Q_2+\cdots+Q_d = Q_1^T+Q_2^T+\cdots+Q_d^T.$$ Hence, the quantum walk on $Y$ is $$U(Y) = (P_1\otimes E_{11} + P_2\otimes E_{22} +\cdots + P_d\otimes E_{d}) C_Y,$$ while the quantum walk on $\Phi(X)$is $$U(\Phi(X)) = (Q_1\otimes E_{11} + Q_2\otimes E_{22}+\cdots+Q_d\otimes E_{dd}) C_{\Phi(X)}.$$ With a proper choice of the coins, we will have $$U(\Phi(X)) = {\begin{pmatrix} 0 & S_Y C_Y\\ S_Y^T C_Y & 0 \end{pmatrix}}.$$
Rotation Systems
================
In either of the coined models, we need a set of linear orders $\{f_u: u\in V(X)\}$ of the neighbors to implement the coin operator. If we convert these linear orders into cyclic permutations, we obtain a rotation system, which determines an orientable embedding of $X$. We introduce some basic concepts about rotation systems. For more background, see Gross and Tucker [@Gross2001].
Formally, a *rotation system* is a set $\{\pi_u: u\in V(X)\}$ where each $\pi_u$ is a cyclic permutation on the neighbors of the vertex $u$. For any arc $(u_1, u_2)$, consider the walk $$(u_1, u_2), (u_2, u_3), (u_3, u_4), \cdots, (u_{k-1}, u_k), \cdots$$ where $$u_{j+1} = \pi_{u_j} (u_{j-1}).$$ Since the graph is finite, eventually the walk will meet an arc that is already taken. Moreover, the first arc that is used twice must be $(u_1, u_2)$, as the preimage $\pi_u^{-1} (v)$ is uniquely determined for each $u$. We will call the closed walk $$(u_1, u_2), (u_2, u_3), (u_3, u_4),\cdots, (u_k, u_1)$$ a *facial walk*. The facial walks partition the arcs of $X$. Clearly, each edge occurs either once in two facial walks, or twice in the same facial walk. On the other hand, a facial walk can use a vertex many times without using an arc twice. The following three embeddings illustrate the difference. In Figure \[face\_cyc\], every face is bounded by a cycle. In Figure \[face\_edge\], the outer face uses the pendent edge twice. In Figure \[face\_vx\], the outer face uses the central vertex four times but no edge twice.
\[every node/.style=[circle, draw]{}\]
\(1) at (1,0); (2) at (0,-1); (3) at (0,1); (4) at (-1,0);
/in [1/2, 2/3, 1/3, 2/4, 3/4]{} () to ();
\[face\_cyc\]
\[every node/.style=[circle, draw]{}\]
\(1) at (0,1); (2) at (0,0); (3) at (-1,-1); (4) at (1,-1);
/in [1/2, 2/3, 3/4, 2/4]{} () to ();
\[face\_edge\]
\[every node/.style=[circle, draw]{}\]
\(1) at (0,0); (2) at (-1,1); (3) at (-1,-1); (4) at (1,1); (5) at (1,-1);
/in [1/2,1/3,1/4,1/5,2/3,4/5]{} () to ();
\[face\_vx\]
Given a facial walk of length $k$, we associate it with a polygon with $k$ sides, labeled by the arcs in the same order as they appear in the facial walk. We then “glue” every two sides of these polygons labeled by the same edge. The result is an embedding of the graph onto an orientable surface.
Now let $X$ be a $d$-regular graph and try to construct a quantum walk from a rotation system of $X$. For simplicity, we will choose the arc-reversal model. The first problem is that there are $d$ ways to convert a cyclic permutation into a linear order, so we wish to unify the coin matrix with respect to all possible conversions. To be more specific, if $f_u$ and $g_u$ are two linear orders obtained from the cyclic order $\pi_u$, then the amplitude from arc $(u, f_u(j))$ to arc $(u, f_u(k))$ is required to equal the amplitude from arc $(u, g_u(j))$ to arc $(u, g_u(k))$.
Let $X$ be a $d$-regular graph. Let $f_u$ and $g_u$ be two linear orders of the neighbors of $u$. Let $C_u$ be a unitary coin indexed by the outgoing arcs of $u$. The following are equivalent.
(i) For $j,k=1,2,\cdots,d$, the amplitude from arc $(u, f_u(j))$ to arc $(u, f_u(k))$ is equal to the amplitude from arc $(u, g_u(j))$ to arc $(u, g_u(k))$.
(ii) $C_u$ commutes with the permutation matrix $$P_u = {\begin{pmatrix} e_{g_uf_u^{-1}(1)} & e_{g_uf_u^{-1}(2)} & \cdots e_{g_uf_u^{-1}(d)} \end{pmatrix}}.$$
[*Proof.* ]{}For notational ease, let us replace $C_u$, $f_u$ and $g_u$ by $C$, $f$ and $g$, respectively. Statement (i) holds if and only if for each $j$ and $k$, $$e^T_{f(k)} C e_{f(j)} = e^T_{g(k)} C e_{g(j)}.$$ On the other hand, since $f$ and $g$ are bijections, statement (ii) holds if and only if for each $j$ and $k$, $$e^T_{g(k)} (CP) e_{f(j)} = e^T_{g(k)} (PC) e_{f(j)}.$$ It follows from $$P e_{f(j)} = e_{g(j)}$$ and $$P^T e_{g(k)} = P^{-1} e_{g(k)} = e_{f(j)}$$ that the above two statements are equivalent. [ ]{}
Now, if the linear orders $f_u$ and $g_u$ arise from the same cyclic order $\pi_u$, then they differ in only a cyclic permutation. An immediate consequence is that each coin $C_u$ must be cyclic so that the quantum walk for a rotation system is well-defined. This answers our first question. In addition, for different rotation systems, we wish to obtain different transition matrices, so the coin $C_u$ should not commute with any non-cyclic permutation. We show that this is guaranteed as long as $C_u$ has simple eigenvalues.
Let $C_u$ be a circulant unitary matrix with simple eigenvalues. Then $C_u$ commutes with a permutation matrix $P_u$ if and only if $P_u$ is cyclic.
[*Proof.* ]{}Since $C_u$ has simple eigenvalues, its commutant consists of precisely the polynomials in $C_u$, which are all cyclic. [ ]{}
In Section \[data\_arcrev\], we will pick a $3\times 3$ circulant coin with simple eigenvalues, and investigate the arc-reversal quantum walks for all rotation systems of cubic graphs on up to 10 vertices.
Hitting Times
=============
The hitting time measures how fast a quantum walk reaches a specific state. Let $x$ be the initial state, and $y$ the target state. A typical choice for $x$ and $y$ would be the characteristic vectors of two arcs in a graph, respectively. Since the state of the system is in a superposition, it is not on an arc until we perform a measurement. However, if we do a complete measurement after each step, the quantum walk will collapse to a classical random walk. Thus, we need a way to define “reaching" without killing the quantum coherence.
One choice is to let the quantum walk evolve and do a complete measurement only once. The *one-shot hitting time*, given by Kempe [@Kempe2002], is the time $k$ when a complete measurement returns the target state with high probability: $$\abs{y^*U^kx}^2\ge 1-\epsilon.$$ For this definition we need to know when to measure the walk. If the chosen probability $1-\epsilon$ is too high, the hitting time may not exist.
Alternatively, we can do a partial measurement with respect to $y$ after each transition, and determine whether the walk hits $y$ and stops, or is in a state orthogonal to $y$ and continues. The partial measurement consists of the projection $yy^*$ onto the space spanned by $y$, and the projection $I-yy^*$ onto the space orthogonal to $y$. Two notions of hitting time based on this measured walk are given in Kempe [@Kempe2002] and Krovi et al [@Krovi2006]. The *concurrent hitting time*, defined in [@Kempe2002], is the earliest time before which the walks stops with a fixed high probability $1-\epsilon$: $$h_{\epsilon} = \min\left\{K: \sum_{k=1}^{K} \abs{ y^* U \left((I-yy^*)U\right)^{k-1} x}^2 \ge 1-\epsilon \right\}.$$ Or, we can follow [@Krovi2006] and consider the *expected hitting time*, the average number of steps the walk takes to reach $x$: $$h = \sum_{k=1}^{\infty} k\abs{ y^* U \left((I-yy^*)U\right)^{k-1} x}^2.$$
Average Limiting Distributions
==============================
In a classical random walk, the probability distribution converges to a stationary distribution under only mild conditions. In a quantum walk, however, the unitary matrix $U$ preserves the difference between the states at two consecutive steps: $$\norm{Ux-x} = \norm{U^2x-Ux}.$$ Hence the probability distribution does not converge unless $Ux=x$. We may also define the probability on a vertex to be the sum over the probabilities on its outgoing arcs. Using Kronecker’s theorem, Aharonov et al [@Aharonov2000] showed that the probability distribution over the vertices does not converge either. Nonetheless, the time-averaged probability distribution converges, under both notions.
In this section, we extend the discussion to the probability on any subset $S$ of the arcs. Suppose the arcs in $S$ are indexed by $\{{j_{1},j_{2},\ldots,j_{m}}\}$. At time $k$, the entries of the following vector $${\begin{pmatrix} e_{j_1} & e_{j_2} & \cdots & e_{j_m} \end{pmatrix}}^T U^kx$$ are the amplitudes on the arcs in $S$. Let $$D_S = {\begin{pmatrix} e_{j_1} & e_{j_2} & \cdots & e_{j_m} \end{pmatrix}}{\begin{pmatrix} e_{j_1} & e_{j_2} & \cdots & e_{j_m} \end{pmatrix}}^T$$ be the characteristic matrix of $S$, that is, the diagonal matrix whose $(a,a)$-entry is $1$ if $a$ is in $S$. Then given initial state $x$, the probability that the quantum walk is on $S$ at time $k$ is $$P_{x,S}(k) := x^* (U^k)^* D_S U^k x.$$ While this probability does not converge as $k$ goes to infinity, the *average probability* over time $$\frac{1}{K} \sum_{k=0}^{K-1} P_{x,S}(k)$$ does as $K$ goes to infinity, for any initial state $x$. Before we prove it, we note that every unitary matrix $U$ can be written as $$U = \sum_r e^{i\theta_r} F_r$$ where $e^{i\theta_r}$ is an eigenvalue of $U$, and $F_r$ is a Hermitian matrix representing the projection onto the eigenspace of $e^{i\theta_r}$. This is usually called the *spectral decomposition* of $U$.
\[lim\_xS\] Let ${F_{1},F_{2},\ldots,F_{m}}$ be the spectral idempotents of the transition matrix $U$. Let $x$ be the initial state. For any subset $S$ of the arcs, the average probability that the quantum walk is on some arc of $S$ converges to $$\sum_r x^* F_r D_S F_r x.$$ \[limit\]
[*Proof.* ]{}Consider the spectral decomposition of $U$ $$U=\sum_r e^{i\theta_r} F_r.$$ It suffices to show that $$\frac{1}{K} \sum_{k=0}^{K-1} (U^k)^*D_SU^k$$ converges to $$\sum_r F_rD_S F_r$$ as $K$ goes to infinity. We have $$\begin{aligned}
(U^k)^* D_S U^k & =\left(\sum_r e^{-ik\theta_r}F_r\right)D_S \left(\sum_s e^{ik\theta_s}F_r\right)\\
&=\sum_r F_r D_S F_r+ \sum_{r\ne s} e^{ik(\theta_s-\theta_r)}F_rD_SF_s.\end{aligned}$$ Note that for all $r$ and $s$, the entries in $F_r D_S F_r$ and $F_rD_SF_s$ are constants, and remain unchanged when we take the average and the limit. Further $$\frac{1}{K}\abs{\sum_{k=0}^{K-1} e^{ik(\theta_s-\theta_r)}} = \frac{1}{K} \abs{\frac{1 - e^{iK(\theta_s-\theta_r)}}{1-e^{i(\theta_s-\theta_r)}}} \le \frac{1}{K} \frac{2}{\abs{1-e^{i(\theta_s-\theta_r)}}}$$ which converges to zero as $K$ goes to infinity. Hence the only term that survives in $$\lim_{K\to \infty}\frac{1}{K} \sum_{k=0}^{K-1} (U^k)^*D_SU^k$$ is $$\sum_r F_r D_S F_r. \tag*{{{\vbox{\hrule height.3pt
\hbox{\vrule width.3pt height5pt \kern5pt
\vrule width.3pt}\hrule height.3pt}}}}$$
Channels
========
From the previous section, we see that the positive semidefinite matrices $F_r D_S F_r$ are crucial in computing the average limiting distribution of a quantum walk. Here is another way to interpret the relation between these matrices and the limit. The spectral idempotents $F_r$ satisfy that $$\sum_r F_r^* F_r = I,$$ so the mapping on density matrices $\rho$ given by $$\rho \mapsto \sum_r F_r \rho F_r^*$$ is a quantum channel. As $K$ goes to infinity, the average state $$\frac{1}{K} \sum_{k=0}^{K-1} (U^k)^*\rho U^k$$ converges to the density matrix after the channel: $$\sum_r F_r \rho F_r.$$ Now, let $\abs{S}$ be the size of the subset $S$ and consider $$\rho := \frac{1}{\abs{S}} D_S = \frac{1}{\abs{S}} e_{j_1}e_{j_1}^T + \cdots \frac{1}{\abs{S}} e_{j_m}e_{j_m}^T.$$ This is a density matrix of a mixed state—a convex combination of the pure states associated with the arcs in $S$. Hence, if the quantum walk starts with state $x$, then the average probability of the walk being on $S$ is the inner product of the density matrices $xx^*$ and $\sum_r F_r \rho F_r$, scaled by the size of $S$: $$\abs{S}\Big<xx^*, \sum_r F_r \rho F_r\Big> = \sum_r x^* F_r D_S F_r x.$$
Mixing Times
============
The mixing time of a quantum walk measures how far the average probability distribution is from the stationary distribution. Since the average probability on every subset converges, it does not make a big difference whether the distribution is on the arcs or on the vertices. In [@Aharonov2000], Aharonov et al studied the mixing time on the vertices and obtained an upper bound for a general graph. They further showed that the mixing time of a quantum walk on an $n$-cycle with the Hadamard coin is bounded above by $O(n\log n)$, giving a quadratic speedup over the classical walk. We now extend some of their results to the arc probabilities.
Given $\epsilon$, define the *mixing time* $M_{\epsilon}$ to be the smallest $K$ such that for all $L\ge K$ and all initial states, the average distribution over $L$ steps on the arcs is $\epsilon$-close to the limiting distribution. More explicitly, let $P_j(k)$ denote the probability on the $j$-th arc, and let $$D_j=e_je_j^T$$ be its characteristic matrix. The mixing time is is the minimum positive integer $K$ such that for all time $L\ge K$ and all initial states $x$, $$\sum_j \Bigg|\frac{1}{L}\sum_{k=0}^{L-1}P_{x,j}(k)- \sum_r x^* F_r D_j F_r x\Bigg| \le \epsilon.$$
\[bound\] For a quantum walk with spectral decomposition $$U = \sum_r \lambda_r F_r,$$ we have $$\sum_j \Bigg|\frac{1}{K} \sum_{k=0}^{K-1} P_{x,j}(k)- \sum_r x^* F_r D_j F_r x\Bigg|\le \frac{2}{K} \sum_{r\ne s} \sum_j\frac{\sqrt{(F_r)_{jj} (F_s)_{jj}}}{\abs{\lambda_r - \lambda_s}}.$$
[*Proof.* ]{}First note that for any $r$ and $s$, $$\begin{aligned}
\abs{x^* F_r D_j F_s x}
&= \abs{{\langle F_r e_j, x\rangle}{\langle F_s e_j, x\rangle}}\\
&\le \sqrt{(F_r)_{jj}} \norm{x} \sqrt{(F_s)_{jj}} \norm{x}\\
&= \sqrt{(F_r)_{jj} (F_s)_{jj}}.\end{aligned}$$ Let $\lambda_r = e^{i\theta_r}$ for some $\theta_r$. By Theorem \[limit\], for the $j$-th arc, $$\begin{aligned}
\Bigg|\frac{1}{K} \sum_{k=0}^{K-1} P_{x,j}(k) - \sum_r x^* F_r D_j F_r x\Bigg|
&= \Bigg| \frac{1}{K}\sum_{k=0}^{K-1} \sum_{r\ne s} e^{ik(\theta_s-\theta_r)} x^* F_r D_j F_s x \Bigg|\\
&=\frac{1}{K} \Bigg| \sum_{r\ne s} \left(\sum_{k=0}^{K-1} e^{ik(\theta_s-\theta_r)}\right) x^*F_r D_j F_s x \Bigg|\\
&=\frac{1}{K} \Bigg| \sum_{r\ne s} \frac{1- e^{iK(\theta_s - \theta_r)}}{1-e^{i(\theta_s-\theta_r)}} x^*F_r D_j F_s x\Bigg|\\
&\le \frac{1}{K} \sum_{r\ne s} \Bigg| \frac{1- e^{iK(\theta_s - \theta_r)}}{1-e^{i(\theta_s-\theta_r)}}\Bigg| \abs{x^*F_r D_j F_s x}\\
&\le \frac{1}{K} \sum_{r\ne s} \frac{2}{\abs{e^{i\theta_s} - e^{i\theta_r}}} \abs{x^*F_r D_j F_s x}\\
&\le \frac{2}{K} \sum_{r \ne s} \frac{\sqrt{(F_r)_{jj} (F_s)_{jj}}}{\abs{\lambda_r - \lambda_s}}.\end{aligned}$$ Summing over all arcs yields the result. [ ]{}
One immediate consequence is that we can bound the mixing time of a quantum walk by its eigenvalue differences. This is an analogy to Lemma 4.3 in Aharonov [@Aharonov2000].
For a $\ell\times \ell$ transition matrix $U$ with spectral decomposition $$U = \sum_r \lambda_r F_r,$$ we have $$M_{\epsilon} \le \frac{2\ell}{\epsilon} \sum_{r\ne s} \frac{1}{\abs{\lambda_r - \lambda_s}}.$$
[*Proof.* ]{}Since $$\sum_r F_r = I,$$ for all $r$ and $j$ we have $$0\le (F_r)_{jj} \le 1.$$ Lemma \[bound\] reduces to $$\sum_j \Bigg|\frac{1}{K} \sum_{k=0}^{K-1} P_{x,j}(k)- \sum_r x^* F_r D_j F_r x\Bigg| \le \frac{2\ell}{K} \sum_{r\ne s} \frac{1}{\abs{\lambda_r - \lambda_s}}.$$ Thus for all $K$ such that $$K\ge \frac{2\ell}{\epsilon} \sum_{r\ne s} \frac{1}{\abs{\lambda_r - \lambda_s}},$$ the right hand side is bounded above by $\epsilon$. [ ]{}
Average Mixing Matrix \[avgmix\]
================================
Let $U$ be the transition matrix of a discrete-time quantum walk. In this section, we consider the probabilities on the arcs only. They are given by the entries of $$U^k \circ {{\mkern2mu\overline{\mkern-2muU^k}}}.$$ Theorem \[limit\] implies that the average probability distribution $$\frac{1}{K} \sum_{k=0}^K U^k \circ {{\mkern2mu\overline{\mkern-2muU^k}}}$$ converges to the following matrix $$\sum_r F_r \circ {{\mkern2mu\overline{\mkern-2muF_r}}}.$$ Note that this is a real symmetric matrix. Following Godsil’s notion for continuous-time quantum walks [@Godsil2011], we will denote this limit by $\widehat{M}$, and refer to it as the *average mixing matrix*. In [@Godsil2011], Godsil established several properties of the continuous-time average mixing. We extend some of his results to discrete-time quantum walks.
The first observation is that $\widehat{M}$ is doubly-stochastic, in both the continuous and discrete cases. Thus each column of $\widehat{M}$ represents a probability distribution. For the $j$-th column of $\widehat{M}$, we define its *entropy* to be the negative expectation of the logarithm of its entries, that is, $$-\sum_{\ell}\widehat{M}_{\ell j} \log\left(\widehat{M}_{\ell j}\right)$$ The entropy reaches its maximum if and only if the probability distribution with respect to the initial state $e_j$ is uniform. We also define the *total entropy* to be the sum of the entropies over all columns, although it is not an entropy. This invariant has been applied in structural pattern recognition. For example, Bai et al [@Bai2016] proposed a graph signature based on the entropy of the average mixing matrix of a graph. According to the experimental results, this entropic measure allows us to distinguish different structures.
In the continuous case, the average mixing matrix is positive semidefinite with eigenvalues no greater than one [@Godsil2011]. We show that the same statement holds for the discrete average mixing matrix.
The average mixing matrix $\widehat{M}$ is positive semidefinite, and its eigenvalues lie in $[0,1]$.
[*Proof.* ]{}Since $F_r$ is positive semidefinite, its complex conjugate $F_r$ is also positive semidefinite. As a principal submatrix of $F_r \otimes {{\mkern2mu\overline{\mkern-2muF_r}}}$, the Schur product $F_r\circ {{\mkern2mu\overline{\mkern-2muF_s}}}$ is positive semidefinite. Hence the eigenvalues of $\widehat{M}$ are non-negative. It follows from $$I= I\circ I=\left(\sum_r F_r\right) \left(\sum_s {{\mkern2mu\overline{\mkern-2muF_s}}}\right)=\widehat{M}+\sum_{r\ne s}F_r \circ {{\mkern2mu\overline{\mkern-2muF_s}}}$$ that the eigenvalues of $\widehat{M}$ are at most 1. On the other hand, since $\widehat{M}$ is doubly stochastic, ${{\bf1}}$ is an eigenvector of $\widehat{M}$ for the eigenvalue $1$. [ ]{}
If all the entries of $\widehat{M}$ are equal, we say the quantum walk has *uniform average mixing*. For example, the shunt-decomposition walk on $K_{3,3}$ with linear orders $$\begin{gathered}
f_1(1)=4,\quad f_1(2)=5,\quad f_1(3)=6;\\
f_2(1)=5,\quad f_2(2)=6,\quad f_2(3)=4;\\
f_3(1)=6,\quad f_3(2)=4,\quad f_3(3)=5;\\
f_4(1)=2,\quad f_4(2)=1,\quad f_4(3)=3;\\
f_5(1)=3,\quad f_5(2)=2,\quad f_5(3)=1;\\
f_6(1)=1,\quad f_6(2)=3,\quad f_6(3)=2,\end{gathered}$$ and the following coin $$C = \frac{1}{7}{\begin{pmatrix}
-2 & 3 & 6\\
6 & -2 & 3\\
3 & 6 & -2 \end{pmatrix}}$$ admits uniform average mixing. It indicates that in the limit, regardless of the initial arc, the quantum walk may be found on any arc with the same probability. Our next goal is to establish necessary and sufficient conditions for uniform average mixing to occur.
Suppose the spectral decomposition of $U$ is $$U = \sum_r \theta_r F_r.$$ We say $U$ is *walk-regular* if $F_r$ has constant diagonal for each $r$.
Let $U$ be an $\ell\times \ell$ unitary matrix. If $m_r$ is the multiplicity of the $r$-th eigenvalue of $U$, then $$\operatorname{tr}(\widehat{M})\ge \frac{1}{\ell} \sum_r m_r^2.$$ Further, equality holds if and only if $U$ is walk-regular. \[tr\]
[*Proof.* ]{}Consider the spectral decomposition $$U = \sum_r \lambda_r F_r.$$ Since $F_r$ is positive semidefinite, its diagonal entries are non-negative and $$\operatorname{tr}(F_r) = m_r.$$ By the Cauchy-Schwarz inequality, $$\operatorname{tr}(F_r \circ {{\mkern2mu\overline{\mkern-2muF_r}}}) \ge \frac{1}{\ell} \operatorname{tr}(F_r)^2 = \frac{1}{\ell} m_r^2.$$ Hence $$\operatorname{tr}(\widehat{M}) \ge \frac{1}{\ell} \sum_r m_r ^2.$$ Equality holds if and only if each $F_r$ has constant diagonal $m_r/\ell $. [ ]{}
\[secd\] For any unitary matrix $U$ we have $\operatorname{tr}(\widehat{M}) \ge 1$. Equality holds if and only if $U$ is walk-regular and has simple eigenvalues.
[*Proof.* ]{}The inequality follows from Lemma \[tr\] and $$\sum_r m_r = \ell.$$ If equality holds, then $m_r=1$ for all $r$, and $U$ is walk-regular. [ ]{}
The following statements are equivalent.
(i) The quantum walk admits uniform average mixing.
(ii) $\operatorname{tr}(\widehat{M})=1$.
(iii) $U$ is walk-regular with simple eigenvalues.
[*Proof.* ]{}Let $U$ be an $\ell \times \ell$ unitary matrix. To see that (i) implies (ii), note that if uniform average mixing occurs, then all entries of $\widehat{M}$ are $1/\ell$, so $\operatorname{tr}(\widehat{M})=1$. Corollary \[secd\] shows that (ii) implies (iii). Now suppose (iii) holds. Then the spectral decomposition of $U$ is $$U = \sum_r z_r z_r^*,$$ where $z_r$ is an normalized eigenvector of $U$ for the eigenvalue $\theta_r$. Since $U$ is walk-regular, for each $r$, the entries of $x_r$ have the same absolute value. Thus $$\widehat{M} = \sum_r (z_r z_r^*) \circ (z_r z_r^*) = \sum_r (z_r \circ z_r) (z_r \circ z_r)^*$$ is flat. Therefore (iii) implies (i). [ ]{}
The average mixing matrix $\widehat{M}$ records the limiting probability on each arc, given that the walk started with an arc. One may also compute $$\lim_{K\to \infty} \frac{1}{K} \sum_{k=0}^{K-1} P_{x, S}(k),$$ that is, the limiting probability on a set $S$ of arcs, given initial state $x$. In [@Aharonov2000], Aharonov et al discussed when the limiting probability distribution is uniform over all vertices. We show that this is guaranteed whenever $\widehat{M}$ is flat, regardless of the initial state.
Suppose $U$ has simple eigenvalues with corresponding eigenvectors ${z_{1},z_{2},\ldots,z_{\ell}}$. Let $$x = \sum_r a_r z_r,$$ where $\sum_r \abs{a_r}^2=1$. Then the limiting probability that the quantum walk is on $S$, given initial state $x$, is $$\lim_{K\to \infty} \frac{1}{K} \sum_{k=0}^{K-1} P_{x, S}(k) = \sum_r \abs{a_r}^2 z_r^* D_S z_r.$$ Moreover, if $U$ is walk-regular, then the limiting probability distribution is constant over $V(X)$, regardless of the initial state.
[*Proof.* ]{}The first statement follows from Theorem \[lim\_xS\]. Now, if $U$ is walk-regular with simple eigenvalues, then each eigenvector $z_r$ is flat, and $z_r^* D_S z_r$ depends only on the size of $S$. [ ]{}
Finally, we prove some algebraic properties of the average mixing matrix. They rely on the well-known fact that a commutative semisimple matrix algebra with identity has a basis of orthogonal idempotents. The following is a useful lemma in proving the existence of some phenomenon of quantum walks.
If the entries of $U$ are algebraic over ${{\mathbb Q}}$, then the entries of $\widehat{M}$ are algebraic over ${{\mathbb Q}}$.
[*Proof.* ]{}Suppose $U$ has algebraic entries. Then its eigenvalues are all algebraic. Let ${{\mathbb F}}$ be the smallest field containing the eigenvalues of $U$. Let $\operatorname{\mathcal{B}}$ be the matrix algebra generated by $U$ over ${{\mathbb F}}$. To show that $\operatorname{\mathcal{B}}$ is semisimple, pick $N\in \operatorname{\mathcal{B}}$ with $N^2=0$. Since $U$ is unitary, the algebra $\operatorname{\mathcal{B}}$ is closed under conjugate transpose and contains the identity. It follows from $(N^*)^2=0$ that $$\begin{aligned}
0&= \operatorname{tr}((N^*)^2N^2)\\
&= \operatorname{tr}(N^*N N^*N)\\
&= \operatorname{tr}((N^*N)^*(N^*N)).\end{aligned}$$ Thus $N^*N=0$. Applying the trace again to $N^*N$, we see that $N=0$. Therefore, the spectral idempotents $F_r$ of $U$ are polynomials in $U$ with algebraic coefficients. Hence the entries in $$\widehat{M} = \sum_r F_r \circ {{\mkern2mu\overline{\mkern-2muF_r}}}$$ are algebraic over ${{\mathbb Q}}$. [ ]{}
In continuous-time quantum walks, rationality has been used in different ways to determine which graphs admit instantaneous uniform mixing [@Godsil2014a]. It is also known that the entries of a continuous average mixing matrix are all rational [@Godsil2011]. We show that the discrete average mixing matrix enjoys the same property, given that all entries of the transition matrix are rational.
If the entries of $U$ are rational, then the entries of $\widehat{M}$ are rational.
[*Proof.* ]{}Let ${{\mathbb F}}$ be the smallest field containing the eigenvalues of $U$. Let $\sigma$ be an automorphism of ${{\mathbb F}}$. Since $U$ is rational, we have $$U=U^{\sigma}=\sum_r \lambda_r^{\sigma} F_r^{\sigma}.$$ Moreover, since $\lambda_r^{\sigma}$ is also an eigenvalue of $U$, the set of idempotents $\{F_r\}$ is closed under field automorphisms. Thus $$\widehat{M}=\sum_r F_r\circ F_r^T$$ is fixed by all automorphisms of ${{\mathbb F}}$ and must be rational. [ ]{}
Traces and Total Entropies
==========================
In this section, we present some numerical evidence for how the structure of a graph affects the limiting distribution of a quantum walk. The graphs we test are all cubic on up to 8 vertices. The parameter of the walk we choose is $\operatorname{tr}(\widehat{M})$, as every diagonal entry of the average mixing matrix represents the probability of a walk returning to the arc it started with. Thus, a higher trace indicates a higher tendency for the walk to stay at home.
Rotation System\[data\_arcrev\]
-------------------------------
For each graph, we enumerate all the rotation systems, and investigate the associated arc-reversal walks. Since these graphs are all cubic, the following circulant with simple eigenvalues is a candidate for the coin: $$C = \frac{1}{7}{\begin{pmatrix}
-2 & 3 & 6\\
6 & -2 & 3\\
3 & 6 & -2 \end{pmatrix}}.$$ We then compute the average mixing matrix and its trace for each rotation system.
There are $2^n$ rotation systems for a cubic graph on $n$ vertices. The following table provides the number of rotation systems with the same genus and the same trace. For cubic graphs that do not have a name, we put their graph6 strings. It appears that, for the same graph, embeddings on higher-genus surfaces give lower traces. We omit the data for cubic graphs on 12 vertices. The same pattern is found with only few exceptions.
graph genus trace number of rotation systems
------------------ ------- ------------ ----------------------------
$K_4$ $0$ $3.000000$ $2$
$K_4$ $1$ $1.753644$ $1$
$K_4$ $1$ $1.694295$ $1$
$K_{3,3}$ $1$ $2.201010$ $36$
$K_{3,3}$ $1$ $2.111111$ $4$
$K_{3,3}$ $2$ $1.052644$ $24$
$K_2\square K_3$ $0$ $3.255278$ $2$
$K_2\square K_3$ $1$ $2.105870$ $12$
$K_2\square K_3$ $1$ $2.089084$ $6$
$K_2\square K_3$ $1$ $1.932964$ $2$
$K_2\square K_3$ $1$ $1.918699$ $12$
$K_2\square K_3$ $1$ $1.866536$ $6$
$K_2\square K_3$ $2$ $1.340085$ $12$
$K_2\square K_3$ $2$ $1.187163$ $12$
$Q_3$ $0$ $4.500000$ $2$
$Q_3$ $1$ $3.000000$ $8$
$Q_3$ $1$ $2.744344$ $16$
$Q_3$ $1$ $2.625302$ $24$
$Q_3$ $1$ $2.446501$ $6$
$Q_3$ $2$ $1.980844$ $8$
$Q_3$ $2$ $1.746199$ $48$
$Q_3$ $2$ $1.694728$ $24$
$Q_3$ $2$ $1.694295$ $8$
$Q_3$ $2$ $1.688522$ $16$
$Q_3$ $2$ $1.680899$ $48$
$Q_3$ $2$ $1.679098$ $24$
$Q_3$ $2$ $1.524169$ $24$
GCrb\`o $1$ $2.520693$ $16$
GCrb\`o $1$ $2.513717$ $8$
GCrb\`o $1$ $2.447996$ $8$
GCrb\`o $1$ $2.077332$ $16$
GCrb\`o $1$ $2.00942$ $8$
GCrb\`o $2$ $1.914302$ $6$
GCrb\`o $2$ $1.886742$ $16$
GCrb\`o $2$ $1.752042$ $32$
GCrb\`o $2$ $1.658131$ $2$
GCrb\`o $2$ $1.650867$ $16$
GCrb\`o $2$ $1.615907$ $16$
GCrb\`o $2$ $1.599010$ $16$
GCrb\`o $2$ $1.598366$ $16$
GCrb\`o $2$ $1.586057$ $16$
GCrb\`o $2$ $1.566265$ $16$
GCrb\`o $2$ $1.557058$ $32$
GCrb\`o $2$ $1.460324$ $16$
GCZJd\_ $0$ $3.019811$ $2$
GCZJd\_ $1$ $2.542603$ $2$
GCZJd\_ $1$ $2.470562$ $4$
GCZJd\_ $1$ $2.429130$ $4$
GCZJd\_ $1$ $2.362862$ $4$
GCZJd\_ $1$ $2.354247$ $8$
GCZJd\_ $1$ $2.326463$ $2$
GCZJd\_ $1$ $2.321465$ $8$
GCZJd\_ $1$ $2.299543$ $4$
GCZJd\_ $1$ $2.271257$ $4$
GCZJd\_ $1$ $2.259590$ $8$
GCZJd\_ $1$ $2.250146$ $4$
GCZJd\_ $1$ $2.214541$ $4$
GCZJd\_ $1$ $2.178556$ $4$
GCZJd\_ $1$ $2.115562$ $2$
GCZJd\_ $1$ $2.079810$ $8$
GCZJd\_ $1$ $1.75710$ $8$
GCZJd\_ $2$ $1.714738$ $8$
GCZJd\_ $2$ $1.713346$ $4$
GCZJd\_ $2$ $1.709795$ $8$
GCZJd\_ $2$ $1.690956$ $4$
GCZJd\_ $2$ $1.685911$ $8$
GCZJd\_ $2$ $1.649712$ $8$
GCZJd\_ $2$ $1.622273$ $8$
GCZJd\_ $2$ $1.618072$ $8$
GCZJd\_ $2$ $1.605598$ $8$
GCZJd\_ $2$ $1.600225$ $8$
GCZJd\_ $2$ $1.599082$ $8$
GCZJd\_ $2$ $1.571617$ $8$
GCZJd\_ $2$ $1.571597$ $8$
GCZJd\_ $2$ $1.563497$ $8$
GCZJd\_ $2$ $1.560172$ $8$
GCZJd\_ $2$ $1.553949$ $4$
GCZJd\_ $2$ $1.541723$ $4$
GCZJd\_ $2$ $1.535939$ $8$
GCZJd\_ $2$ $1.531083$ $8$
GCZJd\_ $2$ $1.521709$ $4$
GCZJd\_ $2$ $1.505587$ $4$
GCZJd\_ $2$ $1.505097$ $8$
GCZJd\_ $2$ $1.486416$ $4$
GCZJd\_ $2$ $1.468886$ $4$
GCZJd\_ $2$ $1.466287$ $4$
GCZJd\_ $2$ $1.457042$ $8$
GCZJd\_ $2$ $1.421791$ $4$
GCXmd\_ $0$ $3.681253$ $4$
GCXmd\_ $1$ $2.560923$ $16$
GCXmd\_ $1$ $2.553190$ $4$
GCXmd\_ $1$ $2.535164$ $16$
GCXmd\_ $1$ $2.502552$ $8$
GCXmd\_ $1$ $2.351188$ $16$
GCXmd\_ $1$ $2.331725$ $16$
GCXmd\_ $1$ $2.278575$ $8$
GCXmd\_ $1$ $2.150255$ $8$
GCXmd\_ $2$ $1.769929$ $16$
GCXmd\_ $2$ $1.746908$ $16$
GCXmd\_ $2$ $1.742805$ $32$
GCXmd\_ $2$ $1.660490$ $32$
GCXmd\_ $2$ $1.619199$ $8$
GCXmd\_ $2$ $1.544993$ $16$
GCXmd\_ $2$ $1.541696$ $8$
GCXmd\_ $2$ $1.534019$ $16$
GCXmd\_ $2$ $1.466143$ $8$
GCXmd\_ $2$ $1.450838$ $8$
GCY\^B\_ $1$ $2.402758$ $24$
GCY\^B\_ $1$ $2.351016$ $24$
GCY\^B\_ $1$ $2.263351$ $12$
GCY\^B\_ $1$ $2.018540$ $4$
GCY\^B\_ $2$ $1.849363$ $12$
GCY\^B\_ $2$ $1.805881$ $24$
GCY\^B\_ $2$ $1.775098$ $24$
GCY\^B\_ $2$ $1.766025$ $24$
GCY\^B\_ $2$ $1.736332$ $12$
GCY\^B\_ $2$ $1.714296$ $8$
GCY\^B\_ $2$ $1.699936$ $4$
GCY\^B\_ $2$ $1.676129$ $24$
GCY\^B\_ $2$ $1.656324$ $12$
GCY\^B\_ $2$ $1.562093$ $12$
GCY\^B\_ $2$ $1.519301$ $12$
GCY\^B\_ $2$ $1.455918$ $24$
We also compute the total entropies and present some data in the following table. The total entropy seems to distinguish different embeddings, and for most of the time it increases as the genus increases.
graph rotation system genus entropy
------------------ ---------------------------------------------------------- ------- -------------
$K_4$ $0: (1, 2, 3), 1: (0, 3, 2), 2: (0, 1, 3), 3: (0, 2, 1)$ $0$ $25.364055$
$K_4$ $0: (1, 2, 3), 1: (0, 2, 3), 2: (0, 1, 3), 3: (0, 1, 2)$ $1$ $27.490608$
$K_4$ $0: (1, 2, 3), 1: (0, 3, 2), 2: (0, 1, 3), 3: (0, 1, 2)$ $1$ $27.763049$
$K_{3,3}$ $1$ $47.42653$
$K_{3,3}$ $1$ $47.862470$
$K_{3,3}$ $2$ $52.001943$
$K_2\square K_3$ $0$ $45.68992$
$K_2\square K_3$ $1$ $48.861877$
$K_2\square K_3$ $1$ $48.864165$
$K_2\square K_3$ $1$ $48.981188$
$K_2\square K_3$ $1$ $49.081280$
$K_2\square K_3$ $1$ $49.692402$
$K_2\square K_3$ $2$ $51.692584$
$K_2\square K_3$ $2$ $51.928224$
$Q_3$ 0 $62.411249$
$Q_3$ $1$ $67.363643$
$Q_3$ $1$ $68.359584$
$Q_3$ $1$ $69.625653$
$Q_3$ $1$ $69.919303$
$Q_3$ $1$ $70.331970$
$Q_3$ $2$ $71.633476$
$Q_3$ $2$ $72.161631$
$Q_3$ $2$ $72.270262$
$Q_3$ $2$ $72.758589$
$Q_3$ $2$ $73.548157$
$Q_3$ $2$ $73.595086$
$Q_3$ $2$ $75.176378$
\[entropy\]
Shunt-Decompositions \[data\_shunt\]
------------------------------------
We now switch to the shunt-decomposition model with the following $d\times d$ coin: $$B = \left(\frac{1}{\sqrt{d}} e^{2(j-k)^2\pi i /d}\right)_{jk}.$$ This is a unitary circulant for odd $d$. When $d=3$, it is self-congruent under all permutations, so we do not need to specify an ordering of the shunts to run a quantum walk. Again, for each graph, we enumerate all shunt-decompositions, and compute the average mixing matrices and their traces.
The following table lists only one shunt-decomposition for each cycle structure. The data indicates that for the same graph, the symmetric shunt-decompositions, if any, always gives the highest trace.
graph shunt-decomposition trace
------------------- ------------------------------------------------------------------------ ------------
$K_4$ $\{(0,1)(2,3), (0,2,1,3), (0,3,1,2)\}$ $2.3333$
$K_4$ $\{(0,1)(2,3), (0,2)(1,3), (0,3)(1,2)\}$ $2.6667$
$K_{3,3}$ $\{(0,3)(1,5,2,4), (0,4,2,5)(1,3), (0,5,1,4)(2,3)\}$ $2.3333$
$K_{3,3}$ $\{(0,3)(1,4)(2,5), (0,4,2,3,1,5), (0,5,1,3,2,4)\}$ $2.3333$
$K_{3,3}$ $\{(0,3,2,5,1,4), (0,4,2,3,1,5), (0,5,2,4,1,3)\}$ $3.6667$
$K_{3,3}$ $\{(0,3)(1,5)(2,4), (0,4)(1,3)(2,5), (0,5)(1,4)(2,3)\}$ $3.6667$
$K_2 \square K_3$ $\{(0,2,4)(1,3,5), (0,3)(1,4,2,5), (0,4,1,3,5,2)\}$ $1.6630$
$K_2 \square K_3$ $\{(0,2)(1,4)(3,5), (0,3,1,5,2,4), (0,4,2,5,1,3)\}$ $1.7482$
$K_2 \square K_3$ $\{(0,2,4)(1,3,5), (0,4,2)(1,5,3), (0,3)(1,4)(2,5)\}$ $2.3665$
$K_2 \square K_3$ $\{(0,2)(1,4)(3,5), (0,3)(1,5)(2,4), (0,4)(1,3)(2,5)\}$ $2.6458$
$Q_3$ $\{(0,4,1,5)(2,7)(3,6), (0,5,3,7,1,4)(2,6), (0,6)(2,4)(1,5,3,7)\}$ $1.7987$
$Q_3$ $\{(0,4,2,6)(1,5,3,7), (0,5,1,4)(2,7)(3,6), (0,6,2,4,1,7,3,5)\}$ $1.8738$
$Q_3$ $\{(0,4,1,5)(2,6,3,7), (0,5,1,7,3,6)(2,4), (0,6,2,7,1,4)(3,5)\}$ $1.9065$
$Q_3$ $\{(0,4,1,7,2,6,3,5), (0,5,3,6,2,7,1,4), (0,6)(1,5)(2,4)(3,7)\}$ $2.3241$
$Q_3$ $\{(0,4,2,6)(1,7,3,5), (0,5,3,6,2,7,1,4), (0,6,3,7,2,4,1,5)\}$ $2.4939$
$Q_3$ $\{(0,4)(1,5)(2,6)(3,7), (0,5,3,6)(1,7)(2,4), (0,6,3,5)(1,4)(2,7)\}$ $2.5232$
$Q_3$ $\{(0,4)(1,5)(2,6)(3,7), (0,5,3,6)(1,4,2,7), (0,6,3,5)(1,7,2,4)\}$ $2.6670$
$Q_3$ $\{(0,4,1,5)(2,6,3,7), (0,5,3,6)(1,4,2,7), (0,6,2,4)(1,7,3,5)\}$ $3.1667$
$Q_3$ $\{(0,4)(1,5)(2,7)(3,6), (0,5)(1,4)(2,6)(3,7), (0,6)(1,7)(2,4)(3,5)\}$ $3.6667$
GCrb\`o $\{(0,3)(1,5,2,6)(4,7), (0,4)(1,6,3,7,2,5)\}$ $1.8266$
GCrb\`o $\{(0,3,7,4,1,6,2,5),(0,4)(1,5)(2,7)(3,6),(0,5,2,6,1,4,7,3)\}$ $1.829313$
GCrb\`o $\{(0,3,7,4)(1,5,2,6),(0,4,7,2,5,1,6,3),(0,5)(1,4)(2,7,3,6)\}$ $1.888459$
GCrb\`o $\{(0,3)(1,6)(2,5)(4,7), (0,4,1,5)(2,6)(3,7), (0,3)(1,6)(2,5)(4,7)\}$ $2.098040$
GCrb\`o $\{(0,3,6,1,4,7,2,5),(0,4)(1,5)(2,6)(3,7),(0,5,2,7,4,1,6,3)\}$ $2.218323$
GCrb\`o $\{(0,3,6,2,7,4,1,5),(0,4)(1,6,3,7,2,5), (0,5,2,6,1,4,7,3)\}$ $2.286421$
GCrb\`o $\{(0,3)(1,6)(2,5)(4,7),(0,4,1,5)(2,6,3,7),(0,5,1,4)(2,7,3,6)\}$ $2.4376$
GCrb\`o $\{(0,3)(1,6)(2,5)(4,7), (0,4)(1,5)(2,6)(3,7), (0,5)(1,4)(2,7)(3,6)\}$ $3.045749$
GCZJd\_ $\{(0,3)(2,6,1,5)(4,7), (0,4)(1,6,3,7,2,5), (0,5)(1,4)(2,7,3,6)\}$ $1.826585$
GCZJd\_ $\{((0,3,7,4,1,6,2,5),(0,4)(1,5)(2,7)(3,6),(0,5,2,6,1,4,7,3)\}$ $1.829313$
GCZJd\_ $\{(0,3,7,4)(1,5,2,6),(0,4,7,2,5,1,6,3),(0,5)(1,4)(2,7,3,6)\}$ $1.888459$
GCZJd\_ $\{(0,3)(1,6)(2,5)(4,7), (0,4,1,5)(2,6)(3,7), (0,5,1,4)(2,7)(3,6)\}$ $2.098040$
GCZJd\_ $\{(0,3,6,1,4,7,2,5), (0,4)(1,5)(2,6)(3,7), (0,5,2,7,4,1,6,3)\}$ $2.218323$
GCZJd\_ $\{(0,3,6,2,7,4,1,5),(0,4)(1,6,3,7,2,5)\}$ $2.286421$
GCZJd\_ $\{(0,3)(1,6)(2,5)(4,7), (0,4,1,5)(2,6,3,7),(0,5,1,4)(2,7,3,6)\}$ $2.437614$
GCZJd\_ $\{(0,3)(1,6)(2,5)(4,7), (0,4)(1,5)(2,6)(3,7), (0,5)(1,4)(2,7)(3,6)\}$ $3.045749$
GCXmd\_ $\{(0,3,6)(2,7)(1,4,5), (0,6,3,7)(1,5,2,4), (0,7,3)(1,6)(2,5,4)\}$ $1.769242$
GCXmd\_ $\{(0,3,7,2,4,5,1,6), (0,6,1,5,4,2,7,3), (0,7)(1,4)(2,5)(3,6)\}$ $2.186914$
GCXmd\_ $\{(0,3)(1,6)(2,7)(4,5), (0,6,3,7)(1,4)(2,5), (0,7,3,6)(1,5)(2,4)\}$ $2.676998$
GCXmd\_ $\{(0,3)(1,6)(2,7)(4,5), (0,6,3,7)(1,4,2,5), (0,7,3,6)(1,5,2,4)\}$ $2.983961$
GCXmd\_ $\{(0,3)(1,6)(2,7)(4,5), (0,6)(1,4)(2,5)(3,7), (0,7)(1,5)(2,4)(3,6)\}$ $3.579491$
GCY\^B\_ $\{(0,3,7,1,4,5)(2,6), (0,5,4,2,7,3)(1,6), (0,6)(1,7,2,4)(3,5)\}$ $1.819247$
GCY\^B\_ $\{(0,3)(1,6)(2,7)(4,5), (0,5,3,7,1,4,2,6), (0,6,2,4,1,7,3,5)\}$ $1.983112$
GCY\^B\_ $\{(0,3)(1,7,2,6)(4,5), (0,5)(1,6,2,4)(3,7), (0,6)(2,7,1,4)(3,5)\}$ $2.418023$
GCY\^B\_ $\{(0,3)(1,6)(2,7)(4,5), (0,5)(1,4)(2,6)(3,7), (0,6)(17)(2,4)(3,5)\}$ $3.046722$
Open Problems
=============
Here is a list of problems we would like to answer.
1. When is the transition matrix $U$ walk-regular? When is the average mixing $\widehat{M}$ rational?
2. There are many graph products, such as the Cartesian product, the tensor product and the strong product. Is there a relation between the quantum walks on two graphs and the quantum walk on their products, given properly chosen coins?
3. Using $3\times 3$ coins, we can implement arc-reversal quantum walks on the following chain of gems $$\Phi(X), \Phi(\Phi(X))), \Phi(\Phi(\Phi(X))), \cdots$$ The sizes of these graphs increase quickly. Is there a way to assign the coins, such that the quantum walks on these graphs are determined by the quantum walk on $\Phi(X)$?
4. For a regular graph $X$, is there a coin for which the entropy of the average mixing matrix distinguishes non-isomorphic rotation systems?
5. Which shunt-decompositions/embeddings give the maximum trace/entropy of the average mixing matrix?
6. Which shunt-decompositions/embeddings give the minimum hitting time/mixing time?
7. Which shunt-decompositions/embeddings give perfect state transfer?
8. In the shunt-decomposition model, is there a relation between the cycle-structures of the shunts and the limiting distribution?
9. Can we extend the shunt-decomposition model to directed graphs where all vertices have the same out-degree?
[^1]: Department of Combinatorics and Optimization, University of Waterloo. `{cgodsil, h3zhan}@uwaterloo.ca`
[^2]: Research supported by Natural Sciences and Engineering Council of Canada (NSERC), Grant Number RGPIN-9439.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This survey paper is written with the intention of giving a mathematical introduction to filtering techniques for intermittent data assimilation, and to survey some recent advances in the field. The paper is divided into three parts. The first part introduces Bayesian statistics and its application to statistical inference and estimation. Basic aspects of Markov processes, as they typically arise from scientific models in the form of stochastic differential and/or difference equations, are covered in the second part. The third and final part describes the filtering approach to estimation of model states by assimilation of observational data into scientific models. While most of the material is of survey type, very recent advances in the field of nonlinear data assimilation covered in this paper include a discussion of Bayesian inference in the context of optimal transportation and coupling of random variables, as well as a discussion of recent advances in ensemble transform filters. References and sources for further reading material will be listed at the end of each section.'
author:
- |
Sebastian Reich\
Department of Mathematics, University of Potsdam\
\
and\
\
Colin Cotter\
Department of Aeronautics, Imperial College London
bibliography:
- 'survey.bib'
title: 'Ensemble filter techniques for intermittent data assimilation - a survey'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We establish an essentially optimal estimate for the ninth moment of the exponential sum having argument $\alp x^3+\bet x$. The first substantial advance in this topic for over $60$ years, this leads to improvements in Heath-Brown’s variant of Weyl’s inequality, and other applications of Diophantine type.'
address: 'School of Mathematics, University of Bristol, University Walk, Clifton, Bristol BS8 1TW, United Kingdom'
author:
- 'Trevor D. Wooley'
title: Mean value estimates for odd cubic Weyl sums
---
Introduction
============
This memoir concerns the mean values $$I_s(X)=\int_0^1\int_0^1\Bigl| \sum_{1\le x\le X}e(\alp x^3+\bet x)\Bigr|^s\d\alp\d\bet ,$$ where $e(z)=e^{2\pi iz}$. Estimates for $I_s(X)$ make an appearance in the literature as early as 1947, when L.-K. Hua [@Hua1947 Lemma 4.3 and Theorem 6] showed that $$\label{1.1}
I_6(X)\ll X^3(\log 2X)^9\quad \text{and}\quad I_{10}(X)\ll_\eps X^{6+\eps}.$$ These mean values have more recently been applied to obtain improvements in Weyl’s inequality and Waring’s problem (see [@Bok1994; @HB1988]), and also in investigations concerning the integral solubility of diagonal cubic equations subject to a linear slice (see [@BR2014]). Presumably, one should in general have the upper bound $I_s(X)\ll X^{s/2}+X^{s-4}$, but hitherto, the best available estimates for $I_s(X)$ are little better than those obtained from Hua’s bounds (\[1.1\]) via Hölder’s inequality. By applying the cubic case of the main conjecture in Vinogradov’s mean value theorem, recently established in [@Woo2014b], we are now able to obtain estimates for $I_s(X)$ substantially sharper than these earlier bounds.
\[theorem1.1\] For each $\eps>0$, one has $I_8(X)\ll X^{13/3+\eps}$ and $I_9(X)\ll X^{5+\eps}$.
By orthogonality, the mean value $I_6(X)$ counts the number of integral solutions of the system $$\sum_{i=1}^3(x_i^3-y_i^3)=\sum_{i=1}^3(x_i-y_i)=0,$$ with $1\le x_i,y_i\le X$ $(1\le i\le 3)$. These simultaneous equations, defining the so-called Segre cubic (see [@Seg1887]) has been the focus of vigorous investigation in recent years. Vaughan and Wooley [@VW1995] showed that $$\label{1.2}
I_6(X)=6X^3+U(X),$$ where $U(X)\asymp X^2(\log X)^5$, and de la Bretèche [@dlB2007] has obtained an asymptotic formula for $U(X)$ of the shape $U(X)\sim CX^2(\log X)^5$, for a suitable positive constant $C$. By interpolating between (\[1.2\]) and the $10^{\text{th}}$-moment of Brüdern and Robert [@BR2014 Theorem 2], one would obtain the estimates $$I_8(X)\ll X^{9/2}(\log X)^{-1}\quad \text{and}\quad I_9(X)\ll X^{21/4}(\log X)^{-3/2}.$$ These estimates are sharper by a factor $X^\eps$ than the estimates that would stem from Hua’s bounds (\[1.1\]), whereas our new estimates save $X^{1/6-\eps}$ and $X^{1/4-\eps}$ in the respective cases. Indeed, our new bound $I_9(X)\ll X^{5+\eps}$ falls short of the best possible bound $I_9(X)\ll X^5$ only by a factor $X^\eps$.
The estimates recorded in Theorem \[theorem1.1\] are consequences of a minor arc bound that will likely be of greater utility than the former in applications of the Hardy-Littlewood method. In order to describe bounds of this type, we must introduce some additional notation. When $Q$ is a real number with $1\le Q\le X^{3/2}$, we define the major arcs $\grM(Q)$ to be the union of the intervals $$\grM(q,a)=\{ \alp \in [0,1): |q\alp-a|\le QX^{-3}\},$$ with $0\le a\le q\le Q$ and $(a,q)=1$. We then define the complementary set of minor arcs $\grm(Q)$ by putting $\grm(Q)=[0,1)\setminus \grM(Q)$. Finally, we define the exponential sum $g(\alp,\bet)=g(\alp,\bet;X)$ by $$\label{1.3}
g(\alp,\bet;X)=\sum_{1\le x\le X}e(\alp x^3+\bet x),$$ and define $I_s^*(X;Q)$ for $s\in \dbN$ by putting $$\label{1.4}
I_s^*(X;Q)=\int_0^1 \int_{\grm(Q)}|g(\alp,\bet)|^s\d\alp \d\bet .$$
\[theorem1.2\] Let $Q$ be a real number with $1\le Q\le X$. Then for each $\eps>0$, one has the estimates $$I_{10}^*(X;Q)\ll X^{6+\eps}Q^{-1/3}\quad \text{and}\quad I_{12}^*(X;Q)\ll X^{8+\eps}Q^{-1}.$$
When $X^{3/4}\le Q\le X^{4/5}$, one finds from Brüdern and Robert [@BR2014 Theorem 2] that $I_{10}^*(X;Q)\ll X^6(\log X)^{-2}$, which saves a factor $(\log X)^2$ over the lower bound of order $X^6$ for the corresponding major arc estimate. Theorem \[theorem1.2\], meanwhile, would save a power of $X$. Indeed, since $I_9(X)\gg X^5$, the bound $I_{12}^*(X;X)\ll X^{7+\eps}$, that stems from Theorem \[theorem1.2\], can be construed as supplying a Weyl estimate $g(\alp,\bet)\ll X^{2/3+\eps}$ on average for $\alp\in \grm(X)$. A direct application of Weyl’s inequality (see [@Vau1997 Lemma 2.4]) would show only that $g(\alp,\bet)\ll X^{3/4+\eps}$.
We would argue that the progress represented in our improved estimates for moments of $g(\alp,\bet)$ justifies an account based on its merit alone. However, we take this opportunity to record an application of Theorem \[theorem1.2\] to Heath-Brown’s variant of Weyl’s inequality. In this context, when $k$ is a natural number, we consider the exponential sum $f(\alp)=f_k(\alp;X)$ defined by $$f_k(\alp;X)=\sum_{1\le x\le X}e(\alp x^k).$$
\[theorem1.3\] Let $k\ge 6$, and suppose that $\alp\in \dbR$, $a\in \dbZ$ and $q\in \dbN$ satisfy $(a,q)=1$ and $|\alp-a/q|\le q^{-2}$. Then for each $\eps>0$, one has $$f_k(\alp;X)\ll X^{1+\eps}\Tet^{2^{-k}}+X^{1+\eps}(\Tet/X)^{\frac{2}{3}2^{-k}},$$ where $\Tet=q^{-1}+X^{-3}+qX^{-k}$.
The conclusion of [@HB1988 Theorem 1] delivers a bound analogous to that of Theorem \[theorem1.3\] of the shape $$\label{1.5}
f_k(\alp;X)\ll X^{1+\eps}(X\Tet)^{\frac{4}{3}2^{-k}}.$$ We note that Boklan [@Bok1994] has applied Hooley $\Del$-functions to replace the factor $X^\eps$ here by a power of $\log X$. A comparison between these estimates is perhaps not so transparent. Suppose then that $\tet$ is a real number with $0\le \tet\le k/2$, and that $a\in \dbZ$ and $q\in \dbN$ satisfy $(a,q)=1$ and $q+X^k|q\alp-a|\asymp X^\tet$. It is a consequence of Dirichlet’s theorem on Diophantine approximation that, given $\alp \in \dbR$, one can choose $a$ and $q$ in such a manner for some $\tet\le k/2$. One finds that the conclusion of Theorem \[theorem1.3\] has strength equal to that of Heath-Brown’s bound for $3\le \tet\le k/2$. When $2<\tet<3$, meanwhile, Theorem \[theorem1.3\] delivers the bound $f_k(\alp;X)\ll X^{1+\eps-\frac{2}{3}(1+\tet)2^{-k}}$, which is superior both to the bound $f_k(\alp;X)\ll X^{1+\eps-\frac{4}{3}(\tet-1)2^{-k}}$ stemming from Heath-Brown’s bound (\[1.5\]), and also to the classical version of Weyl’s inequality, which yields $f_k(\alp;X)\ll X^{1+\eps-2^{1-k}}$ (see [@Vau1997 Lemma 2.4]). Both Theorem \[theorem1.3\] and (\[1.5\]) are weaker than the classical version of Weyl’s inequality for $0<\tet<2$, though Theorem \[theorem1.3\] remains non-trivial throughout this range.
By a standard transference principle (see Exercise 2 of [@Vau1997 §2.8]), the conclusion of Theorem \[theorem1.3\] may be extended to a superficially more general conclusion which improves the first assertion of [@HB1988 Theorem 1] for ranges of parameters analogous to those discussed above.
\[corollary1.4\] Let $k\ge 6$, and suppose that $\alp\in \dbR$, $a\in \dbZ$ and $q\in \dbN$ satisfy $(a,q)=1$. Then one has $$f_k(\alp;X)\ll X^{1+\eps}\Phi^{2^{-k}}+X^{1+\eps}(\Phi/X)^{\frac{2}{3}2^{-k}},$$ where $$\Phi=(q+X^k|q\alp-a|)^{-1}+X^{-3}+(q+X^k|q\alp-a|)X^{-k}.$$
We finish by directing the reader to a couple of immediate applications of Theorems \[theorem1.1\] and \[theorem1.2\], the proofs of which, amounting to routine applications of the circle method, we omit. First we consider the solubility of diagonal cubic equations constrained by a linear slice. When $s\in \dbN$, consider fixed integers $a_j,b_j$ $(1\le j\le s)$. Define $N(B)=N(B;\bfa,\bfb)$ to be the number of integral solutions of the simultaneous equations $$\label{1.6}
\sum_{j=1}^sa_jx_j^3=\sum_{j=1}^sb_jx_j=0,$$ with $|x_j|\le B$ $(1\le j\le s)$. Then by incorporating the $10^{\rm{th}}$-moment estimate of Theorem \[theorem1.2\] into the methods described in Brüdern and Robert [@BR2014 §8], one obtains the following conclusion.
\[theorem1.5\] Let $s\ge 10$ and suppose that $a_j\ne 0$ $(1\le j\le s)$. Suppose in addition that the pair of equations (\[1.6\]) has non-singular solutions both in $\dbR$ and in $\dbQ_p$ for each prime number $p$. Then there are positive numbers $\calC(\bfa,\bfb)$ and $\del$ for which $$N(B;\bfa,\bfb)=\calC(\bfa,\bfb)B^{s-4}+O(B^{s-4-\del}).$$
Brüdern and Robert [@BR2014 Theorem 1] establish precisely this conclusion as the cubic case of a more general result, though with the error term $O(B^{s-4-\del})$ replaced by $O(B^{s-4}(\log B)^{-2})$. We offer no details of the proof of Theorem \[theorem1.5\], since the first estimate of Theorem 1.2 may be substituted for [@BR2014 Theorem 2] in the argument of [@BR2014 §8], without complication[^1].
Next, consider a fixed natural number $k$, and fixed coefficients $a_0,\ldots ,a_s\in \dbZ\setminus \{0\}$ and $b_1,\ldots ,b_s\in \dbZ$. By more fully exploiting the potential of the $9^{\rm{th}}$ moment estimate of Theorem \[theorem1.1\], it would be possible to apply the circle method to the problem of representing large positive integers $n$ in the shape $$\label{1.7}
F(x_1,\ldots ,x_s)+w^k=n,$$ for the class of non-degenerate cubic forms $F$ of the shape $$F(\bfx)=a_0(b_1x_1+\ldots +b_sx_s)^3+a_1x_1^3+\ldots +a_sx_s^3.$$ Thus, provided only that $s\ge 8$, for any $k\ge 1$, one can show that all sufficiently large natural numbers $n$ subject to the necessary congruence conditions are represented in the form (\[1.7\]).
The strategy for proving this assertion is to replace (\[1.7\]) by the equivalent system of equations $$\left. \begin{aligned}
a_0x_0^3+a_1x_1^3+\ldots +a_sx_s^3&=n-w^k\\
x_0-b_1x_1-\ldots -b_sx_s&=0
\end{aligned} \right\}.$$ The analysis of this system is achieved by Hölderising the associated exponential sums in order to utilise the mean value estimate $$\int_0^1\int_0^1 |g(\alp,\bet)|^{s+1}\d\alp\d\bet \ll X^{s-3+\eps},$$ valid for $s\ge 8$, together with a pedestrian application of Weyl’s inequality for the exponential sum over the $k$th power $w^k$. A routine treatment of the major arc contribution completes the analysis.
Throughout this paper, whenever $\eps$ appears in a statement, we assert that the statement holds for each $\eps>0$. Implicit constants in Vinogradov’s notation $\ll$ and $\gg$ may depend on $\eps$, and other ambient exponents such as $k$, but not on the main parameter $X$. Finally, we write $\|\tet\|$ for $\underset{m\in \dbZ}{\min}|\tet -m|$.
The basic mean value estimate
=============================
Our starting point for the proof of Theorems \[theorem1.1\] and \[theorem1.2\] is the mean value estimate supplied by the cubic case of the main conjecture in Vinogradov’s mean value theorem, established in our very recent work [@Woo2014b Theorem 1.1]. When $k$ and $s$ are natural numbers, and $X$ is a large real number, denote by $J_{s,k}(X)$ the number of integral solutions of the system $$x_1^j+\ldots +x_s^j=y_1^j+\ldots +y_s^j\quad (1\le j\le k),$$ with $1\le x_i,y_i\le X$ $(1\le i\le s)$. Then [@Woo2014b Theorem 1.1] shows that $$\label{2.1}
J_{s,3}(X)\ll X^\eps (X^s+X^{2s-6}).$$ We transform this estimate into a bound for the $12$-th moment of $g(\alp,\bet)$ restricted to the set of minor arcs $\grm(Q)$ defined in the preamble to the statement of Theorem \[theorem1.2\]. In this section we prove a number of mean value estimates for the exponential sum $g(\alp,\bet)$ defined in (\[1.3\]), beginning with a mean value of the type (\[1.4\]).
\[theorem2.1\] Suppose that $Q$ is a positive number with $Q\asymp X$. Then for each $\eps>0$, one has $I_{12}^*(X;Q)\ll X^{7+\eps}$.
When $k\in \dbN$, write $$f(\bfalp)=\sum_{1\le x\le X}e(\alp_1x+\ldots +\alp_kx^k)$$ and $$F(\bfbet ,\tet)=\sum_{1\le x\le X}e(\bet_1 x+\ldots +\bet_{k-2}x^{k-2}+\tet x^k).$$ Then it follows from orthogonality that $$J_{s,k}(X)=\int_{[0,1)^k}|f(\bfalp)|^{2s}\d\bfalp .$$ In addition, write $\grn_k(Q)$ for the set of real numbers $\alp\in [0,1)$ having the property that, whenever $q\in \dbN$ and $\| q\alp\|\le QX^{-k}$, then $q>Q$. Then the argument of the proof of [@Woo2012b Theorem 2.1] leading to the penultimate display of that proof yields the estimate $$\label{2.2}
\int_{\grn_k(Q)}\int_{[0,1)^{k-2}}|F(\bfbet,\tet)|^{2s}\d\bfbet \d\tet \ll
X^{k-2}(\log X)^{2s+1}J_{s,k}(2X).$$ By specialising to the case $k=3$ and $s=6$, we therefore deduce from (\[2.1\]) that $$\int_0^1\int_{\grm (Q)}|g(\alp,\bet)|^{12}\d\alp \d\bet \ll X(\log X)^{13}J_{6,3}(2X)\ll X^{7+\eps}.$$ This completes the proof of the lemma.
We remark that a more careful analysis of the proof of [@Woo2012b Theorem 2.1] would reveal that, without restriction on $Q$, one may replace the estimate (\[2.2\]) by the bound $$\int_{\grn_k(Q)}\int_{[0,1)^{k-2}}|F(\bfbet,\tet)|^{2s}\d\bfbet \d\tet \ll
X^{k-1+\eps}(Q^{-1}+X^{-1}+QX^{-k})J_{s,k}(2X).$$ Such an estimate would suffice to establish the bound $I_{12}^*(X;Q)\ll X^{8+\eps}Q^{-1}$. We will recover this estimate from Theorem \[theorem2.1\] and Lemma \[lemma2.3\] below in a manner that will likely prove more transparent for the reader.
By way of comparison, it follows from [@Hua1947 Theorem 6] that $I_{10}(X)\ll X^{6+\eps}$. Applying this estimate in combination with Weyl’s inequality (see [@Vau1997 Lemma 2.4]) when $Q\asymp X$, one would obtain the upper bound $$I_{12}^*(X;Q)\ll \Bigl( \sup_{\alp\in \grm(Q)}|g(\alp,\bet)|\Bigr)^2I_{10}(X)\ll X^{15/2+\eps},$$ in place of the conclusion of Theorem \[theorem2.1\]. The superiority of our new estimate is clear.
We next establish some auxiliary major arc estimates. It is useful in this context to introduce some additional notation. We define the function $\Psi(\alp)$ for $\alp\in [0,1)$ by putting $$\Psi(\alp)=(q+X^3|q\alp -a|)^{-1},$$ when $\alp\in \grM(q,a)\subseteq \grM(\tfrac{1}{2}X^{3/2})$, and otherwise by taking $\Psi(\alp)=0$.
\[lemma2.2\] Let $Q$ be a positive number with $Q\asymp X$, and suppose that $\alp \in \grM(Q)$. Then for each $\eps>0$, one has $$\int_0^1|g(\alp,\bet)|^4\d\bet \ll X^{3+\eps}\Psi(\alp).$$
By orthogonality, one has $$\begin{aligned}
\int_0^1|g(\alp,\bet)|^4\d\bet &=\sum_{\substack{1\le x_1,x_2,x_3,x_4\le X\\ x_1+x_2=x_3+x_4}}
e((x_1^3+x_2^3-x_3^3-x_4^3)\alp )\\
&=\sum_{\substack{1\le x_1,x_2,x_3\le X\\ 1\le x_1+x_2-x_3\le X}}
e(-3(x_1+x_2)(x_1-x_3)(x_2-x_3)\alp).\end{aligned}$$ The change of variables $$u_1=x_2-x_3,\quad u_2=x_1-x_3,\quad u_3=x_1+x_2$$ therefore reveals that $$\int_0^1|g(\alp,\bet)|^4\d\bet =\sum_{-X<u_1,u_2,u_3\le 2X}e(-3u_1u_2u_3\alp),$$ in which the summation over $\bfu$ is subject to the condition that each of $$u_2+u_3-u_1,\quad u_3+u_1-u_2,\quad u_3-u_1-u_2\quad \text{and}\quad u_1+u_2+u_3$$ is even, and lies in the interval $[1,2X]$. For a fixed choice of $u_1$ and $u_2$, the sum over $u_3$ consequently amounts either to an empty sum, or else to a sum over an arithmetic progression modulo $2$ lying in an interval of length at most $3X$. Thus we deduce that $$\label{2.3}
\int_0^1|g(\alp,\bet)|^4\d\bet \ll \sum_{1\le u_1,u_2\le 2X}\min \{ X,\|6\alp u_1u_2\|^{-1}\}.$$
Suppose that $a\in \dbZ$ and $q\in \dbN$ satisfy the conditions $(a,q)=1$ and $|\alp-a/q|\le q^{-2}$. That such a rational approximation exists is a consequence of Dirichlet’s theorem. Then by making use of a divisor function estimate together with a standard reciprocal sums lemma (see, for example [@Vau1997 Lemma 2.2]), one deduces from (\[2.3\]) that $$\begin{aligned}
\int_0^1|g(\alp,\bet)|^4&\ll X^\eps \sum_{1\le y\le 24X^2}\min\{X^3/y,\|\alp y\|^{-1}\}\\
&\ll X^{3+\eps}(q^{-1}+X^{-1}+qX^{-3}).\end{aligned}$$ Hence, by a standard transference principle (see Exercise 2 of [@Vau1997 §2.8]), one finds that whenever $\alp\in [0,1)$, $b\in \dbZ$ and $r\in \dbN$ satisfy $(b,r)=1$, then $$\label{2.4}
\int_0^1|g(\alp,\bet)|^4\d\bet \ll X^{3+\eps}(\lam^{-1}+X^{-1}+\lam X^{-3}),$$ where $\lam=r+X^3|r\alp-b|$.
Suppose now that $\alp \in \grM(q,a)\subseteq \grM$. Then we have $$q+X^3|q\alp -a|\ll X,$$ and thus it follows from (\[2.4\]) that $$\int_0^1|g(\alp,\bet)|^4\d\bet \ll X^{3+\eps}(X^{-1}+\Psi(\alp))\ll X^{3+\eps}\Psi(\alp).$$ This completes the proof of the lemma.
\[lemma2.3\] Suppose that $Q$ is a positive number with $Q\asymp X$. Then for each $\eps>0$, one has $$\int_0^1\int_{\grM(Q)}|g(\alp,\bet)|^8\d\alp\d\bet \ll X^{4+\eps}.$$
Suppose that $(\alp,\bet)\in [0,1)^2$, and that $a\in \dbZ$ and $q\in \dbN$ satisfy $(a,q)=1$ and $|\alp-a/q|\le q^{-2}$. Then it follows from Weyl’s inequality (see [@Vau1997 Lemma 2.4]) that $$|g(\alp,\bet)|\ll X^{1+\eps}(q^{-1}+X^{-1}+qX^{-3})^{1/4}.$$ By applying the same transference principle that delivered (\[2.4\]), we therefore deduce that when $\alp\in \grM(q,a)\subseteq \grM(Q)$, one has $$\label{2.5}
|g(\alp,\bet)|^4\ll X^{4+\eps}\Psi(\alp).$$ By combining this estimate with the conclusion of Lemma \[lemma2.2\], therefore, we find that $$\int_0^1|g(\alp,\bet)|^8\d\bet \ll X^{4+\eps}\Psi(\alp)\int_0^1|g(\alp,\bet)|^4\d\bet
\ll X^{7+2\eps}\Psi(\alp)^2.$$ Consequently, we obtain the estimate $$\begin{aligned}
\int_0^1\int_{\grM(Q)}|g(\alp,\bet)|^8\d\alp\d\bet &\ll X^{7+\eps}
\sum_{1\le q\le Q}\sum_{a=1}^qq^{-2}\int_{-1/2}^{1/2}(1+X^3|\gam|)^{-2}\d\gam \\
&\ll X^{4+\eps}\sum_{1\le q\le Q}q^{-1}\ll X^{4+2\eps}.\end{aligned}$$ This completes the proof of the lemma.
By utilising the conclusions of Lemma \[lemma2.3\] and Theorem \[theorem2.1\], we obtain the mean value estimates recorded in Theorem \[theorem1.1\].
The estimate $I_6(X)\ll X^{3+\eps}$ is essentially classical (see [@Hua1947 Lemma 5.2]). By combining this estimate with Theorem \[theorem2.1\] via Schwarz’s inequality, one finds that $$\label{2.6}
I_9^*(X;X)\le (I_{12}^*(X;X))^{1/2}(I_6(X))^{1/2}\ll (X^{7+\eps})^{1/2}
(X^{3+\eps})^{1/2}=X^{5+\eps}.$$ Meanwhile, the trivial estimate $|g(\alp,\bet)|\le X$ combines with Lemma \[lemma2.3\] to deliver the bound $$\label{2.7}
\int_0^1\int_{\grM(X)}|g(\alp,\bet)|^9\d\alp\d\bet \le X\int_0^1\int_{\grM(X)}|g(\alp,\bet)|^8\d\alp\d\bet
\ll X^{5+\eps}.$$ Since $[0,1)$ is the union of $\grM(X)$ and $\grm(X)$, the upper bound $I_9(X)\ll X^{5+\eps}$ follows by combining (\[2.6\]) and (\[2.7\]). Finally, by interpolating between the bound just obtained and Hua’s estimate $I_6(X)\ll X^{3+\eps}$ via Hölder’s inequality, one obtains $$I_8(X)\le (I_6(X))^{1/3}(I_9(X))^{2/3}\ll (X^{3+\eps})^{1/3}(X^{5+\eps})^{2/3}=X^{13/3+\eps}.$$ This completes the proof of Theorem \[theorem1.1\].
Our last task in this section is that of establishing the minor arc bounds recorded in Theorem \[theorem1.2\].
Suppose that $Q$ is a real number with $1\le Q\le X$, and write $\grK(Q)=\grM(X)\setminus \grM(Q)$. Then since $\grm(Q)$ is the union of $\grm(X)$ and $\grK(Q)$, one finds that $$I_{12}^*(X;Q)\le I_{12}^*(X;X)+\Bigl( \sup_{\alp \in \grK(Q)}|g(\alp,\bet)|\Bigr)^4\int_0^1
\int_{\grM(X)}|g(\alp,\bet)|^8\d\alp \d\bet .$$ When $\alp\in \grM(q,a)\cap \grK(Q)$, it follows that $q+X^3|q\alp-a|>Q$. Thus we deduce from (\[2.5\]) that $$\sup_{\alp \in \grK(Q)}|g(\alp,\bet)|\ll X^{1+\eps}(Q^{-1}+X^{-1})^{1/4}\ll
X^{1+\eps}Q^{-1/4}.$$ Consequently, we find from Theorem \[theorem2.1\] and Lemma \[lemma2.3\] that $$I_{12}^*(X;Q)\ll X^{7+\eps}+(X^{4+\eps}Q^{-1})(X^{4+\eps})\ll X^{8+2\eps}Q^{-1}.$$ This confirms the second estimate recorded in Theorem \[theorem1.2\]. For the first, we apply Hölder’s inequality to interpolate between the bound just obtained, and the second estimate asserted by Theorem \[theorem1.1\]. Thus one has $$\begin{aligned}
I_{10}^*(X;Q)&\le (I_9(X))^{2/3}(I_{12}^*(X;Q))^{1/3}\\
&\ll (X^{5+\eps})^{2/3}(X^{8+\eps}Q^{-1})^{1/3}
=X^{6+\eps}Q^{-1/3}.\end{aligned}$$ This completes the proof of Theorem \[theorem1.2\].
A variant of Weyl’s inequality
==============================
We turn in this section to the proof of Theorem \[theorem1.3\], and begin by recalling the key elements of the work of Heath-Brown [@HB1988] concerning a hybrid of the methods of Weyl and of Vinogradov. For the present, suppose that $k\ge 4$, and consider the exponential sum $f(\alp)=f_k(\alp;X)$. For each integer $m$, let $\grI(m)$ denote the real interval $[mX^{-3},(m+1)X^{-3})$. Given a real number $x$, we then denote by $m=m(x)$ the integer for which $x\in \grI(m)$, and we put $\calI(x)=\grI(m(x))$. Finally, we define $$T(x)=\max_{I\subseteq [1,X]}\sup_{\alp \in \calI(x)}\max_{\bet\in [0,1]}\Bigl|
\sum_{n\in I}e(\alp n^3+\bet n)\Bigr|,$$ in which the first maximum is taken over subintervals of $[1,X]$.
Write $\kap=\tfrac{1}{6}k!2^{k-3}$. Then [@HB1988 Lemma 1] asserts that $$\label{3.1}
|f(\alp)|^{2^{k-3}}\ll X^{2^{k-3}-1}+X^{2^{k-3}-k+2+\eps}\sum_{h=1}^{\kap X^{k-3}}T(\alp h).$$ Moreover, the discussion of [@HB1988] leading just beyond [@HB1988 Lemma 4] reveals that for some real number $\bet=\bet(x)$, one has $$\label{3.2}
T(x)\ll (\log X)\sum_{l=0}^4X^{4-l}\int_{\calI(x)}\int_\bet^{\bet +X^{-1}}
\Bigl| \sum_{1\le n\le X}n^le(\xi n^3+\eta n)\Bigr| \d\eta \d\xi .$$
The relation (\[3.2\]) is the starting point for the main discussion of this section. For ease of discussion, and without loss of generality, we may suppose that $X$ is an integer. Our first step is to remove the weight $n^l$ from the innermost sum of (\[3.2\]). On recalling (\[1.3\]), we find by applying partial summation that $$\begin{aligned}
\sum_{1\le n\le X}n^le(\xi n^3+\eta n)&=\sum_{1\le n\le X}n^l\left( g(\xi,\eta;n)-g(\xi,\eta;n-1)\right) \\
&=X^lg(\xi,\eta;X)-\sum_{1\le n\le X-1}((n+1)^l-n^l)g(\xi,\eta;n).\end{aligned}$$ On substituting this relation into (\[3.2\]), we deduce that $$\begin{aligned}
T(x)\ll &\, (\log X)\sum_{l=0}^4 X^{4-l}\sum_{1\le n\le X}n^{l-1}\int_{\calI(x)}
\int_\bet^{\bet +X^{-1}}|g(\xi,\eta;n)|\d\eta \d\xi \\
&\, +(\log X)\sum_{l=0}^4X^4\int_{\calI(x)}\int_\bet^{\bet+X^{-1}}|g(\xi,\eta;X)|\d\eta \d\xi ,\end{aligned}$$ and hence $$\begin{aligned}
T(x)\ll &\,X^{3+\eps}\sum_{1\le P\le X}\int_{\calI(x)}\int_\bet^{\bet+X^{-1}}|g(\xi,\eta;P)|\d\eta \d\xi
\\
&\, +X^{4+\eps}\int_{\calI(x)}\int_\bet^{\bet+X^{-1}}|g(\xi,\eta;X)|\d\eta \d\xi .\end{aligned}$$ Thus we conclude that $$\label{3.3}
\sum_{h=1}^{\kap X^{k-3}}T(\alp h)\ll X^{4+\eps}\max_{1\le P\le X}
\sum_{h=1}^{\kap X^{k-3}}\Ttil(\alp h;P),$$ where $$\Ttil(x;P)=\int_{\calI(x)}\int_\bet^{\bet+X^{-1}}|g(\xi,\eta;P)|\d\eta \d\xi .$$
We must now consider the double integral $\Ttil(x;P)$, though we pause first to discuss some basic properties of the set $\calI(x)$. Let $x\in \dbR$, and suppose that $\calI(x)$ contains a point $\xi$ lying in $\grM(\frac{1}{12}P)$. Then there exists $a\in \dbZ$ and $q\in \dbN$ with $0\le a\le q\le \tfrac{1}{12}P$, $(a,q)=1$ and $|q\xi-a|\le \tfrac{1}{12}P^{-2}$. For all other points $\xi'\in \calI(x)$, one has $$|q\xi'-a|\le q|\xi'-\xi|+\tfrac{1}{12}P^{-2}\le qX^{-3}+\tfrac{1}{12}P^{-2}\le \tfrac{1}{6}P^{-2}.$$ Hence we obtain the relation $\calI(x)\subseteq \grM(\tfrac{1}{6}P)$. We record for future reference also the bound $$\label{3.4}
\int_{\calI(x)}\int_\bet^{\bet+X^{-1}}\d\eta \d\xi \ll X^{-4}.$$
Denote by $\grA(P)$ the set of integers $m$ with $1\le m\le X^3$ for which one has $\calI(mX^{-3})\cap \grM(\tfrac{1}{12}P)\ne \emptyset$, and define $$G_m(\bet;P)=\int_{\grI(m)}\int_\bet^{\bet+X^{-1}}|g(\xi,\eta;P)|\d\eta\d\xi .$$ Thus, we have $$\label{3.5}
G_m(\bet;P)=\Ttil(mX^{-3};P).$$ Then on recalling (\[3.4\]), we find that an application of Hölder’s inequality delivers the bound $$\sum_{m\in \grA(P)}G_m(\bet;P)^8\ll X^{-28}
\sum_{m\in \grA(P)}\int_{\grI(m)}\int_\bet^{\bet+X^{-1}}|g(\xi,\eta;P)|^8\d\eta\d\xi .$$ But $\calI(mX^{-3})\subseteq \grM(\tfrac{1}{6}P)$ whenever $m\in \grA(P)$, and hence we obtain the relation $$\sum_{m\in \grA(P)}G_m(\bet;P)^8\ll X^{-28}\int_{\grM(\frac{1}{6}P)}
\int_0^1|g(\xi,\eta;P)|^8\d\eta \d\xi .$$ We thus conclude from Lemma \[lemma2.3\] and (\[3.5\]) that $$\label{3.6}
\sum_{m\in \grA(P)}\Ttil(mX^{-3};P)^8=\sum_{m\in \grA(P)}G_m(\bet;P)^8\ll X^{\eps-24}.$$
Meanwhile, when $\calI(mX^{-3})\cap \grM(\tfrac{1}{12}P)=\emptyset$, one has $\calI(mX^{-3})\subseteq \grm(\tfrac{1}{12}P)$. Then we find in a similar manner that $$\begin{aligned}
\sum_{\substack{1\le m\le X^3\\ m\not\in \grA(P)}}G_m(\bet;P)^{12}&\ll X^{-44}
\sum_{\substack{1\le m\le X^3\\ m\not\in \grA(P)}}\int_{\grI(m)}
\int_\bet^{\bet+X^{-1}}|g(\xi,\eta;P)|^{12}\d\eta \d\xi \\
&\ll X^{-44}\int_{\grm (\frac{1}{12}P)}\int_0^1|g(\xi,\eta;P)|^{12}\d\eta \d\xi .\end{aligned}$$ We thus conclude from Theorem \[theorem2.1\] and (\[3.5\]) that $$\label{3.7}
\sum_{\substack{1\le m\le X^3\\ m\not\in \grA(P)}}\Ttil(mX^{-3};P)^{12}=
\sum_{\substack{1\le m\le X^3\\ m\not\in \grA(P)}}G_m(\bet;P)^{12}\ll X^{\eps-37}.$$
Next, define $$\calT(m)=\bigcup_{l=-\infty}^\infty \grI(m+X^3l),$$ write $\calS(m)$ for the set of integers $h$ with $1\le h\le \kap X^{k-3}$ for which one has $\alp h\in \calT(m)$, and denote by $K(m)$ the cardinality of $\calS(m)$. We then take $\calS_1(P)$ to be the union of the sets $\calS(m)$ over integers $m$ with $1\le m\le X^3$ satisfying $\calI(mX^{-3})\cap \grM(\tfrac{1}{12}P)\ne \emptyset$, and $\calS_2(P)$ the corresponding union where instead $m$ satisfies $\calI(mX^{-3})\cap \grM(\tfrac{1}{12}P)=\emptyset$.
An application of Hölder’s inequality reveals that $$\begin{aligned}
\Bigl( \sum_{h\in \calS_1(P)}\Ttil(\alp h;P)\Bigr)^8&\ll (X^{k-3})^7\sum_{h\in \calS_1(P)}
\Ttil(\alp h;P)^8\\
&\ll X^{7k-21}\sum_{m\in \grA(P)}K(m)\Ttil(mX^{-3};P)^8.\end{aligned}$$ Thus, by (\[3.6\]), we see that $$\begin{aligned}
\Bigl( \sum_{h\in \calS_1(P)}\Ttil(\alp h;P)\Bigr)^8&\ll X^{7k-21}\Bigl( \max_{1\le m\le X^3}K(m)\Bigr)
\sum_{m\in \grA(P)}\Ttil(mX^{-3};P)^8\\
&\ll X^{7k-45+\eps}\max_{1\le m\le X^3}K(m),\end{aligned}$$ whence $$\label{3.8}
\sum_{h\in \calS_1(P)}\Ttil(\alp h;P)\ll \left( X^{7k-45+\eps}\max_{1\le m\le X^3}K(m)\right)^{1/8}.$$ Similarly, one finds that $$\begin{aligned}
\Bigl( \sum_{h\in \calS_2(P)}\Ttil(\alp h;P)\Bigr)^{12}&\ll (X^{k-3})^{11}
\sum_{h\in \calS_2(P)}\Ttil(\alp h;P)^{12}\\
&\ll X^{11k-33}\sum_{\substack{1\le m\le X^3\\ m\not\in \grA(P)}}K(m)\Ttil(mX^{-3};P)^{12}.\end{aligned}$$ Thus, by (\[3.7\]), we obtain $$\begin{aligned}
\Bigl( \sum_{h\in \calS_2(P)}\Ttil(\alp h;P)\Bigr)^{12}&\ll X^{11k-33}
\Bigl( \max_{1\le m\le X^3}K(m)\Bigr)
\sum_{\substack{1\le m\le X^3\\ m\not\in \grA(P)}}\Ttil(mX^{-3};P)^{12}\\
&\ll X^{11k-70+\eps}\max_{1\le m\le X^3}K(m),\end{aligned}$$ so that $$\label{3.9}
\sum_{h\in \calS_2(P)}\Ttil(\alp h;P)\ll \left( X^{11k-70+\eps}\max_{1\le m\le X^3}K(m)\right)^{1/12}.$$
Suppose now that $k\ge 6$, and that $\alp\in \dbR$, $a\in \dbZ$ and $q\in \dbN$ satisfy $(a,q)=1$ and $|\alp-a/q|\le q^{-2}$. Then one finds from [@HB1988 Lemma 6] that $K(m)\ll \Tet X^{k-3}$, where $\Tet=q^{-1}+X^{-3}+qX^{-k}$. Thus, on combining (\[3.3\]), (\[3.8\]) and (\[3.9\]), we deduce that $$\begin{aligned}
\sum_{h=1}^{\kap X^{k-3}}T(\alp h)&\ll X^{4+\eps}
\left( (X^{8k-48}\Tet)^{1/8}+(X^{12k-73}\Tet)^{1/12}\right) \\
&\ll X^{k-2+\eps}\left( \Tet^{1/8}+(\Tet/X)^{1/12}\right) .\end{aligned}$$ Finally, on substituting this estimate into (\[3.1\]), we conclude that $$f(\alp)\ll X^{1-2^{3-k}}+X^{1+\eps}\left( \Tet^{2^{-k}}+(\Tet/X)^{\frac{2}{3}2^{-k}}\right).$$ Since $\Tet\ge X^{-3}$, the conclusion of Theorem \[1.2\] now follows.
[18]{}
K. D. Boklan, *The asymptotic formula in Waring’s problem*, Mathematika **41** (1994), no. 2, 329–347.
R. de la Bretèche, *Répartition des points rationnels sur la cubique de Segre*, Proc. London Math. Soc. (3) **95** (2007), no. 1, 69–155.
J. Brüdern and O. Robert, *Rational points on linear slices of diagonal hypersurfaces*, submitted.
D. R. Heath-Brown, *Weyl’s inequality, Hua’s inequality, and Waring’s problem*, J. London Math. Soc. (2) **38** (1988), no. 2, 216–230.
L.-K. Hua, *The additive prime number theory*, Trav. Inst. Math. Stekloff, **22**, Acad. Sci. USSR, Moscow-Leningrad, 1947.
C. Segre, *Sulla varietà cubica con dieci punti doppii dello spazio a quattro dimensioni*, Atti R. Acc. Sci. Torino **22** (1887), 791–801.
R. C. Vaughan, *The Hardy-Littlewood method*, Cambridge University Press, Cambridge, 1997.
R. C. Vaughan and T. D. Wooley, *On a certain nonary cubic form and related equations*, Duke Math. J. **80** (1995), no. 3, 669–735.
T. D. Wooley, *The asymptotic formula in Waring’s problem*, Internat. Math. Res. Notices (2012), no. 7, 1485–1504.
T. D. Wooley, *The cubic case of the main conjecture in Vinogradov’s mean value theorem*, submitted; arXiv:1401.3150.
[^1]: The author is very grateful to Jörg Brüdern and Olivier Robert for supplying an advance copy of their joint paper [@BR2014], reference to which provides an excellent framework for the proof of this result.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Feature maps contain rich information about image intensity and spatial correlation. However, previous online knowledge distillation methods only utilize the class probabilities. Thus in this paper, we propose an online knowledge distillation method that transfers not only the knowledge of the class probabilities but also that of the feature map using the adversarial training framework. We train multiple networks simultaneously by employing discriminators to distinguish the feature map distributions of different networks. Each network has its corresponding discriminator which discriminates the feature map from its own as fake while classifying that of the other network as real. By training a network to fool the corresponding discriminator, it can learn the other network’s feature map distribution. We show that our method performs better than the conventional direct alignment method such as L1 and is more suitable for online distillation. Also, we propose a novel cyclic learning scheme for training more than two networks together. We have applied our method to various network architectures on the classification task and discovered a significant improvement of performance especially in the case of training a pair of a small network and a large one.'
bibliography:
- 'afd.bib'
---
Introduction {#intro}
============
\[fig:concept\]
With the advent of Alexnet [@krizhevsky2012imagenet], deep convolution neural networks have achieved remarkable success in a variety of tasks. However, high-performance of deep neural network is often gained by increasing the depth or the width of a network. Deep and wide networks cost a large number of computation as well as memory storage which is not suitable for a resource-limited environment such as mobile or embedded systems. To overcome this issue, many researches have been conducted to develop smaller but more accurate neural networks. Some of the well-known methods in this line of research are *parameter quantization or binarization* [@rastegari2016xnor], *pruning* [@li2016pruning] and *knowledge distillation* (KD) [@hinton2015distilling].
KD has been an active area of research as a solution to improve the performance of a light-weight network by transferring the knowledge of a large pre-trained network (or an ensemble of small networks) as a teacher network. KD sets the teacher network’s class probabilities as a target which a small student network tries to mimic. By aligning the student’s predictions to those of the teacher, the student can improve its performance. Recently, some studies have shown that rather than using a pre-trained teacher, simultaneously training networks to learn from each other in a peer-teaching manner is also possible. This approach is called online distillation. *Deep mutual learning* (DML) [@zhang2018deep] and *on-the-fly native ensemble* (ONE) [@lan2018knowledge] are [the representative]{} online distillation methods that show appealing results in the image classification tasks. Conventional distillation method requires pre-training a powerful teacher network and performs an one-way transfer to a relatively small and untrained student network. On the other hand, in online mutual distillation, there is no specific teacher-student role. All the networks learn simultaneously by teaching each other from the start of training. It trains with the conventional cross-entropy loss from the ground truth label along with the mimicry loss to learn from its peers. Networks trained in such an online distillation way achieve results superior not only to the networks trained with the cross-entropy loss alone but also to those trained in a conventional offline distillation manner from a pre-trained teacher network.
However, aforementioned online distillation methods make use of only the [logit information.]{} While the logit contains the probabilistic information over classes, the feature map, the output of convolution layer, has more meaningful and abundant feature information on image intensity and spatial correlation. [In offline distillation which utilizes a pre-trained model as a teacher network,]{} many methods such as *FitNet* [@romero2014fitnets], *attention transfer* (AT) [@zagoruyko2016paying] and *factor transfer* (FT) [@kim2018paraphrasing] [make use of]{} this intermediate feature representation as a target to learn for the student network. On the other hand, in online distillation, to the best of our knowledge, no feature map-based knowledge distillation method has been proposed.\
This is due to some challenges. Unlike the offline methods that have a clear target to mimic, there is no static target to follow in an online method. At every training iteration, the feature maps of the co-trained network change, thus in online feature map-level distillation, the problem turns into mimicking the moving target properly. While each node of the logit is confined to represent its assigned class probability which does not change drastically over iterations, at the feature map-level, much more flexibility comes into play, which makes the problem more challenging. Therefore, the direct aligning method such as using L1 or L2 distance is not suitable for online mutual feature map distillation because it updates the network parameters to generate a feature map trying to mimic the current output feature map of the other network. In other words, the direct alignment method only tries to minimize the distance between the two feature map points (one for each network), hence it ignores the distributional difference between the two feature maps (Fig. \[fig:concept\](a)).\
To alleviate this problem, in this paper, we propose a novel online distillation method that transfers the knowledge of feature map[s]{} adversarially as well as a cyclic learning framework for training more than two networks simultaneously. Unlike the direct aligning method, our adversarial distillation method enables a network to learn the overall feature map distribution of the co-trained network (Fig. \[fig:concept\](b)). Since the discriminator is trained to distinguish the difference between the networks’ feature map distributions ([containing]{} the history of feature maps for different input images) at every training iteration, by fooling the discriminator, the network learns the co-trained network’s changing feature map distribution. Exchanging the knowledge of feature map distribution facilitates the networks to converge to a better feature map manifold that generalizes better and yields more accurate results. Moreover, since it does not care about from which image a specific feature map originated, it is fitted to a secure federated learning environment [@li2019federated].
The contributions of this paper can be summarized as follows: 1) we propose an online knowledge distillation method that utilizes not only the logit but also the feature map from the convolution layer. 2) Our method transfers the knowledge of feature map[s]{} not by directly aligning [them]{} using the distance loss but by learning [their]{} distribution[s]{} using the adversarial training via discriminators. 3) We propose a novel cyclic learning scheme for training more than two networks simultaneously.
Related work {#rl}
============
The idea of *model compression* by transferring the knowledge of a high performing model to a smaller model was originally proposed by @bucilua2006model. Then in recent years, this research area got invigorated due to the work of *knowledge distillation* (KD) by @hinton2015distilling. The main contribution of KD is to use the softened logit of pre-trained teacher network that has higher entropy as an extra supervision to train a student network. KD trains a compact student network to learn not only by the conventional cross-entropy (CE) loss subjected to the labeled data but also by the final outputs of the teacher network. While KD only utilizes the logit, method such as FitNet [@romero2014fitnets], [AT [@zagoruyko2016paying], FT [@kim2018paraphrasing]]{} and KTAN [@liu2018ktan] use the intermediate feature representation to transfer the knowledge of a teacher network.
**Online Knowledge Distillation:** Conventional offline methods require training a teacher model in advance while online methods do not require any pre-trained model. [Instead,]{} the networks teach each other mutually by sharing their knowledge throughout the training process. Some examples of recent online methods are DML [@zhang2018deep] and ONE [@lan2018knowledge] which demonstrate promising results. DML simply applies KD losses mutually, treating each other as teachers, and it achieves results that is even better than the offline KD method. The drawback of DML is that it lacks an appropriate teacher role, hence provide[s]{} only limited information to each network. ONE pointed out this defect of DML. [R]{}ather than mutually distilling between the networks, ONE generates a gated ensemble logit of the training networks and uses it as a target to align for each network. ONE tries to create [a]{} powerful teacher logit that can provide more generalized information. The flaw of ONE is that it can not train different network architectures at the same time due to its architecture of sharing the low-level layers [for the gating module]{}. The common limitation of existing online methods is that [they are]{} dependent only on the logit and do not make any use of the feature map [information]{}. Considering that KD loss term is only applicable to the classification task, transferring knowledge at feature map-level [can enlarge the applicability to other tasks]{}. Therefore, our method proposes a distillation method that utilizes not only the logit but also the feature map via adversarial training, moreover, our method can be applied in case where the co-trained networks have different architectures.
**Generative Adversarial Network (GAN):** GAN [@goodfellow2014generative] is a generative model framework that is proposed with [an]{} adversarial training scheme, using a generator network $G$ and a discriminator network $D$. $G$ learns to generate the real data distribution while $D$ is trained to distinguish the real samples of the dataset from the fake results generated by $G$. The goal of $G$ is to trick $D$ to make [a]{} mistake of determining the fake results as the real samples. Though it was initially proposed [for]{} generative models, its adversarial training scheme is not limited to data generation. Adversarial training has been adapted to various tasks such as image translation [@isola2017image; @zhu2017unpaired], captioning [@dai2017towards], semi-supervised learning [@miyato2016adversarial; @springenberg2015unsupervised], reinforcement learning [@pfau2016connecting], and many others. In this paper, we utilize GAN’s adversarial training strategy to transfer the knowledge at feature map-level in an online manner. The networks learn the other networks’ feature map distributions by trying to deceive the discriminators while the discriminators are trained to distinguish the different distributions of each network.
Proposed Method
===============
In this section, we describe the overall process of our proposed *Online Adversarial Feature map Distillation* (AFD). Our method consists of two major losses: 1) logit-based loss and 2) feature map-based loss. Logit-based loss is defined by two different loss terms which are conventional cross-entropy [(CE)]{} loss and the mutual distillation loss using the Kullback-Leibler divergence (KLD). [Our newly proposed feature map-based]{} loss is to distill the feature map indirectly via discriminators. We use the feature map from the last convolution layer since deeper convolution layer generates more meaningful features with a high-level abstraction [[@kim2018paraphrasing]]{}. [The]{} adversarial training scheme of *generative adversarial networks* (GAN) [@goodfellow2014generative] [is utilized]{} to transfer the knowledge at feature map-level.\
As can be seen in Figure \[fig:method\], when training two different networks, [$\Theta_1$]{} and [$\Theta_2$]{}, in an online manner, we employ two discriminators, $D_1$ and $D_2$. We train $D_1$ such that the feature map of $\Theta_2$ is regarded as a real and that of $\Theta_1$ is classified as a fake and do vice versa for discriminator $D_2$. Then, each network $\Theta_1$ and $\Theta_2$ are trained to fool its corresponding discriminator so that it can generate a feature map that mimics the other network’s feature map. Throughout this adversarial training, each network learns the feature map distribution of the other network. By exploiting both logit-based distillation loss and feature map-based adversarial loss together, we could observe a significant improvement of performance in various [pairs of network architectures]{} especially when training small and large networks together.\
Also we introduce a cyclic learning scheme for training more than two networks simultaneously. It reduces the number of required discriminators from $2\times_2C_K$(when employing discriminators bidirectionally between every network pairs.) to $K$ where $K$ is the number of networks participating. This cyclic learning framework not only requires less computation than the bidirectional way but also achieves better results compared to other online training schemes for multiple networks. First, we explain the conventional mutual knowledge distillation method conducted among the networks at the logit-level. Then we introduce our novel online feature map distillation method using the adversarial training scheme in addition to the cyclic learning framework for training more than two networks at the same time.
Logit-based Mutual Knowledge Distillation
-----------------------------------------
We use two loss terms for logit-based learning, one is the conventional cross-entropy (CE) loss and the other is mutual distillation loss between networks based on Kullback Leibler(KL) divergence. We formulate our proposed method assuming training two networks. Training scheme for more than two networks will be explained in Sec \[cycle\]. Below is the overall logit-based loss for two networks: $$\begin{aligned}
\centering
\mathcal{L}_{logit}^1 &= \mathcal{L}_{ce}(y,\sigma(z_1)) + T^2\times \mathcal{L}_{kl}( \sigma(z_2/T),\sigma(z_1/T) )\\
\mathcal{L}_{logit}^2 &= \mathcal{L}_{ce}(y,\sigma(z_2)) + T^2\times \mathcal{L}_{kl}( \sigma(z_1/T),\sigma(z_2/T) )
\vspace{-5mm}\end{aligned}$$ Here, $\sigma(\cdot)$ refers to softmax function and $z \in \mathbb{R}^C$ is the logit produced from a network for $C$-class classification problem. The temperature term $T$ is used to control the level of smoothness in probabilities. As the temperature term $T$ goes up, it creates a more softened probability distribution. We use $T=3$ for every experiment. $\mathcal{L}_{ce}$ is the CE loss between the ground truth label $y$ and the softmax output $\sigma(z)$ that is commonly used in image classification. $\mathcal{L}_{kl}$ is the KL loss between the softened logit of each network. We multiply the KL loss term with $T^2$ because the gradients produced by the soft targets are scaled by $1/T^2$. While the CE loss is between the correct labels and the outputs of the model, the KL loss is the KL distance between the outputs of two training networks. The KL loss provides an extra information from the peer network so that the network can improve its generalization performance. The difference with DML is that while DML updates asynchronously which means that it updates one network first and then the other network, our AFD updates the networks synchronously, not alternately. The CE loss trains the networks to predict the correct truth label while the mutual distillation loss tries to match the outputs of the peer-networks, enabling the networks to share the knowledge at logit-level.
Adversarial Training for Feature-map-based KD
---------------------------------------------
Our AFD uses adversarial training to transfer knowledge at feature map-level. We formulate our adversarial feature map distillation [for]{} two networks [which will be extended for more networks later]{}. We divide a network into two parts, one is the feature extractor part that generates a feature map and the other is the classifier part which is the FC layer that transforms the feature map into a logit. Each network also has a corresponding discriminator which distinguishes different feature map distributions. The architecture of the discriminator is simply a series of Conv-Batch\_Normalization-Leaky\_ReLU-Conv-Sigmoid. It takes a feature map of the last layer and it reduces the spatial size and the number of channel of the feature map as it goes through the convolution operation so that it can produce a single scalar value at the end. Then we apply the sigmoid function of the value to normalize it between 0 and 1.\
We utilize the feature extractor part to enable feature map-level distillation. For the convenience of mathematical notation, we name the feature extractor part as $G_k$ and its discriminator as $D_k$, $k$ indicates the network number. As depicted in Figure \[fig:method\], each network has to fool its discriminator to mimic the peer network’s feature map and the discriminator has to discriminate from which network the feature map is [originated]{}. Following LSGAN [@mao2017least], our overall adversarial loss for discriminator and the feature extractor can be written as below: $$\begin{aligned}
\centering
\mathcal{L}_{D_1} &= [1-D_1(G_2(x))]^2 + [D_1(G_1(x))]^2 \\
\mathcal{L}_{G_1} &= [1-D_1(G_1(x))]^2\end{aligned}$$ The feature extractors $G_1$ and $G_2$ take input $x$ and generate feature maps. The discriminator $D_1$ takes a feature map and yields a scalar between 0 (fake) and 1 (real). It is trained to output 1 if the feature map came from the co-trained network (in this case, $G_2$) or 0 if the feature map is produced from the network it belongs to ($G_1$ in this case). The goal of $D_1$ is to minimize the discriminator loss term $\mathcal{L}_{D1}$ by correctly distinguishing the two different feature map distributions while $G_1$’s goal is to minimize the loss term $\mathcal{L}_{G_1}$ by fooling $D_1$ to make mistake of determining $G_1$’s feature map as real and yield 1. Each training network’s object is to minimize $\mathcal{L}_{G_k}$ to mimic the peer network’s feature map distribution. This adversarial scheme works exactly the same by changing the role of two networks. In case where the two networks’ feature map outputs have different channel sizes, for example a pair like (WRN-16-2, WRN-16-4) [@zagoruyko2016wide], we use a transfer layer that is composed of a convolution layer, a batch normalization and a ReLU which converts the number of channels to that of peer network. The above loss terms change as $$\begin{aligned}
\centering
\mathcal{L}_{D_1} &= [1-D_1(T_2(G_2(x)))]^2 + [D_1(T_1(G_1(x)))]^2\\
\mathcal{L}_{G_1} &= [1-D_1(T_1(G_1(x)))]^2\end{aligned}$$ when using the transfer layer $T_k$.
**Optimization:** Combining both logit-based loss and the adversarial feature map-based loss, the overall loss for each network $\Theta_1$ and $\Theta_2$ are as follows: $$\begin{gathered}
\mathcal{L}_{\Theta_1} = \mathcal{L}_{logit}^1 + \mathcal{L}_{G_1}, \qquad \mathcal{L}_{\Theta_2} = \mathcal{L}_{logit}^2 + \mathcal{L}_{G_2} \end{gathered}$$ However, the logit-based loss term $\mathcal{L}_{logit}^k$ and the feature map-based loss term $\mathcal{L}_{G_k}$ are not optimized by the same optimizer. In fact, they are optimized successively in a same mini-batch. At every mini-batch iteration, we infer an image into a model and it computes a logit and a feature map. Then we calculate the two loss terms (logit-based and feature-map-based) and optimize the networks based on the two losses separately, meaning that we update the parameters by the logit-based loss once and then update again by the feature map-based loss. The reason we optimize separately for each loss term is because they use [different learning rates]{}. The adversarial loss requires much slower learning rate thus if we use the same optimizer with the same learning rate, the networks would not be optimized. Note that we do not infer for each loss term, inference is conducted only once, only the optimization is conducted twice, one for each loss term.
Cyclic Learning Framework {#cycle}
-------------------------
In case when we want to train more than two networks simultaneously, our method proposes a novel cyclic peer-learning scheme . Our cyclic peer-learning scheme transfers each network’s knowledge to its next peer network in an one-way cyclic manner. If we train $K$ number of networks together, each network distills its knowledge to its next network except the last network transfers its knowledge to the first network, creating a cyclic knowledge transfer flow as $1\rightarrow2, 2\rightarrow3, \cdots, {(K-1)}\rightarrow{K}, K\rightarrow1$. Refer to Appendix A for detailed explanation and figure. The main contribution of using this cyclic learning framework is to avoid employing too many discriminators. If we apply our adversarial loss for every pair of networks, it would demand two times the amount of every possible pair of $K$ networks which would cost enormous computation. In Sec \[multi\], we empirically show that our cyclic training scheme is better than other online methods for training multiple networks.
Experiment
==========
In this section, to show the adequacy of our method, we first present comparison experiment with distance method and ablation study to analyze our method. Then we compare our approach with existing online knowledge distillation methods under different settings. We demonstrate the comparison experiment results of using the same network architectures in Sec \[same-arch\] and then apply our method on networks with different architectures in Sec \[diff-arch\]. In Sec \[multi\], we also show the results of training more than two networks to demonstrate that our method generalizes well even when the number of networks increases.\
In most of experiments, we use the CIFAR-100 [@cifar100] dataset. It consists of 50K training images and 10K test images over 100 classes, accordingly it has 600 images per each class. All the reported results on CIFAR-100 are average of 5 experiments. Since our method uses two loss terms, logit-based loss and feature map-based loss, we use different learning details for each loss term. For overall learning schedule, we follow the learning schedule of ONE[@lan2018knowledge] which is 300 epochs of training to conduct fair comparison . In terms of logit-based loss, the learning rate starts at 0.1 and is multiplied by 0.1 at 150, 225 epoch. We optimize the logit-based loss using SGD with mini-batch size of 128, momentum 0.9 and weight decay of 1e-4. This learning details for logit-based loss is equally applied to other compared online distillation methods. For feature map-based loss, the learning rate starts at 2e-5 for both discriminators and feature extractors and is decayed by 0.1 at 75, 150 epoch. The feature map-based loss is optimized by ADAM[@kingma2014adam] with the same mini-batch size of 128 and weight decay of 1e-1. In tables, ‘2 Net Avg’ and ‘Ens’ represents the average accuracy of the two networks and the ensemble accuracy respectively. The average ensemble is used for AFD, DML and KD while ONE uses gated ensemble of the networks according to its methodology.
Comparison with direct feature map alignment {#comp_dist}
--------------------------------------------
[max width=]{}
-- ------- ------- ------- ------- ------- ------- ------- ------- ------- ----------- ------- ----------- ----------- -----------
Ens Ens Ens Ens
70.69 72.44 71.91 69.79 72.07 **75.64**
72.57 69.09 73.11 71.52 74.75 **76.34**
Ens Ens Ens Ens
71.07 73.50 69.84 73.41 74.63 72.35 74.82 75.10 73.94 73.62 76.56 **75.88** **77.08** **77.82**
-- ------- ------- ------- ------- ------- ------- ------- ------- ------- ----------- ------- ----------- ----------- -----------
\[distance\_comparison\]
Since our goal is to distill feature map information that suits for online distillation, we briefly compare our method with conventional direct alignment method in Table \[distance\_comparison\]. We train two networks together using the same and different types of network architecture. For $L_1$, each network is trained not only to follow the ground-truth label by CE loss, but also to mimic the other network’s feature map using the $L_1$ distance loss. For $L_1 +$ KD, KD [@hinton2015distilling] loss is applied at the logit level along with the $L_1$ loss between the feature maps. We also compare our results with offline method, $L1 +$ KD (offline), it employs a pre-trained network as a teacher network and distills its knowledge to an untrained student network by $L1$ loss at the feature map-level and KD loss at the logit level. We employ ResNet-32 [@he2016deep], WRN-16-2 [@zagoruyko2016wide] and WRN-28-2 that shows 69.79%, 71.52%, and 73.62% accuracy as the teacher networks.\
The results clearly show that learning the distributions of feature maps by adversarial loss performs better than direct alignment method in both online and offline distillation. We could observe that using $L_1$ distance loss actually disturbs the networks to learn good features in online environment. The accuracy of ResNet-32 has dropped more than 2% compared to its vanilla version accuracy (69.38%) and the accuracy of WRN-16-2 is also lower than its vanilla network (71.07%). Even when combined with KD loss ($L1$ + KD), direct alignment method shows poor performance compared to ours in both online and offline manner. Though distance loss is used in many conventional offline methods, it disturbs proper learning in online environment. Also in case of different architecture types, we observe a severe degradation of performance when using the direct alignment method. It indicates that when it comes to online feature map distillation, transferring feature map information with direct alignment method such as $L1$ distance extremely worsen the performance while our method advances it.
Ablation study {#abl_study}
--------------
Table \[ablation\] shows the ablation study of our proposed method. We conduct experiments using the same and different network architectures. The experiments are conducted under three different training settings for each model case. The three settings are full model, without mutual knowledge distillation at logit-level and without adversarial feature map distillation. When trained without the adversarial feature map distillation, the accuracy decreases in all three model cases. The accuracy of both ResNet-32 and WRN-16-2 dropped by 0.65% and 0.52% respectively, and those of (WRN-16-2, WRN-28-2) pair declined by 0.89% and 0.44% compared to the full model. Ensemble results are also lower than those of the full models. When only the adversarial feature map distillation is applied, the accuracy has increased by 0.71% and 0.87% compared to the vanilla versions of ResNet-32 and WRN-16-2 respectively. Especially in case of different sub-network architecture, the accuracy of WRN-16-2 has increased by almost 1%. Based on these experiments, we could confirm that adversarial feature map distillation has some efficacy of improving the performance in online environment.
[width=1]{}
----------------- ------- ------- ------- ------- ------- ------- ------- ------- -------
Ens Ens Ens
74.77 75.21 75.64
75.92 76.20 76.34
Ens Ens Ens
WRN-(16-2,28-2) 72.05 73.80 76.82 74.99 76.64 77.28 75.88 77.08 77.82
----------------- ------- ------- ------- ------- ------- ------- ------- ------- -------
: Ablation study of AFD. Top-1 accuracy(%) on CIFAR-100 dataset.
\[ablation\]
same architecture {#same-arch}
-----------------
[width=0.7]{}
Model Type Vanila
------------ -------- --------------- ------- --------------- ------- ------------------- -----------
Same Arch. Net Ens Ens Ens
ResNet-20 67.48 70.90(+3.42%) 72.08 70.56(+3.08%) 72.26 **71.72(+4.24%)** **72.98**
ResNet-32 69.38 73.40(+4.02%) 74.89 72.61(+3.23%) 74.07 **74.03(+4.65%)** **75.64**
ResNet-56 73.84 75.48(+1.64%) 76.73 76.45(+2.61%) 77.16 **77.25(+3.41%)** **78.35**
WRN-16-2 71.07 74.68(+3.61%) 75.81 73.85(+2.78%) 74.84 **75.33(+4.26%)** **76.34**
WRN-16-4 75.38 78.17(+2.79%) 79.06 77.32(+1.94%) 77.79 **78.55(+3.17%)** **79.28**
WRN-28-2 73.50 77.02(+3.52%) 78.64 76.67(+3.17%) 77.40 **77.22(+3.72%)** **78.72**
WRN-28-4 76.60 79.16(+2.56%) 80.56 79.25(+2.65%) 79.73 **79.46(+2.86%)** **80.65**
\[same\]
We compare our method with DML and ONE for training two networks with the same architecture. The vanilla network refers to the original network trained without any distillation method. As shown in Table \[same\], in both ResNet and WRN serises, DML, ONE and AFD all improves the networks’ accuracy compared to the vanilla networks. However, AFD shows the highest improvement of performance in both 2 Net average and ensemble accuracy among the compared distillation methods. Especially in case of ResNet-20, ResNet-32 and WRN-16-2, our method significantly improves the accuracy by more than 4% compared to the vanilla version while other distillation methods improve around 3% on average except the ResNet-32 of DML.
different architecture {#diff-arch}
----------------------
[width=0.7]{}
----------- ----------- ------- ----------- ------- ------- ----------- ----------- ----------- ----------- -----------
Ens Ens Ens
ResNet-32 ResNet-56 72.92 76.27 76.71 73.48 76.35 76.74 **74.13** **76.69** **77.11**
ResNet-32 WRN-16-4 72.67 77.26 76.94 73.48 77.43 77.01 **74.43** **77.82** **77.67**
ResNet-56 WRN-28-4 75.48 78.91 79.23 76.03 **79.32** 79.38 **77.95** 79.21 **80.01**
ResNet-20 WRN-28-10 70.08 **78.17** 76.12 71.03 77.70 75.78 **72.62** 77.83 **76.70**
WRN-16-2 WRN-16-4 74.87 77.42 77.30 74.87 77.17 76.96 **75.81** **78.00** **77.84**
WRN-16-2 WRN-28-2 74.86 76.45 77.29 75.11 76.91 77.24 **75.88** **77.08** **77.82**
WRN-16-2 WRN-28-4 74.51 78.18 77.60 74.95 78.23 77.67 **76.23** **78.26** **78.28**
73.63 77.52 77.31 74.14 77.59 77.25 **75.29** **77.84** **77.92**
----------- ----------- ------- ----------- ------- ------- ----------- ----------- ----------- ----------- -----------
\[different\]
In this section, we compare our method with DML and KD using different network architectures. We set Net2 as the higher capacity network. For KD, we use the ensemble of the two networks as a teacher to mimic at every iteration. The difference with original KD [@hinton2015distilling] is that it is an online learning method, not offline. We did not include ONE because ONE can not be applied in case where the two networks have different model types due to its architecture of sharing the low-level layers. In table \[different\], we could observe that our method shows better performance improvement than other methods in both Net1 and Net2 except for a couple of cases. The interesting result is that when AFD is applied, the performance of Net1 (smaller network) is improved significantly compared to other online distillation methods. This is because AFD can transfer the higher capacity network’s meaningful knowledge (feature map distribution) to the lower capacity one better than other online methods.When compared with KD and DML, AFD’s Net1 accuracy is higher by 1.66% and 1.15% and the ensemble accuracy is better by 0.61% and 0.67% on average respectively. In case of (WRN-16-2, WRN-28-4) pair, the Net1’s parameter size (0.70M) is more than 8 times smaller than Net2 (5.87M). Despite the large size difference, our method improves both networks’ accuracy, particularly our Net1 performance is better than KD and DML by 1.72% and 1.28% respectively. The performance of KD and DML seems to decline as the difference between the two model sizes gets larger. Throughout this experiment, we have shown that our method also works properly for different architectures of networks even when two networks have large difference in their model sizes. Using our method, smaller network considerably benefits from the larger network.
[width=1]{}
Model Type Vanilla
------------ --------- ------- ------- ------- ------- ----------- -----------
Same arch. Net Ens Ens Ens
ResNet-32 69.38 73.43 76.11 73.25 74.94 **74.14** **76.64**
ResNet-56 73.84 76.11 77.83 76.49 77.38 **77.37** **79.18**
WRN-16-2 71.07 75.15 76.93 73.87 75.26 **75.65** **77.54**
WRN-28-2 73.50 77.12 79.41 76.66 77.53 **77.20** **79.78**
: Top-1 accuracy(%) comparison with other online distillation methods using 3 networks on CIFAR-100 dataset. ’3 Net Avg’ represents the average accuracy of the 3 networks.
\[3branch\]
[width=1]{}
----------- ----------- ------- ------- ------- ------- ------- ------- ------- -------
Ens Ens
ResNet-18 ResNet-34 69.76 73.27 70.19 73.57 73.33 70.39 74.00 74.47
----------- ----------- ------- ------- ------- ------- ------- ------- ------- -------
: Top-1 accuracy(%) comparison with DML on ImageNet dataset.
\[imagenet\]
Expansion to 3 Networks {#multi}
-----------------------
To show our method’s expandability for training more than two networks, we conduct experiment of training 3 networks in this section. As proposed in Sec \[cycle\], our method uses a cyclic learning framework rather than employing adversarial loss between every network pairs in order to reduce the amount of computation and memory. DML calculates the mutual knowledge distillation loss between every network pairs and uses the average of the losses. ONE generates a gated ensemble of the sub-networks and transfers the knowledge of the ensemble logit to each network. As it can be seen in Table \[3branch\], AFD outperforms the compared online distillation methods on both 3 Net average and ensemble accuracy in every model types. Comparing the results of Table \[3branch\] to that of Table \[same\], AFD’s overall tendency of performance gain is maintained.
ImageNet Experiment {#img_net}
-------------------
We evaluate our method on ImageNet dataset to show that our method can also be applicable to a large scale image dataset. We use ImageNet LSVRC 2015 [@ILSVRC15] which has 1.2M training images and 50K validation images over 1,000 classes. We compare our method with DML using two pre-trained networks ResNet-18 and ResNet-34 as a pair. The results are after 30 epochs of training. As shown in Table \[imagenet\], networks trained using our method shows higher performance than the networks trained with DML. This indicates that our method is applicable in large-scale dataset.
Analysis
========
In this section, we analyze how each distillation method affects the formation of feature maps. We first show the similarity of feature maps from two networks trained with different distillation methods by measuring the $L1/L2$ distance and the cosine similarity. Then we visualize the feature maps to see which part of the feature map is activated differently for each method. We expect that even though our method shows higher performance, since our method does not directly align the feature maps but rather distill the distribution by adversarial loss, the feature maps of different networks would contain different features and information.
Quantitative analysis on feature map similarity {#quantitative}
-----------------------------------------------
[width=1]{}
Model Type Method $L1$ $L2$ Cosine
----------------- ------------ ------------ ------------ ------------
$L1+$KD 0.0841 0.1943 0.9959
ResNet-32 DML 1.2662 6.0103 **0.2421**
AFD (Ours) **1.4113** **6.9310** 0.2550
$L1+$KD 0.0291 0.0383 0.9959
WRN-16-2 DML 0.9893 3.5271 0.2280
AFD (Ours) **1.0013** **3.6386** **0.2143**
$L1+$KD 0.0662 0.2921 0.9822
WRN-(16-2-28-2) DML 0.7999 2.4893 0.2003
AFD (Ours) **0.8644** **2.8908** **0.1972**
: Comparison of $L1/L2$ distance and cosine similarity of two trained networks’ feature maps between three different distillation methods.
\[famp\_anal\]
In this experiment, we examine the similarity between the feature maps of networks trained with three different online distillation methods using $L1/L2$ distance and cosine similarity. The three methods are $L1+$KD from Section \[comp\_dist\], DML [@zhang2018deep] and our AFD. We use pre-trained networks from previous experiments in Table \[distance\_comparison\],\[same\] and \[different\]. As expected, the $L1+$KD method shows the shortest $L1/L2$ distance and the highest cosine similarity in all three different types of model, meaning that two networks have become identical networks that yield the same feature map. To successfully transfer knowledge between networks, it is important that the network learns different features from the other network. It is shown that using a direct alignment loss such as $L1$ at feature map-level in online environments makes the network identical, not being able to transfer any new knowledge which severely impairs the performance. What is more interesting is the results of DML and AFD. Even though our method transfers knowledge at both logit and feature map-level while DML transfers knowledge only at logit-level, our feature map distance (similarity) is larger (lower) than DML except for ResNet-32. It is obvious that DML has low similarity because it transfers knowledge only at logit but it is noteworthy that our feature-map-based KD method has lower feature map similarity. It means that our method guides the network to successfully transfer knowledge at feature map-level while maintaining the genuine features it learned. Since our method tries to transfer indirect knowledge via discriminator, it does not change the feature map directly and drastically. Our method can transfer knowledge at feature-map-level while preserving the features which have been learned by each individual network without damaging the knowledge it has gained on its own.
Qualitative analysis on feature map similarity {#qulitative}
----------------------------------------------
Figure \[fig:grad\_cam\] illustrates the Grad-cam [@selvaraju2017grad] visualization of feature maps. Grad-cam visualizes the regions where the network has considered important. We compare the results of the two networks, Net1 and Net2 trained by different distillation methods. The activation maps of $L1+$KD are exactly the same which means that the two networks see the exact same area of the image. This indicates that the two networks have become indistinguishable networks that outputs the same feature maps. Purpose of knowledge distillation is not making the networks the same, it is to transfer useful knowledge it has earned to the other network. However, $L1+$KD method ignores the knowledge or features each network learned and just force the networks to copy each other, eventually disturbing them to learn proper features. On the other hand, our method, AFD’s Net1 and Net2 activation maps are not the same but rather try to find different features. This shows that even though AFD transfers knowledge at feature map-level, it does not harm the knowledge each network has learned and yet still distills useful knowledge that leads to performance gain.
Conclusion
==========
We proposed an online knowledge distillation method that transfers knowledge both at logit and feature map-level using the adversarial training scheme. Unlike existing online distillation methods, our method utilizes the feature map distribution and showed that online knowledge distillation at feature map-level is possible. Through extensive experiments, we demonstrated the adequacy of transferring the distribution via adversarial training for online feature map distillation and could achieve higher performance than existing online methods and conventional direct alignment methods. We also confirmed that our method is broadly suitable to various architecture types from a very small network (ResNet-20) to a large (WRN-28-4) network. We hope that our work broadens the area of knowledge distillation and be further advanced and studied by many researchers.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present activity demographics and host-galaxy properties of infrared-selected galaxies in the local Universe, using the representative Star Formation Reference Survey (SFRS). Our classification scheme is based on a combination of optical emission-line diagrams (BPT) and IR-color diagnostics. Using the weights assigned to the SFRS galaxies based on its parent sample, a far-infrared-selected sample comprises 71% H<span style="font-variant:small-caps;">ii</span> galaxies, 13% Seyferts, 3% Transition Objects (TOs), and 13% Low-Ionization Nuclear Emission-Line Regions (LINERs). For the SFRS H<span style="font-variant:small-caps;">ii</span> galaxies we derive nuclear star-formation rates and gas-phase metallicities. We measure host-galaxy metallicities for all galaxies with available long-slit spectroscopy and abundance gradients for a subset of 12 face-on galaxies. The majority of H<span style="font-variant:small-caps;">ii</span> galaxies show a narrow range of metallicities, close to solar, and flat metallicity profiles. Based on their host-galaxy and nuclear properties, the dominant ionizing source in the far-infrared selected TOs is star-forming activity. LINERs are found mostly in massive hosts (median of $10^{10.5}$ M$ _{\sun} $), median $L(60\micron) =
10^{9}$ L$_{\sun}$, median dust temperatures of $ F60/F100 = 0.36
$, and median $L_{\textrm{H}\alpha}$ surface density of $ 10^{40.2}
$ erg s$ ^{-1} $kpc$ ^{-2} $, indicating older stellar populations as their main ionizing source rather than AGN activity.
author:
- |
A. Maragkoudakis,$^1$ $^2$ $^3$ A. Zezas,$^1$ $^2$ $^3$ M. L. N. Ashby,$^2$ S. P. Willner$^2$\
$^1$University of Crete, Department of Physics, Heraklion 71003, Greece\
$^2$Harvard-Smithsonian Center for Astrophysics, Cambridge, MA 02138\
$^3$Foundation for Research and Technology - Hellas (FORTH), Heraklion 71003, Greece
bibliography:
- 'Bibliography.bib'
title: 'The Star Formation Reference Survey. II. Activity demographics and host-galaxy properties for Infrared-selected galaxies'
---
galaxies: nuclei – galaxies: Seyfert – galaxies: starburst – galaxies: star formation – galaxies: abundances
INTRODUCTION
============
Distinguishing between the processes of star-formation and active galactic nuclei (AGN) is crucial in the study of energy production in galaxies. The ratio of these processes and their interplay is important on both whole-galaxy and galaxy nucleus scales for understanding their part in galaxy evolution. Several scaling relations between black hole mass ($M_{\textrm{BH}}$), bulge luminosity ($L_{\textrm{bulge}}$), bulge mass ($M_{\textrm{bulge}}$), velocity dispersion ($ \sigma $) e.g., $M_{\textrm{BH}}-L_{\textrm{bulge}}$: [@Kormendy95], $M_{\textrm{BH}}-M_{\textrm{bulge}}$ [@Magorrian98], $M_{\textrm{BH}}-\sigma$ [@Ferrarese00], average black hole growth rate, and mean total stellar mass [@Mullaney12] describe empirical relations between the central regions of galaxies and their hosts. However, the origin and degree of influence of such relations are still debated. (For a review see @KH13).
The close similarities between the cosmic evolution of the star-formation rate density and the supermassive black hole (SMBH) mass accretion rate density (e.g., @Silverman08; @Aird10) point towards a connection between the evolution of star formation and AGN activity. [@Davies07] provided strong evidence for a starburst–AGN connection by examining the nuclei of nine Seyfert galaxies and showing that the peak of AGN activity occurs 50–100 Myr after the onset of star formation. The delay between star formation and AGN activity indicates that the starburst had a significant impact on fueling the central black hole. Hydrodynamical simulations (e.g., @Hobbs11) support this scenario by showing that stellar winds and supernovae enhance the black hole mass accretion by injecting turbulence into the surrounding gas disc. However, a complete picture of these processes has not yet been established, and many questions remain open regarding the ways star formation and AGN feedback mechanisms impact and regulate one another considering the different physical conditions at different cosmic epochs.
A complete study of the starburst–AGN connection requires a representative sample of galaxies probing the wide range of different physical conditions that these phenomena engage. The nearby Universe offers the opportunity to study the above processes and overall galaxy properties in observationally resolved conditions. A number of nearby galaxy surveys (e.g., *Spitzer* Infrared Nearby Galaxies Sample$ - $SINGS; @Kenn03, Great Observatories All-sky LIRG Survey$ - $GOALS; @GOALS, *Herschel* Reference Survey$ - $HRS; @Boselli10) have approached the subject by focusing mainly on specific aspects of the star formation or AGN processes. For example, SINGS galaxies constitute a relatively small number of nearby and extended objects at the low-luminosity end of FIR luminosity function and therefore do not depict star-formation in its whole extent. Furthermore, powerful AGNs were intentionally excluded from the SINGS sample, thus limiting the diversity of AGN systems available for a comprehensive study of the starburst–AGN connection. GOALS on the other hand consists predominantly of Luminous Infrared Galaxies (LIRGs, $10^{11}L_{\sun} < L < 10^{12}L_{\sun} $) and a few Ultra-Luminous Infrared Galaxies (ULIRGs $L \geq 10^{12}L_{\sun}$), sampling the higher ends of star formation and galactic activity in general, while the HRS is particularly oriented toward high-density environments.
For a comprehensive study of the star formation and AGN phenomenon in a wide range of environments and physical conditions in the local universe, it is imperative to employ a well defined and representative sample of galaxies. To this cause the “Star Formation Reference Survey" (SFRS; @SFRS) was established. An initial result from the SFRS sample indicating a connection between the central regions of star-forming galaxies and their hosts [@Maragkoudakis17] showed a correlation between the nuclear star-formation rate (SFR) and the total stellar mass of the galaxies. This was referred to as the Nuclear Main Sequence of star-forming galaxies.
The current paper presents the activity demographics and the host galaxy properties for the SFRS galaxies and is organized as follows. Section \[SFRS\] describes the SFRS and gives basic information on the sample properties. Section \[Observations\] describes the observations, the initial reduction of the spectra, and the measurement of the fluxes of the strong optical emission lines. The activity classification is presented in Section \[Classification\]. Gas-phase metallicities, metallicity gradients, and nuclear SFR measurements are presented in Section \[Z-SFR\]. Section \[Discussion\] provides a detailed discussion of the activity demographics in the local Universe, the different classification methods used, the galaxy properties based on activity type, and the dominant source of ionization in LINERs and transition objects. Finally, we summarize our results in Section \[Conclusions\]. Throughout the paper the activity class H<span style="font-variant:small-caps;">ii</span> (or H<span style="font-variant:small-caps;">ii</span> galaxies) is used to describe pure star-forming galaxies (SFGs) that exhibit no fraction of AGN activity.
The SFRS sample {#SFRS}
===============
The SFRS galaxies were chosen in order to build a representative sample spanning the full range of properties of local galaxies hosting primarily star-formation activity. The parent sample from which SFRS was drawn is the PSC$ z $ [@Saunders00], a database of 15,411 nearby star-forming galaxies brighter than 0.6 Jy at 60$\mu m$ across 84 per cent of the sky at redshifts $ z \leq 0.2 $. Selection from the PSC$ z $ into the SFRS was based on observational proxies for SFR, specific SFR (sSFR), and dust temperature. Specifically, [@SFRS] used the IRAS 60 $\mu$m luminosity as a SFR tracer, the flux ratio of $F_{60}$ to 2MASS $K_{S}$ ($F_{60}/K_{S}$) as sSFR proxy, and the far-IR flux density ratio of $F_{100}/F_{60}$ as a measure of dust temperature to create a sample of galaxies spanning this three-dimensional parameter space. 369[^1] galaxies were selected [@SFRS] to explore all existing combinations of these parameters over their entire range. Each SFRS galaxy was assigned a weight reflecting its relative prevalence with respect to the parent population from which the SFRS was defined [@SFRS]. The overall distribution of SFRS’s total infrared (TIR) luminosities is $10^{7.95}-10^{12.25} \textrm{L}_{\sun} $, the $ F_{60}/F_{100} $ flux-density ratio distribution is between 0.14–1.64, and the stellar masses range between $10^{7.79}-10^{12.14} \textrm{M}_{\sun} $ [@Maragkoudakis17]. Thus, the parameter space covered by the SFRS captures the full range of star-forming and AGN conditions and environments, making it an ideal survey for the study of these processes in the local Universe.
Exploring all conditions of star formation requires an uncensored sample of galaxies, and therefore galaxies hosting AGN activity were not discarded from the SFRS. This also provides the opportunity to examine the AGN contribution to the overall energy output of galaxies and simultaneously investigate the interplay between star formation and AGN mechanisms (i.e., quenching or boosting of star formation). Therefore, we initiated a long-slit spectroscopic campaign to acquire spectra for the SFRS galaxies lacking SDSS spectroscopic coverage and used a combination of optical emission-line and IR diagnostics to identify AGN hosts.
The SFRS has an impressive range of multiband photometric data from UV to radio bands: GALEX FUV–NUV, SDSS $ugriz$, 2MASS $JHKs$, *Spitzer*/IRAC $ 3.6, 4.5, 5.8, 8.0\,\micron$, *Spitzer*/MIPS $ 24\,\micron$, *AKARI FIR All-Sky Survey*, *IRAS*, VLA/NVSS. Furthermore, there is an ongoing H$ \alpha $ imaging campaign at the 1.3-m Ritchey-Crétien telescope at the Skinakas Observatory in Crete. This complete set of multi-wavelength photometry combined with the spectroscopically defined activity classification will allow us to derive fundamental galaxy properties (i.e., SFRs, stellar masses, dust-luminosities) for each galaxy activity class using both luminosity-dependent calibrations as well as spectral energy distribution (SED) fitting techniques.
OBSERVATIONS AND ANALYSIS {#Observations}
=========================
SDSS spectra and aperture effects
---------------------------------
210 SFRS galaxies had available SDSS spectra as of data release 7 (DR7; @SDSSDR7). These cover a spectral range from 3800 Å to 9200 Å with a resolution $R = 1500$ at 3800 Å and $R = 2500$ at 9000 Å. The spectra refer to a 3-diameter region of the nucleus of each galaxy. Taking into account that at the mean SFRS redshift of 0.024 the fiber diameter correspond to a physical size of approximately 1.5 kpc, the obtained SDSS spectra provide a reasonable representation of the nuclear region of the SFRS galaxies. Because the SDSS fibers are located on the nucleus, and the region of spectral extraction can not be adjusted, we refer to all SDSS classification as nuclear region classification. Thirty six SFRS galaxies are located at redshifts between $0.05<z<0.2$, 21 of which have SDSS fiber spectra. The projected aperture diameter at these redshifts is between 3 kpc and 12 kpc, which are closer to global spectra of galaxies rather than nuclear. [@Maragkoudakis14] showed that line ratios and consequently activity classification obtained from SDSS $3\arcsec$ fibers can be affected by aperture size. Besides AGN spectral features being veiled by the host-galaxy starlight, extranuclear star-forming activity encompassed in the fiber can influence spectral lines, often increasing lower ionization emission-line intensities. Although starlight removal techniques can mitigate these effects they do not fully remove them, especially the contamination by emission lines from the ionized interstellar medium (ISM) in the host galaxy (@Maragkoudakis14). However, these galaxies constitute only a very small fraction of the SFRS sample, and they do not change the conclusions of the analysis.
Long-slit spectra {#spectra}
-----------------
Long-slit spectra were acquired at the Fred Lawrence Whipple Observatory in Arizona using the 60-inch Tillinghast telescope with the FAST spectrograph [@FAST]. The target list included 159 galaxies without SDSS spectra plus 19 galaxies observed by SDSS but with insufficient signal-to-noise ratio (S/N) or bad fiber centering. Each galaxy was observed with a $3\arcsec$ wide $\times6\arcmin$ long slit visually centered on the nucleus and oriented along the galaxy major axis. The data were collected using the 2688 $\times$ 512 pixel FAST3 CCD camera. We used two configurations: a) a 600 l mm$^{-1}$ grating giving a spectral coverage of the 3800 – 5700 Å or 5500 – 7800Å wavelength range in the blue and red part of the spectrum respectively with a dispersion of 0.75Å pixel$^{-1}$ b) a 300 l mm$^{-1}$ grating covering the spectral range of 3400 – 7200 Å with a dispersion of 1.47 Å pixel$ ^{-1} $. The first configuration was used for the observation of 39 galaxies of which 23 also have SDSS spectra. The second configuration provided simultaneous coverage of all diagnostic lines of interest with sufficient resolution and was used for most of the galaxies. The resolution of the 300 l mm$^{-1}$ grating was 5.993 Å as measured from the full width at half maximum (FWHM) of the \[O<span style="font-variant:small-caps;">i</span>\] 5577 Å sky line. Two to three exposures were acquired for each object to correct for the presence of cosmic rays on the images. The individual exposure times varied between 600 and 1800 seconds in order to achieve a S/N of 40 at the H$\alpha$ emission line. For large galaxies covering a major portion of the slit, separate sky exposures were obtained as close as possible to the location of the galaxies in the sky to allow proper sky subtraction. Spectrophotometric standard-star exposures from the list of [@Massey88] were also obtained during each night and used for flux calibration of the spectra.
The initial reduction of the long-slit spectra was performed with <span style="font-variant:small-caps;">iraf</span> (Image Reduction and Analysis Facility). Standard procedures of bias subtraction, flat fielding, and wavelength calibration were applied to the two-dimensional CCD images. The different exposures of each galaxy were combined in order to remove cosmic ray imprints on the data. Two types of one-dimensional spectra were extracted with <span style="font-variant:small-caps;">iraf</span>’s <span style="font-variant:small-caps;">apall</span> task: integrated spectra with aperture size matched to the point where the galaxy’s flux reached the level of the background and nuclear spectra, extracted from a $3.5\arcsec \times 3\arcsec$ aperture, determined from the average seeing at the time of the observations. The standard star spectra were processed in the same way as the galactic spectra. We identified the strongest lines on the extracted spectra and measured their location with <span style="font-variant:small-caps;">iraf</span>’s <span style="font-variant:small-caps;">splot</span> task. The measured emission-line centroids were used to establish galaxy redshifts and shift all spectra to the rest frame. This step was also performed on the SDSS spectra. During the SFRS long-slit campaign (2010$-$2014), 45 galaxies with no previous SDSS data were added to the SDSS data releases. We have analyzed these galaxies to examine the agreement, or possible deviations, between long-slit and fiber spectroscopy.
Starlight Subtraction
---------------------
The stellar continuum was removed from both the long-slit (integrated and nuclear) and the SDSS spectra using the <span style="font-variant:small-caps;">starlight</span> v.04 code [@STARLIGHT] to model the galaxy spectra. <span style="font-variant:small-caps;">starlight</span> fits the stellar continuum on emission-line-free regions of the spectrum using linear combinations of simple stellar populations (SSPs) from the Bruzual & Charlot ([-@BC03]) (hereafter BC03) libraries. We used a base of 138 SSPs with 23 ages ranging between 1 Myr and 13 Gyr and 6 metallicities from 0.005 to 2.5$Z\sun$. The extinction ($A_{V}$) was also fitted by <span style="font-variant:small-caps;">starlight</span>, assuming the [@CCM89] reddening law with the ratio of total to selective extinction $R_{V} \equiv A_{V}/E(B-V) = 3.1 $. In order to handle any age mixture biases in the fitting process for galaxies having an excessive blueness when $A_{V}$ is close to 0, we allowed $A_{V}$ to take negative values as discussed by [@STARLIGHT_TEST]. Negative $A_{V}$ was also allowed in studies using different fitting methods (e.g., [@Kauffmann03]). A statistical justification for this unphysical condition is that when $A_{V}$ is truly = 0, unbiased estimates should oscillate around 0, including both negative and positive values. There are only 14 cases showing negative extinctions but very close to zero, with a median of $-0.08$.
An initial assessment of starlight subtraction results can be done by comparing the estimated $A_{V}$ extinction with the Two-Micron All-Sky Survey (2MASS) $H-K$ colors. It has been recognized for some time that infrared colors could be used to directly measure extinction (e.g., @Hyland80; @Lada94). Fig. \[Ext1\] shows the theoretical $H-K$ reddening curve for K-type and O-type stars and the fit to <span style="font-variant:small-caps;">starlight</span> estimated extinction. The observed correlation between the two quantities indicates reasonably good fitting results. Based on the fitted spectra, the range of equivalent widths of H$\alpha$ and H$\beta$ absorption lines are 2.49Å $<$ EW(H$\beta$) $<$ 4.70Å and 2.05Å $<$ EW(H$\alpha$) $<$ 3.04Å. Appendix \[Balmer-test\] shows that the uncertainties in the calculated EWs are $\la$0.2 Å and should have negligible effect on the results.
![Comparison between the $A_{V}$ extinction calculated by <span style="font-variant:small-caps;">starlight</span> and the 2MASS $H-K$ colors in Vega magnitudes. The black solid line is the fit to the data. The red dashed line is the theoretical $H-K$ reddening curve [@CCM89], assuming intrinsic $ (H-K)_{o} = 0.1 $ for K-type stars [@Blum00]. Similarly, the green dash-dot line is the theoretical $H-K$ reddening curve, assuming intrinsic $ (H-K)_{o} = -0.1 $ based on the O-type stars calibrations by [@Martins06], representing the bluest possible stellar populations. NGC 4688 is the galaxy with $A_{V}=-0.68$ also reported with $E(B-V)=0$ by [@Ho97c].[]{data-label="Ext1"}](PLOTS/H-K_AV_v6.pdf)
<span style="font-variant:small-caps;">starlight</span> also provides physical parameters such as the dominant stellar populations in the host galaxies, their corresponding metallicities, stellar ages, and masses. However adopting these parameters as the actual or most representative parameters of the input galaxies is not a trivial matter, as many degeneracies are involved in the fitting process and cannot be neglected. (See Appendix \[ST-SIMs\].) Despite the uncertainties in the parameters of the stellar populations, <span style="font-variant:small-caps;">starlight</span> appears to be robust in measuring, fitting, and removing the stellar features (continuum and absorption lines) in our spectra. We reached this result by fitting the same spectra using different sets of stellar population.
Figure \[spectrum\] shows an example of the observed, <span style="font-variant:small-caps;">starlight</span>-fitted, and starlight-subtracted spectra of NGC 2712. All the SFRS galaxy spectra are available online from MNRAS.
![The observed (black line), <span style="font-variant:small-caps;">starlight</span>-fitted (red line), and starlight-subtracted (blue line) spectra of NGC 2712 observed with the FAST spectrograph.[]{data-label="spectrum"}](PLOTS/NGC2712.pdf)
Emission Line Measurements {#emission-lines}
--------------------------
The emission lines of interest were fitted with <span style="font-variant:small-caps;">sherpa</span> [@SHERPA], a general-purpose fitting environment designed for CIAO, the Chandra (X-Ray Observatory) Interactive Analysis of Observations software package [@CIAO]. <span style="font-variant:small-caps;">sherpa</span> can handle multi-component models (Fig. \[sherpa\]) such as emission lines with broad components (e.g., Seyfert 1 galaxies) or doubled-peaked line profiles resulting from galaxy rotation. One of the most important advantages of <span style="font-variant:small-caps;">sherpa</span> is its capability of fitting spectra accounting for the flux density uncertainties. We fitted the intensity, full width at half-maximum (FWHM), and central wavelength of each emission line using Gaussian profiles and a constant for the local continuum. The latter is a good approximation given that we are fitting the starlight-subtracted spectra. Uncertainties for the parameters were calculated at the 1$\sigma$ confidence intervals. This was done by varying a parameter’s value in a grid of values while at the same time the values of the other model parameters were allowed to float to new best-fit values.
We measured the intensities of the forbidden lines \[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$5007, \[N<span style="font-variant:small-caps;">ii</span>\] $\lambda$6583, \[S<span style="font-variant:small-caps;">ii</span>\] $\lambda\lambda$6716, 6731, \[O<span style="font-variant:small-caps;">i</span>\] $\lambda$6300, and the Balmer H$\beta$ and H$\alpha$ lines. For five galaxies (NGC 3758, UGC 8058, IRAS 13218+0552, MK 268, and UGC 9412) having spectra with extremely broad lines (Fig. \[Type1AGN\]) no emission line fitting was performed, since from the profile of the lines they could be clearly classified as AGN. Therefore, they were assigned a Sy-1 classification (see Section \[Final\_Class\]). The emission-line intensity ratios with respect to the H$\alpha$ or H$\beta$ lines used in the three activity classification diagnostic diagrams (see Section \[sec:BPT\]), along with the intensity of the H$\alpha$ and the <span style="font-variant:small-caps;">starlight</span> fitted extinction, are presented in Table \[table1\] along with the corresponding H$\alpha$/H$\beta$ ratio.
![Example of a multi-component emission-line fit with <span style="font-variant:small-caps;">sherpa</span> of the H$ \alpha $ and \[N<span style="font-variant:small-caps;">ii</span>\] $\lambda\lambda$6548,6583 doublet in the nuclear spectrum of Seyfert IC 910. An additional broad component is used in this case. Points show starlight-subtracted data. Orange lines show four individual emission line components, and red line shows the sum which corresponds to the best fit.[]{data-label="sherpa"}](PLOTS/NUC_IC910-Ha_v1.pdf)
![An example of a Type-1 AGN spectrum (NGC 3758) at the H$ \alpha $ line wavelength region, where no emission line fitting was performed. For five such galaxies a Sy-1 classification was automatically assigned.[]{data-label="Type1AGN"}](PLOTS/NGC3758.pdf)
H$ \alpha $ emission line extinction correction {#ext-cor}
-----------------------------------------------
Because the emission-line ratios used in our analysis are very close in wavelength, they are almost completely insensitive to reddening and therefore we do not perform extinction correction to the respective emission lines. However, because we measured nuclear SFR (Section \[Z-SFR\]) based on L$_{\textrm{H}\alpha} $, we did correct the H$ \alpha $ flux. The extinction was derived from the relative observed strengths of the H$ \alpha $ and H$ \beta $ Balmer lines. The intrinsic Balmer decrement remains roughly constant for typical gas conditions, and in our analysis we assumed an intrinsic flux ratio value (FH$ \alpha/$FH$ \beta $)$ _{0} $ of 2.86 for H<span style="font-variant:small-caps;">ii</span> regions, corresponding to a temperature T = 10$ ^{4} $ K and an electron density n$ _{e} $ = 10$ ^{2} $ cm$ ^{-3} $ for Case B recombination [@Osterbrock89]. For the narrow line region of AGN a value of 3.1 is generally adopted, where H$ \alpha $ emission is slightly enhanced by collisional excitation due to higher density gas and a much higher ionizing continuum.
We adopted the [@CCM89] extinction law and calculated the gas extinction A$ _{V}$ based on the following relation: $$\label{ext-gas}
A_{V} = \frac{\log_{10}[(FH\alpha/FH\beta)/(FH\alpha/FH\beta)_{0}]}{0.4 \times (A_{H\beta} - A_{H\alpha})},$$ where $A_{H\beta}$ and $A_{H\alpha}$ are the magnitude attenuations at the H$ \alpha $ and H$ \beta $ wavelengths. The extinction-corrected H$ \alpha $ fluxes and the corresponding gas reddening E(B-V) values calculated from Eq. \[ext-gas\] and $R_{V} \equiv A_{V}/E(B-V) = 3.1 $, are given in Table \[table1\].
Fiber and nuclear long-slit spectra emission-line comparison {#FAST-SDSS}
------------------------------------------------------------
Fig. \[F-S-Flux\] shows a comparison between the nuclear fluxes measured from long-slit spectroscopy and SDSS fibers for the 45 galaxies observed with both methods. The measured scatter between the two methods is 0.45 dex for H$\alpha $ and 0.37 dex for \[O<span style="font-variant:small-caps;">iii</span>\] $\lambda5007$, and the largest deviations appear for galaxies in interacting groups. Depending on distance, the $3.5\arcsec \times 3\arcsec$ long-slit aperture and the 3-radius circular aperture may encompass different fractions of galaxy light as well as companion galaxy light. Fig. \[F-S-Lines\] compares the four line ratios used in the three BPT classification diagrams. There is an overall good agreement between the two observing methods especially in the cases of the most commonly used [\[O<span style="font-variant:small-caps;">iii</span>\]$/$H$\beta$]{} and [\[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{} ratios. Part of the observed scatter can be attributed to sensitivity to the exact co-alignment of the fiber and the long slit, which translates to a difference between the areas covered by the two methods. In addition, in the case of the [\[O<span style="font-variant:small-caps;">i</span>\]$/$H$\alpha$]{} and [\[S<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{} ratios, the scatter is also attributed to the generally weaker [\[O<span style="font-variant:small-caps;">i</span>\]]{} and [\[S<span style="font-variant:small-caps;">ii</span>\]]{} emission lines. Indeed, the largest deviations are observed in the few cases of low S/N long-slit spectra, where the fitting of the emission-line profiles is more uncertain. In certain cases, galaxies observed with the 600 l mm$^{-1}$ grating configuration had the blue and red part of their spectrum observed on different nights. As a result, differences in the flux calibration of the two spectral regions can occur depending on the observational conditions and/or the (mis)alignment of the long slit at the exact same location on the galaxy between observing nights. For these cases, we used the emission lines measured from the long-slit observations for the nuclear activity classification because the line ratios used in the optical diagrams are very close in wavelength and in the same spectrum (blue or red). However, for the nuclear SFR measurements (Section \[Z-SFR\]), we used the SDSS spectra, where available, to calculate the Balmer decrement and obtain extinction corrected H$
\alpha $ fluxes. Seven star-forming galaxies observed with the 600 l mm$^{-1}$ grating configuration and having mismatches between their red and blue continua have no available SDSS spectra, and Table \[Nuc-Host-Z-SFR\] flags their SFR measurements as uncertain.
![Comparison between the diagnostic line ratios derived from the nuclear long-slit and fiber spectroscopy. FAST measurements are shown on the ordinate and SDSS measurements on the abscissa. All measurements represent the logarithmic values of the line ratios. One-to-one lines are shown.[]{data-label="F-S-Lines"}](PLOTS/F-S_Line_Ratio_Comparison_v4.pdf)
Activity Classification {#Classification}
=======================
Optical Diagnostics {#sec:BPT}
-------------------
The standard optical activity diagnostic diagrams of \[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$5007$/$H$\beta$ versus \[N<span style="font-variant:small-caps;">ii</span>\] $\lambda$6583$/$H$\alpha$, \[S<span style="font-variant:small-caps;">ii</span>\] $\lambda\lambda$6716, 6731$/$H$\alpha$, and \[O<span style="font-variant:small-caps;">i</span>\] $\lambda$6300$/$H$\alpha$ (@BPT (hereafter BPT); @VO87) are the primary classifiers for the galactic nuclear activity in our sample (Fig. \[ALL\_BPT\]). The standard BPT diagrams are probably the most widely used and best calibrated activity classification diagnostics for galaxies in the nearby universe ($z < 0.4$). These diagrams are able to distinguish star formation from AGN powered galaxies because the relative intensity of spectral lines with different excitation energies depends on the hardness of the ionizing continuum. The diagrams were refined by Kewley et al. ([-@Kewley01]) and Kauffmann et al. ([-@Kauffmann03]), with the former defining a theoretical upper bound to the location of the SFGs using H<span style="font-variant:small-caps;">ii</span> region model spectra and the latter empirically separating the pure star forming galaxies in the \[O<span style="font-variant:small-caps;">iii</span>\]$/$H$\beta$ versus \[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$ diagnostic based on a sample of 122,808 galaxies from the SDSS. Objects that host both starburst and AGN activity lie between these lines and are known as composite or transition objects (TO; e.g., @Ho93). Another refinement on the \[S<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$ and \[O<span style="font-variant:small-caps;">i</span>\]$/$H$\alpha$ diagrams was introduced by [@Kewley06], who calculated an empirical separating line in the AGN plane distinguishing the Seyfert populations from the Low-Ionization Nuclear Emission-Line Regions (LINER) [@Heckman80]. Lastly, [@Schawinski07] defined an empirical line in the \[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$ BPT diagnostic separating Seyfert from LINER classes, similarly to the [@Kewley06] empirical lines.
Infrared Diagnostics {#IR-diagnostics}
--------------------
While the BPT method performs well in distinguishing the different type of energy mechanisms in the majority of nearby galaxies, it is insensitive to faint or highly obscured AGN where the characteristic diagnostic lines may be faint or veiled by the gaseous and dusty torus. For these cases, diagnostics employing infrared lines and/or continuum measurements appear more sensitive than optical methods. Stern et al. [-@Stern05] used mid-infrared photometry to distinguish AGN dominated from star-forming galaxy SEDs. Galaxies that lie within the following region in the IRAC color–color space are defined as AGN: $$\label{stern-eq}
\begin{split}
([5.8]-[8.0])&>0.6, \\
([3.6]-[4.5])&>0.2\cdot([5.8]-[8.0])+0.18 , ~ \textrm{and} \\
([3.6]-[4.5])&>2.5\cdot([5.8]-[8.0])-3.5
\end{split}$$ where \[3.6\], \[4.5\], \[5.8\], \[8.0\] are the IRAC magnitudes at 3.6, 4.5, 5.8, and 8.0 correspondingly in the Vega system.
In an effort to uncover even the most “hidden" cases of galaxies hosting AGN activity, we applied the [@Stern05] criteria to the SFRS galaxies, using both integrated IRAC colors as well as colors from their nuclear region. The use of nuclear colors has the advantage of decreasing host-galaxy light contamination that can potentially mask the presence of AGN. In order to measure nuclear colors in a consistent way avoiding aperture effects due to varying galaxy distances, we performed matched aperture photometry from physical sizes of $1 \times 1$ kpc$^{2}$ to all galaxies having robust IRAC data in all bands. The apertures were centered on the nucleus, based on the source coordinates (@SFRS), and the angular aperture sizes in each galaxy were adjusted according to galaxy distance to ensure consistent sampling of linear scales, following the method described by [@Maragkoudakis17]. Fig. \[Stern-plot\] shows the comparison of the nuclear and integrated color–color plots. Galaxies that are BPT-classified as H<span style="font-variant:small-caps;">ii</span>, TO, or LINERs and lie within the AGN region defined by Eq \[stern-eq\] in both nuclear and total IRAC colors are considered obscured AGN and assigned a Sy classification.
Another criterion to diagnose AGN is based on IR continuum measurements such as the flux-density ratio of the 25 and 60 IRAS bandpasses: “Warm" sources with $F25 /F60 > 0.2$ are indicative of AGN signatures [@Sanders88]. Because many SFRS galaxies are not reliably detected by IRAS in the 25 band, we used MIPS 24 flux measurements, which on average closely track the IRAS 25 [@Dale09]. There are 6 galaxies above the $F25 /F60 > 0.2$ threshold that are not recovered by the other diagnostic methods (3 BPT diagrams or the IRAC color-color plot). However, half of them were reported as starbursts by [@Balzano83], and therefore we do not adopt those 6 galaxies as AGN. Ashby et al. [-@SFRS] used the [@Stern05] integrated IRAC color diagram, the $F_{24}/F_{60}$ ratio criterion, along with a preliminary application of the BPT method (using the [\[N<span style="font-variant:small-caps;">ii</span>\] $\lambda$6583$/$H$\alpha$]{} diagram) for the available at the time SDSS spectra to obtain an initial sense of the AGN content for the SFRS sample. Of the 165 SDSS sources, 30 were identified as AGN with the BPT method, while 19 and 22 galaxies from the entire sample where assigned an AGN classification using the [@Stern05] wedge and the $F_{24}/F_{60}$ ratio respectively, giving a total of 52 AGN-powered galaxies when accounting for overlaps among the methods.
Final Activity Classification {#Final_Class}
-----------------------------
The final adopted activity classification for the SFRS galaxies – H<span style="font-variant:small-caps;">ii</span>, Sy, LINER, TO – was based on a combination of the three variations of the optical emission-line diagrams (BPT), supplemented by the [@Stern05] IR color–color diagnostics to account for obscured AGN. Specifically, galaxies classified as H<span style="font-variant:small-caps;">ii</span> in all three diagrams or classified as TO in the [\[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{} BPT and H<span style="font-variant:small-caps;">ii</span> in the remaining two diagrams where assigned H<span style="font-variant:small-caps;">ii</span> class. In general, the TO class is identified explicitly on the [\[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{} diagram. However, by definition TOs are considered to have a composite contribution from both star forming and AGN activity. Therefore, we assigned TO classification when all of the following three conditions apply: (i) defined as TO in the [\[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{} diagnostic; (ii) have Sy or LINER classification in one of the other BPTs; and (iii) have H<span style="font-variant:small-caps;">ii</span> classification in the remaining diagram. Sy and LINERs were defined when presenting an AGN or TO class in the [\[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{} BPT and Sy or LINER classification accordingly in both [\[S<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{} and [\[O<span style="font-variant:small-caps;">i</span>\]$/$H$\alpha$]{} diagrams. Cases of non-unanimous classification were visually inspected in all BPT diagrams (see Appendix \[Notes\]), based on both their positions and their uncertainties in the BPTs, to derive their final classification. To quantify the BPT activity classification uncertainties, especially for the ambiguous cases, we performed Monte Carlo sampling to all 4 diagnostic emission line ratios based on their corresponding ratio uncertainties and derived the probability of a galaxy to belong to a certain class in each BPT diagram (see Appendix \[Probabilistic-class\]). As mentioned in Section \[emission-lines\], five galaxies were classified Sy-1 on the basis of extremely broad lines. Finally, 6 H<span style="font-variant:small-caps;">ii</span> galaxies, 1 LINER, and 3 TOs that were IR-diagnosed as AGN under the nuclear or total IRAC-color criteria were reassigned a Sy classification. As a result, based on our combined analysis the SFRS sample consists of: 269 (73%) star-forming galaxies, 50 (13%) Seyferts (including 3C 273 and OJ 287), 33 (9%) LINERs, and 17 (5%) TOs.
MEx - CEx Diagrams {#MEx-CEx}
------------------
Recently, several other diagnostics were proposed to distinguish between the different activity types for intermediate or high redshift galaxies. These include the Mass-Excitation (MEx) diagram [@MEx], which substitutes for the standard [\[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{} ratio of the BPT diagram and plots the \[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$5007/H$\beta$ against stellar mass, and the Color-Excitation (CEx) diagram [@CEx] of \[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$5007$/$H$\beta$ versus the rest-frame $U - B$ color. These can prove beneficial in the case of high-redshift galaxies, as [\[N<span style="font-variant:small-caps;">ii</span>\] $\lambda$6583]{} and H$\alpha$ can only be observed in the optical out to $z \la$ 0.45. These methods also avoid blending of lines in low-resolution spectra. In the MEx diagram, two empirically determined dividing lines [@MEx2] separate the star-forming, TO, and AGN regions. In the CEx diagram, [@CEx] provided a demarcation line to distinguish the star-forming and AGN populations, and [@MEx] further added a line to mark the region of TO populations.
We used the MEx and CEx diagrams to compare the activity classification derived with the other diagnostics. The SFRS stellar masses were calculated using the asymptotic $K-$band fluxes measured from fitting the galaxy profiles (Bonfini et al. in preparation) and using the $ M/L $ calibrations described by [@Bell03]: $$\label{eq:MassKs} \frac{M_{\star}}{M_{\sun}} = 10^{-0.273+(0.091)(u-r)} \times \frac{L_{K_{s}}}{L_{K_{\sun}}}$$ where $u-r$ are the Petrosian SDSS colors in AB magnitudes. Because the MEx diagram was calibrated based on a Chabrier [-@Chabrier03] initial mass function (IMF) and the [@Bell03] mass-to-light ratio calibration assumes a Salpeter IMF, we converted the stellar masses to the Chabrier IMF following [@Longhetti09].
Fig. \[MEx\] and Fig. \[CEx\] show the SFRS galaxies on the MEx and CEx diagram respectively. While the plots visualize the comparison between the BPT and MEx-CEx results, the classification from MEx and CEx methods shown in Table \[table2\] was obtained using the IDL routines provided by @MEx ([-@MEx]; [-@MEx2]). These routines give the probabilities for a given object to belong to each activity class depending on its position on the MEx and CEx diagrams and were calibrated based on the bivariate distribution of a $ z \sim 0 $ SDSS DR7 sample of galaxies with known activity types.
![Mass-excitation [@MEx] diagram for the SFRS galaxies. Lines [@MEx2] divide the classes with AGN to the upper right, SFGs to the lower left, and TO in the narrow middle region. The point colors are based on the BPT-only classification; blue: star-forming galaxies, red: Seyfert, orange: TO, green: LINER.[]{data-label="MEx"}](PLOTS/SFRS_MEx_Diagram_v9.pdf)
![Color-excitation diagram [@CEx] for the SFRS galaxies. The red line separates AGN (above the line) from SFGs. The black curve was added by [@MEx] and follows the transition where the AGN probability is P(AGN) $> 30\%$. The new region between this curve and the straight lines is analogous to the MEx-intermediate region of the MEx diagram. The point colors are based on the BPT-only classification; blue: star-forming galaxies, red: Seyfert, orange: TO, green: LINER.[]{data-label="CEx"}](PLOTS/SFRS_CEx_Diagram_RF_v4.pdf)
Gas-Phase Metallicities And Nuclear Star Formation Rates {#Z-SFR}
========================================================
The most direct way of measuring gas-phase metallicities is based on temperature-sensitive line ratios such as the \[O<span style="font-variant:small-caps;">iii</span>\] $\lambda\lambda$4959,5007$/$[\[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$4363]{} because the flux ratio of lines originating from different excitation states is temperature-dependent, and metals are the primary coolants of H <span style="font-variant:small-caps;">ii</span> regions. Higher metallicity increases the rate of collisional excitation followed by radiative decay resulting in lower nebular temperatures. Unfortunately, this method cannot be used when [\[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$4363]{} is undetectable or extremely weak as it often is. To overcome this difficulty, several empirical or theoretical calibrations involving strong emission lines were developed (e.g. @Pagel, @KD02). Here we derive metallicities using two different calibrations as provided by Pettini & Pagel ([-@PP04], hereafter PPO4): $$\label{eq:N2l}
12 + \log(\textrm{O/H}) = 8.90 + 0.57 \times \textrm{N2}$$ $$\label{eq:O3N2}
12 + \log(\textrm{O/H}) = 8.73 - 0.32 \times \textrm{O3N2}$$ where $\textrm{N2}\equiv\log([\textrm{N}\,\textsc{ii}]\,\lambda6583/\textrm{H}\alpha)$, and $\textrm{O3N2}\equiv\log(($[\[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$5007]{}$/\textrm{H}\beta)/($[\[N<span style="font-variant:small-caps;">ii</span>\] $\lambda$6583$/$H$\alpha$]{}$))$. Equation (\[eq:O3N2\]) is valid in the range of $-1 < \textrm{O3N2} < 1.9$. These relations were calibrated using the electron temperature (T$_{e}$)-based metallicity for a sample of 137 H<span style="font-variant:small-caps;">ii</span> regions. Because the \[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$4363 line was not available in all of our spectra, we used these calibrations to derive the nuclear-region metallicities for the SFGs and the host-galaxy metallicities for all galaxies with available long-slit spectra.
Table \[Nuc-Host-Z-SFR\] presents the circumnuclear and host-galaxy gas-phase metallicities of the star-forming SFRS galaxies as derived from both N2 and O3N2 calibrations. For galaxies with available long-slit spectra, we also measured host-galaxy elemental abundances, regardless of activity type, by subtracting the nucleus contribution from the integrated spectrum (Table \[non-HII-Z\]). While in certain cases the long-slit or SDSS fibers don’t cover the entire galaxy surface, and the intensity of the individual lines does depend on the area of the galaxy covered by the slit, their ratio on the other hand does not provided the slit or fiber samples a representative section of the galaxy.
Because the emission line spectrum, characterized in the BPT diagrams, is dependent on both the shape of the ionizing radiation as well as on the metallicity, we derived abundance gradients for 12 large, face-on galaxies observed with the FAST spectrograph in order to measure any systematic effects or variations on the average metallicity measured from the integrated long-slit spectra. We extracted spectra from 3.5-wide sub-apertures from successive regions of the galactic disks, starting from the nucleus and moving outwards on both sides along the slit. We performed the standard analysis and starlight subtraction procedure on each individual sub-spectrum and measured the [\[N<span style="font-variant:small-caps;">ii</span>\] $\lambda$6583]{}, [\[O<span style="font-variant:small-caps;">iii</span>\] $\lambda$5007]{}, H$\alpha$, and H$\beta$ emission lines on the corresponding regions. Using the PP04 calibrations, we measured each region’s metallicity with respect to the galactocentric radius, normalized to the disk radius ($R_{25}$) at the $B = 25$ mag arcsec$^{-2}$ isophote (Fig. \[Z-Gradients\]). The abundance gradients were derived using linear regression analysis on the metallicity values calculated with calibrations (\[eq:N2l\]) and (\[eq:O3N2\]). The basic parameters of the 12 galaxies used to derive metallicity gradients are summarized in Table \[z-grad\], along with the gradient slopes and the abundances at the central and characteristic radius $r = 0.4R_{25}$.
Nuclear SFRs were calculated for the SFRS SFGs (Table \[Nuc-Host-Z-SFR\]) based on the extinction-corrected H$\alpha$ emission line luminosity. The H$\alpha$ luminosity was measured from the $3.5\arcsec$-wide spectral apertures in long-slit data or the SDSS 3-diameter fibers. These correspond to 1.3 and 1.1 kpc respectively at the median distance of the SFRS galaxies. The SFRs were derived based on the [@KE12] calibration: $$\label{eq:SFRHa}
\frac{\textrm{SFR}_{\textrm{H}\alpha}}{(\textrm{M}_{\sun}/\textrm{yr)}} = \frac{L_{\textrm{H}\alpha}}{\textrm{(erg/s)}} \times 10^{-41.27}$$ The nuclear SFRs range between $10^{-5}$ and 6.15 M$_{\sun}$ yr$^{-1}$ with mean 0.26 M$_{\sun}$ yr$^{-1}$ and median 0.08 M$_{\sun}$ yr$^{-1}$.
![Abundance gradients for the sub-sample of 12 SFRS galaxies. Red points represent metallicities calculated based on the O3N2 calibration, and blue points using the N2 calibration. The red and blue lines are the fit to the O3N2 and N2 calibrations correspondingly. Circles and triangles represent metallicities measured from regions on the left and on the right side of the slit respectively.[]{data-label="Z-Gradients"}](PLOTS/Z-Gradients_v2.pdf)
Results And Discussion {#Discussion}
======================
Activity Demographics of IR-selected galaxies
---------------------------------------------
The weights assigned to the SFRS galaxies project the SFRS sample back to the parent PSC$z$ population. Therefore, given the SFRS activity classifications we are able to deliver activity classification fractions for the parent sample (PSC$z$). Specifically, the PSC$z$ comprises 71% H<span style="font-variant:small-caps;">ii</span>, 13% Seyferts, 3% TOs, and 13% LINERs. The close similarity between the SFRS demographics and its parent sample demonstrates the representative nature of the SFRS with respect to local IR-selected galaxies.
The fraction of all AGNs (Sy, TO, broad-line LINERs) in the PSC$z$ is 20%. 13% of the PSC$z$ galaxies are Seyferts, of which 39% show broad-line spectral features. The mean FWHM of SFRS galaxies fitted with a broad H$ \alpha $ component is $ \sim$2923 km s$ ^{-1} $. The landmark galaxy demographics study of [@Ho97] (H97), which used the optical magnitude-limited ($B_{T} \leq 12.5$ mag) sample of the Palomar survey, found 10% Seyfert galaxies of which 20% showed broad lines.The PSC$z$ 39% broad-line fraction is substantially higher than the 20% of H97. The PSC$z$ broad-line LINERs constitute 25%, comparable to the 23% found by H97. In contrast to the PSC$z$, H97 identified a relatively large number of low-luminosity AGNs (LLAGN) with a higher fraction of TOs and LINERs (13% and 19% correspondingly) as opposed to the PSC$z$ (3% and 13% correspondingly). In the case of TOs, this can be attributed to the strict conditions we applied to define them based on all three BPT diagnostics (Section \[Final\_Class\]), contrary to the common identification based solely on the [\[N<span style="font-variant:small-caps;">ii</span>\] $\lambda$6583$/$H$\alpha$]{} BPT. Furthermore, the higher fraction of LLAGN in the H97 sample can be attributed to the larger number of less-active and early-type systems present in the H97 sample.
Star-forming galaxies are the dominant galaxy class in both the SFRS (73%) and PSC$z$ (71%). Despite the broad range in the three-dimensional parameter space covered by the SFRS (Section \[SFRS\]), far IR selection is oriented towards selecting star-forming populations, ensuring a small AGN “contamination". This result is of particular importance considering that IR-selection methods are expected to capture both the star-forming phenomenon, as stars are born embedded in dense concentrations of dust and gas, as well as obscured AGN activity. However, the cosmic decrease in the AGN bolometric energy density (e.g., @AH12) indicating a downsizing of AGN activity and a subsequent increase of LLAGN in the low-redshift Universe is in agreement with the demographics of H97 but seems at odds with our results. The discrepancy arises because optical selection criteria such as those of H97, are not biased against passive or elliptical galaxies and are able to capture a fair number of LLAGN systems, while IR-selection criteria show a strong preference for SFGs over LLAGN populations.
Classification Comparison {#sec:classification-comparison}
-------------------------
The BPT and MEx diagrams provide consistent classification for 245 out of the 355 galaxies which lie within the predefined bound of the parameter space covered by the SDSS calibration sample (Section \[MEx-CEx\]) and have available stellar-mass measurements. However, 57 BPT-defined SFGs are classified as TOs by MEx, and likewise 6 BPT-SFGs fall within the Sy region of the MEx diagram. These represent 66% and 15% of the TO and Sy galaxies defined with the MEx method. In reverse 7 and 6 BPT-classified Sy and TOs have an H<span style="font-variant:small-caps;">ii</span> classification in the MEx diagram. For BPT-classified LINERs, 12 have a TO classification, 10 have an H<span style="font-variant:small-caps;">ii</span> classification, and 5 are classified as Sy in the MEx diagnostic. An important drawback of the MEx diagram is the fact that it is mass dependent and does not rely on direct observable measurements such as flux ratios. This is important because different prescriptions in the $ M/L $ calibrations or SED fitting parameters used to determine stellar masses will generally yield different classification results.
The CEx method classifies 300 galaxies that fall within the calibration regions of the CEx diagram. The activity types for 190 of those galaxies are in agreement with the BPT method. Similarly to the MEx classification, the CEx method tends to classify more BPT-SFGs as TO (64 galaxies representing 75% of the CEx TO class). In addition, 10 BPT-Sy and 4 BPT-TOs are classified as LINERs (33% of the CEx LINER class) with the CEx method. The CEx diagram is expected to be biased against broad-line AGNs where the broadband color is not dominated by the host galaxy. Furthermore, as reported by [@CEx], the emphasis given in the empirical definition of the demarcation line in the CEx diagram was based on limiting contamination of the AGN calibration sample.
The IRAC color classification scheme does not provide detailed demographics on the different activity sub-classes. The separation is done purely on the basis whether a galaxy is classified as an SFG or contains AGN activity and is classified as AGN, regardless its type (Sy, LINER, TO) or intensity. This is evident in Figure \[Stern-plot\], where objects of both Type-1 (broad-line) and Type-2 (narrow-line) classes are found within the AGN “wedge”. Specifically, taking into account both the nuclear and integrated color approaches of the IRAC color diagrams (Section \[IR-diagnostics\]), 28 out of the 369 SFRS galaxies are classified as AGN based on the empirical [@Stern05] criteria, 10 of which had a non-Sy classification in the BPT diagrams (6 BPT-H<span style="font-variant:small-caps;">ii</span>, 3 BPT-TOs, and 1 BPT-LINER). Furthermore, the BPT and IRAC colors diagnostics classify unanimously 72% (265 objects) of the SFRS galaxies as SFGs and 5% (17 objects) as AGN (BPT-Sy).
There are 45 SFRS galaxies which have both nuclear long-slit and SDSS spectra. There is a good agreement between the two observing methods with 38, 39, and 27 galaxies having the same classification in the [\[N<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{}, [\[S<span style="font-variant:small-caps;">ii</span>\]$/$H$\alpha$]{}, and [\[O<span style="font-variant:small-caps;">i</span>\]$/$H$\alpha$]{} BPT diagrams respectively. Fig. \[F-S-Comp\] show the galaxies with discrepant classifications in one or more BPT diagrams. Most cases presenting a difference between the two methods agree within 3$\sigma$ of the line ratios.
Galaxy Properties With Respect To Activity Type {#galaxy_properties}
-----------------------------------------------
Being an IR-selected sample, the PSC$z$ contains mostly star-forming galaxies. SFGs have a broad range in 60$\micron$ luminosity between $6.53 < \log( L(60_{\micron})/L_{\sun}) < 11.18$ as seen in Fig. \[SFRSHist\], highlighting the fact that the SFRS selection criteria capture all amplitudes of star-forming activity. AGN hosts (Sy and TO) are preferentially found at $L(60\micron) > 10^{9} L_{\sun}$ while LINERs can be seen at lower $L(60\micron)$ with a drop at luminosities higher than $\sim$10$^{9} L_{\sun}$, indicative of hosts without intense star formation. Interestingly, Type-1 and Type-2 Sy and LINERs show similar $L(60\micron)$ distributions with $p$-values of 0.86 and 0.26 respectively returned from a Kolmogorov-Smirnov (K-S) test.
The $F60/F100$ flux ratio distribution of SFGs which is a proxy for the dust temperature (Fig. \[SFRSHist\]) peaks around 0.45, which is close to the limit ($\sim$0.5) above which starburst galaxies are typically found (@RC89). Seyfert galaxies have a broad distribution of $F60/F100$ and present a peak around $F60/100 \sim 0.85 $ attributed to AGN radiation heating dust to high temperatures. TOs have a dust-temperature distribution with median 0.46, close to the median of SFGs at 0.49, implying that the major contribution to dust heating is from star formation. On the other hand $F60/F100$ in LINERs is confined to lower values, implying both a low radiation density and less dusty environments, the latter typical of early-type galaxies. Again, Type-1 and Type-2 Sy and LINERs have similar $F60/F100$ distributions (K-S $p$-values 0.93 and 0.56 respectively).
The nuclear H$\alpha$ luminosity distribution, serving also as a proxy of SFR for non-AGN hosts, ranges over 7 orders of magnitudes ($10^{36} - 10^{44}$ erg s$ ^{-1} $) in the case of SFGs. Fig. \[SFRSHist\] shows that Sy show a similar broad range of $L_{\textrm{H}\alpha}/\textrm{kpc}^{2} $ distribution as SFGs ($p$-value = 0.87). In a different manner, TOs and LINERs present lower $ L_{\textrm{H}\alpha}/\textrm{kpc}^{2} $ values as opposed to Seyferts, pointing towards gas-deficient hosts. The opposite view is seen in the case of Seyferts, where the AGN enhances gas ionization, and the $ L_{\textrm{H}\alpha}/\textrm{kpc}^{2} $ distribution peaks at $10^{40.75}$ erg s$ ^{-1} $.
The SFGs cover a broad range of stellar masses between $10^{7.64}-10^{11.56}$ M$ _{\sun} $ as shown in Fig. \[MassHist\]. The [@Bell03] relations on which mass estimations were based should ideally be applied exclusively to star-forming populations because the presence of non-stellar emission from AGN contaminates the NIR band fluxes, resulting in an overestimation of stellar mass. Therefore, the stellar masses of non-SFG populations should be considered upper limits. Fig. \[MassHist\] shows that all galaxy types other than SFGs are preferentially located in massive hosts. With the exception of broad-line LINERs that are clearly related to the AGN phenomenon, LINER emission in local galaxies is now considered to result from photoionisation by hot evolved stars and not AGN (@CidFernandes11; @Belfiore16). Therefore, mass estimates of narrow-line LINERs using $ M/L $ calibrations should be accurate provided correct stellar populations are assumed. LINER nuclei are predominantly found at higher stellar masses with a median of $10^{10.52}$ M$ _{\sun} $. Mass estimates of TOs where the ionization continuum stems from both star-forming and AGN processes should in most cases be treated as upper limits similarly to Sy mass estimations. The median host-galaxy properties of the SFRS sample and the parent population (an unbiased subset of the PSC$z$) from which SFRS was defined are summarized in Table \[Median\_Prop\].
![Stellar mass distribution for the different activity types. Blue lines show SFGs (labeled as H<span style="font-variant:small-caps;">ii</span>), red lines show Seyfert galaxies, orange lines TO, and green lines LINERs. Masses refer to Chabrier [-@Chabrier03] IMF. Galaxy numbers refer to the SFRS parent sample.[]{data-label="MassHist"}](PLOTS/SFRS_Mass_Hist_v6.pdf)
-------------------- ------------ ---------------- -------------------- --------------------------------------- ------------------- -----------
$\log L_{60} $ $\log \Sigma_{L_{\textrm{H}\alpha}} $ $\log M_{\star} $ Distance
Activity Class No. (L$ _{\sun} $) $ F_{60}/F_{100} $ (erg s$ ^{-1} $kpc$ ^{-2} $) (M$ _{\sun} $) (Mpc)
All galaxies 369 (2560) 9.55 (9.55) 0.49 (0.49) 40.08 (40.10) 10.49 (10.51) 77 (71)
SFGs 269 (1815) 9.41 (9.47) 0.48 (0.49) 40.15 (40.17) 10.35 (10.43) 69 (69)
Seyferts 50 (345) 10.07 (10.24) 0.71 (0.80) 40.24 (40.26) 10.87 (10.82) 117 (105)
Transition Objects 17 (79) 9.72 (9.71) 0.41 (0.46) 39.62 (39.60) 10.82 (10.80) 114 (108)
LINERs 33 (321) 9.26 (9.04) 0.45 (0.36) 39.50 (39.53) 10.55 (10.52) 54 (34)
LIRGs / ULIRGs 117 / 11 10.14 (10.18) 0.51 (0.55) 40.08 (40.15) 11.16 (11.07) 147 (131)
-------------------- ------------ ---------------- -------------------- --------------------------------------- ------------------- -----------
Dominant Central Ionizing Mechanism in LINERs and TOs
-----------------------------------------------------
TOs are considered to be galaxies with composite contributions from star-forming and AGN processes in their central energy output. LINERs on the other hand have a much more controversial status in the literature regarding their main ionizing energy source. They are sometimes considered as the low-luminosity, low-accretion-rate branch of Seyfert galaxies (e.g., @Ho93, @Ho03), which is still a viable assertion at least for LINERs showing broad H$ \alpha $ features. However, this view is disputed in the case of narrow-line LINERs considering that post asymptotic giant branch (post-AGB) stars can produce the required ionizing spectrum necessary to excite LINER emission (e.g., @Binette94, @Stasinska08, @Belfiore16).
The $L(60\micron)$ distribution of TOs follows that of SFGs ($p$-value = 0.08) with a median of $\log L(60\micron) = 9.7$ L$ _{\sun} $ and 9.5 L$ _{\sun} $ respectively, indicative of a similar star-forming activity profile in their host galaxies. Similarly, the median $ F60/F100 $ ratio of both TOs and SFGs is 0.46 and is distinct from the median of Seyferts at 0.80, suggesting star-forming activity as the main source of dust heating in TOs. The $L_{\textrm{H}\alpha}$ surface density distribution of TOs has a smaller range than Sy and SFGs, with median $ L_{\textrm{H}\alpha}/\textrm{kpc}^{2} = 10^{40.30} $ erg s$ ^{-1} $kpc$ ^{-2} $. While the host-galaxy stellar mass is not a property directly linked to the main source of ionization, it gives a sense of the environment where the TOs reside. TOs are allocated predominately at masses above $10^{10.5} $. However, as discussed in Section \[galaxy\_properties\], Sy-host masses should be considered only as upper limits, and this would apply to TOs depending on their dominant ionizing mechanism. Nevertheless, given the previously discussed similarities of TOs host-galaxy properties with those of SFGs, star-forming activity should be considered as the main source of ionization in the SFRS (PSC$z$) TOs. Given this assertion, TOs reside in massive hosts.
LINERs in the SFRS sample show a sharp drop in their number densities above $L(60_{\micron}) \sim 10^{9} L_{\sun}$, illustrating weak star-forming activity in their hosts. The bulk of LINERs are concentrated mostly at lower values of $ F60/F100 $ with a median of 0.36, with just 12 LINERs lying above the 0.5 starburst threshold as opposed to 118, 38, and 3 SFGs, Sy, and TOs respectively. Furthermore, the $L_{\textrm{H}\alpha}$ surface density distribution of LINERs is confined in the narrow range of $ 10^{38.5} - 10^{40.5} $ erg s$ ^{-1} $kpc$ ^{-2} $ and drops significantly above that, as opposed to Sy and SFGs. Lastly, LINERs are found predominantly in high stellar-mass hosts at a median of $ 10^{10.5}\ \textrm{M}_{\sun}$. Given their generally low H$ \alpha $ and 60 luminosity it is unlikely that this is the result of bias due to AGN contamination. The above properties demonstrate that the host-galaxy and nuclear properties of LINER bear close resemblance to those of passive early-type systems, hinting that their ionizing continuum stems mostly from older stellar populations (i.e., post-AGB stars) rather than AGN activity. Nevertheless, if LINERs should be considered as AGN powered systems, given the fact that they reside in more massive hosts they should be associated with more massive central black holes.
LIRGs and ULIRGs {#LIRGs}
----------------
The class of LIRGs is characterized by total IR luminosities ($L(\textrm{TIR})$) between $ 10^{11} - 10^{12}$ L$ _{\sun} $, while galaxies with $L(\textrm{TIR}) > 10^{12} $ L$ _{\sun} $ are described as ULIRGs. The SFRS sample contains 117 LIRGs and 11 ULIRGs. Previous studies have demonstrated that a varying fraction of local LIRGs and ULIRGs, ranging from 25% [@Veilleux97] to 70% [@Nardini10], host an AGN. In the case of SFRS (U)LIRGs, 25% are Sy hosts, while this percentage increases to 43% when including LINERs and TOs. [@Lee11] using a sample of 115 ULIRGs showed that Type-2 AGNs are more frequently encountered (49 galaxies) compared to Type-1 AGNs (8 galaxies), and the percentage of type 2 ULIRGs increases with infrared luminosity. The SFRS LIRGs consist of 45 AGN, considering all non-SFG classes, out of which 13 are Type-1. Similarly 10 out of 11 SFRS ULIRGs are AGN, with 4 being Type-1, showing an agreement with previous studies. The joined properties of SFRS LIRGs and ULIRGs are summarized in Table \[Median\_Prop\].
Metallicities {#sec:Metallicity-Disc}
-------------
The SFG’s metallicities derived from the O3N2 and N2 calibrations agree for galaxies with sub-solar metallicities but differ for metallicities larger than solar. However, as discussed by PP04, the O3N2 calibration is particularly useful at solar and super-solar metallicities where \[N<span style="font-variant:small-caps;">ii</span>\] saturates, but the strength of \[O<span style="font-variant:small-caps;">iii</span>\] continues to decrease with increasing metallicity. Furthermore, the abundance gradient slopes derived from the O3N2 calibration are generally comparable to the N2 gradients (Fig. \[Z-Gradients\]). However, the O3N2 calibration indicates in general higher metallicities for the central galactic regions but shows similar values at the characteristic radii of $r = 0.4R_{25}$ (Table \[z-grad\]). There is also a good agreement between the nuclear abundances measured directly from the central aperture and the ones calculated from the metallicity gradients at $r = 0$ in both calibrations.
The SFG’s abundances have a narrow range close to solar (12 + $\log$(O/H) = 8.66), with a median value of 8.72 and 8.67 in the O3N2 and N2 calibrations respectively. The sub-sample of 12 galaxies used to derive metallicity gradients show overall flat metallicity profiles as a function of galactocentric radius (Table \[z-grad\] and Fig. \[Z-Gradients\]). The relatively flat profiles derived from our analysis seem to be at odds with similar studies in nearby galaxies (e.g., @Pilyugin04; @Moustakas10). However, flat average metallicity profiles were also observed by [@Moran12] (M12) in a sample of 174 local star-forming galaxies using the O3N2 calibration by PP04. As discussed by M12, galaxies from their local sample with $\log(M_{\star}) > 10.2$ appear to have flatter metallicity profiles and metallicities close to solar. Furthermore, M12 argued that inner metallicity profiles which decline steadily with radius are observed only at the lowest masses. In addition, M12 re-examined the 21 galaxies from the [@Moustakas10] sample and estimated the stellar masses of these galaxies, noting that five out of eight galaxies with gradients have $\log(M_{\star}) < 10.2$. Nine out of 12 galaxies SFRS galaxies used to derive metallicity gradients in this paper have $\log(M_{\star}) > 10.2$. Similarly, 178 out of the 261 SFRS SFGs have $\log(M_{\star}) > 10.2$ and metallicities close to solar, indicating that they are consistent with the metallicity gradients observed in local galaxies. Furthermore, the small abundance gradients with respect to galactocentric radius ensures that on average the emission lines measured from the central extraction aperture are not affected by metallicity, regardless the portion of host-galaxy light encompassed, and therefore the resulting classification is robust, capturing the hardness of the central ionizing radiation spectrum.
Conclusions {#Conclusions}
===========
Using a fully representative sample of FIR-selected galaxies in the local Universe, we have presented the activity demographics, nuclear metallicities, and host-galaxy properties of the 369 galaxies in the SFRS and derived metallicity gradients for a sub-sample of 12 galaxies. Using the weights assigned to the SFRS galaxies that project the SFRS sample back to the parent PSCz population, we derived the activity demographics and host-galaxy properties for the parent population, summarized as follows.
\(i) BPT classification of FIR-selected galaxies gives 71% SFGs, 13% Seyferts, 13% LINERs, and 3% TOs.
\(ii) NIR classification [@Stern05] reveals 6 H<span style="font-variant:small-caps;">ii</span> galaxies, 1 LINER, and 3 TOs as defined with the BPT method to be AGN galaxies.
\(iii) The MEx method is in agreement with the BPT classification for 69% of galaxies but tends to classify 21% of BPT-SFGs as TO and 2% as Sy.
\(iv) The CEx method agrees with BPT classification in 63% of the cases but classifies 24% of BPT-SFGs as TO and 21% as LINER.
\(v) The inferred fraction of LLAGN galaxies in the local Universe using IR sample-selection criteria is small compared to SFGs.
\(vi) In terms of host-galaxy properties, Sy and TOs are preferentially found at $L(60\micron) > 10^{9}\textrm{L}_{\sun}$ while LINERs have lower $L(60\micron)$ values with few at luminosities higher than $10^{9.5}\textrm{L}_{\sun}$, indicative of hosts with no intense star-formation. SFGs cover a broad range in L(60$\mu$m) between $6.5 < \log( L(60_{\micron})/\textrm{L}_{\sun}) < 11.18$.
\(vii) Seyfert galaxies have the highest $F60/F100$ ratios. These can be attributed to AGN radiation heating dust to high temperatures. TOs have the same median dust-temperature as SFGs implying star-formation as the dominant ionizing mechanism, while LINERs are confined to lower dust-temperatures indicative of low-ionization mechanisms and dust-deprived environments.
\(viii) The stellar mass distribution of SFGs covers a broad range ($10^{7.79}-10^{12.14}$ M$ _{\sun} $), while TOs and Seyfert are found exclusively in high stellar-mass hosts, albeit their mass estimations are considered mostly as upper limits. LINERs, which are now strongly considered as galaxies powered by non-AGN processes, are found in high stellar masses galaxies with a median of 10$ ^{10.5} \textrm{M}_{\sun} $.
\(ix) SFGs have a wide distribution of nuclear $L_{\textrm{H}\alpha}$ surface densities, ranging between $10^{36.8} < L_{\textrm{H}\alpha}/\textrm{kpc}^{2} < 10^{43.0}$ erg s$ ^{-1} $kpc$ ^{-2} $, and TOs show a similar distribution. Sy nuclear $ L_{\textrm{H}\alpha}/\textrm{kpc}^{2} $ is concentrated above $ 10^{39.5} $ erg s$ ^{-1} $kpc$ ^{-2} $, while LINERs on the other hand are mainly distributed at lower values with median $ L_{\textrm{H}\alpha}/\textrm{kpc}^{2} = 10^{40.18} $ erg s$ ^{-1} $kpc$ ^{-2} $.
\(x) Based on their host-galaxy and nuclear properties, the dominant ionizing source in the SFRS (PSC$z$) TOs is star-forming activity. Similarly, LINER host-galaxy characteristics resemble those of passive early-type systems, indicating older stellar populations as their main ionizing source, rather than AGN activity.
\(xi) The SFRS SFGs have a narrow range of abundances close to solar and have flat metallicity radial profiles, as found from a sub-sample of 12 galaxies. This is evident for galaxies with $\log(M_{\star}) > 10.2$, as discussed by M12, which is the case for 68% of the SFRS SFGs.
ACKNOWLEDGMENTS {#acknowledgments .unnumbered}
===============
We would like to thank the referee for the constructive comments and suggestions that have improved the clarity of this paper. AM and AZ acknowledge funding from the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement n. 617001. Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/. This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. 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. For this research, we have made extensive use of the Tool for Operations on Catalogues And Tables (TOPCAT) software package (@TOPCAT).
Probabilistic Activity Classification {#Probabilistic-class}
=====================================
Galaxy activity classification based on the BPT diagrams often stumbles upon the complication of deriving an unambiguous or consistent classification from all three diagnostic diagrams, especially in the occasions where galaxies fall near the demarcation lines. Therefore, obtaining a sense of the uncertainty in a galaxy’s classification is important, especially for the ambiguous cases. The uncertainties of the intensities of the diagnostic lines can be translated into an uncertainty in a galaxy’s activity classification. We quantified this activity classification uncertainty by calculating the probability that a galaxy falls in the locus of a given class. For each galaxy we generated 1000 samples of their line ratios drawn from a Gaussian distribution, based on their line-ratio uncertainties in all three BPT diagrams, and examined their placement and resulting activity classification on each diagnostic. Thus we derived the probability for each galaxy to be classified as H<span style="font-variant:small-caps;">ii</span>, TO, or AGN (Seyfert or LINER) in each of the three diagnostic diagrams. The probabilistic activity classifications of the SFRS sample are presented in Table \[Prob-AC\].
STARLIGHT Code Simulation Study {#ST-SIMs}
===============================
Modeling the input spectrum
---------------------------
### Statistical fluctuations {#ST-Flux_Dist}
A primary concern when measuring emission lines in galaxy spectra is the efficient subtraction of the underlying stellar continuum and absorption features at the wavelengths where the emission lines of interest reside. While spectral synthesis codes suffer from known degeneracies where different combinations of SSP properties (e.g., ages, metallicities) and galaxy properties (e.g., extinction, AGN component) can fairly reproduce the observed spectrum, to a first level, any combination reproducing the input spectrum is considered adequate for starlight subtraction. However, due to <span style="font-variant:small-caps;">starlight</span>’s complicated architecture where the fit is carried out using a mixture of simulated annealing, Metropolis algorithm, and Markov Chain Monte Carlo techniques, intrinsic uncertainties are introduced in the process, and even when making identical-parameter runs the results can slightly vary.
To quantify the above uncertainties in <span style="font-variant:small-caps;">starlight</span>, we performed 100 identical-parameter runs for two galaxies observed with SDSS (IC 4395) and FAST (NGC 5147), using a two-set grid of base spectra, one with 138 SSPs and one with 45 SSPs. In each run we subtracted the model spectrum from the observed and fitted the H$ \beta $ and H$ \alpha $ emission lines using <span style="font-variant:small-caps;">sherpa</span>. The distributions of the resulting emission line fluxes F(H$ \beta $) and F(H$ \alpha $) for the two sets of 100 runs (using the 138 and 45 SSP grids respectively) are shown in Figure \[STARLIGHT-Hist\]. While the spread of F(H$ \beta $) is somewhat larger compared to F(H$ \alpha $), especially when using the 45 SSPs grid, the variance between the values is within the measured line-flux uncertainties. Specifically, in all cases the standard deviation of the F(H$ \beta $) and F(H$ \alpha $) distributions is smaller by a factor of 10 compared to the uncertainties of the measured fluxes. These results are independent of the number of runs performed. We verified this by repeating the tests for 500 identical-parameter runs and recovering a factor of 10 difference between the standard deviation of the fluxes and their measured uncertainties. Therefore, the intrinsic uncertainties introduced by <span style="font-variant:small-caps;">starlight</span> have no impact on our results.
  
### The Balmer regions {#Balmer-test}
<span style="font-variant:small-caps;">starlight</span> masks the most notorious optical emission lines in the spectrum including the Balmer series before performing a fit. While Balmer lines can be valuable in characterizing the ages of the underlying stellar populations, the presence of an emission component at the same wavelength of the absorption feature makes their usage impractical. Therefore the fit is based on modeling the emission-free regions of the galaxy spectrum. Assessing the goodness of the fit in the Balmer regions is not feasible when modeling emission-line galaxies, but it is important because an inadequate fit can lead to the over- or under-subtraction of the absorption components, which will have an impact on the measured emission lines.
In order to determine the accuracy of <span style="font-variant:small-caps;">starlight</span> in modeling the Balmer absorption lines, we performed simulations using a modified version of the SED modeling code CIGALE [@CIGALE2]. We used the high-resolution BC03 libraries (instead of the default low-resolution BC03 libraries) with a predefined set of star-formation histories and dust attenuations. Specifically, we generated galaxies with a single star-formation episode decaying with an e-folding time fixed at 50 Myr. Onsets were modeled at 50, 100, 200, 400, 500, 1000, 1500, and 2000 Myr before now. These ages sample the range of Balmer equivalent widths for both young and old populations. Each of the 8 base SEDs was reddened by one of three $E(B-V)$ color excess values: 0.16, 0.32, or 0.65 mag. CIGALE applied these reddening values only to stars $<$10 Myr old with older stars being reddened by 0.001 times as much. No gas or nebular emission was included to ensure the creation of emission-free galaxy SEDs. The models gave a total of 24 galaxy spectra referred to as “reference” spectra (Table \[Balmer-Table\]). The model results are as expected with Balmer equivalent widths peaking at ages near 500 Myr. Because reddening was applied mainly to the youngest populations, it reduces their contribution to the emerging light and slightly increases the equivalent width in the resulting light output at each age. This effect becomes negligible for populations $\ga$400 Myr old.
In order to simulate observations, we generated a set of 100 spectra for each of the 24 reference SEDs adding Gaussian noise. The noise amplitude was the median uncertainty of the measured H$ \alpha $ emission line of the SFRS galaxies. We ran <span style="font-variant:small-caps;">starlight</span> for the 2400 spectra and obtained the best-fit model for each, referred to as “model” spectra. We then computed the mean and standard deviation of $\rm EW_{mod} - EW_{ref} $ for both H$ \beta $ and H$ \alpha $. Table \[Balmer-Table\] summarizes the results.
The models reproduce the absorption EWs reasonably well, showing standard deviation in H$\beta$ EW $\sim$0.1 Å for the youngest populations and less than that for populations older than 400 Myr. Differences from the reference values are of the same order and always $\le$0.18 Å except for the very youngest populations. H$\beta$ is by far the more important line because the stellar absorption EW is about twice that of H$ \alpha $, and the emission line EW is typically 1/5 of that for H$ \alpha $. The lowest emission EW of H$\beta $ measured in the SFRS sample is 0.19 Å for a single galaxy, and the second lowest value is 0.45 Å. Therefore, <span style="font-variant:small-caps;">starlight</span> over- or under-fitting the absorption lines should not have a significant impact on the emission line measurements themselves or on the results for either the classifications or the H$\alpha$/H$\beta$ extinction estimates (Sec. \[ext-cor\]).
Global galaxy parameters and AGN component
------------------------------------------
[@STARLIGHT_TEST] (F14) made an extensive study of the <span style="font-variant:small-caps;">starlight</span> code using 1638-zone spectra of the Sb NGC 2916, observed with an Integral Field Unit (IFU). They performed simulations using observed and synthetic spectra (built from <span style="font-variant:small-caps;">starlight</span> fits to the observed data) to explore uncertainties related to the data and the spectral synthesis method. They added random noise and shape-changing perturbations in both observed and synthetic spectra in order to address noise fluctuations and continuum shape calibrations. These simulations were then used to evaluate uncertainties in the global parameters of the stellar populations derived by <span style="font-variant:small-caps;">starlight</span> such as the mean ages, masses, metallicities, extinction, and star formation histories (SFHs). Despite the fact that <span style="font-variant:small-caps;">starlight</span> solutions depend on the seed for the random-number generator, given the pseudo-random nature of its Markov chains, variations of this kind have a tiny effect on the derived properties. Furthermore, synthetic unperturbed spectra produce higher dispersions of the $\Delta$ (simulation minus original) value of each derived quantity, compared to the observed unperturbed spectra, attributed to the fact that observed spectra have access to a smaller subspace of acceptable solutions than the synthetic spectra. When introducing random noise, F14 found that the mass-weighted ages and metallicities have broader $\Delta$ distributions ($\sigma\sim$0.15 dex) than the luminosity-weighted counterparts ($\sigma\sim$0.1 dex), while stellar masses are good to $\sim$0.1 dex. Overall, the SFHs appear less constrained than other global galaxy parameters, being a higher-order product of the spectral synthesis. Averaging over spatial regions comprising many zones reduces the uncertainties.
Similar but simpler simulations were also performed by [@STARLIGHT] to 65 test galaxies created from the mean derived properties of fits to a sample of SDSS galaxies. The results showed that the population vectors recovered by <span style="font-variant:small-caps;">starlight</span> are subject to large uncertainties, but mean derived properties such as stellar ages and metallicities were better recovered.
As a supplementary analysis to those studies, we performed a set of simulations using only individual sets of SSPs or, at most, two-component composite stellar populations (CSPs) from the BC03 libraries. We used populations of different ages (5 Myr$-$10 Gyr) at a given metallicity (0.2 Z$_{\sun}$) with a 10% uncertainty on the flux as input in <span style="font-variant:small-caps;">starlight</span>. The top panels in Figs. \[ST-SIMS1\], \[ST-SIMS2\], \[ST-SIMS3\], and \[ST-SIMS4\] shows the current mass fraction against age of the input and output spectra that contributes more than 10% to current stellar mass. <span style="font-variant:small-caps;">starlight</span> is more sensitive in identifying the actual contribution of the younger populations, while it systematically underestimates the contribution of older populations. In the latter case, while successfully recovering the actual population, it uses supplementary SSPs of similar ages to match the input spectrum. Therefore, in the simplest case of a SSP, <span style="font-variant:small-caps;">starlight</span> is able to identify both the age and metallicity of the input spectrum but not necessarily its actual contribution (100% in the case of a SSP).
---------------------------- ----------------------------
 
 
---------------------------- ----------------------------
To further assess <span style="font-variant:small-caps;">starlight</span>’s results, we created synthetic spectra, adding noise and extinction separately on the same SSPs used previously, in an effort to simulate an actual spectrum in a fully controlled environment. With the inclusion of Gaussian noise (2% of the input flux, Fig. \[NOISE\]), there is a deviation from the actual contribution to the current mass for all populations, and in some cases inclusion of extra SSPs, but the dominant component describes the correct age and metallicity (Fig. \[ST-SIMS2\], second panel).
![Example of a SSP spectrum (blue line) and the same SSP spectrum perturbed with 2% Gaussian noise (red line)[]{data-label="NOISE"}](PLOTS/ST_SIM_NOISE_v1.pdf)
----------------------------- -----------------------------
 
 
----------------------------- -----------------------------
Adding extinction (A$_{V}$ = 0.85) without noise to the input spectrum will force <span style="font-variant:small-caps;">starlight</span> in almost every case to split the total contribution using extra SSPs for the fit (Fig. \[ST-SIMS3\], third panel). While for populations between 100 Myr and 5 Gyr the highest contribution comes from the correct SSP component, in most cases that component is modeled below 50% of the total. However, <span style="font-variant:small-caps;">starlight</span> will correctly fit the extinction of the input spectrum.
------------------------------ ------------------------------
 
 
------------------------------ ------------------------------
Another test we applied was to limit the spectral area <span style="font-variant:small-caps;">starlight</span> uses to fit the input spectrum by masking most spectral regions except those containing Lick indices, which are characteristic and representative of the stellar populations. In most cases <span style="font-variant:small-caps;">starlight</span> was able to attribute more accurately the actual contribution to the input spectrum or even limit the number of extra SSP used in the fit, but the overall improvement was not significant over the default configuration.
Finally, we tested <span style="font-variant:small-caps;">starlight</span>’s sensitivity in uncovering the AGN component of a galaxy spectrum using power-law AGN templates of different indices along with the standard SSPs. We created a synthetic spectrum from an SSP adding an AGN component (Fig. \[AGN-spectrum\]) and provided STARLIGHT with the same AGN template. Although <span style="font-variant:small-caps;">starlight</span> revealed the correct AGN component, the stellar populations used in the fit varied significantly from the input SSP used to create the synthetic spectrum (Fig. \[ST-SIMS4\], bottom panel). Furthermore, if more AGN templates of different power-law indices are allowed to be used in the fit, <span style="font-variant:small-caps;">starlight</span> will use most of them to reproduce the input spectrum. Additionally, by fitting a SSP without AGN component, but allowing <span style="font-variant:small-caps;">starlight</span> to use AGN templates in the fitting process (which illustrates the case of fitting an actual galaxy spectrum without a prior knowledge for the presence of an AGN or not), <span style="font-variant:small-caps;">starlight</span> has a tendency to use the AGN templates present.
![Example of a composite AGN spectrum (red line), composed by a SSP spectrum (green line) and an AGN power-law spectrum (blue line).[]{data-label="AGN-spectrum"}](PLOTS/ST_SIM_AGN_v3.pdf)
------------------------------ ------------------------------
 
 
------------------------------ ------------------------------
As a verdict, <span style="font-variant:small-caps;">starlight</span> heads towards the right direction in uncovering the actual fitted stellar population, but even in ideal environments, absent noise and extinction, older populations are harder to uncover. The presence of an AGN is even trickier to handle, as in the absence of AGN templates the AGN continuum is attributed to and fitted with stellar populations. With the inclusion of AGN templates, <span style="font-variant:small-caps;">starlight</span> tends to over-assign an AGN contribution to the fitted spectrum even when the galaxy lacks AGN activity. In reality, an actual galactic spectrum has a much higher complexity than independent SSPs, but <span style="font-variant:small-caps;">starlight</span> remains robust in fitting and allowing the SED of the stellar component to be subtracted.
Notes On Individual Galaxies {#Notes}
============================
SFRS 30 ( = *OJ 287*) is a blazar and is assigned a Sy classification to match the terminology adopted for the activity types in this paper.\
SFRS 42 ( = *IC 2434*) is an ambiguous case and is assigned a TO classification based on the \[N<span style="font-variant:small-caps;">ii</span>\] BPT diagram. The \[S<span style="font-variant:small-caps;">ii</span>\] $\lambda\lambda6716, 6731$ and \[O<span style="font-variant:small-caps;">i</span>\] $\lambda6300$ detections are considered upper limits.\
SFRS 57 ( = *CGCG 238-066*) is an ambiguous case in the BPT diagrams but is recognized as AGN in the IRAC color-color diagram and is assigned a Sy classification.\
SFRS 61 ( = *CGCG 181-068*) is an ambiguous case and is assigned a TO classification based on the \[N<span style="font-variant:small-caps;">ii</span>\] BPT diagnostic. The \[S<span style="font-variant:small-caps;">ii</span>\] $\lambda\lambda6716, 6731$ and \[O<span style="font-variant:small-caps;">i</span>\] $\lambda6300$ measurements are considered upper limits.\
SFRS 80 ( = *NGC 3190*) has no detectable H$\beta$ emission in the nuclear spectrum. The integrated spectrum was used to obtain activity classification (LINER).\
SFRS 93 ( = *UGC 5720*) H$\alpha$ is absent in the SDSS spectrum. The activity type (H<span style="font-variant:small-caps;">ii</span>) was determined solely from the [\[O<span style="font-variant:small-caps;">iii</span>\]$/$H$\beta$]{} value.\
SFRS 139 ( = *NGC 3758*) is a Type-1 AGN with very broad line profiles and is automatically classified as Sy.\
SFRS 148 ( = *NGC 3822*) has no \[S<span style="font-variant:small-caps;">ii</span>\] BPT classification because the \[S<span style="font-variant:small-caps;">ii</span>\] $\lambda\lambda6716, 6731$ doublet falls inside a telluric line region. It is assigned a Sy classification because in the \[O<span style="font-variant:small-caps;">i</span>\] BPT diagram the galaxy is located inside the Sy region, further away from the Sy/LINER demarcation line compared to the \[N<span style="font-variant:small-caps;">ii</span>\] diagnostic demarcation line.\
SFRS 182 ( = *NGC 4194*) H$\alpha$ is absent in the SDSS spectrum. The activity type (H<span style="font-variant:small-caps;">ii</span>) was determined solely from the [\[O<span style="font-variant:small-caps;">iii</span>\]$/$H$\beta$]{} value.\
SFRS 187 ( = *NGC 4237*) is an ambiguous case, classified as TO/H<span style="font-variant:small-caps;">ii</span> in the N<span style="font-variant:small-caps;">ii</span>/S<span style="font-variant:small-caps;">ii</span>-BPT diagnostics. Because the \[OI\] line is unreliable, this galaxy is assigned an H<span style="font-variant:small-caps;">ii</span> classification.\
SFRS 201 ( = *NGC 4435*) is an ambiguous case and is assigned a LINER classification because in the \[O<span style="font-variant:small-caps;">i</span>\] BPT is located in the middle of the LINER region, while there is no \[S<span style="font-variant:small-caps;">ii</span>\] $\lambda\lambda6716, 6731$ detection.\
SFRS 204 ( = *3C 273*) is a quasar and is assigned a Sy classification to match the terminology adopted for the activity types in this paper.\
SFRS 227 ( = *NGC 4689*) is assigned an H<span style="font-variant:small-caps;">ii</span> classification because in the \[S<span style="font-variant:small-caps;">ii</span>\] BPT diagram the galaxy is located inside the H<span style="font-variant:small-caps;">ii</span> region further away from the demarcation line. There is no \[O<span style="font-variant:small-caps;">i</span>\] $\lambda6300$ detection.\
SFRS 228 ( = *NGC 4688*) is assigned an H<span style="font-variant:small-caps;">ii</span> classification because it is located further away from the demarcation line and into the H<span style="font-variant:small-caps;">ii</span> region in the \[N<span style="font-variant:small-caps;">ii</span>\] BPT diagram, while it falls right on the line in the \[S<span style="font-variant:small-caps;">ii</span>\] BPT diagnostic. There is no \[O<span style="font-variant:small-caps;">i</span>\] $\lambda6300$ detection.\
SFRS 233 ( = *MCG 8-23-097*) is an ambiguous case and is assigned a LINER classification.\
SFRS 239 ( = *UGC 8058*) is a Type-1 AGN with very broad line profiles and is automatically classified as Sy.\
SFRS 259 ( = *NGC 5104*) is an ambiguous case and is assigned a Sy classification.\
SFRS 261 ( = *NGC 5112*) is an ambiguous case and is assigned a TO classification because it falls on the \[S<span style="font-variant:small-caps;">ii</span>\] BPT demarcation line, while in the \[O<span style="font-variant:small-caps;">i</span>\] BPT diagram the emission-line uncertainties can place the galaxy further inside the Sy region.\
SFRS 262 ( = *NGC 5123*) is assigned an H<span style="font-variant:small-caps;">ii</span> classification. There is no \[O<span style="font-variant:small-caps;">i</span>\] $\lambda6300$ detection.\
SFRS 263 ( = *IRAS 13218+0552*) is a Type-1 AGN with very broad line profiles and is automatically classified as Sy.\
SFRS 266 ( = *NGC 5204*) is an ambiguous case, and is assigned an H<span style="font-variant:small-caps;">ii</span> classification. There is no \[O<span style="font-variant:small-caps;">i</span>\] $\lambda6300$ detection. Visual inspection of the SDSS image shows bright blue colors, characteristic of star-forming galaxies.\
SFRS 270 ( = *IRAS 13349+2438*) is a literature verified QSO (e.g. @Lee13) and was assigned a Sy classification.\
SFRS 276 ( = *MK 268*) is a Type-1 AGN with very broad line profiles and is automatically classified as Sy.\
SFRS 305 ( = *NGC 5515*) is an ambiguous case and is assigned a Sy classification. It is also a literature verified Sy (e.g. @Veron06)\
SFRS 322 ( = *UGC 9412*) is a Type-1 AGN with very broad line profiles and is automatically classified as Sy.\
SFRS 331 ( = *UGC 9618\_NED02*) has a TO/H<span style="font-variant:small-caps;">ii</span>/H<span style="font-variant:small-caps;">ii</span> class in the corresponding \[N<span style="font-variant:small-caps;">ii</span>\]/\[S<span style="font-variant:small-caps;">ii</span>\]/\[O<span style="font-variant:small-caps;">i</span>\] BPT diagnostics. However, because in the O<span style="font-variant:small-caps;">i</span> BPT diagram it falls on top of the demarcation line, it is assigned a TO classification.\
SFRS 350 ( = *UGC 10120*) has broad-line profiles and is fitted with an additional broad component along with the narrow ones when measuring its emission lines. However, it is classified as H<span style="font-variant:small-caps;">ii</span> in all three BPTs but is recognized as AGN in the IRAC color-color diagram and is reported as Sy in the literature (@Contini98). Therefore it is assigned a Sy classification.\
[^1]: Quasar 3C 273 and blazar OJ 287 are excluded from the spectroscopic studies related to star formation but are included in the rest of the SFRS analysis.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present novel hysteretic behaviour of a three-level ladder atomic system exhibiting double-cavity optical bistability in the mean-field limit. The two fields coupling the atomic system experience feedback via two independent, unidirectional, single mode ring cavities and exhibit cooperative phenomena, simultaneously. The system displays a range of rich dynamical features varying from normal switching to self pulsing and a period-doubling route to chaos for both the fields. We focus our attention to a new hump like feature in the bistable curve arising purely due to cavity induced inversion, which eventually leads to negative hysteresis in the bistable response. This is probably the only all-optical bistable system that exhibits positive as well as negative bistable hysteresis in different input field intensity regimes. For both the fields, the switching times, the associated critical slowing down, the self-pulsing characteristics, and the chaotic behaviour can be controlled to a fair degree, moreover, all these effects occur at low input light levels.'
author:
- 'H. Aswath Babu and Harshawardhan Wanare'
title: 'Negative and positive hysteresis in double-cavity optical bistability in three-level atom '
---
All-optical bistability has remained a focus of research for more than four decades [@szoke; @gibbs; @lugiato1984; @joshi_modern] and apart from its potential application as a switch [@gibbs_book] in optical communication technology it has continued to remain a test bed for fundamental research related to cooperative phenomena [@bonifacio; @vengalattore] as well as to the study of nonlinear dynamical aspects such as self-pulsing, instabilities and chaos [@gibbs1981; @ikeda1; @ikeda2; @taki]. A deeper understanding of issues related to quantum aspects like entanglement and cooperative behavior in presence or absence of instabilities would be critical in realizing a functional quantum computer. Due to recent developments related to cold-atoms in optical lattices [@zoller1] and atomic chips [@folman], the aspects related to cooperative phenomena have become vital [@pritchard]. Design of smaller trap size ($< \lambda$) demands a treatment that allows for cooperative effects, and possibly at multiple frequencies. Cooperative effects at multiple frequencies can be realized in multi-level atoms, and we show the onset of instabilities in such systems even at low light levels. Multilevel atoms have been recently used to create and control cooperative effects, such as in the multiparticle dark states [@schempp] and Rydberg blockade effects [@pfau] in cold atoms in a trap. In order to understand interplay of multi-colored cooperative effects we explore the semi-classical dynamics of three-level atoms interacting with two fields coupling two adjacent transitions at which they exhibit cooperative behavior simultaneously. The atomic level structure itself provides the coupling between the two distinct cooperative branches.
In an early work, one of the authors (HW with G.S. Agarwal) had shown the possibility of control of optical bistability [@harsha_ob] in three-level atomic medium. In that configuration, the optical bistability exhibited by one (probe) field is controlled by another (control) field coupling an adjacent transition. The field exhibiting bistability experiences conventional cavity feedback, whereas the control field is held constant without feedback. In order for the probe field to exhibit optical bistability (OB) it not only needs to interact strongly with the active media but it must experience sufficient feedback (thus satisfying the conditions necessary to exhibit cooperative behavior), which is obtained through an external unidirectional ring-cavity. Such configuration allows effective engineering of various characteristics of OB, including tailoring of thresholds, changing On/Off intensities of the output field and obtaining multistability [@joshi]. These effects were also experimentally realized [@xiao_ob]. Furthermore, three-level atom with a control field (without feedback) has been shown to exhibit instability in the context of OB [@yang].
![(Color online) (Top) Schematic of the three-level ladder system interacting with the two fields $E_1$ and $E_2$ at frequencies $\omega_1$ and $\omega_2$, respectively. (Bottom) The OB set up with the collection of atoms (within the length $L$) that interact with the two fields in two independent unidirectional, single mode ring cavities.[]{data-label="schatomfig"}](schatom.eps){width="45.00000%"}
In this work, we again have two fields that couple two adjacent transitions in a three-level atom, however, both the fields experience feedback via two independent, single mode, unidirectional ring cavities, as indicated in Fig. \[schatomfig\]. This system provides independent control over the two cooperative parameters. The fields are chosen to be counter-propagating within the active medium in order to minimize the two-photon Doppler broadening in the ladder-system, however our calculation is undertaken for a homogeneously broadened atomic gas. Such a configuration leads to [*negative*]{} hysteresis apart from the conventional positive bistable hysteresis. As mentioned above this [*single mode*]{} two cavity system leads to a variety of nonlinear dynamical effects including self-pulsing of the two fields at their distinct frequencies and a period-doubling route to chaos in the lower cooperative branch.
Nonlinear dynamics related to OB has traditionally been associated with inclusion of multi-modal treatment of the cavity field [@lugiato1984; @segard], here, we have just two modes corresponding to the two monochromatic fields coupling the adjacent transitions. The theory of two-photon amplifiers and absorbers forms the traditional basis of studying these systems, which traditionally reduce the problem to an effective two-level model and has been extensively studied [@narducci; @galatola]. An early work describes the possibility of obtaining chaos for atoms interacting with a single mode, where chaos occurs in the upper-branch for extremely large cooperative parameters accompanied by large atomic and cavity detunings [@lugiato-chaos]. There have been other studies relating to the dispersive regime of two-photon OB which deals with the effects of cavity detuning that controls both the fields simultaneously [@grangier]. In our model, the system exhibits chaos at sufficiently low-light levels in the lower branch of the OB response without being restricted to any special regime. Moreover, the independent cavities allow for an effective control of the feedback of the fields, thus tailoring the cooperative character. It should be noted that all the rich nonlinear dynamical features arise purely due to the [*interplay*]{} of the two cooperative branches.
Apart from a plethora of nonlinear dynamical features, the fields exhibit cavity induced inversion and positive as well as negative hysteresis OB, as one varies the input intensity. To the best of our knowledge, we are not aware of any other [*all-optical*]{} bistable system that exhibits chaos at low-light levels or even the negative and positive hysteresis OB. Our results are also in contravention with those reported earlier, in particular we observe enhanced output field resulting from creation of population inversion between the states $|1\rangle$ and $|2\rangle$, in presence of concurrent feedback for fields coupling [*both*]{} the transitions. The effects are particularly significant as they occur at low light levels and thus could be used as a nonlinear component in optical circuitry. The details of the nonlinear dynamical studies are presented in detail in the companion paper [@aswath_2]. The atom-field density matrix equations and the field equations governing the dynamics of the OB system in the mean-field limit are given as $$\begin{aligned}
\frac{\partial\rho}{\partial t} &=&-\frac{i}{\hbar} \left[{\hat H},\rho \right]+\cal{\hat L}
\rho,\nonumber\\
\frac{\partial x_{1}}{\partial t}& = & \kappa_{1}\left[-x_{1}(1+i
\theta_{1})+y_{1} + 2 i C_{1} \rho_{12},\right] ~\label{finaleqn1}\nonumber\\
\frac{\partial x_{2}}{\partial t}& = & \kappa_{2}\left[-x_{2}(1+i
\theta_{2})+y_{2} + 2 i C_{2} \rho_{23}.\right]
~\label{diffeqn}\end{aligned}$$ These equations describe the three-level atom coupled to the two fields at frequencies $\omega_1$ and $\omega_2$, which experience feedback through two independent cavities characterized by the cavity decay $\kappa_1$ and $\kappa_2$, with cavity detuning $\delta_1$ and $\delta_2$ (scaled as $\theta_i=\delta_i/\kappa_i$ where $i=1,2$) respectively. The total Hamiltonian in the dipole approximation after undertaking the rotating wave approximation is given as: $\hat H = \hbar(\Delta_1 + \Delta_2) |1\rangle \langle 1| + \hbar \Delta_2 |2\rangle \langle 2| -(d_{12}.E_1 |1\rangle \langle 2| + d_{23}.E_2 |2\rangle \langle 3| + \rm{h.c.})$, where $\Delta_1$ and $\Delta_2$ are the atomic detunings. The detailed density matrix equations and the scaled variables along the lines of Ref. [@lugiato1984] are all explicitly given in Ref. [@aswath_2].
We consider the mean-field limit wherein a single pass through the ring cavity only marginally affects the fields, and the strong cooperative nature arises due to the extremely large photon lifetime of the fields within the cavity, as the transmission coefficient of the cavity mirrors is chosen to be negligibly small. The last two equations in Eq. (\[diffeqn\]) arise due to the cavity feedback and the fields are determined self-consistently by the coupling with the three-level atom via the cooperative parameters ($C_1$ and $C_2$). Under steady state conditions one can obtain the various domains of stability for different input field strengths, we describe in detail such a stability domain map in Ref. [@aswath_2], wherein the regions of stable switching, unstable regions that exhibit self-pulsing and chaotic dynamics are clearly identified. The bistable behavior can be obtained for different combinations of the input/output fields while either keeping the other input field constant ([*i.e.*]{} OB response of $|E_1^{in}|$ versus $|E_1^{out}|$ as well as $|E_2^{out}|$, while holding $|E_2^{in}|$ constant and other such combinations) or, varying the other input field with a specific functional dependence.
The output levels, the threshold of switching, the range of input fields exhibiting bistability are all dependent on the parameters like the decay rates (both atomic and the cavity), the detunings (both atomic and the cavity) and the cooperative parameters corresponding to the two transitions. The novel effects arise due to the inclusion of feedback for both the fields that couple the three-level atoms. The two fields being complex (on interaction) their relative phases and strengths become critical factors in determining the dynamics. In the context of nonlinear dynamics the corresponding phase space, as well as, the parameter space is exceedingly large owing to fourteen different physical parameters ($\gamma_{1,2}$, $\Delta_{1,2}$, $\theta_{1,2}$, $\kappa_{1,2}$, $C_{1,2}$, $E^{in}_{1,2}$) all of which can be varied independently. Unlike in the OB with a control field (without feedback, as in Ref. [@harsha_ob]) it is the simultaneous interaction of the atom with both the fields and their [*independent*]{} feedback from the cavities that self-consistently determine the output fields (both their phase and amplitude). Hence, with regard to the numerical implementation we express a note of caution as this setup does not permit an [*apriori*]{} choice of both the (complex) output fields. In the conventional computation of OB (in steady state and in the mean field limit) involving feedback for one field, one usually specifies the output field (which could be chosen to be a real value) and calculate the requisite unique (complex) input field. Such strategy is conventionally adopted due to the multivalued nature of the output field, because of the input-output relationship in form of the $S-$ shaped OB curve; however a given output field uniquely determines the input field. A generalization of a such a strategy fails because it is impossible to choose the amplitude and phase of both the output fields apriori. The input fields, the cavity fields and their interaction with the medium self-consistently determine the amplitudes and phases of output fields. Without taking into account the complex nature of the fields within the cavity one would miss out the hump like feature (discussed below) at low light levels [@xiao-joshi] which is crucial and eventually transforms into the negative hysteresis as shown in Fig. \[enha\](a). In order to deal with such numerical constraints we used the Newton-Raphson method to obtain the steady state solution of the nonlinear atom-field equations along with the boundary conditions. As expected an [*arbitrary*]{} choice of relative strength and phases of the output field does not necessarily correspond to a physically viable input field variation.
![(Color online) (a) The OB response for the $\omega_2$ field for various values of the input field at $\omega_1$. (b) Enhancement of the population inversion on the $|1\rangle \leftrightarrow |2\rangle$ transition with (red) and without (blue) the feedback for the field at $\omega_1$. Similarly, (c) and (d) indicate the absorption $Im\{\rho_{23}\}$ and $Im\{\rho_{12}\}$, corresponding parameter values $|y_1|=25$ and $C_1 =1000$ (red curve), for blue dashed curve $|y_1| = 0.5$ and $C_1 =0$. The other relevant parameters are $C_1=1000,C_2=100,\kappa_i=1,\gamma_i=1,\Delta_i=0,\theta_i=0$ for $i=1,2$[]{data-label="enha"}](conversion_rij_aug.eps){width="50.00000%"}
In order to mimic a typical experimental situation we use real values for the input fields and compute the resultant complex output fields self-consistently along the different branches of the S-shaped OB response. We also note that in order to focus our attention to the new low input light OB regime in this paper, we have avoided the regimes involving multi-stability which is easily obtained in this double cavity OB system. There are two important features that we highlight here. The first one is the enhancement of the output field at $\omega_2$ coupling the $|2\rangle \leftrightarrow |3\rangle $ transition arising due to the creation of inversion in the $|1\rangle \leftrightarrow |2\rangle $ transition which leads to the hump like response and eventually to negative hysteresis OB. The second feature is the novel switching characteristics wherein the output fields exhibit self-pulsing, moreover the two fields completely mimic each other in their temporal response. The hump like feature in the OB response indicates an enhancement of the field at $\omega_2$ arising due to suppressed absorption along the $|2\rangle \leftrightarrow |3\rangle $ transition. This occurs in the lower cooperative branch at low intensities of the $\omega_2$ input field, where initially the population is dominantly in the ground state $|3\rangle$. A large cooperative parameter $C_1$ along the upper transition $|1\rangle \leftrightarrow |2\rangle$ leads to enhanced interaction with the $\omega_1$ field resulting in the extraction of a significant fraction of the population into the excited states. Furthermore, as the upper state ($|1\rangle $) population builds up there is a lowering of the influence of the field at $\omega_2$ on these atoms. This leads to the creation of inversion along the upper transition $|1\rangle \leftrightarrow |2\rangle $ [*i.e. $\rho_{11} > \rho_{22}$*]{}. This dynamics occurs due to the asymmetric choice of the cooperative parameters $C_1 > C_2$ at the cavity resonant condition $\theta_1=\theta_2 = 0$. However, similar dynamics can be obtained for comparable values of $C_1$ and $C_2$ in the bad cavity limit, [*i.e.*]{} with finite cavity detuning $\theta_2$ [@aswath_2]. The decrease in absorption ($Im\{\rho_{23}\}$) of the $\omega_2$ field, accompanied by enhanced absorption ($Im\{\rho_{12}\}$) at the $\omega_1$ field [*with*]{} and [*without*]{} the feedback for the $\omega_1$ field clearly demonstrates the reliance of the enhancement of the field at $\omega_2$ on cavity assisted inversion \[see Figs. \[enha\](b),(c) and (d)\]. With increasing incident field $y_1$, the above effects are enhanced and the hump like feature becomes quite exaggerated and eventually results in OB with a negative-hysteresis loop as indicated in Fig. \[enha\](a). A similar scenario is observed with increasing cooperative parameter $C_1$ of the upper transition $|1\rangle \leftrightarrow |2\rangle$.
We describe in detail the negative hysteresis loop arising in this system (Fig. \[crit\]). As described above the field at $\omega_2$ is enhanced, however, on further increase in the input field $y_2$ the population from the upper states is drawn back into the lower levels $|2\rangle$ and $|3\rangle$ ultimately leading to large absorption of the $\omega_2$ field, and the output switches to the [*off-*]{} state (indicated by $N_1$ in Fig. \[crit\](a)). In the reverse direction as the input intensity $y_2$ is decreased the output field switches from an [*off-*]{} state to an [*on-*]{} state along $N_2$ (different from $N_1$) due to the large $\omega_1$ field already circulating in the other cavity, thus encompassing within it a negative-hysteresis loop. This hysteresis is exactly opposite to the usual bistability (we denote as the positive hysteresis) wherein low input intensities leads to low output intensity and only for larger input intensities the transition saturates (for zero atomic detuning) leading to the large output field and the corresponding reverse loop that encloses a hysteresis (indicated as $P_1$ and $P_2$ in Fig. \[crit\](a)). This conventional positive hysteresis OB occurs at higher input field intensities where larger input field results in large output field and vice-versa.
![(Color online) (a) The OB response for the $\omega_2$ field which exhibits negative as well positive hysteresis (for $C_1=5000,C_2=20,|y_1|=50$). The color red (black) indicate stable(unstable) steady state response, and the arrows indicate the associated switching. (b) and (c) Critical slowing down associated with the negative hysteresis along $N_1$ and $N_2$, respectively. The time evolution of fields are indicated by dashed lines for the operating points $|y_{2}|_{op} = 6.5(N_1)$ and $3.5(N_2)$ and the solid lines indicate operating points $|y_{2}|_{op} = 6.2(N_1)$ and $3.6(N_2)$. The corresponding threshold points are $|y_{2}|_{th} = 6.13(N_1)$ and $4.05(N_2)$. The other relevant parameters are same as in Fig. \[enha\][]{data-label="crit"}](double_cslow.eps){width="50.00000%"}
The nature of threshold points associated with the negative hysteresis are similar to the positive hysteresis of the conventional OB and displays critical slowing down. The study of the time-dependent switching indicates that, as one chooses the operating point $|y_2|_{op}$ closer to the threshold of switching one observes an increase in the time required to switch to the steady state \[Fig. \[crit\](b-c)\]. Note that this behavior occurs at all the four threshold points $|y_2|_{th}$ associated with the switching transitions $N_{1,2}$ and $P_{1,2}$. The variation of any other parameter results in shifting of the threshold point itself and thus the associated change in the switching times due to critical slowing down.
Apart from such multicolored stable switching there is a wealth of dynamics this system can display. It exhibits periodic self-pulsing in the steady state for both the fields. A constant input intensity at both the frequencies $\omega_1$ and $\omega_2$ results in periodic output that largely mimics each other, and the periodicity can be controlled using the cavity parameters (such as $\kappa_{1,2}$ and $\theta_{1,2}$). These detailed nonlinear dynamical aspects are discussed in the companion paper [@aswath_2]. Before we conclude, we would like to point out that there is nothing special about the two-photon resonant condition that we have used in the calculations presented here, similar results are obtained even in the non-resonant cases. The nonlinear dynamics can be obtained in a robust manner in a wide variety of regimes with appropriately chosen parameters.
In conclusion, we have demonstrated a simple all optical double cavity OB system that exhibits negative as well as positive hysteresis. We self-consistently determine the amplitude and phases of both the output fields. A novel region of response involving a hump like feature in the S-shaped OB curves, as well as negative hysteresis is demonstrated at low input light levels. These effects are a consequence of the cavity induced inversion arising from the simultaneous cooperative coupling at two different frequencies. The system also exhibits a range of nonlinear dynamical features such as self-pulsing and chaos, again at low input light levels.
[99]{}
A. Szoke, V. Daneu, J. Goldhar and N.A. Kurnit, Appl. Phys. Lett. [**15**]{}, 376 (1969). H. M. Gibbs, S. L. McCall, and T. N. C. Venkatesan Phys. Rev. Lett. [**36**]{}, 1135 (1976). L.A. Lugiato, [*Progress in Optics*]{}, edited by E.Wolf (North-Holland. Amsterdam) [**Vol.XXI**]{}, 69(1984).
Joshi, Amitabh and Xiao, Min, Journal of Modern Optics. [**57: 14**]{}, 1196 (2010).
H. M. Gibbs, [*Optical Bistability: Controlling Light with Light*]{}, (Academic, New York, 1985). R. Bonifacio and L. A. Lugiato, Phys. Rev. A [**18**]{}, 1129 (1978) M. Vengalattore, M. Hafezi, M. D. Lukin, and M. Prentiss Phys. Rev. Lett. [**101**]{}, 063901 (2008). H.M. Gibbs, F.A. Hopf, D.L. Kaplan and R.L. Shoemaker, Phys. Rev. Lett. [**46**]{}, 474 (1981). K. Ikeda, H. Daido, and O. Akimoto, Phys. Rev. Lett. [**45**]{}, 709 (1980). H. Nakatsuka, S. Asaka, H. Itoh, K. Ikeda, and M. Matsuoka, Phys. Rev. Lett. [**50**]{}, 109 (1983). M. Taki, Phys. Rev. E [**56**]{}, 6033 (1997). D. Jaksch, C. Bruder, J. I. Cirac, C. W. Gardiner, and P. Zoller Phys. Rev. Lett. [**81**]{}, 3108 (1998). R. Folman, P. Kruger, D. Cassettari, B. Hessmo, T. Maier, and J. Schmiedmayer, Phys. Rev. Lett. [**84**]{}, 4749 (2000). J. D. Pritchard, D. Maxwell, A. Gauguet, K. J. Weatherill, M. P. A. Jones and C. S. Adams, arXiv:1006.4087v1 \[quant-ph\]. H. Schempp, G. Gunter, C. S. Hofmann, C. Giese, S. D. Saliba, B. D. DePaola, T. Amthor, and M. Weidemuller, S. Sevincli, and T. Pohl, Phys. Rev. Lett. [**104**]{}, 173602 (2010). R. Heidemann, U. Raitzsch, V. Bendkowsky, B. Butscher, R. Low, L. Santos, and T. Pfau, Phys. Rev. Lett. [**99**]{}, 163601 (2007). W. Harshawardhan and G.S. Agarwal, Phys. Rev. A, [**53**]{}, 1812 (1996). A. Joshi and M. Xiao, [*Progress in Optics*]{},edited by E.Wolf(North-Holland. Amsterdam) [**Vol. 49**]{}, 97 (2006). A. Joshi, A. Brown, H. Wang and M. Xiao, Phys. Rev. A, [**67**]{}, 041801 (2003). W. Yang, A. Joshi, and M. Xiao, Phys. Rev. Lett., [**95**]{}, 093902 (2005). B. Segard, B. Macke, L.A. Lugiato, F. Prati and M. Brambilla, Phys. Rev. A [**39**]{}, 703 (1989). L. M. Narducci, W. W. Eidson, P. Furcinitti and D. C. Eteson, Phys. Rev. A [**16**]{}, 1665 (1977). P. Galatola, L.A. Lugiato, M. Vadacchino and N.B. Abraham, Optics Commun., [**69**]{} 414 (1989). L.A. Lugiato, L.M. Narducci, D.K. Bandy, C.A. Pennise, Optics Commun., [**43**]{}, 281(1982). P. Grangier, J. F. Roch, J. Roger, L. A. Lugiato, E. M. Pessina, G. Scandroglio and P. Galatola, Phys. Rev. A [**46**]{}, 2735 (1992).
H. Aswath Babu, and Harshawardhan Wanare, arXiv:1009.1550v1 \[nlin.CD\]
A. Joshi, and M. Xiao, Appl. Phys. B, [**79**]{}, 65 (2004).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
**Abstract**—The charge topology of coherent-dissociation events is presented for $^{11}$C and $^{12}$N nuclei of energy 1.2 *A* GeV per nucleon bombarding nuclear track emulsions. This topology is compared with respective data for $^{7}$Be, $^{8,10}$B, $^{9,10}$C and $^{14}$N nuclei.
**DOI:** 10.1134$/$S1063778815060022
author:
- '**D. A. Artemenkov$^{\textbf{1)}}$, V. Bradnova$^{\textbf{1)}}$, A. A. Zaitsev$^{\textbf{1)}}$, P. I. Zarubin$^{\textbf{1)~*}}$, I. G. Zarubina$^{\textbf{1)}}$, R. R. Kattabekov$^{\textbf{1), 2)}}$, N. K. Kornegrutsa$^{\textbf{1)}}$, K. Z. Mamatkulov$^{\textbf{1), 3)}}$, P. A. Rukoyatkin$^{\textbf{1)}}$, V. V. Rusakova$^{\textbf{1)}}$, R. Stanoeva$^{\textbf{4)}}$.**'
title: |
Charge Topology of the Coherent Dissociation\
of Relativistic $^{11}$C and $^{12}$N Nuclei
---
Introduction
============
Light nuclei can be represented as various superpositions of bound states of lighter nuclear cores, extremely light nuclear clusters (alpha particles, tritons, $^3$He nuclei, and deuterons), and nucleons that coexist in dynamical equilibrium. Owing to this variety, the group of nuclei at the beginning of the table of isotopes provides a laboratory for studying the coexistence and evolution of cluster and shell degrees of freedom. The $^{11}$C nucleus, which exhibits a remarkable combination of cluster and shell features of the ground state, exemplify such nuclei. The isotope $^{11}$C is a connecting link between light stable nuclei featuring a pronounced alpha-particle clustering of nucleons and light nuclei at the proton drip line, where clustering that involves the isotope $^3$He is of importance. The interaction of virtual helium isotopes and neutron exchange between them in the $^{11}$C nucleus lead to the formation of cluster structures, including the 2$^{4}$He + $^{3}$He configuration. Among them, configurations characterized by low binding energies, such as $^{7}$Be + $\alpha$ (7.6 MeV), $^{10}$B + *p* (8.7 MeV) and $^{3}$He + $^{8}$Be (9.2 MeV), are expected to be more probable than the 9$^{9}$Be + 2*p* (15.3 MeV) and $^{8}$B + *t* (27.2 MeV) configurations.
Because of a wide variety of virtual cluster configurations that may exist in the structure of the $^{11}$C nucleus, investigation of this nucleus becomes an interesting problem in and of itself.
A balanced coexistence of these virtual cluster modes determines not only the ground-state properties of the $^{11}$C nucleus but also the fact that it is bound, which is of importance for obtaining deeper insight into the abundances of light isotopes. Nuclear astrophysical synthesis of the isotope $^{11}$C may proceed in a mixture of the isotopes ${}^{3}$He and ${}^{4}$He via the formation of the stable isotope $^{7}$Be or the unstable isotope $^{8}$Be, and a partial clustering into a $^{10}$B + *p* pair may follow it. The decay of the $^{11}$C nucleus leads to the formation of the stable isotope $^{11}$B, which one can observe in cosmic rays. This scenario of nucleosynthesis is not commonly recognized– the isotopes $^{10,11}$B are assumed to be products of the bombardment of carbon-star surfaces with high-energy protons. If observations reveal the dissociation of $^{11}$C nuclei through the ${}^{7}$Be + $\alpha $ and ${}^{3}$He + ${}^{8}$Be channels, this would confirm the existence of cluster modes in this nucleus that are genetically related to its synthesis.
Knowledge of the structure of $^{11}$C is necessary for interpreting data on the next isotope $^{12}$N and, in prospects, on the isotope $^{13}$O, the $^{11}$C nucleus playing the role of a core in both of them. In fast nucleosynthesis processes (hot breakout cycles), these three isotopes are genetically related waiting stations. The formation of the isotope $^{12}$C and heavier nuclei may proceed through them via the addition of protons. It is necessary to know fundamental properties of the relativistic fragmentation of the $^{11}$C nucleus in order to apply intense beams of these nuclei in nuclear medicine.
POTENTIAL OF THE NUCLEAR-EMULSION METHOD
========================================
Within the BECQUEREL project at the nuclotron of the Joint Institute for Nuclear Research (JINR) \[1\], the cluster structure of light nuclei is being studied in relativistic-fragmentation processes on the basis of the nuclear-emulsion method \[2–11\]. The development of these investigations and their illustrations are presented in the review article of Zarubin \[12\]. Among events of the fragmentation of relativistic nuclei, those of their coherent dissociation to narrow jets of fragments are especially important for studying nucleon clustering. They do not feature tracks of either slow fragments of emulsion nuclei or charged mesons. This special feature reflects the fact that the excitation of the relativistic nucleus under investigation is minimal in the case of a tangential collision with a heavy track-emulsion nucleus. Nuclear diffraction interaction \[13\] not accompanied by angular-momentum transfer is a basic mechanism of excitation of coherent dissociation in nuclear track emulsions.
The experimental method in question is based on record spatial resolution and sensitivity of nuclear track emulsions whose layers are exposed longitudinally to beams of relativistic nuclei. It has already furnished unique information about cluster aspects of the structure of the whole family of light nuclei, including radioactive ones. For practical reasons, the $^{11}$C nucleus, which is among the key ones, was skipped. A new series of investigations of the BECQUEREL Collaboration was motivated by the need for filling this gap.
Because of the absence of tracks of strongly ionizing particles, events of coherent dissociation were called white stars. The term white stars reflects successfully a sharp “breakdown” of the ionization density at the interaction vertex upon going over from the primary-nucleus track to a narrow cone of secondary tracks. This special feature generates a fundamental problem for electronic methods because more difficulties should be overcome in detecting events where the degree of dissociation is higher. On the contrary, such events in nuclear-track emulsions are observed and interpreted in the most straightforward way, and their distribution among interaction channels characterized by different compositions of charged fragments is determined exhaustively. This probabilistic distribution is a basic feature that is observed for the virtual cluster structure of the nucleus under consideration.
The probability distribution of the final configurations of fragments in white stars makes it possible to reveal their contributions to the structure of nuclei under consideration. We assumed that, in the case of dissociation, specific configurations arise at random (random-phase approximation) without sampling and that the dissociation mechanism itself does not lead to the sampling of such states via angular-momentum or isospin exchange. By and large, available results confirm the assumption that cluster features of light nuclei determine the picture of their relativistic dissociation. At the same time, events that involve the dissociation of deeply bound cluster states and which cannot arise at low collision energies are detected.
We mentioned above that, for the ${}^{11}$C nucleus, one expects the ${}^{7}$Be + $\alpha $, ${}^{10}$B + *p*, and ${}^{3}$He + ${}^{8}$Be dissociation channels, where the binding energies have low values. From the experimental point of view, the last channel is a three-body one and can involve the decays of both the ground state (0${}^{+}$) of the ${}^{8}$Be nucleus and its 2${}^{+}$ excited state. Moreover, channels that, in charge topology, correspond to the dissociation of the ${}^{7}$Be and ${}^{10}$B core nuclei should appear. Thereby, one expects that the role of multi-particle channels in the coherent cluster dissociation of $^{11}$C nuclei should be significant, so that the application of the nuclear-track-emulsion method is reasonable.
In addition, the nuclear-emulsion method should reveal multi-particle channels corresponding in charge topology to the coherent dissociation of the $^{7}$Be \[3,11\] and $^{10}$B \[14\] core nuclei in $^{11}$C. An approximate equality of the probabilities for the He + He and He + 2H dominant channels of coherent dissociation is a special feature of the $^{7}$Be nucleus (see Table 1). The respective branching ratio is 1 $\pm$ 0.2 according to data of a group from the Lebedev Physical Institute Moscow \[3\] and 0.7 ± 0.1 according to vaster data sample obtained at JINR \[11\]. The 2He + H three-body channels are leading ones(of weight about 75$\%$) among $^{10}$B white stars (see Table 2). Events corresponding to the He + 3H channel saturate 12%. Lithium and helium fragments appear simultaneously in 10% of events. Beryllium and hydrogen fragments appear only in 2% of events; this indicates that, in the structure of the $^{10}$B nucleus, the probability for the $^{9}$Be + *p* configuration is insignificant. On the contrary, the contribution of the Be + Í channel to the coherent dissociation of the $^{8}$B nucleus is dominant, which indicates that the $^{8}$B nucleus features the $^{7}$Be + *p* configuration, which contains a proton halo. The contribution of configurations that involve only helium and hydrogen clusters is estimated at a level of 50%.
Channel ${}^{7}$Be \[3\] ${}^{7}$Be \[11\]
------------- ---------------------- ----------------------
2He 41 (44 %) 115 (40 %)
He + 2H 42 (45 %) 157 (54 %)
Li + H 9 (10 %) 14 (5 %)
4H 2 (2 %) 3 (1 %)
: \[Table:1\] Distribution of white stars produced by $^{7}$Be nuclei of energy 1.2 GeV per nucleon among charge channels of nuclear dissociation.
EXPOSURES OF TRACK EMULSIONS
============================
In December 2013, a set of test samples of nuclear-track emulsions produced at the MICRON workshop of the OJSC Slavich Company was exposed to a secondary beam of relativistic $^{11}$C nuclei at the JINR nuclotron \[15\]. The samples were prepared by pouring emulsion layers about 200 $\mu$m in thickness onto glass substrates 9$\times$12 cm in size. This nuclear-track emulsion is close in properties to the BR-2 nuclear-track emulsion,which provided sensitivity up to relativistic particles.
Nuclei of $^{11}$C were produced in the fragmentation of $^{12}$C nuclei at an energy of 1.2 GeV per nucleon on a polyethylene target 1.5 g cm$^{-2}$ thick. A secondary beam of $^{11}$C nuclei is formed via separation in the magneto-optical channel for beam transportation, the momentum acceptance being about 2%. At an intensity of the $^{12}$C primary beam on the order of $10^{7}$ nuclei per cycle, the intensity of the $^{11}$C beam is $10^{4}$; this is an optimum value for a controlled exposure of a track-emulsion stack. The beam profile was formed in such a way as to ensure the highest possible degree of uniformity of irradiation along the narrower side.
The flux of nuclei that was directed to the irradiated track-emulsion stack was controlled by means of a scintillation monitor. The presence of accompanying nuclei in the composition of the main beam makes it possible to assess the potential of the magneto-optical channel used for the separation of $^{11}$C nuclei \[16\]. The momentum acceptance of the channel was about 2%. Figure 1 shows the spectrum of the charge-to-digital converter in transmitting $^{12}$C nuclei. One can observe a contribution of lighter nuclei produced on a generating target at the beginning of the separation channel, the charge-to-mass number ratio being identical to that for the $^{12}$C nucleus. The shoulder on the left of the main peak corresponds to $^{10}$B nuclei, Be nuclei do not manifest themselves ($^{8}$Be is an unbound nucleus), the contribution of lithium nuclei is distinguishable, and helium nuclei manifest themselves quite distinctly.
Figure 2 presents the analogous spectrum for the case of tuning the channel to the separation of $^{11}$C nuclei that have the same energy per nucleon as $^{12}$C nuclei. Signals from boron, beryllium ($^{7}$Be), lithium, and helium are only slightly seen and are associated with the fragmentation of $^{11}$C nuclei. The disappearance of helium nuclei is worthy of special note. Upon a decrease in the magnetic rigidity of the channel, $^{4}$He nuclei have already disappeared, while $^{3}$He nuclei have not yet appeared. All these facts indicate that the degree of the separation of $^{11}$C nuclei is quite high, which gives sufficient grounds to disregard the contribution of other isotopes. An irradiation of 40 track-emulsion layers was performed in a beam of this composition. In order to vary the irradiation density, the layers in question were combined into seven stacks irradiated successively.
A reduced thickness and glass substrates of the test track-emulsion series turned out to be factors that prevented an analysis that would involve tracing beam and secondary tracks without sampling. Therefore, we scanned track-emulsion layers along transverse bands with the aim of finding tracks of relativistic fragments whose total charge is not less than three and subsequently tracing them up to interaction vertices. Tracks corresponding to singly and doubly charged relativistic fragments were determined visually. The fact that carbon nuclei were dominant in the beam made it possible to establish the charges of heavier fragments in white stars as the values that were needed to reach six charge units.
STATUS OF INVESTIGATIONS OF THE $^{11}$C NUCLEUS
================================================
In six scanned track-emulsion layers, 144 white stars where the total charge of relativistic fragments is six charge units have been found to date. Their distribution in charge states is given in Table 3. This table also present data on the isotopes $^{10}$C \[9\] and $^{9}$C \[5\]. These data indicate that white stars have an individual character for each isotope and that the exposures in our present experiments correspond to the mass numbers of the aforementioned carbon isotopes. In the investigation of the coherent dissociation of relativistic $^{12}$C nuclei in \[17\], all of the 100 white stars found there arose in the single channel $^{12}$C $\rightarrow ^{3}$He, clearly reflecting a virtual alpha-particle clustering of this nucleus. The discovery of the decay of relativistic $^{8}$Âå nuclei, which made a contribution of magnitude not less than 20%, was a key observation.
Events featuring only relativistic helium and hydrogen isotopes, especially 2He + 2H, were dominant among $^{11}$C white stars, their weight being as high as 77%. The branching ratio for the 2He + 2H and He + 4H channels is 6 ± 3, which is at odds with the above idea that only the $^{7}$Be core nucleus undergoes dissociation.
![Photograph of the spectrum of the charge-to-digital converter of the scintillation monitor for beam nuclei in transmitting $^{12}$C nuclei (arbitrary units). The inset shows the lower left section of the spectrum on an enlarged logarithmic scale.[]{data-label="fig:1"}](image1){width="3.2in"}
In contrast to what was found for the neutron-deficient nuclei studied earlier, we observed here a significant fraction of Li + He + H events, which could correspond to the $^{6}$Li + $^{4}$He + *p* structure. There were no Be + 2H events ( $^{9}$Be + 2*p* cluster structure). At the same time, we observed a significant fraction of Be + He events. In the case of identifying the isotope $^{4}$He in them, one determines unambiguously the isotope $^{7}$Be. Most probably, the 3He channel corresponds to the 2$^{4}$He + $^{3}$He configuration, which may arise both from the decay of the $^{8}$Be and $^{7}$Be core nuclei and from the decay of three-body states. An additional contribution to multi-particle channels may come from the dissociation of the $^{6}$Li cluster as an individual element of the $^{11}$C nucleus in accordance with its virtual $\alpha + d$ structure \[18\]. Figuratively speaking, we can state that the charge-topology distributions presented above have an individual character for $^{11}$C, which distinguishes it among other isotopes, appearing to be some kind of a signature of this nucleus.
Channel $^{10}$B \[14\] $^{8}$B \[4\]
------------- --------------------- -------------------
Be + H 1 (2 %) 25 (48 %)
2He + H 30 (73 %) 14 (27 %)
He + 3H 5 (12 %) 12 (23 %)
Li + He 5 (13 %)
: \[Table:2\] Distribution of white stars produced by ${}^{10}$B (of energy 1.0 GeV per nucleon) and ${}^{8}$B (of energy 1.2 GeV per nucleon) nuclei among charge channels of the dissociation of nuclei
![As in Fig. 1, but in transmitting $^{11}$C nuclei.[]{data-label="fig:2"}](image2){width="3.2in"}
By and large, the structure of $^{11}$C can be thought to be a superposition of cluster states featuring $^{10}$B and $^{7}$Be core nuclei. This prescribes further lines of our investigations. The above aspects furnish a motivation for a new series of studies on the basis of irradiations already performed. They will be aimed at enlarging the statistics of $^{11}$C white stars, as well as performing measurements of multiple scattering in order to identify helium and hydrogen isotopes and angular measurements in order to determine the fraction of 8 Be decays and to explore dissociation dynamics. A selection of values of the total transverse momentum of relativistic fragments within the range characteristic of diffractive dissociation would make it possible to compensate indirectly for the impossibility of directly identifying isotopes heavier than helium.
INTERPRETATION OF THE CLUSTER STRUCTURE OF THE $^{12}$N NUCLEUS
===============================================================
The analysis of the experimental data for $^{11}$C nuclei in the preceding section creates preconditions for giving a more justified interpretation of the charge topology of white stars generated by relativistic $^{12}$N nuclei \[10\]. In Table 4, their statistics are presented along with comparable data on $^{14}$N white stars \[19\]. The $^{11}$C + *p* (0.6 MeV), $^{8}$B + $^{4}$He (8.0 MeV), and *p* + $^{7}$Be + $^{4}$He channels can make substantial contributions to $^{12}$N white stars. Also, a multi-particle dissociation through the $^{3}$He + $^{9}$B (10 MeV) channel involving an unbound nucleus is possible. An interpretation of the B + 2H channel is complicated by the presence of the $^{10}$B + 2*p* (9.2 MeV) channel. In just the same way as in the case of $^{11}$C, multi-particle channels may emerge owing both to the dissociation of the $^{10}$B core nucleus and to the dissociation of $^{7}$Be. Possibly, a leading character of the 2He + 3H channel (see Table 3) reflects the dissociation of $^{11}$C via a process involving $^{10}$B. A small probability for the dissociation of $^{10}$B to a $^{9}$Be + *p* pair makes it possible to specify Be in Table 4 as $^{7}$Be. Owing to the limit on the mass number, we can specify B from the Be + He channel in Table 4 as $^{8}$B.
Channel ${}^{11}$C ${}^{10}$C \[9\] ${}^{9}$C \[5\]
----------------- ---------------- ---------------------- -------------------
B + H 6 (5 %) 1 (0.4 %) 15 (14 %)
Be + He 17 (14 %) 6 (2.6 %)
Be + 2H 16 (15 %)
3He 22 (18 %) 12 (5.3 %) 16 (15 %)
2He + 2H 60 (48 %) 186 (82 %) 24 (23 %)
He + 4H 14 (11 %) 12 (5.3 %) 28 (27 %)
Li + He + H 4 (3 %)
Li + 3H 1 (0.4 %) 2 (2 %)
6H 3 (2 %) 9 (4 %) 6 (6 %)
: \[Table:3\] Distribution of white stars produced by carbon isotopes (of energy 1.2 GeV per nucleon) among charge channels of the dissociation of nuclei.
Channel ${}^{12}$N \[10\] ${}^{14}$N \[19\]
------------------ ----------------------- ---------------------
C + H 4 (6 %) 13 (28 %)
B + He 3 (4 %) 4 (9 %)
B + 2H 11 (15 %) 3 (7 %)
Be + He + H 9 (13 %) 1 (2 %)
Be + 3H 10 (14 %)
Li + He + 2H 1 (2 %)
Li + 4H 1 (2 %)
3He + H 2 (3 %) 17 (37 %)
2He + 3H 24 (33 %) 6 (13 %)
He + 5H 9 (13 %)
: \[Table:4\] Distribution of white stars produced by $^{14}$N (of energy 1.2 GeV per nucleon) and $^{14}$N (of energy 2 GeV per nucleon) nuclei among charge channels of the dissociation of nuclei
A limited volume of the statistical sample of $^{12}$N white stars \[10\] (see Table 4) is due to forming a $^{12}$N beam with the aid of the charge-exchange reaction involving relativistic $^{12}$C nuclei. This way was aimed primarily at simplifying the identification of $^{12}$N white stars on the basis of the total relativistic-fragment charge equal to seven units against a more intense background of events associated with accompanying carbon isotopes. However, a cumbersome determination of charges of beam tracks as those that correspond to seven charge units becomes necessary because of a sizable contribution of events of coherent carbon-isotope dissociation involving meson production within a narrow cone of relativistic fragmentation. This fact drastically reduces the efficiency of the charge-exchange reaction. At the same time, a good separation of carbon isotopes at the JINR nuclotron was confirmed by data on their coherent dissociation. This indicates that, in order to increase sharply statistics of white stars, one can use irradiations of track emulsions with nuclei of the isotopes $^{12,13}$N produced in the fragmentation of relativistic $^{14}$N nuclei.
CONCLUDING REMARKS
==================
It has been shown that, in relation to what we have for $^{7}$Be, $^{8,10}$B, $^{9,10}$C, and $^{14}$N nuclei, the track-emulsion method is more efficient in identifying events of the coherent dissociation of relativistic $^{11}$C and $^{12}$N neutron-deficient nuclei. A complete determination of the branching ratios for channels of the breakup of these nuclei makes it possible to reconstruct their virtual cluster structure. The data that we obtained on charge topology is an important first step in this field.
New possibilities of the track-emulsion method may open upon accelerating $^{16}$O nuclei in order to perform similar investigations in beams of the neutron-deficient isotopes $^{13,14,15}$O. A further advancement toward heavier neutron-deficient isotopes by means of the track-emulsion method remains promising but runs into more serious problems. On this way, the variety of the *p*–-$^{3}$He–-$\alpha$ ensembles under study may become ever wider.
Starting from the $^{11}$C and $^{12}$N nuclei, there arise problems associated with the limitations of the approach based on the coherent dissociation of relativistic nuclei in nuclear-track emulsions—specifically, we mean here the impossibility of a direct identification in mass number for relativistic fragments heavier than helium. The fraction of events involving such fragments increases sharply as the mass number of the nuclei under study grows. At energies of incident nuclei (Å) around a few GeV units per nucleon, one can identify them in electronic experiments involving a magnetic analysis. In the future, such an identification may become possible for *Å* in the region of several tens of GeV units per nucleon in experiments with hadron calorimeters. Investigations on the basis of the track-emulsion method remain valuable since they may provide guidelines for electronic experiments aimed at studying the coherent dissociation of relativistic neutron-deficient nuclei.
ACKNOWLEDGMENTS
We are grateful to A.I. Malakhov (JINR), N.G. Polukhina, and S.P. Kharlamov (Lebedev Physical Institute), and N.S. Zelenskaya (Skobeltsyn Institute of Nuclear Physics, Moscow State University) for their support of this study and critical comments on its results.
This work was supported by the Russian Foundation for Basic Research (project nos. 12-02-00067 and 15-02-01073) and by grants from the plenipotentiaries of Bulgaria, Egypt, Romania, and the Czech Republic at JINR.
**REFERENCES**
1. The BECQUEREL Project, http://becquerel.jinr.ru/.
2. D. A. Artemenkov *et al.*, Phys. Atom. Nucl. 70, 1222 (2007); nucl-ex/ 0605018.
3. N. G. Peresadko *et al.*, Phys. Atom. Nucl. 70, 1266 (2007); nucl-ex/0605014.
4. R. Stanoeva *et al.*, Phys. Atom. Nucl. 72, 690 (2009); arXiv: 0906.4220.
5. D. O. Krivenkov *et al.*, Phys. Atom. Nucl. 73, 2103 (2010); arXiv: 1104.2439.
6. R. R. Kattabekov *et al.*, Phys. Atom. Nucl. ** 73, 2110 (2010); arXiv: 1104.5320.
7. D. A. Artemenkov *et al.*, Few Body Syst. 50, 259 (2011); arXiv: 1105.2374.
8. D. A. Artemenkov *et al*., ** Int. J. Mod. Phys. E 20, 993 (2011); arXiv: 1106.1749.
9. K. Z. Mamatkulov *et al.*, Phys. At. Nucl. 76, 1224(2013); arXiv: 1309.4241.
10. R. R. Kattabekov *et al.*, Phys. At. Nucl. 76, 1219(2013); arXiv: 1310.2080.
11. N. K. Kornegrutsa *et al.* Few Body Syst. 55, 1021 (2014); arXiv: 1410.5162.
12. P. I. Zarubin, Lect. Notes in Phys., 875, 51(2013) Springer Int. Publ.; arXiv: 1309.4881.
13. N. G. Peresadko *et al.*, JETP Lett. 88, 75 (2008); arXiv: 1110.2881.
14. M. I. Adamovich, Phys. At. Nucl. 67, 514 (2004); nucl-ex/0301003.
15. The Slavich Company, www.slavich.ru, www.newslavich.com.
16. P. A. Rukoyatkin *et al.*, EPJ ST 162, 267(2008); nucl-ex/1210.1540.
17. V. V. Belaga *et al.*, Phys. At. Nucl. **58**, 1905 (1995)\]; nucl-ex/1109.0817.
18. M. I. Adamovich *et al.*, Phys. At. Nucl. **62**, 1378 (1999); nucl-ex/1109.6422.
19. T. V. Shchedrina *et al.*, Phys. Atom. Nucl. **70**, 1230 (2007); nucl-ex/0605022.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'A. Khlystova'
bibliography:
- 'ref.bib'
date: 'Received ; accepted '
title: 'Center-limb dependence of photospheric velocities in regions of emerging magnetic fields on the Sun'
---
[We investigate the ratio of the vertical and horizontal velocities of the photospheric plasma flows in the region of emerging magnetic fields on the Sun.]{} [We carried out a study of photospheric velocities during the first hours of the appearance of 83 active regions with magnetic flux more than 10$^{21}$ Mx with data acquired by the Michelson Doppler Imager (MDI) on board the Solar and Heliospheric Observatory (SOHO). The emerging magnetic fluxes under investigation were isolated from extended concentrations of existing magnetic fields; they have different spatial scales and are located at different distances from the solar disk center.]{} [We found that the values of maximum negative Doppler velocities that accompany the emergence of active region magnetic fields during the first 12 hours increase nonlinearly with the heliocentric angle. This result shows that the horizontal photospheric velocities of plasma outflows are higher than the vertical ones of the plasma upflows during the first hours of the emergence of active regions. The horizontal velocity component at the site of emerging active regions exceeds that of convective flows in the quiet Sun. A comparison between the velocities and the maximum value of the total magnetic flux has not revealed any relation.]{}
Introduction
============
The emerging magnetic flux of active regions on the solar surface appears in the form of separate loops; they then coalesce to form opposite polarity poles that separate and increase in size [@str99]. Earlier studies of velocities of plasma motions were carried out in the region of emerging magnetic fields at different height levels of the Sun. Investigations into the chromospheric level have been made by @bru67 [@bru69] and others [see review by @cho93]. Velocities at the photospheric level were measured with different methods (discussed in more detail below). In the last decade, helioseismological methods have been applied to study the subphotospheric level [@cha99; @kos00; @zha08; @kos09; @kom08; @kom09].
Direct measurements of photospheric velocities revealed negative values (blue Doppler shift) on the polarity inversion line of emerging magnetic fields. The events under consideration were located in the central part of the solar disk, and, therefore, vertical motions were measured. Negative velocities or plasma upflows up to 1 km s$^{-1}$ were observed in separate magnetic loops inside emerging active regions [@tar90; @lit98; @str99; @kub03; @gri09]. @gri07 revealed high velocities of about 1.7 km s$^{-1}$ at the beginning of the powerful active region emergence NOAA 10488 at heliographic coordinates N08 E31 ($B_{0}$ = + 4.9).
Indirect measurements of the horizontal velocities of the photospheric flow that accompanies the emergence of active regions were made by tracking the displacement of individual magnetic elements. The velocities obtained ranged from 0.1 to 1.4 km s$^{-1}$ [@fra72; @sch73; @str99]. @bar90 studied 45 bipolar pairs in the emerging active region. The opposite polarity poles were separating from each other with velocities of 0.5–3.5 km s$^{-1}$, decreasing with time (consequently, the drift with regard to the polarity inversion line is only half that value). @gri09 calculated the separation velocities of the external boundaries of the photospheric magnetic flux in the active region NOAA 10488. The velocities decreased as the magnetic fields emerged: they were 2–2.5 kms$^{-1}$ by the end of the first hour and 0.3 km s$^{-1}$ in two hours and a half.
It should be noted that there are several papers dealing with the study of vertical [@gug06] and horizontal [@har73; @cho87; @hag01] velocities at photospheric level during the appearance of ephemeral active regions. We do not discuss these works in detail, since they concern another spatial scale of emerging magnetic fields.
The papers above are the researches into photospheric velocities that accompany the emergence of magnetic fields in active regions located only in the central zone of the solar disk. Plasma motions in the emerging magnetic fields distant from the disk center have never been studied before.
Data processing and investigated objects
========================================
{width="\textwidth"}
We used full solar disk magnetograms and Dopplergrams in the photospheric line Ni I 6768 $\mbox{\AA}$ and continuum images obtained on board the space observatory SOHO/MDI [@sch95]. The temporal resolution of the magnetograms and Dopplergrams is 1 minute, that of the continuum, 96 minutes. The spatial resolution of the data is 4$\arcsec$, the pixel size is approximately 2$\arcsec$. Magnetograms with a 1.8.2 calibration level were used [@ulr09]. Besides photospheric velocities of solar plasma motions, the Dopplergrams include the contribution of 1) the velocity of the differential solar rotation; 2) the velocity of the SOHO satellite with regard to the Sun; 3) the instrumental distortions caused by nonuniform transmission of MDI’s filter systems on the field of view. The technique described in @gri07 was used to separate photospheric velocities. The negative velocity on SOHO/MDI Dopplergrams corresponds to the blue Doppler shift (motion of matter toward the observer); the positive one, to the red Doppler shift (motion of matter away from the observer). There are regular errors in velocity measurements at different distances from the disk center. For example, Doppler velocities near the limb are underestimated by up to 40 m s$^{-1}$ because of limb darkening [@alb85]. Additionally, we deal with different height levels above the solar surface in measurements of velocities at the disk center and near the limb because of the different optical thickness of the atmosphere.
We study the velocities that accompany the emerging magnetic fields. Therefore, the precise spatial superimposition of the used data is vary important. For this purpose, the region of the emerging magnetic fields was taken from the time sequences of magnetograms and Dopplergrams, taking into account solar rotation. An approximate value of the displacement of the region was calculated with the differential rotation law for photospheric magnetic fields [@sno83]. The exact tracking of the region under investigation was performed by applying two magnetograms adjacent in time, with the use of cross–correlation analysis. This procedure requires the existence of the magnetic poles which slightly vary in time on the magnetograms. The heliographic coordinates obtained were used to crop identical fragments with a size of 160$\arcsec\times$160$\arcsec$ (or 80$\times$80 pixels) from the magnetogram and Dopplergram obtained at the same time. Thus, we achieved the precise spatial superimposition of the data. It has allowed us to draw reliable conclusions about the processes that take place. For the active region emerging near the limb, we selected the cropped region in a way that it excludes the area outside the limb.
Time variations of the total magnetic flux of the active region and the maximum negative velocities were calculated with the use of a sequence of the magnetogram and Dopplergram fragments corresponding to the emerging active region. The calculation area was limited to the region of the emerging magnetic fields (Fig. \[fig1\]). The position of the limits was visually controlled.
The plasma motions connected with the emerging magnetic fields are observed on the background of the convective flows. If the average velocity or the velocity flux is considered as a characteristic of motions of matter, the center–limb dependence of convective velocities will make a substantial contribution. Besides, the active regions under consideration have different spatial scales and are located at different distances from the solar disk center; therefore the solar surface regions where the calculation is performed have different sizes. This complicates a comparison of the emerging active regions that are located at different distances from the disk center. The maximum negative velocity (motion of matter toward the observer) is taken as a characteristic of plasma motion in the regions of the emerging magnetic fields during the first 12 hours (Fig. \[fig1\]b). The magnetic field emergence begins with the appearance of the loop apex where the magnetic field is horizontal; therefore, we began the calculation of the maximum velocity values 30 minutes before the beginning of the appearance of active region. The magnetic flux of the active region emerges as separate loops, and the maximum negative velocity corresponds to the appearance of a single magnetic loop. Therefore, the chosen parameter does not characterize the emerging active region as a whole, it only shows the highest velocity. For example, two peaks corresponding to the emergence of two different magnetic loops can be distinguished in the plot of variations in the maximum negative velocity during the first hours of the NOAA9456 appearance (Fig. \[fig1\]e). The residual magnetic flux emerges at lower velocities. To exclude Evershed flows we also took care of the time of the penumbra formation in sunspots, using continuum images. The Evershed flows are horizontal photospheric outflows in a sunspot penumbra. Their maximum velocity values can reach 2 km s$^{-1}$ in SOHO/MDI low spatial resolution data [@bai98].
The total magnetic flux is calculated inside isolines $\pm$60 G taking into account the projective effect (Fig. \[fig1\]c)
$$\begin{aligned}
\Phi & = & |\sum_{i=1}^{n_{+}}(B_{i} \cdot S_{i})| + |\sum_{i=1}^{n_{-}}(B_{i}
\cdot S_{i})|,\end{aligned}$$
where $\Phi$ is the total magnetic flux in Mx, $B_{i}$ is the magnetic field induction in G, $S_{i}$ is the area on the solar surface corresponding to $i-th$ pixel, $n_{+}$ is the number of pixels inside isoline +60G (positive polarity) and $n_{-}$ is the number of pixels inside isoline -60G (negative polarity). The isoline level of 60G was chosen to exclude both the magnetogram noise (about 30G) and the contribution of the magnetic field with a short lifetime. Sometimes small magnetic loops appear in the place of the future active region several hours prior to the main emergence of the magnetic fields. The beginning of the active region emergence was taken as a time moment corresponding to the beginning of the continuous growth of magnetic flux. We determined the maximum value of the total magnetic flux by the first inflection point of the curve where the flux increase was followed by its decrease (Fig. \[fig1\]f) or by the last value of the plot (for active regions passing beyond the west–limb or with insufficiently downloaded data sequences). Note that powerful active regions significantly increase in size as they develop, and the extended concentrations of the magnetic fields existing on the surface enter into the region of new flux emergence (Fig. \[fig1\]a, \[fig1\]b, and \[fig1\]c). Interaction between existing and emerging magnetic fields is accompanied by integration or cancellation processes and, hence, by increase or decrease in the magnetic flux. In order to take this into account to some extent, the signal background that existed before the emergence of magnetic fields was subtracted from the total magnetic flux maximum in all active regions.
![Total magnetic flux of the active regions under investigation versus the heliocentric angle that corresponds to the beginning of the magnetic field emergence. The total magnetic flux maximum was determined by the inflection point in the increase of the flux curve (marked by diamonds) or by the last measurement (marked by asterisks). Active regions with magnetic saturation are also marked by asterisks. In active regions marked by asterisks, the maximum of total magnetic flux may be higher.[]{data-label="fig2"}](f2){width="8.86cm"}
{width="\textwidth"}
The selection of active regions for this investigation was based on the following criteria. Active regions have to emerge on the visible side of the solar disk. Emerging magnetic fields have to be isolated from the extended concentrations of existing magnetic fields. The presence of single poles with a total magnetic flux of less than 0.5$\times$10$^{21}$ Mx was allowed in the area of the direct emergence of an active region during the first 12 hours. The objects under investigation have to have complete data series with a one–minute time resolution during the first 12 hours of appearance. With this in mind, the 83 active regions were selected for the period of 1999–2008. Among them, there were 80 active regions with a normal magnetic field configuration and three with an inverse magnetic field configuration according to Hale’s polarity law. The events had different spatial scales and were located at different distances from the solar disk center (Fig. \[fig2\]). The total magnetic flux of each event exceeded 10$^{21}$ Mx. Magnetic saturation is possible in SOHO/MDI measurements of the magnetic field strength. This occurs when the spectral line is shifted out of the filter system passband. @liu07 modeled MDI measurements and demonstrated that magnetic saturation takes place only in strong magnetic fields with high velocities. According to their calculations, magnetic saturation is reached at the magnetic field strength of 2800 G, when the velocity is $\pm$2000 m s$^{-1}$. In our study, the maximum magnetic field strength exceeds 2800 G only in two events of the active regions under consideration. These events are also marked by asterisks in Fig. \[fig2\]. The morphology of the active regions (i.e., the degree of fragmentation of magnetic fields, the rate and monotony of magnetic flux increase, and the orientation of the axis connecting opposite polarity poles with regard to the line of sight) was not taken into account.
Results
=======
Figure \[fig3\]a presents the maximum negative velocities observed during the first 12 hours emergence of active regions according to their position with regard to the disk center. The distance from the disk center is expressed in the heliocentric angle $\theta$ – it is the angle between the normal to the surface and the line–of–sight to the emerging magnetic flux. The maximum negative velocities obtained are observed, as a rule, on the polarity inversion line of the emerging magnetic fields; however, they are related to the convection in adjacent regions for some events in the central part of the disk ($\theta<25^\circ$). Figure \[fig3\]a shows that values of maximum negative velocities increase nonlinearly with the heliocentric angle. The mean velocity for $\theta=20^\circ$ is -673 m s$^{-1}$; the deviation from the mean does not exceed 200 m s$^{-1}$. The mean velocity for $\theta=60^\circ$ is -1277 m s$^{-1}$; the deviation from the mean reaches up to 550 m s$^{-1}$. One can see a growth tendency of deviation from mean with increasing $\theta$ (Fig. \[fig3\]a). The high deviation for large $\theta$ indicates the existence of other relations that determe the value of the horizontal velocity of plasma flow at the site of emerging magnetic fields at the photospheric level. Our statistics include three inverse active regions whose parameters do not deviate from the dependence formed by normal active regions. Thus we see that the horizontal velocities of plasma outflows exceed the vertical ones of plasma upflows at the beginning of the emergence of active regions.
Figure \[fig3\]b shows the maximum negative velocities for the quiet Sun connected with convective flows. To calculate this dependence, the region on the surface was traced taking into account solar rotation. The region with a size of 40$\arcsec\times40\arcsec$ under study was going through the longitude range of W00–W67 at the latitude of S04 from 3 to 8 May 1999. The tilt of the solar North rotational axis toward the observer $B_{0}$ was -3.5$^\circ$. Data with a one-minute resolution were used; they were processed with the technique described in section 2 of this paper. In Fig. \[fig3\]b one can see that the velocities increase with the heliocentric angle; these values do not exceed 1200 m s$^{-1}$. The comparison between the plots in Fig. \[fig3\]a and Fig. \[fig3\]b has revealed that the horizontal velocity component at the sites of the emergence of active regions is higher than that of the convective flow in the quiet Sun.
We performed the regression analysis with the second degree polynom. As a result, we obtained the equation
$$\begin{aligned}
\upsilon & = & 6.01 - 40.18 \theta + 0.31 \theta^2,\end{aligned}$$
where $\upsilon$ is the maximum negative Doppler velocity during the first 12 hours of the emergence of active regions in ms$^{-1}$ and $\theta$ is the heliocentric angle in degrees. According to the F–statistic this regression model is significant. The correlation ratio (estimate of the closeness of the nonlinear relation) between the value of the maximum negative velocity and the heliocentric angle $\theta$ (or the angle of view to the emerging magnetic flux) is equal to -0.74; this implies a high relation between the parameters under consideration. The confidence intervals for the means were calculated with a confidence probability of 99$\%$. They are marked by dotted lines in Fig. \[fig3\]a. The confidence intervals have close limits for $10^\circ<\theta<70^\circ$ with high statistics and show that equation (2) approximates the data well. The limits of the confidence intervals essentially increase for $\theta<10^\circ$ and $70^\circ<\theta$ with low statistics. It follows from equation (2) that the mean vertical velocity for $\theta=0^\circ$ is 6 m s$^{-1}$ with a confidence interval of $\pm$399 ms$^{-1}$. Positive velocity values are false in this confidence range. Obviously, velocities corresponding to $\theta=0^\circ$ at the sites of emerging active regions will be not lower than the velocities of convective flows of the quiet Sun ($\sim$-150 m s$^{-1}$). The absence of statistics for $\theta<10^\circ$ at the emerging active regions does not allow one to take it into account in regression equation (2). The mean horizontal velocity for $\theta=90^\circ$ from equation (2) is -1081.44 m s$^{-1}$ with a confidence interval $\pm$601 m s$^{-1}$. The wide confidential interval shows that this value of the mean horizontal velocity is ambiguous.
A comparison of the maximum negative velocities with the total magnetic flux did not reveal any relation (Fig. \[fig4\]). The comparison between the plots in Fig. \[fig2\] and Fig. \[fig3\]a also showed no connection, because high velocities were not observed in emerging active regions with high magnetic flux in the central part of the solar disk, whereas near the limb there may be high velocity values even in active regions with low magnetic flux. From the author’s point of view the absence of a dependence between velocity and total magnetic flux can be explained by the fact that active region magnetic fields emerge by separate fragments on time scales from several hours to 5-7 days and therefore the maximum negative velocity during the first 12 hours does not characterize an active region emergence as a whole.
![Maximum negative velocities during the first 12 hours of the magnetic field emergence versus the maximum value of the total magnetic flux in the active regions. The designations are the same as for the Fig. \[fig2\].[]{data-label="fig4"}](f4){width="8.86cm"}
Conclusions
===========
We presented a statistical study of the velocities of plasma flows that accompany the emergence of active regions during the first 12 hours at the photospheric level with data with high temporal resolution. We found that the values of maximum negative Doppler velocities increase nonlinearly with the heliocentric angle. This shows that the horizontal velocities of plasma outflows exceed the vertical ones of plasma upflows during the first hours of active region emergence. Horizontal velocities at the sites of emerging active regions are higher than those of convective flows in the quiet Sun. This result is a direct confirmation of theoretical models that showed that magnetic fields emerging in the solar atmosphere expand faster in the horizontal direction than in the vertical direction [e.g. @shi89; @fan01; @mag03; @arc04].
The author is grateful to the referee for helpful comments that improved the manuscript. This work used data obtained by the SOHO/MDI instrument. SOHO is a mission of international cooperation between ESA and NASA. The Michelson Doppler Imager is a project of the Stanford–Lockheed Institute for Space Research. I am grateful to my supervisors V.M. Grigor’ev and L.V. Ermakova for their helpful suggestions on this scientific research, and to V.G. Fainshtein for very useful discussions. I would like to thank Tom Duvall and Alexander Kosovichev for discussions that helped me to understand peculiarities of SOHO/MDI Dopplergrams better. This study was supported by RFBR grants 08-02-00027-a, 09-02-00165-a, 10-02-00607-a, 10-02-00960-a, the NNSFC-RFBR grant 08-02-92211, state contract 02.740.11.0576 and by the Russian Academy of Sciences (the program 4 of the RAS Presidium and the joint project 4 of its Siberian Branch).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Many of the current bio-inspired delivery networks set their focus on search, e.g., by using artificial ants. If the network size and, therefore, the search space gets too large, the users experience high delays until the requested content can be consumed. In previous work, we proposed different replication strategies to reduce the search space. In this report we further evaluate measures for storage load balancing, because peers are most likely limited in space. We periodically apply clean-ups if a certain storage level is reached. For our evaluations we combine the already introduced replication measures with least recently used (LRU), least frequently used (LFU) and a hormone-based clean-up. The goal is to elaborate a combination that leads to low delays while the replica utilization is high.'
author:
- Anita Sobe
- Wilfried Elmenreich
- Laszlo Böszörmenyi
bibliography:
- 'TRSaso.bib'
title: 'Storage Balancing in Self-organizing Multimedia Delivery Systems'
---
=1
Introduction
============
The increased complexity of modern networks and the increasingly dynamic access patterns in multimedia consumption have led to new challenges for content delivery. Bio-inspired networks promise a robust and adaptive behavior to cope with this task [@harsh:10]. We rely on a bio-inspired delivery algorithm introduced in [@Sobe2010a] and its extension in [@Sobe2011], providing the possibility of sharing small multimedia units during and after a social event. A multimedia unit can be either a photo or a short video sequence, e.g., generated at a live sports event by a visitor. With the model of small units two typical problems of delivery do not exist anymore: (1) The transport of large files, as typical for multimedia content; (2) Content consistency, since units do not change after creation. However, we assume dynamic access patterns. Users can “compose” their presentations consisting of a number of different units. E.g., a visitor of a sports event wants to see an overview of the highlights of the last 20 minutes presented as 4 units in parallel in a split-screen. Such access patterns allow flexibility, but also introduce complexity, because the typical video sequence as known from movies is not existing anymore.\
Our delivery algorithm is capable of handling this complexity relying on simple local information. The algorithm is inspired by two existing bio-inspired approaches. The first approach is Ant Colony Optimization as discussed in [@Dressler2010]; a specific application for search in P2P networks was introduced as SemAnt [@Michlmayr2007a]. The second approach is an artificial hormone-based agreement for task allocation introduced by Brinkschulte et al. in [@Brinkschulte2007].\
We adopted the keyword search from SemAnt and introduce one type of hormone by keyword. The hormone value expresses the current demand (the *goodness*, as adopted from Brinkschulte et al.) for the corresponding keyword. SemAnt assumes the content distribution to be fixed, which keeps the search space very large. To reduce the search space we apply replication and exploit the unstructured overlay by letting the content travel towards higher levels of its corresponding hormone. This allows intermediate nodes to decide if the traveling unit is needed for future requests, thus predicting places to reduce search space and delivery costs.\
Since peers are not likely to provide unlimited storage for other peers and in a dynamic system the popularity of units changes, we investigate different measures to efficiently balance the storage of the peers. We periodically apply different strategies if a certain storage level is reached. We compare LRU (Least Recently Used), LFU (Least Frequently Used) and a hormone-based clean-up. We show that the chosen clean-up mechanism has an impact on the delivery performance and that the system is still robust against peer churn. We performed simulations for evaluating the replication and clean-up mechanisms in scale-free and random overlay networks.\
Hormone Algorithm Description
=============================
![Interplay of diffusion and unit movement in hormone-based delivery[]{data-label="fig:hormone"}](fig/hormonetrail2 "fig:")\
The hormone-based delivery approach introduced in [@Sobe2010a] involves three components: hormone creation, hormone diffusion, and the behavior of units in presence of a corresponding hormone. The distributed, self-organizing nature of the approach allows to handle the complexity of requests and the search for units in the network with comparably simple decision algorithms based on local knowledge.\
Upon request of a particular unit, a corresponding hormone is issued at the requesting node. The hormone is represented as a real value; the current value of a hormone is called *hormone level*. The hormone is diffused via the neighboring nodes to the network, creating a hormone gradient towards the requesting node.\
The diffusion of hormones depends on the network structure, e.g., a node that is only a few hops away from the requesting node will be flooded earlier with hormones and will get more hormones than a more distant node.\
Figure \[fig:hormone\] depicts the diffusion of hormones from a requesting node $P_1$. A part from the total amount of deployed hormones is diffused to the neighboring nodes $P_2$, $P_3$, and $P_4$. Note that the amount of diffused hormone is spread unevenly, based on the condition of the network links to the respective nodes (a thicker line indicates a better/faster connection).\
As long as a request is not fulfilled, the hormone level is raised periodically, thus increasing the *pull* for a unit fulfilling the particular request.\
A unit that corresponds to the given hormone will react by moving towards the network path with the steepest gradient, eventually arriving at the requesting node. In order to reduce attracting multiple copies of a unit, the diffusion of a hormone is stopped, if the respective unit is residing already on the same node.\
The hormone-based delivery creates a feedback loop between network conditions. The network conditions influence hormone diffusion, the hormone diffusion influences unit movement, which in turn creates network traffic and changes the network conditions.\
Multiple requests for different units lead to a set of different hormones being handled in parallel by the network. Requests for the same unit result in a superimposed hormone landscape for that unit. In this case, a unit might be attracted by two hormone trails. Without replication the unit has to move first to the first requester and afterwards to the second requester. Which requester gets the unit first depends on the strength of hormone reaching the unit. In order to avoid such detours, an intelligent replication mechanism has to take care of this issue.
Parameter Settings {#sec:params}
==================
ID Explanation
------------ -----------------------------------------------------------
$\eta_0$ Hormone strength of a unit at new request
$\eta$ Increase of hormone after each time step by the requester
$\alpha$ Percentage of hormones to be forwarded to the neighbors
$\epsilon$ Hormone evaporation value
$t$ Minimum hormone strength
$m$ Minimum hormone difference to move unit
: Parameters to configure at system startup[]{data-label="tab:parameters"}
The proposed algorithm is self-organizing and has only a few configuration parameters. In Table \[tab:parameters\] the necessary parameters are shown. These parameters are dependent on each other. E.g., if $\eta_0$ and $\eta$ are low and the evaporation value $\epsilon$ is high, the movement of units can be limited. The more hormones are created and forwarded, the more hops the hormones can travel and therefore increase the search space. The difference to move a unit $m$ controls the mobility of units. If $m$ is high, the units need a higher hormone concentration to move to a neighbor, leading to a longer waiting time for the requester. In general, the parameter settings are essential for the algorithm to work. Therefore, we decided not to tweak the parameters manually, but optimize them by using a genetic algorithm. In this work a genetic algorithm such as in [@Elmenreich2007] is used.\
Initially, the algorithm creates a random population. Then, it uses elite selection for building the next generations. The candidates are sorted according to their fitness and the best $x$ candidates are chosen. These candidates propagate to the next generation. To reach the same population size as the last generation, the rest of the slots are reserved for mutation, crossover and new candidates. For mutation and crossover random elite candidates are chosen. Finally, random new candidates are added to the population.\
The fitness function targets client satisfaction, therefore, it optimizes the number of consumed units. The genetic algorithm is part of a simulator that also runs the artificial hormone-system. Depending on the number of generations the genetic algorithm starts the simulation of the artificial hormone-system and records the resulting parameters. The parameters have to be generated only once and can be used as input for the artificial hormone-system.
Clean-up/Replacement Strategies
===============================
We concluded in [@Sobe2011] that efficient replacement or clean-up has to be done in order to avoid blocking the transport of units. A clean-up is triggered if a certain storage level is reached, which leads to a balanced storage load of the system.\
However, if we want to guarantee that always at least one instance of the unit is available in the system, the decision of what unit to delete becomes more complicated. Therefore, the goal is to find an efficient strategy that neither influences the delay, but on the other side increases the utilization of replicas. We apply the basic principle, where none of the strategies deletes a unit if there is no copy on one of the neighbor nodes.\
We compare three mechanisms least recently used (LRU), least frequently used (LFU) and hormone-based clean-up. LRU takes care of popularity changes of units. If a unit is not popular anymore its replicas can be removed. LFU targets units of low popularity. Hormone clean-up exploits local knowledge about hormone concentration. A unit is deleted if there are no hormones for it on the neighbors, thus there is currently no demand for it. So, units currently in delivery are not deleted.\
We combine the results of the clean-up/replacement strategies with the replication strategies evaluated in [@Sobe2011]. In the following a short description of these strategies is provided:
- **Owner Replication.** The content is replicated at the requester’s node [@Lv2002]
- **Path Replication.** In a multi-hop network where content is not transported directly such as in Freenet [@Clarke2001], it is possible to cache a replica of the content in transmission in each intermediate node. Since the intermediate nodes are acting as caches, the path replication is also called cache-based replication. It is assumed that intermediate nodes provide storage space for replicas even if they are not interested in the content. Path replication leads to a high number of unused replicas.
- **Path Adaptive Replication.** An alternative to path replication is to specify a node specific replication probability, where nodes decide ad-hoc if a file is replicated or not. The replication probability is dependent on the peer’s resource status and optionally refers to the replication rate, too [@Yamamoto2005].
- **Simple Hormone**. If a unit is requested by peers from opposite parts of the network, the unit has to move first to one requester and afterwards to the other requester. This can lead to long traveling paths, which can be avoided by replicating a unit if more than one neighbor holds hormones for it. Note that it is not possible to differentiate if hormones on the neighbor are created by different peers. Thus, it is possible that unnecessary replications are made.
- **Local Popularity**. Each node uses the local request history of the corresponding content to decide if it is likely to be requested again in the future. If the rank of a content is among the best 30 % the corresponding unit is kept. So popular units are more likely to be replicated, but popularity information from neighbors is ignored. The communication effort is minimized.
- **Neighbor Popularity Ranking**. After collecting the popularity ranks for a content from the neighbors, the peer decides if it is worth to replicate the corresponding unit. The ranks are aggregated to a region rank (see [@Sobe2009b]), which is calculated as follows: $$R=\frac{1}{n} \sum_1^n{\ln(r_i)}$$ If the region rank $R$ is lower than a given threshold (e.g., the best 30 % at all neighbors) the unit is replicated. $n$ represents the number of neighbors and $r$ is the rank of the specific unit at this neighbor. To reduce the impact of peak ranks (e.g., one unit is best ranked at two nodes, but worst ranked on the third node) the logarithm is used. The cooperation of neighbors is advantageous if their taste diverges.
- **Neighbor Hormone Ranking**. Analogue to the popularity ranking the units can also be ranked by their hormone values at the neighbors. The higher the hormone value for a unit on a neighbor is, the better is the unit’s rank. The collected ranks can be aggregated as before and if the region rank is lower than a given threshold (e.g., the best 30 %), the unit is replicated.
Simulation Settings
===================
We implemented a simulator that allows for the definition of network topology, storage behavior, client behavior, etc. and performs at certain time steps hormone management, request generation, unit movement and replacement. In the following the specific settings of the current evaluation are described.
Network Topology
----------------
We assume for small overlay networks of 50 nodes a connected Erdős-Rényi random graph with a diameter of 6. For larger networks, e.g., with 1,000 nodes, we assume a scale-free network topology. To generate such a network the Eppstein Power Law Algorithm [@Eppstein2002] is used. The algorithm gets as input a random graph and by repetitively removing and adding edges a power law distribution is reached. The network diameter of the scale-free graph is 13. The bandwidth was set to 100 Mbit/s. Note that further bandwidth scenarios and parameter studies are target of future work.
Initial Storage
---------------
Each node creates units until 30 % of each node’s storage is filled, where each node supports the system with 900 MB. At the beginning only one instance of each unit exists. We expect that in a scenario with 50 motivated persons, each person is contributing with equal probability. In a scenario with 1,000 visitors we expect that there are few highly motivated persons and a high number of less motivated persons. We further assume that each person is represented by one peer. We generate 5,000 units for the 50 peers scenario, and 15,000 units for the 1,000 peers scenario.\
The average size of a unit is 2.6 MB, whereas the maximum size is 16 MB and the minimum size is 190 KB, with a playback bit rate of 1 Mbit/s. These unit sizes are the result of the third SOMA use case “The long night of research”[^1].
Request Generation
------------------
We provide keyword search, where a number of units are supposed to match the same keyword. The number of units per keyword follows a Zipf-like distribution.\
One request consists of a number of keywords. The request is fulfilled if for each keyword at least one unit is stored on the node. We further implemented a taste change, i.e., if a user likes the content just watched, her taste for future requests might be similar to the currently watched unit.\
In this paper we do not consider any order of the units, thus, if a requested unit arrives, it is presented to the user. Sequential and parallel dependencies have been handled in [@Sobe2010a].\
We further introduce a deadline for each unit, until which it has to be delivered. The deadline is dependent on the size, the link bandwidth and the maximum number of hops a unit can travel $maxhops$. If a deadline is missed, no further hormones for that unit are created to stop attracting content.\
A request is considered as failed if none of the requested units could fulfill their deadline. A user can only submit one request at a time. If this request is fulfilled or failed, a new request will be generated.
Simulation Parameters
---------------------
We used the genetic algorithm as described in Section \[sec:params\]. We change the fitness function to maximize the number of successful requests (i.e., consumed units within the deadline). The optimized parameter set is used for both random and scale-free network scenarios.\
In Table \[tab:pars\] the resulting parameters are shown. On creation the hormone value is high, leading to a wider travel range. The diffusion of 45 % of the hormones supports longer travel distances, too. The evaporation rate is in comparison to the creation number rather low, which means the hormones last for some time. The evaporation hormones are a fixed value subtracted from the current hormone level. The migration threshold describes the minimum hormone difference between two nodes to make a unit move. In this case the difference is very low in comparison to the creation amount of hormones. This leads to a very dynamic behavior of the units. The clean-up is triggered by a node if its current storage level exceeds 60 %.
------------ ------
$\eta_0$ 3.95
$\eta$ 4.39
$\alpha$ 0.45
$\epsilon$ 0.16
$m$ 0.23
$c$ 60 %
$t$ 0.23
$maxhops$ 10
------------ ------
: Parameter settings[]{data-label="tab:pars"}
Metrics
-------
We want to evaluate the request fulfillment on the one hand and the utilization of replicas on the other hand. The fulfillment of requests is represented by the *delay*. The delay is measured from the request time of a unit until the arrival of that unit on the node. A delay of $0$s means that the unit was already on the node. The delay is presented as cumulative distribution function (CDF) over the simulation time. The *deadline missed rate* represents the rate of units (not requests), for which the deadline is missed. If a unit missed its deadline, the delay is calculated as deadline minus request time (max. delay). The *request failed rate* indicates requests from which all units missed their deadline.\
A unit is presented for some time, and we measure the rate of units that currently started with presentation. The more unit presentations started in comparison to the number of their replicas, the better the *unit utilization*. The utilization and the request failed rate will be depicted as box plot with 1.5 interquartile range whiskers.\
Evaluation
==========
Scenario 1: 50 nodes network
----------------------------
![Delay Comparison of replication without clean-up [@Sobe2011][]{data-label="fig:nocleanup"}](fig/delay_none.png)
![Hormone, LRU and LFU utilization comparison[]{data-label="fig:utilization"}](fig/utilization_cleanup_boxplot.png)
![Hormone, LRU and LFU request failed rate comparison[]{data-label="fig:failedReq"}](fig/reqfailed_cleanup.png)
In this section we concentrate on the evaluation of clean-up mechanisms if applied to the replication strategies discussed above. In Figure \[fig:nocleanup\] we show again the delay distribution of the replication strategies without clean-upIn comparison to the pure replication shown in Figure \[fig:nocleanup\] in Figures \[fig:hcleanup\], \[fig:lrucleanup\] and \[fig:lfucleanup\] hormone clean-up, LRU and LFU are applied. One can see immediately that some combinations of clean-up and replication are positive, whereas some combinations lead to performance drops. The most conspicuous example is the combination of local popularity replication and hormone clean-up as shown in Figure \[fig:hcleanup\]. The reason for this is that the hormone clean-up only considers current demands for a unit and therefore does not fit the replication mechanism.\
It is further interesting to see that path replication in combination with hormone clean-up results in a lower delay than hormone ranking. Hormone ranking initially places units more efficiently, leading to less replicas than path replication would create. The downside of hormone ranking is that popular nodes evolve, which are filled first. This leads to blocked transport paths, which in turn increases the delay.\
The LRU clean-up depicted in Figure \[fig:lrucleanup\] has the most positive impact on path-random replication since the delay is stable in comparison to the best effort scenario, while the other replication mechanisms experience higher delay.\
LFU (Figure \[fig:lfucleanup\]) leads to a high number of wrong clean-up decisions, because the delay increases remarkably for all replication mechanisms. Especially the hormone replication mechanism suffers from wrong decisions; the curve gets flatter, indicating that the delay jitter increases. The local popularity mechanism experiences positive impact, since LFU also prioritizes popular units.\
The utilization measures the improvement in storage efficiency of the single replication mechanisms (see Figure \[fig:utilization\]). All clean-up strategies lead to an increase of utilization. Hormone replication does not take advantage of a clean-up, because it already creates before only a low number of replicas. Hormone ranking has the worst utilization if combined with LRU, but has even in that case a higher utilization than path replication. Local popularity replication results in a high variance, which also shows that some improvement is needed.\
A high utilization does not necessarily indicate that the replication mechanism is best suited for the delivery. A high utilization can also be reached if deadlines are missed and therefore the delivery did not take place, which leads to a lower number of replicas. Therefore, it is important to aggregate the information of delay, utilization and the requests failed statistics.\
In Figure \[fig:failedReq\] one can see immediately the low performance of local popularity replication if combined with hormone clean-up. The lowest variance is reached by combining it with LFU. The other replication strategies do not profit from the combination with LFU, the request failed rate is the highest. Most replication strategies work best if combined with LRU, except hormone ranking.\
The decision of what clean-up strategy to choose should be made by combining the results of delay, utilization and request failed rate.
Impact of Peer Churn
--------------------
![Delay distribution of Hormone Ranking with Hormone clean-up if 5,10, 20 nodes fail[]{data-label="fig:delaychurn"}](fig/delay_hchurn.png)
![Delay distribution of path adaptive with LRU if 5, 10, 20 nodes fail[]{data-label="fig:delaypathchurn"}](fig/delay_pchurn.png)
![Failed request rate in case of peer churn[]{data-label="fig:failedchurn"}](fig/reqfailed_churn.png)
We simulate churn as nodes being removed regularly one-by-one. We do not handle isolated nodes after peer deletion. Thus, there is a performance gain potential if an overlay algorithm takes care of reconnecting such nodes. We randomly chose 5, 10 and 20 nodes to be removed.\
We compare hormone ranking with hormone clean-up and path adaptive replication with LRU.\
Figure \[fig:delaychurn\] shows the delay distribution of the hormone ranking algorithm. One can see that the replication algorithm and the delivery algorithm are capable of handling loss. Overall the delay increased a bit, but interestingly the delay for 5, 10 and 20 removed nodes is very similar. However, the failed requests increase slightly in both hormone ranking and path adaptive scenarios, as shown in Figure \[fig:failedchurn\]. In general if a keyword is matched by a number of units, a system wide loss of a unit does not have a major impact.\
The path adaptive algorithm leads to interesting results as shown in Figure \[fig:delaypathchurn\]. Here, the delay of the peer churn scenarios is lower than in the original case. This can be explained by referencing the clean-up failures of the original scenario. A clean-up fails if on the current node all units are currently in use or there is no copy of the current unit on a neighbor. A disadvantageous of replica distribution might be that at every second hop a replica is placed, which means that a high number of replicas exist in the system, but because the nodes only see their neighbors, the units get not deleted. If a peer fails, a unit has to move an alternative way. Therefore, the unit movement is increased and more requests can be fulfilled. Path adaptive replication might need an alternative clean-up policy taking a larger neighborhood into account.
Scenario 2: 1000 nodes scale-free network
-----------------------------------------
![Delay Distribution of Hormone Ranking with Hormone clean-up if 100, 200, 500 nodes fail[]{data-label="fig:1000delayhranking"}](fig/delay_1000hranking.png)
![Delay distribution of path adaptive with LRU if 100,200, 500 nodes fail[]{data-label="fig:1000delaypathrand"}](fig/delay_1000pathrand.png)
![Failed request rate in case of peer churn[]{data-label="fig:reqfailure1000"}](fig/reqfailed_1000.png)
In this section we evaluate the applicability of our delivery algorithm for scale-free networks. We reduce our scenarios to hormone ranking with hormone clean-up and path adaptive replication with LRU. It is shown that the parameters for the 50 node network also work for the 1,000 peer network. Specific optimization using the genetic algorithm could lead to even better results.\
In Figure \[fig:1000delayhranking\] it is depicted that the delay is increased by around 500 ms in comparison to the small network for the hormone ranking algorithm. Furthermore, if 100, 200 and even 500 nodes fail over time the delay does not increase considerably. Note that also high degree nodes may fail, because the nodes leaving the network are chosen randomly. Figure \[fig:1000delaypathrand\] shows that the problem with clean-up failures is not experienced as in the 50 nodes network, which can be explained by the network structure and its rather low diameter. Therefore, the delay of path adaptive replication is similar to the hormone ranking algorithm. Both algorithms show slight increases of request failures also in the presence of peer churn (see Figure \[fig:reqfailure1000\]).\
Conclusion
==========
In this paper we compared different clean-up strategies to be combined with replication mechanisms. We evaluated the delay impact of the clean-up mechanisms, as well as the impact on the utilization of replicas. Furthermore, although reducing the number of replicas, the robustness of the system is still very high. Similar results are reached if the system is applied to a scale-free network of 1,000 nodes.\
Although the results of the replication mechanisms are promising, the clean-up has a negative impact on the delay, which means that the goals of the clean-up introduction are not reached. Ideally, a combination of replication and clean-up leads to low delay and high utilization. If the settings are as strict as in this paper, path replication with hormone clean-up, although inefficient, performs best. Alternatives could be path adaptive replication with LRU and hormone ranking replication with hormone clean-up. The node failure scenarios showed, that there are still nodes, which block the transport of units. To solve this issue the settings could be less strict regarding the deletion policy. Instead of deleting a unit only if there is a copy of it in the neighborhood, it could be weakened to delete a unit if another unit covering the same keyword is in the neighborhood.
[^1]: SOMA Web-page http://soma.lakeside-labs.com
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For the inverse source problem with the two-dimensional Helmholtz equation, the singular values of the ’source-to-near field’ forward operator reveal a sharp frequency cut-off in the stably recoverable information on the source. We prove and numerically validate an explicit, tight lower bound for the spectral location of this cut-off. We also conjecture and support numerically a tight upper bound for the cut-off. The bounds are expressed in terms of zeros of Bessel functions of the first and second kind.'
address: 'Department of Applied Mathematics and Computer Science, Technical University of Denmark, Matematiktorvet build. 303 B, DK-2800 Kgs. Lyngby, Denmark'
author:
- M Karamehmedović
title: |
Explicit tight bounds on the\
stably recoverable information for the\
inverse source problem
---
Introduction {#section:introduction}
============
We treat the single-frequency inverse source problem for the Helmholtz equation in the plane, illustrated in .
Fix a positive constant wavenumber $k=2\pi/\lambda$, where $\lambda$ is the operating frequency, and let $D_0$ and $D$ be open disks in $\mathbb{R}^2$ centered at the origin and with radii $R_0$ and $R\ge R_0$, respectively. Write $\Delta=\partial_{x_1}^2+\partial_{x_2}^2$ for the Laplacian, and consider the Helmholtz problem $$\label{equation:IHP}
\left\{\begin{array}{rcl}
(\Delta+k^2)u&=&s\quad{\rm in}\,\,\,\mathbb{R}^2,\\
\lim_{|x|\rightarrow\infty}\sqrt{|x|}(\partial_{|x|}-\rmi k)u(x)&=&0,\quad{\rm uniformly}\,\,{\rm for}\,x/|x|\in S^1,
\end{array}
\right.$$ for some source $s\in L^2(D_0)$ extended by zero to the whole plane. The second condition in is the outgoing Sommerfeld radiation condition in the plane. The inverse source problem, ISP, is now
*given a single measurement $U\in L^2(\partial D)$, find a source $s\in L^2(D_0)$ such that\
there is a function (’radiated field’) $u$ satisfying $u|_{\partial D}=U$ and satisfying the system *.
The ISP arises naturally in inverse acoustic and electromagnetic scattering, and has been devoted a substantial body of literature. The ISP is treated, e.g., in the multi-frequency regime by @Bao-2010, and with far-field measurement data by [@Griesmaier-2012]; see also [@ElBadia-2011]. It occurs in antenna synthesis and diagnostics [@Persson-2005; @Jorgensen-2010], the analytic continuation of solutions of exterior scattering problems [@SS; @Zaridze-1998; @Bliznyuk-2005; @oac], and in linearized inverse obstacle scattering problems.
In terms of the forward operator $F:s\mapsto U$, described in detail in , solving the ISP amounts to solving $$\label{equation:Fs=U}
Fs=U\quad\text{for}\,\,s\in L^2(D_0).$$ This problem is ill-posed, since $\ker F=(\Delta+k^2)H^2(D_0)$, where $H^2(D_0)$ is the Sobolev space $\{\partial^{\alpha}w\in L^2(D_0)\,\,\,{\rm for}\,\,\,\alpha\in\mathbb{N}^2_0\,\,\,{\rm with}\,\,\,|\alpha|\le2\}$. Also, measurements are typically noisy and sampled over a finite set of points. A common regularizing measure is to look for the minimum-$L^2$-norm, or minimum-energy, solution of , which is given by $s^{\dagger}=F^{\dagger}U$; here, $F^{\dagger}=(F^{\ast}F)^{-1}F^{\ast}=F^{\ast}(FF^{\ast})^{-1}$ is the Moore-Penrose pseudoinverse of $F$. Another regularization scheme uses a truncated singular value decomposition (TSVD) of the forward operator $F$. Here, $s$ is approximated by a finite sum of the form $\sum_1^N\sigma_m^{-1}(U,\phi)_m\psi_m$, with $(\sigma_m,\psi_m,\phi_m)$ a singular system of $F$. Our aim is to estimate the maximal amount of information about *any* source $s\in L^2(D_0)$ that can be stably recovered *in principle*, that is, *regardless* of the sampling frequency in the measurement and of the choice of the regularisation scheme. By ’stably recoverable information’ we mean ’information recoverable robustly to noise,’ and we refer to for a more precise definition. A non-asymptotic analysis of the singular values $\sigma_m$ of the forward operator $F$, performed in , reveals a low-pass filter behavior with well-defined passband and stopband. This turns out to be true also when the singular values are ordered according to increasing angular frequency of the right singular vectors of $F$, that is, of the singular vectors defined at the measurement boundary $\partial D$. In this case, the singular values within the passband generally do not increase or decrease monotonically, and the singular values in the stopband, still ordered according to angular frequency $m$, are monotonic functions of $m$.
We call the *bandwidth* $\mathscr{B}$ of the forward operator $F$ the singular value index (angular frequency $m$ of a right singular vector of $F$) at which the singular value spectrum of $F$ becomes strictly decreasing as function of nonnegative $m$: $${\mathscr{B}}={\rm argmin}_{m\in\mathbb{N}_0}\{\sigma_{m+n}>\sigma_{m+n+1}\,\,{\rm for}\,\,{\rm all}\,\,n\in\mathbb{N}_0\}.$$ With this in mind, we define the stably recoverable information on a source $s$ to be the projection of $s$ onto the singular subspace of $F$ defined by $|m|\le{\mathscr{B}}$. Then, finding the maximal amount of stably recoverable information about any source $s$, regardless of measurement sampling quality and of regularization scheme, amounts to estimating the bandwidth ${\mathscr{B}}$ of the forward operator $F$.
To simplify the notation, write $\kappa_0=kR_0$ and $\kappa=kR$ for the size parameters of the source support and of the measurement boundary, respectively. Also, for integer $m$, write $j_{m,1}$ and $y_{m,1}$ for the first positive zero of the Bessel function $J_m$ of the first kind, respectively Bessel function $Y_m$ of the second kind, and order $m$. It is well-known [@Magnus-1966 p. 146] that $j_{m,1}>0$ for all $m\in\mathbb{N}_0$. Our main result, proved in , is
\[theorem:MAIN\] The bandwidth ${\mathscr{B}}$ of the forward operator $F:s\mapsto U$ associated with the Helmholtz problem and measurement at $\partial D$ is bounded from below by $${\mathscr{B}}_-={\rm argmin}_{m\in\mathbb{N}_0}\{j_{m,1}\ge\kappa_0\}.$$
For convenience, in we also show that the bandwidth bound of Theorem \[theorem:MAIN\] can be expressed explicitly in the source size parameter $\kappa_0$:
\[corollary:ceil\] For sufficiently large $\kappa_0$, we have $$\hspace{-25mm}{\mathscr{B}}_-\approx\widetilde{\mathscr{B}}_-\hspace{-1mm}=\hspace{-1mm}\left\lceil\hspace{-1mm}\left(\frac{1}{6}\left(108\kappa_0+12\sqrt{12a_-^3+81\kappa_0^2}\right)^{1/3}\hspace{-1mm}-\frac{2a_-}{\left(108\kappa_0+12\sqrt{12a_-^3+81\kappa_0^2}\right)^{1/3}}\right)^3\,\right\rceil$$ with $a_-=1.855757$.
Finally, the general form of the result in Theorem \[theorem:MAIN\], as well as extensive numerical experimentation, lead us to conjecture a tight upper bound on the bandwitdth $\mathscr{B}$:
\[conjecture:y\] $${\mathscr{B}}_+={\rm argmin}_{m\in\mathbb{N}_0}\{y_{m,1}\ge\kappa_0\}.$$
In we analyze the singular value spectrum of the forward operator $F$. In particular, we prove Theorem \[theorem:MAIN\] and Corollary \[corollary:ceil\] in . We validate the bounds ${\mathscr{B}}_-$ and ${\mathscr{B}}_+$ on the bandwidth ${\mathscr{B}}$ numerically in , and discuss some implications of Theorem \[theorem:MAIN\] in . A conclusion and suggestions for further work are given in .
Spectral analysis of the forward operator {#section:spectralanalysis}
=========================================
The function $(\rmi/4)H_0^{(1)}(k|x|)$, $x\in\mathbb{R}^2$, is the radial outgoing fundamental solution of the Helmholtz operator in the plane, with singularity at the origin. Recall that $H_0^{(1)}=J_0+\rmi Y_0$ is the Hankel function of zero order and of the first kind. As in [@Bao-2010], introduce the forward operator $$Fs(x)=\int_{y\in D_0}H_0^{(1)}(k_0|x-y|)s(y),\quad x\in\partial D,\,s\in L^2(D_0),$$ that maps sources $s$ to the traces at $\partial D$ of the corresponding radiated fields. It is well-known [@Bao-2010] that $F:L^2(D_0)\rightarrow L^2(\partial D)$ is compact. The adjoint $F^{\ast}$ is defined by $$F^{\ast}U(y)=\int_{x\in\partial D}H_0^{(2)}(k|x-y|)U(y),\quad y\in D_0,\,\,U\in L^2(\partial D),$$ where $H_0^{(2)}=J_0-\rmi Y_0$ is the Hankel function of zero order and of the second kind.
A singular system of $F$ {#subsection:asingular}
------------------------
@Bao-2010 derived a singular system of the forward operator $F$. We here slightly improve a part of their Proposition 2.1:
\[lemma:SVD\] The forward operator $F$ admits the singular value decomposition $$F=\sigma_0(\cdot,\psi_0)_{L^2(D)}\phi_0+\sum_{m\in\mathbb{N}}\sigma_m\left[(\cdot,\psi_m)_{L^2(D)}\phi_m+(\cdot,\psi_{-m})_{L^2(D)}\phi_{-m}\right],$$ where $$\label{equation:xi_m}
\sigma_m=\sqrt{2R}\pi R_0|H_m^{(1)}(\kappa)|A_m(\kappa_0),\quad m\in\mathbb{N}_0,$$ and $$\begin{aligned}
\psi_m(y)&=&(\sqrt{\pi}R_0A_m(\kappa_0))^{-1}J_m(k|y|)\rme^{\rmi m\arg y},\\
\phi_m(x)&=&(2\pi R)^{-1/2}\rme^{\rmi\arg H_m^{(1)}(\kappa)}\rme^{\rmi m\arg x},\end{aligned}$$ for $m\in\mathbb{Z}$, $x\in\partial D$ and $y\in D_0$. Here $$\begin{aligned}
A_m(\kappa_0)&=&\sqrt{J_m(\kappa_0)^2-J_{m-1}(\kappa_0)J_{m+1}(\kappa_0)}\\&=&\sqrt{J_m(\kappa_0)^2+J_{m+1}(\kappa_0)^2-\frac{2m}{\kappa_0}J_m(\kappa_0)J_{m+1}(\kappa_0)}\end{aligned}$$ for $m\in\mathbb{Z}$.
Our slight improvement of Proposition 2.1 of @Bao-2010 consists in explicitly evaluating the integral $\int_{\varrho=0}^{R_0}\varrho J_m^2(k\varrho)$, occurring in $\sigma_m$ and $\psi_m$, in terms of $A_m(\kappa_0)$. This explicit evaluation is crucial to our proof of Theorem \[theorem:MAIN\]. We also note that our expressions for the singular vectors $\phi_m$, as well as the singular values $\sigma_m$, differ from [@Bao-2010] in that they are only proportional to those given in that reference.
For $s\in L^2(D_0)$ and $y\in D_0$ we have $$\label{equation:F*F}
F^{\ast}Fs(y)=\int_{z\in D_0}s(z)\int_{x\in\partial D}H_0^{(1)}(k|x-z|)H_0^{(2)}(k|x-y|).$$ A special case of the Graf addition theorem [@AbramowitzStegun Eq. 9.1.79, p. 363] reads $$H_0^{(1)}(k|x-y|)=\sum_{m\in\mathbb{Z}}H_m^{(1)}(\kappa)J_m(k|y|)\rme^{\rmi m(\arg x-\arg y)},\quad x\in\partial D,\, y\in D_0.$$ Similar to [@Bao-2010], inserting this in we get $$\begin{aligned}
\hspace{-1cm}F^{\ast}Fs(y)&=\sum_{m,n\in\mathbb{Z}}H_m^{(1)}(\kappa)H_n^{(2)}(\kappa)J_n(k|y|)\rme^{-\rmi n\arg y}\\&\times\int_{z\in D_0}s(z)J_m(k|z|)\rme^{-\rmi m\arg z}\int_{x\in\partial D}\rme^{\rmi(m+n)\arg x}\\&=2\pi R\sum_{m\in\mathbb{Z}}|H_m^{(1)}(\kappa)|^2J_m(k|y|)\rme^{\rmi m\arg y}\int_{z\in D_0}s(z)J_m(k|z|)\rme^{-\rmi m\arg z},\end{aligned}$$ since $J_{-m}=(-1)^mJ_m$ and $Y_{-m}=(-1)^mY_m$ for all integer $m$. This gives an eigendecomposition of the operator $F^{\ast}F$; to normalize the eigenvectors, we note that @GradRyzh [Eq. 5.54.2, p. 629] gives $$\int\varrho J_m(k\varrho)^2=\frac{\varrho^2}{2}\left(J_m(k\varrho)^2-J_{m-1}(k\varrho)J_{m+1}(k\varrho)\right),\quad m\in\mathbb{Z},$$ and the recursion formula for cylinder functions [@GradRyzh Eq. 8.471.1, p. 926] implies $$\label{equation:recursion}
J_{m-1}(\kappa)+J_{m+1}(\kappa)=\frac{2m}{\kappa}J_m(\kappa),\quad m\in\mathbb{Z}.$$ Thus, $$\begin{aligned}
\hspace{-20mm}\int_{\varrho=0}^{R_0}\varrho J_m(k\varrho)^2&=R_0^2\int_{\varrho=0}^1\varrho J_m(\kappa_0\varrho)^2=\frac{R_0^2}{2}\left(J_m(\kappa_0)^2-J_{m-1}(\kappa_0)J_{m+1}(\kappa_0)\right)\\&=\frac{R_0^2}{2}\left(J_m(\kappa_0)^2+J_{m+1}(\kappa_0)^2-\frac{2m}{\kappa_0}J_m(\kappa_0)J_{m+1}(\kappa_0)\right)=\frac{R_0^2A_m(\kappa_0)^2}{2},\end{aligned}$$ and $F^{\ast}F$ admits the spectral decomposition $$F^{\ast}F=\sigma_0^2(\cdot,\psi_0)_{L^2(D_0)}+\sum_{m\in\mathbb{N}}\sigma_m^2\left[(\cdot,\psi_m)_{L^2(D_0)}\psi_m+(\cdot,\psi_{-m})_{L^2(D_0)}\psi_{-m}\right].$$ Evidently, $\sigma_0^2$ has multiplicity one and all the other eigenvalues $\sigma_m^2$, $m\in\mathbb{N}$, have multiplicity two. The lemma now follows from Theorem 4.7 on p. 100 of @ColtonKress; it here just remains to compute $$\begin{aligned}
\phi_m(x)&=&\sigma_m^{-1}F\psi_m(x)=\frac{\int_{y\in D_0}H_0^{(1)}(k|x-y|)J_m(k|y|)\rme^{\rmi m\arg y}}{\sqrt{2R}\pi^{3/2}R_0^2|H_m^{(1)}(\kappa)|A_m(\kappa_0)^2}\\&=&\frac{\sum_{\nu\in\mathbb{Z}}H_{\nu}^{(1)}(\kappa)\rme^{\rmi\nu\arg x}\int_{\varrho=0}^{R_0}\varrho J_{\nu}(k\varrho)J_m(k\varrho)\int_{\theta=0}^{2\pi}\rme^{\rmi\theta(m-\nu)}}{\sqrt{2R}\pi^{3/2}R_0^2|H_m^{(1)}(\kappa)|A_m(\kappa_0)^2}\\&=&(2\pi R)^{-1/2}\rme^{\rmi\arg H_m^{(1)}(\kappa)}\rme^{\rmi m\arg x},\quad x\in\partial D,\,\,m\in\mathbb{N}_0.\end{aligned}$$
Figure 3 shows the first 71 nonnegative-index singular values of the forward operator $F$ with size parameters $\kappa=\kappa_0=10\pi$.
\[figure:Fig3\]
{width="52.00000%"} {width="52.00000%"}
Clearly, the forward operator is a low-pass filter with respect to the singular values $\sigma_m$, with bandwidth ${\mathscr{B}}=27$. We quantify the frequency response of this filter in .
Proof of Theorem \[theorem:MAIN\] and of Corollary \[corollary:ceil\] {#subsection:proofof}
---------------------------------------------------------------------
In this section we prove the lower bound ${\mathscr{B}}_-$ on the bandwidth ${\mathscr{B}}$ given in Theorem \[theorem:MAIN\], and the approximate value of ${\mathscr{B}}_-$ given in Corollary \[corollary:ceil\]. For completeness, we first prove that the distance between the zeros of the function $0\le\mu\mapsto J_{\mu}(\kappa_0)$ is greater than $1$.
\[lemma:only\_one\_zero\] If $\mu_2>\mu_1\ge0$ and $J_{\mu_1}(\kappa_0)=J_{\mu_2}(\kappa_0)=0$ then $\mu_2-\mu_1>1$.
Let $n\in\mathbb{N}$. The interlacing property of the zeros of Bessel functions (see, e.g., @Palmai-2011) implies $j_{0,n+1}>j_{1,n}$. Since $\mu\mapsto j_{\mu,n}$ is strictly increasing with $\mu$, $j_{\mu,n}=j_{0,n+1}$ implies $j_{\mu,n}>j_{1,n}$, hence $\mu>1$. As illustrated in , to show that $j_{\mu_1,n+1}=j_{\mu_2,n}$ implies $\mu_2-\mu_1>1$, it now suffices to establish that $$\label{equation:ddmu}
\frac{dj_{\mu,n}}{d\mu}<\frac{dj_{\mu,n+1}}{d\mu}\quad\text{for all }\mu\ge0,\,n\in\mathbb{N}.$$
![The zeros of the function $0\le\mu\mapsto J_{\mu}(\kappa_0)$ diverge, see Lemma \[lemma:only\_one\_zero\].[]{data-label="figure:Bessel_zeros"}](bessel_zeros)
For nonnegative order $\mu$, the $n$’th zero $j_{\mu,n}$ of the Bessel function $J_{\mu}$ satisfies [@Watson-1945 pp. 508–510] $$\frac{dj_{\mu,n}}{d\mu}=2j_{\mu,n}\int_{t=0}^{\infty}K_0(2j_{\mu,n}\sinh t)\rme^{-2\mu t}.$$ Substituting $q=2j_{\mu,n}\sinh t$ and using that $j_{\mu,n}>0$, $\exp({\rm arcsinh}\,\tau)=\tau+\sqrt{1+\tau^2}$, as well as that $\cosh{\rm arcsinh}\,\tau=\sqrt{1+\tau^2}$, we get $$\frac{dj_{\mu,n}}{d\mu}=\int_{q=0}^{\infty}K_0(q)\left(q/2j_{\mu,n}+\sqrt{1+q^2/4j_{\mu,n}^2}\right)^{-2\mu}\left(1+q^2/4j_{\mu,n}^2\right)^{-1/2}.$$ Setting $$f(c)=\int_{q=0}^{\infty}K_0(q)(q/2c+\sqrt{1+q^2/4c^2})^{-2\mu}\left(1+q^2/4c^2\right)^{-1/2},\quad c>0,$$ we find $$\frac{\partial f}{\partial c}(c)=2^{2\mu+1}c^{2\mu}\int_{q=0}^{\infty}\frac{qK_0(q)\left(2\mu\sqrt{4c^2+q^2}+q\right)}{(4c^2+q^2)^{3/2}\left( \sqrt {4c^2+q^2}+q \right)^{2\mu}}>0\quad{\rm for}\,\,c>0.$$ Finally, for any $\mu\ge0$ and $n\in\mathbb{N}_0$, we have $j_{\mu,n}<j_{\mu,n+1}$, so indeed holds.
We can now link the variation of the function $m\mapsto A_m(\kappa_0)$ with that of the Bessel function of the first kind. Fix $m\in\mathbb{N}_0$.
\[lemma:A\_variation\] If $J_{\xi}(\kappa_0)=0$ for some $\xi\in[m,m+1]$ then $A_m(\kappa_0)\le A_{m+1}(\kappa_0)$.
The recursion formula implies $$\begin{aligned}
\hspace{-23mm}A_m(\kappa_0)^2-A_{m+1}(\kappa_0)^2&=&J_m(\kappa_0)^2+J_{m+1}(\kappa_0)^2-\frac{2m}{\kappa_0}J_m(\kappa_0)J_{m+1}(\kappa_0)\\&-&J_{m+1}(\kappa_0)^2-J_{m+2}(\kappa_0)^2+\frac{2(m+1)}{\kappa_0}J_{m+1}(\kappa_0)J_{m+2}(\kappa_0)\\&=&\frac{2(m+1)}{\kappa_0}J_{m+1}(\kappa_0)(J_m(\kappa_0)-J_{m+2}(\kappa_0))\\&-&\frac{2m}{\kappa_0}J_m(\kappa_0)J_{m+1}(\kappa_0)+\frac{2(m+1)}{\kappa_0}J_{m+1}(\kappa_0)J_{m+2}(\kappa_0)\\&=&\frac{2}{\kappa_0}J_m(\kappa_0)J_{m+1}(\kappa_0).\end{aligned}$$ For any fixed positive argument $x$, the function $\mathbb{R}\ni\mu\mapsto J_{\mu}(x)$ is differentiable and not identically zero. Thus, by assumption, and by lemma \[lemma:only\_one\_zero\], this function changes sign precisely once in the interval $[m,m+1]$, so $$A_m(\kappa_0)^2-A_{m+1}(\kappa_0)^2=\frac{2}{\kappa_0}J_m(\kappa_0)J_{m+1}(\kappa_0)\le0.$$
\[remark:h2\_increasing\] Clearly, the function $\mathbb{N}_0\ni m\mapsto |H_m^{(1)}(\kappa)|^2$ is positive-valued. It is also strictly increasing, as can be seen from Nicholson’s integral for $|H_m^{(1)}(\kappa)|^2$ [@Watson-1945 pp. 441-444], $$|H^{(1)}_m(\kappa)|^2=\frac{8}{\pi^2}\int_{t=0}^{\infty}K_0(2\kappa\sinh t)\cosh2mt,\quad m\in\mathbb{Z},\,\kappa>0,$$ where $K_0$ is the modified Bessel function of the second kind. This namely implies $$\partial_m(|H_m^{(1)}(\kappa)|^2)=\frac{16m}{\pi^2}\int_{t=0}^{\infty}K_0(2\kappa\sinh t)\sinh2mt>0,\quad m>0,$$ since both $K_0$ and the hyperbolic sine are positive over positive reals.
The above discussion suffices for a proof of the lower bound ${\mathscr{B}}_-$.
\[proof:M-\] Let $m\in\mathbb{N}_0$. If $\kappa_0>j_{m,1}$ then there are $n\in\mathbb{N}_0$ and $\xi\in[m+n,m+n+1]$ satisfying $J_{\xi}(\kappa_0)=0$, and, by Lemma \[lemma:A\_variation\], $A_{m+n}(\kappa_0)\le A_{m+n+1}(\kappa_0)$. Since $\mathbb{N}_0\ni\mu\mapsto|H_{\mu}^{(1)}(\kappa)|$ is strictly increasing, and $\sigma_{\mu}$ is proportional to $|H_{\mu}^{(1)}(\kappa)|A_{\mu}(\kappa_0)$ for $\mu\in\mathbb{N}_0$, we have $\sigma_{m+n}\le\sigma_{m+n+1}$, hence $m<\mathscr{B}$. In conclusion, $\mathscr{B}\ge{\rm argmax}_{m\in\mathbb{N}_0}\{j_{m,1}<\kappa_0\}+1={\rm argmin}_{m\in\mathbb{N}_0}\{j_{m,1}\ge\kappa_0\}$.
We use that [@Watson-1945 p. 516] $j_{m,1}=m+a_-m^{1/3}+O(m^{-1/3})$, with $a_-=1.855757$. The real solution of $n^3+a_-n-\kappa_0=0$ is readily found to be $$\hspace{-4mm}n=\frac{1}{6}\left(108\kappa_0+12\sqrt{12a_-^3+81\kappa_0^2}\right)^{1/3}-\frac{2a_-}{\left(108\kappa_0+12\sqrt{12a_-^3+81\kappa_0^2}\right)^{1/3}},$$ so ${\mathscr{B}}_-\approx\lceil n^3\rceil$ for sufficiently large $\kappa_0$.
For completeness, let us also provide an approximate expression for the conjectured value of the upper bound ${\mathscr{B}}_+$. We have [@Watson-1945 p. 516] $y_{m,1}=m+a_+m^{1/3}+O(m^{-1/3})$, with $a_+=0.931577$. Also, $m<m-2+a_+(m-2)^{1/3}$ for integer $m\ge12$, so, for sufficiently large $\kappa_0$, $m>\kappa_0$ implies $y_{m-2,1}>\kappa_0$ and hence ${\mathscr{B}}_+\approx\widetilde{\mathscr{B}}_+=\lceil\kappa_0\rceil$.
Numerical validation {#section:numericalvalidation}
====================
We here compute the bandwidth ${\mathscr{B}}$, as well as the bandwidth bounds ${\mathscr{B}}_-$ and ${\mathscr{B}}_+$ of Theorem \[theorem:MAIN\] and Conjecture \[conjecture:y\], respectively, for 300 values of the size parameters $\kappa=\kappa_0$ uniformly distributed over the interval $\kappa\in[2,100\pi]$. Recall that $\kappa=kR=2\pi R/\lambda$ and $\kappa_0=kR_0=2\pi R_0/\lambda$, where $R$ is the radius of the sampling circle $\partial D$, $R_0$ is the radius of the source domain, and $\lambda$ is the operating wavelength. Thus, we consider 300 values of the relative wavelength $\lambda/R=\lambda/R_0$ distributed nonuniformly over the interval $\lambda/R\in[1/50,\pi]$. Figure 5 shows the errors $\varepsilon_{\pm}={\mathscr{B}}_{\pm}-{\mathscr{B}}$ and the relative errors $\varepsilon_{\sf rel,\pm}=|{\mathscr{B}}_{\pm}-{\mathscr{B}}|/{\mathscr{B}}$ in the estimated bandwidth as function of the problem size parameter $\kappa$.
\[figure:epsilon\]
![Errors (top) and relative errors (bottom) in the lower and upper bounds on the bandwidth ${\mathscr{B}}$ of Theorem \[theorem:MAIN\], over the range of the source size parameter corresponding to $R/\lambda=R_0/\lambda\in[1/10,50]$.](./eps_- "fig:"){width="49.50000%"} ![Errors (top) and relative errors (bottom) in the lower and upper bounds on the bandwidth ${\mathscr{B}}$ of Theorem \[theorem:MAIN\], over the range of the source size parameter corresponding to $R/\lambda=R_0/\lambda\in[1/10,50]$.](./eps_+ "fig:"){width="49.50000%"} ![Errors (top) and relative errors (bottom) in the lower and upper bounds on the bandwidth ${\mathscr{B}}$ of Theorem \[theorem:MAIN\], over the range of the source size parameter corresponding to $R/\lambda=R_0/\lambda\in[1/10,50]$.](./eps_-_rel "fig:"){width="49.50000%"} ![Errors (top) and relative errors (bottom) in the lower and upper bounds on the bandwidth ${\mathscr{B}}$ of Theorem \[theorem:MAIN\], over the range of the source size parameter corresponding to $R/\lambda=R_0/\lambda\in[1/10,50]$.](./eps_+_rel "fig:"){width="49.50000%"}
\[figure:epsilon\_approx\]
![Errors (top) and relative errors (bottom) in the approximate lower and upper bounds $\widetilde{{\mathscr{B}}}_{\pm}$ on the bandwidth ${\mathscr{B}}$ of Theorem \[theorem:MAIN\], over a range of the problem size parameter $\kappa$.](./eps_approx_- "fig:"){width="49.50000%"} ![Errors (top) and relative errors (bottom) in the approximate lower and upper bounds $\widetilde{{\mathscr{B}}}_{\pm}$ on the bandwidth ${\mathscr{B}}$ of Theorem \[theorem:MAIN\], over a range of the problem size parameter $\kappa$.](./eps_approx_+ "fig:"){width="49.50000%"} ![Errors (top) and relative errors (bottom) in the approximate lower and upper bounds $\widetilde{{\mathscr{B}}}_{\pm}$ on the bandwidth ${\mathscr{B}}$ of Theorem \[theorem:MAIN\], over a range of the problem size parameter $\kappa$.](./eps_approx_-_rel "fig:"){width="49.50000%"} ![Errors (top) and relative errors (bottom) in the approximate lower and upper bounds $\widetilde{{\mathscr{B}}}_{\pm}$ on the bandwidth ${\mathscr{B}}$ of Theorem \[theorem:MAIN\], over a range of the problem size parameter $\kappa$.](./eps_approx_+_rel "fig:"){width="49.50000%"}
For the two lowest considered values of $\kappa$, we find that ${\mathscr{B}}=0$ and ${\mathscr{B}}_-=0$; there, we set $\varepsilon_{\sf rel,-}=0$. Both ${\mathscr{B}}$ and $\mathscr{B}_-$ are positive for higher considered values of $\kappa$. In particular, there is zero bandwidth for $\kappa$ smaller than some threshold value between approx. 1.7 and approx. 2.7, and for such size parameters $\kappa$ the inverse source problem is, from the viewpoint of the bandwidth of the singular values, similar to the inverse heat conduction problem. Over the considered interval for $\kappa$, the mean errors are $\overline{\varepsilon}_-=-1.68$, $\overline{\varepsilon}_+=3.02$, and the maximum absolute errors are $\max|\varepsilon_-|=3$, $\max|\varepsilon_+|=4$. The relative error in ${\mathscr{B}}_-$ is below $5\%$ for $\kappa\ge24.7461$, i.e., for $R/\lambda\ge3.94$, and ${\mathscr{B}}_+$ is below $5\%$ for $\kappa\ge45.7181$, i.e., for $\lambda/R\ge7.28$.
We find both ${\mathscr{B}}$, ${\mathscr{B}}_-$ and ${\mathscr{B}}_+$ to be approximately linear functions of $\kappa$ in the given interval, with least-squares fits summarized in Table \[table:lsf\].
\[table:lsf\] linear interpolant mean absolute error standard deviation
------------------- ----------------------- --------------------- --------------------
${\mathscr{B}}$ $0.9793\kappa-3.9569$ $0.4813$ $3.9\cdot10^{-4}$
${\mathscr{B}}_-$ $0.9736\kappa-4.7394$ $0.5715$ $4.8\cdot10^{-4}$
${\mathscr{B}}_+$ $0.9861\kappa-2.0083$ $0.4052$ $3.4\cdot10^{-4}$
: Linear regression of the computed bandwidth $\mathscr{B}$, and the lower (${\mathscr{B}}_-$) and upper (${\mathscr{B}}_+$) bandwidth bounds. We have here held equal the size parameters $\kappa$ and $\kappa_0$.
Figure 6 shows errors in the approximations $\widetilde{{\mathscr{B}}}_{\pm}$ (for definition of $\widetilde{{\mathscr{B}}}_{\pm}$ see proof of Corollary \[corollary:ceil\] and the paragraph immediately following it, on page 9.) The approximate expression for the lower bound shows almost the same small error as the lower bound itself, and the approximate expression $\widetilde{\mathscr{B}}_+\approx\lceil\kappa_0\rceil$ for the upper bound, while simple, has error below $5\%$ only for problem size parameters of approx. 175 or higher when $\kappa=\kappa_0$ is maintained.
Our bounds ${\mathscr{B}}_{\pm}$ are independent of the radius of the measurement surface, and we next validate this property numerically. Figure 7 shows the first 71 nonnegative-index singular values of the forward operator $F$ with size parameters $\kappa=100\pi$, $\kappa_0=10\pi$. The bandwidth is unchanged at ${\mathscr{B}}=27$ (compare with Figure 3), as predicted by our bounds. The decrease in the numerical stability of the ISP due to the measurement boundary being farther away from the source is instead expressed in terms of the overall lower level of the singular values.
{width="49.50000%"} {width="49.50000%"}
Discussion {#section:discussion}
==========
The bandwidth estimates ${\mathscr{B}}_{\pm}$ are directly applicable as optimal filter estimates in the numerical solution of the inverse source problem in terms of a truncated singular value decomposition (TSVD) of the forward operator. Next, it has been amply observed in the literature concerning the single-frequency inverse source problem that the numerical stability of the solution increases with the operating frequency. Theorem \[theorem:MAIN\] confirms and explicitly quantifies this increase in numerical stability, also for non-asymptotic frequencies.
Theorem \[theorem:MAIN\] of course has direct implications for the maximum achievable stable resolution of the reconstruction in the inverse source problem. Detailed analysis of this resolution requires an investigation of the pointwise behavior of the left singular vectors of the forward operator. While we here do not perform such analysis, we do note that the left singular vectors tend to be supported near the origin for low values of index $m$, and near the measurement boundary for high index values. This means the amplified noise produces is a ’wall of non-information’ near the measurement boundary and blocks faithful reconstruction of the source inside $D$.
As shown in [@Bao-2010] and in Section \[section:spectralanalysis\] here, the right singular vectors (defined over the measurement boundary) of the forward operator are proportional to $\exp(\rmi m\theta)$, $m\in\mathbb{Z}$. This means that the bandwidth index ${\mathscr{B}}$ is approximately the angular frequency of the highest-frequency data component that can be stably inverted. Thus, the sampling theorem [@Shannon-1949] is directly applicable with Theorem \[theorem:MAIN\] to give the following: in case the radiated field $u$ is sampled equidistantly at the boundary $\partial D$, any angular sampling rate greater than approximately $\Delta\theta\approx\pi/{\mathscr{B}}\le\pi/{\mathscr{B}}_-=\pi/{\rm argmin}_{m\in\mathbb{N}_0}\{j_{m,1}\ge\kappa_0\}$ is excessive due to the limited bandwidth of the forward operator.
Bandwidth bounds in Theorem \[theorem:MAIN\] and Conjecture \[conjecture:y\] involve the size parameter of only the source support, and in light of the successful numerical validation of these bounds, we find it justified to say that the bandwidth is generally independent of the radius $R$ of the measurement boundary relative to the radius $R_0$ of the source support (as long as $R\ge R_0$). As illustrated in Section \[section:numericalvalidation\], the decrease in the robustness of the inversion (in the presence of noise) as $R_0/R$ decreases seems instead to be expressed by a lower overall level of the singular values. We therefore briefly analyze the asymptotic behavior of the singular spectrum as $m\rightarrow0$, and as $m\rightarrow\infty$. The standard large-argument approximation of the Bessel functions of the first and second kind, valid for $\kappa_0\gg m^2-1/4$, yields $$\hspace{-10mm}J_m(\kappa_0)\sim\sqrt{\frac{2}{\pi\kappa_0}}\cos\left(\kappa_0-\frac{m\pi}{2}-\frac{\pi}{4}\right),\quad Y_m(\kappa_0)\sim\sqrt{\frac{2}{\pi\kappa_0}}\sin\left(\kappa_0-\frac{m\pi}{2}-\frac{\pi}{4}\right),$$ so $$\begin{aligned}
A_m(\kappa_0)^2&\sim&\frac{2}{\pi\kappa_0}\left(\cos\left(\kappa_0-\frac{m\pi}{2}-\frac{\pi}{4}\right)^2\right.\\&-&\Biggl.\cos\left(\kappa_0-\frac{m\pi}{2}-\frac{\pi}{4}+\frac{\pi}{2}\right)\cos\left(\kappa_0-\frac{m\pi}{2}-\frac{\pi}{4}-\frac{\pi}{2}\right)\Biggr)\\&=&\frac{2}{\pi\kappa_0}\left(\cos\left(\kappa_0-\frac{m\pi}{2}-\frac{\pi}{4}\right)^2+\sin\left(\kappa_0-\frac{m\pi}{2}-\frac{\pi}{4}\right)^2\right)=\frac{2}{\pi\kappa_0}\end{aligned}$$ and, since $\kappa\ge\kappa_0$, we also have $H_m(\kappa)^2\sim2/\pi\kappa$. Thus $$\sigma_m\sim\frac{\sqrt{2}}{\pi}\lambda\sqrt{R_0}$$ for $R_0/\lambda\gg(m^2-1/4)/2\pi$. Forward operators mapping from source spaces with larger supports thus have higher-valued singular values in the bandpass region, regardless of the size $R$ of the measurement boundary relative to the size $R_0$ of source support. However, we also see that the height of the bandpass decreases when the operating wavelength lambda decreases (equivalently, when the operating frequency increases), which may counteract the increase in stably recoverable information gained due to the increase in bandwidth. In the small-argument limit ($0<\kappa^2\ll m+1$) the standard approximation is $$J_m(\kappa_0)\approx\frac{1}{m!}\left(\frac{\kappa_0}{2}\right)^m,\quad Y_m(\kappa_0)\approx-\frac{(m-1)!}{\pi}\left(\frac{2}{\kappa_0}\right)^m,$$ so (since $\kappa_0\le\kappa$) $A_m(\kappa_0)^2\sim(\kappa_0/2)^{2m}m!^{-2}(m+1)^{-1}$ and $H_m(\kappa)^2\sim(\kappa/2)^{2m}m!^{-2}+m!^2m^{-2}\pi^{-2}(\kappa/2)^{-2m}$, resulting in $$\begin{aligned}
A_m(\kappa_0)^2H_m(\kappa)^2&\sim\left(\frac{R_0}{R}\right)^{2m}\frac{1}{\pi^2m^2(m+1)},\end{aligned}$$ and thus $$\sigma_m\sim\frac{1}{m}\sqrt{\frac{2}{m+1}}\left(\frac{R_0}{R}\right)^{m-1/2}R_0^{3/2}.$$ Evidently, the ratio $R_0/R$ of the source support radius to the measurement boundary radius strongly affects the rate of decay of the singular values, the robustness of the inversion to noise generally improving as the source support approaches the measurement boundary.
Conclusion and further work {#section:conclusionand}
===========================
We analyzed the singular values of the forward operator associated with the single-frequency inverse source problem for the Helmholtz equation in the plane. In particular, we considered bounds on the information content that is preserved by the forward operator, proving a tight lower bound and conjecturing a tight upper bound on the singular value index of the highest-frequency data component that is stably recoverable. The bounds were expressed in terms of the zeros of Bessel functions of the first and the second kind. We validated both bounds numerically, establishing concrete estimates on the stably recoverable information in the inverse source problem regardless of the data sampling rate and the choice of regularization. The result can be used directly, e.g., to estimate optimal TSVD filters and data sampling rates.
Proving the statement in Conjecture \[conjecture:y\] is a natural next step. Also, it would complete the picture to supplement the results on the bandwidth with a more precise description of the general levels and decay rates of the singular values as function of the size parameters of the source support and of the measurement boundary, individually or in relation to one another. Finally, a spectral analysis of the forward operator in dimension greater than 2 will be interesting.
[xx]{}
Abramowitz, M. Stegun, I. A., 1972. *Handbook of Mathematical Functions With Formulas, Graphs and Mathematical Tables*. Tenth Printing. United States Department of Commerce: National Bureau of Standards.
El Badia A and Nara T 2011 An inverse source problem for Helmholtz’s equation from the Cauchy data with a single wave number *Inverse Problems* **27**(10) 105001
Bao G, Lin J Triki F 2010 A multi-frequency inverse source problem *Journal of Differential Equations* **249** 3443–3465
Bliznyuk N, Pogorzelski R J Cable V P 2005 Localization of Scattered Field Singularities in Method of Auxiliary Sources. In: IEEE Antennas and Propagation Society, *2005 IEEE Antennas and Propagation Society International Symposium and USNC/URSI National Radio Science Meeting*. Washington DC, 3–8 July 2005
Colton D and Kress R 2013 *Inverse Acoustic and Electromagnetic Scattering Theory* 3ed (New York: Springer)
Gradsteyn I S and Ryzhik I M *Table of Integrals, Series, and Products* 7ed (Burlington: Academic Press)
Griesmaier R, Hanke M and Raasch T 2012 Inverse source problems for the Helmholtz equation and the windowed Fourier transform *SIAM Journal on Scientific Computing* **34**(3) A1544–A1562
Jørgensen E, Meincke P, Cappellin C and Sabbadini M 2010 Improved Source Reconstruction Technique for Antenna Diagnostics *Proceeedings of the 32nd ESA Antenna Workshop*, ESTEC, Noordwijk, The Netherlands
Karamehmedović M 2015 On analytic continuability of the missing Cauchy datum for Helmholtz boundary problems *American Mathematical Society. Proceedings* **143**(4) 1515–1530
Magnus W, Oberhettinger F Soni R P 1966 *Formulas and Theorems for the Special Functions of Mathematical Physics* (Berlin: Springer)
Pálmai T and Apagyi B 2011 Interlacing of positive real zeros of Bessel functions *Journal of Mathematical Analysis and Applications* **375** 320–322
Persson K and Gustafsson M 2005 Reconstruction of Equivalent Currents using a Near-Field Data Transformation – With Radome Applications *Progress in Electromagnetics Research* **PIER 54** 179–198
Shannon C E 1949 Communication in the presence of noise *Proc. Institute of Radio Engineers* **37**(1) 10–21
Sternin B and Shatalov V 1994 *Differential Equations on Complex Manifolds* (Kluwer Academic Publishers)
Watson G N 1945 *A treatise on the theory of Bessel functions* (Cambridge University Press)
Zaridze R S, Jobava R, Bit-Banik G, Karkasbadze D, Economou D P Uzunoglu N K 1998 The Method of Auxiliary Sources and Scattered Field Singularities (Caustics) *Journal of Electromagnetic Waves and Applications* **12** 1491–1507
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we determine the automorphism groups of the profinite braid groups with four or more strings in terms of the profinite Grothendieck-Teichmüller group.'
address:
- 'Arata Minamide: University of Nottingham, School of Mathematical Sciences, University Park Nottingham NG7 2RD, United Kingdom; Mathematical Institute, University of Oxford, Woodstock Road Oxford OX2 6GG, United Kingdom'
- 'Hiroaki Nakamura: Department of Mathematics, Graduate School of Science, Osaka University, Toyonaka, Osaka 560-0043, Japan'
author:
- Arata Minamide and Hiroaki Nakamura
title: |
The automorphism groups\
of the profinite braid groups
---
\#1
[:=]{}
[\_r]{}
[\_i]{}
[\_[>0]{}]{}
[\_[0]{}]{}
[\_l]{}
[\^1]{}
[\_3]{}
[\_4]{}
[\_n]{}
[\_4]{}
[\_3]{}
[\_n]{}
[\_n]{}
[\_n]{}
[\_3]{}
[\_n]{}
[\_3]{}
[B\_n]{}
[P\_n]{}
[\_n]{}
[\_n]{}
[\_n]{}
[\_n]{}
[\_n]{}
[\_n]{}
[\_3]{}
[\_4]{}
[C\_n]{}
[\_n]{}
[\_[0, \[n\]]{}]{}
[\_[0, n]{}]{}
[\_[n+1]{}]{}
[\_[0, \[n+1\]]{}]{}
[\_[0, \[n+1\]]{}]{}
[\_[0, n+1]{}]{}
[\_[0, n+1]{}]{}
[\_[0, 4]{}]{}
[\_[0, 4]{}]{}
[\_[0, n+1]{}]{}
[\_]{}
[\_]{}
[\_]{}
[\_4]{}
[\_4]{}
[\_3]{}
[H\_]{}
[Z\_]{}
[B\_]{}
[14G32; 20F36, 20E18, 14H30, 14H10]{}
Introduction
============
Let $\bn$ be the Artin braid group with $n(\ge 2)$ strings defined by generators $\sigma_1, \sigma_2, \dotso, \sigma_{n-1}$ and relations:
- $\sigma_i \sigma_{i+1} \sigma_i
\ = \ \sigma_{i+1} \sigma_i \sigma_{i+1}$ $\quad (i=1,\dots,n-1)$,
- $\sigma_i \sigma_j \ = \ \sigma_j \sigma_i \ \ \
(|i-j|\ge2)$.
In [@dg], J. L. Dyer and E. K. Grossman studied the automorphism group $\aut(B_n)$ and showed $\out(B_n)\cong\Z/2\Z$ for $n\ge 3$. In this paper, we study the continuous automorphisms of the profinite completion $\bnh$ of $B_n$. We prove
Let $n\ge 4$. There exists a natural isomorphism $$\out(\widehat B_n) \ \cong \ \gt\times (1+n(n-1)\zh)^{\times},$$ where $\gt$ is the profinite Grothendieck-Teichmüller group introduced by V.Drinfeld [@Dr], Y.Ihara [@I90]-[@im] and $(1+n(n-1)\zh)^{\times}$ is the kernel of the natural projection $\zh^\times\to (\Z/n(n-1)\Z)^\times$.
It is well known that the center $\cnh$ of $\bnh$ is (topologically) generated by $
\zeta_n \defeq (\sigma_1\sigma_2 \dotsm \sigma_{n-1})^n
$ and is isomorphic to $\zh$. Write $$\mbnh \ \defeq \ \bnh/\cnh. $$ Since $\cnh$ is a characteristic subgroup of $\bnh$, there is induced the natural homomorphism $\gt \to \out(\mbnh)$. The key fact for the proof of Theorem A is the following isomorphism theorem.
Let $n\ge4$. Then, it holds that $\gt \isom \out(\mbnh)$.
Our proofs of Theorems A and B rely on preceding works by many authors on the Grothendieck-Teichmüller group $\gt$ and the profinite completion $\widehat\Gamma_{0,n}$ of the mapping class group $\Gamma_{0,n}$ of the sphere with $n$ marked points (cf.[@im], [@LS1], [@LS2], [@C12]). The permutation of labels defines a natural inclusion of the symmetric group of degree $n$: $\mathfrak{S}_n\hookrightarrow \Out(\widehat\Gamma_{0,n})$, whose image commutes with the standard action of $\gt$ on $\widehat\Gamma_{0,n}$ ([@im]). D.Harbater and L.Schneps [@HS] remarkably showed that when $n\ge 5$, $\gt$ is characterized as a “special” subgroup of the centralizer of $\mathfrak{S}_n$ in $\Out(\widehat\Gamma_{0,n})$. In a recent work [@hmm], this result has been improved by showing that the focused centralizer is indeed [*full*]{} as large as possible in $\Out(\widehat\Gamma_{0,n})$. In particular,
\[hmm\] There is a natural isomorphism of profinite groups $$\gt \times \sno \ \isom \ \out(\pgnoh)$$ for every integer $n\ge 4$. $\square$
Theorems A and B will be derived by translating the ingredient of Theorem \[hmm\] for $\Out(\pgnoh)$ into the language of $\Out(\mbnh)$ and $\Out(\bnh)$. Arguments given by Dyer-Grossman [@dg] for discrete braid groups generically guide us also in profinite context. However, for the case $n=4$, we elaborate a different treatment in §3 due to the existence of non-standard surjections $B_4\epi\frakS_4$ found in E. Artin’s classic [@a]. Our argument in §3 looks at the “Cardano-Ferrari” homomorphism $B_4\epi B_3$ which has close relations with the universal monodromy representation in once-punctured elliptic curves. Noting that $\cB_4$ is isomorphic to the mapping class group $\Gamma_{1,2}$ of a topological torus with two marked points, we obtain from Theorem B the following remarkable
There is a natural isomorphism $\gt\isom\out(\widehat\Gamma_{1,2})$.
[*Acknowledgement*]{}: The first author would like to thank Prof.Shinichi Mochizuki for helpful discussions and warm encouragements. During preparation of this manuscript, the authors learnt that Yuichiro Hoshi and Seidai Yasuda also had discussions on topics including a similar phase to this paper. This work was supported by the Research Institute for Mathematical Sciences, an International Joint Usage/Research Center located in Kyoto University. Work on this paper was partially supported by EPSRC programme grant “Symmetries and Correspondences” EP/M024830.
Generalities on braid groups
============================
We begin with recalling basic facts on braid groups (cf.e.g.,[@kt]). Let $n\ge3$ be an integer. The pure braid group $\pn$ is the kernel of the epimorphism $$\begin{aligned}
\varpi_n : \bn & \ \epi \ \sn \\
\sigma_i & \ \mapsto \ (i, i+1)
\qquad (i=1,\dots,n-1).\end{aligned}$$ The center $C_n$ of $P_n$ coincides with the center of $B_n$ which is a free cyclic group generated by $$\zeta_n \ \defeq \ (\sigma_1 \sigma_2 \dotsm \sigma_{n-1})^n.$$ Write $\cP_n:=P_n/C_n$ and $\cB_n:=B_n/C_n$. The above $\varpi_n$ factors through $\pi_n:\cB_n\epi\frakS_n$ and there arise the following exact sequences of finitely generated groups: $$\begin{gathered}
\label{exactSn}
\begin{CD}
1 @>>> \mpn @>>> \mbn @>\pi_n>> \sn @>>> 1,
\end{CD}
\\
\label{exactBn}
\begin{CD}
1 @>>> \cn @>>> \bn @>>> \mbn @>>> 1,
\end{CD}
\\
\label{exactPn}
\begin{CD}
1 @>>> \cn @>>> \pn @>>> \mpn @>>> 1.
\end{CD}
$$ We introduce the [*mapping class group of the $n$-times punctured sphere*]{} $\gn$ to be the group generated by $\bar\sigma_1, \bar\sigma_2, \dotso, \bar\sigma_{n-1}$ with the relations
- $\bar\sigma_i \bar\sigma_{i+1} \bar\sigma_i \ = \
\bar\sigma_{i+1} \bar\sigma_i \bar\sigma_{i+1}
\quad (i=1,\dots,n-1)$,
- $\bar\sigma_i \bar\sigma_j \ = \ \bar\sigma_j
\bar\sigma_i \ \ \ (|i-j|\ge2)$,
- $\bar\sigma_1 \dotsm \bar\sigma_{n-2}\, \bar\sigma_{n-1}^2
\bar\sigma_{n-2} \dotsm \bar\sigma_1 \ = \ 1$,
- $(\bar\sigma_1 \bar\sigma_2 \dotsm \bar\sigma_{n-1})^n \ = \ 1$.
Observe that there is a natural epimorphism $$\begin{aligned}
\label{Psi_n}
\Psi_n : B_n & \ \epi \ \gn \\
\sigma_i & \ \mapsto \ \bar\sigma_i \qquad (i=1,\dots,n-1) \notag\end{aligned}$$ which factors through $\mbn=B_n/C_n$. We also write $\pgn$ for the [*pure mapping class group of the $n$-times punctured sphere*]{} which is by definition the kernel of the epimorphism $$\begin{aligned}
\gamma_n : \gn & \ \epi \ \sn \\
\bar\sigma_i & \ \mapsto \ (i, i+1)\qquad (i=1,\dots,n-1)
\notag\end{aligned}$$ fitting in the exact sequence $$\begin{gathered}
\label{exactGn}
\begin{CD}
1 @>>> \pgn @>>> \gn @>>> \sn @>>> 1.
\end{CD}
$$ In this paper, besides the above epimorphism $\Psi_n$ (\[Psi\_n\]), another shifted morphism $$\begin{aligned}
\Phi_n : \bn & \ \to \ \gno \\
\sigma_i & \ \mapsto \ \bar\sigma_i
\qquad (i=1,\dots,n-1) \notag\end{aligned}$$ plays an important role, whose kernel is known to coincide with $\cn$ ([@fm §9.2-3]). The homomorphism $\Phi_n$ induces the following commutative diagram of groups $$\label{discreteCD}
\vcenter{
\xymatrix{
1 \ar[r] &\mpn \ar[r] \ar[d]^{\vertsim}&\mbn \ar[r]^{\mpin}
\ar@{^{(}->}[d] & \sn \ar[r] \ar@{^{(}->}[d]^{\iota_n}&1 \\
1 \ar[r] &\pgno \ar[r] &\gno \ar[r]^{{{\gamma}_{n+1}}} &\sno\ar[r] &1,
}
}$$ where the horizontal sequences are exact; the left-hand (resp. middle; right-hand) vertical arrow is the isomorphism (resp. the injection; the natural injection which trivially extends each permutation of $\{1, 2, \dotso, n\}$ to that of $\{1, 2, \dotso, n+1\}$) induced from $\Phi_n$.
It is well known that the profinite completion functor preserves the (injectivity of the) kernel part of the exact sequences (\[exactSn\])-(\[exactPn\]) and (\[exactGn\]) respectively. If $Z(G)$ denotes the center of a profinite group $G$, then $$\label{various-centers}
\begin{cases}
&Z(\mpn) = Z(\mbn) = Z(\mpnh) = Z(\mbnh) = \{1\}, \\
&\cnh = Z(\pnh) = Z(\bnh) \ (\cong \zh).
\end{cases}$$ hold (cf.e.g.,[@n §1.2-1.3]).
\[moduli\] Let $n\ge3$ be an integer. We shall write $(\ast)$ for the commutative diagram of profinite groups $$\vcenter{
\xymatrix{
1 \ar[r] &\mpnh \ar[r] \ar[d]^{\vertsim}&\mbnh \ar[r]^{\mpinh}\ar@{^{(}->}[d]& \sn\ar[r] \ar@{^{(}->}[d]^{\iota_n}&1 \\
1 \ar[r] &\pgnoh \ar[r] &\gnoh \ar[r]^{{\widehat{\gamma}_{n+1}}} &\sno\ar[r] &1
}
}
\tag{$\ast$}$$ which is obtained as the profinite completion of (\[discreteCD\]). Note that the horizontal sequences are exact as remarked as above.
\[char\] Suppose that $n \neq 4$, $n\ge3$. Then every epimorphism $\mbnh \epi \sn$ has kernel $\mpnh$. In particular, $\mpnh$ is a characteristic subgroup of $\mbnh$.
E.Artin ([@a Theorem 1]) classified all surjective homomorphisms $B_n\epi \frakS_n$ up to equivalence by conjugation in $\frakS_n$: When $n\ne 4,6$, there is a unique equivalence class and when $n=6$ there are two classes mutually equivalent by a nontrivial outer automorphism of $\frakS_6$. This proves the assertion for discrete braid groups. Lemma \[ker\] below with the residual finiteness of $\mbn$ settles the assertion for the profinite braid groups.
\[ker\] Let $G$ be a residually finite group, $N$ a normal subgroup of $G$ with finite quotient $Q \defeq G/N$. Suppose that every epimorphism $G \epi Q$ has the same kernel $N$. Then, every epimorphism $\widehat{G} \epi Q$ has the same kernel $\widehat{N}$.
Note first that, by one-to-one correspondence between the finite index subgroups of $G$ and the open subgroups of $\widehat{G}$, the image of the monomorphism $\widehat{N} \to \widehat{G}$ coincides with the closure of $N$ in $\widehat{G}$. Let $p : \widehat{G} \epi Q$ be a given epimorphism. Then, by [@rz Proposition 3.2.2 (a)], the closure of $H \defeq \Ker(p) \cap G$ in $\widehat{G}$ coincides with $\Ker(p)$. Consider the composite: $$\varphi : G \ \epi \ G/H \ \isom \ \widehat{G}/\Ker(p) \ \isom \ Q,$$ where the first arrow is the projection, the second arrow is the isomorphism induced from the associated morphism $G \hra \widehat{G}$ ([@rz Proposition 3.2.2 (d)]) and the third arrow is the isomorphism induced by $p$. From the assumption, $\varphi$ has the kernel $N$, i.e., $N=\Ker(\varphi) = H$. Thus, $\Ker(p)$ coincides with $\widehat{N}$.
Special case $\mbfh$
====================
The main aim of this section is to provide a proof of the following
\[charaP4B4\] $\mpfh$ is a characteristic subgroup of $\mbfh$.
In the proof of [@dg Theorem 11] claiming that $\cP_n$ is characteristic in $\cB_n$ for $n\ge 3$, we find an inaccurate argument for the case $n=4$: By E.Artin’s classic work ([@a Theorem 1]), each surjective homomorphism $B_4\epi \frakS_4$ is equivalent to one of the following $\epsilon_1,\epsilon_2,\epsilon_3$ up to change of labels in $\{1,2,3,4\}$: $$\begin{aligned}
\epsilon_1:B_4 & \epi \Sf \qquad
(\sigma_1 \mapsto \ (12), \
\sigma_2 \mapsto \ (23) ,\
\sigma_3 \mapsto \ (34)); \\
\epsilon_2:B_4 & \epi \Sf \qquad
(\sigma_1 \mapsto \ (1234), \
\sigma_2 \mapsto \ (2134) ,\
\sigma_3 \mapsto \ (1234)); \\
\epsilon_3:B_4 & \epi \Sf \qquad
(\sigma_1 \mapsto \ (1234), \
\sigma_2 \mapsto \ (2134) ,\
\sigma_3 \mapsto \ (4321)).\end{aligned}$$ Among them, $\ker(\epsilon_1)=P_4$, while neither $\ker(\epsilon_2)$ or $\ker(\epsilon_3)$ equals to $P_4$, for $\sigma_1^2\in P_4$ has non-trivial images in $\frakS_4$: $\epsilon_2(\sigma_1^2)=
\epsilon_3(\sigma_1^2)=(13)(24)$.
Let $\bar\epsilon_1:\cB_4\epi\frakS_4$ be the induced map. Given an arbitrary automorphism $\phi\in \aut(\cB_4)$, consider the composite $$\epsilon_\phi:
B_4\to B_4/C_4=\cB_4
\underset{\phi}{\isom}
\cB_4
\overset{\bar\epsilon_1}{\longrightarrow}
\frakS_4.$$ Dyer-Grossman [@dg p.1159] discusses that $\epsilon_\phi$ cannot be equivalent to $\epsilon_2$, for $(\bar\sigma_1\bar\sigma_2\bar\sigma_3)^2$ has order exactly two in $\cB_4$ hence does not belong to $\cP_4$ (torsion-free), while $\epsilon_2((\sigma_1\sigma_2\sigma_3)^2)=1$. If moreover one knew $\epsilon_\phi\not\sim \epsilon_3$, then one could get $\epsilon_\phi\sim\epsilon_1$ and hence $\phi(\cP_4)=\cP_4$ so as to conclude Proposition \[charaP4B4\]. However, in [@dg], apparently omitted is a discussion about $\epsilon_3$ as the existence of $\epsilon_3$ is already missed in their citation of Artin’s theorem in [@dg Theorem 2]. Since $\epsilon_3((\sigma_1\sigma_2\sigma_3)^2)=(12)(34)\ne 1$, a simple replacement of the above argument for $\epsilon_\phi\not\sim\epsilon_2$ does not work to eliminate another possibility $\epsilon_\phi\sim\epsilon_3$.
The fact that $\mathcal{P}_4$ is a characteristic subgroup of $\mbf$ has followed in a different approach by topologists (see, e.g., [@Ko Theorem 3]) by using finer analysis of the mapping class group action on the complex of curves $C(S)$ on a topological surface $S$. However, a profinite variant of $C(S)$ to derive Proposition \[charaP4B4\] still remains unsettled even to this day. Below, we give an alternative argument looking closely at a family of characteristic subgroups of $\cB_4$. We argue in the profinite context, however, our discussion works also for the discrete case in the obvious interpretation. Our main targets arise from the following epimorphisms $b_{43}:\mbfh \epi \mbth$ and $s_{43}:\Sf \epi \st$ defined by $$\label{orbicurve}
b_{43}:\mbfh \epi \mbth :
\begin{cases}
\bar\sigma_1, \bar\sigma_3 & \mapsto \bar\sigma_1, \\
\quad \bar\sigma_2 & \mapsto \bar\sigma_2;
\end{cases}
\quad s_{43}:\Sf \epi \St :
\begin{cases}
(12), (34) & \mapsto (12) , \\
\quad (23) & \mapsto (23) ,
\end{cases}
$$ and the composition $$\mathfrak{P} := \mpith\circ b_{43} \left(= s_{43}\circ\widehat{\pi}_4\right) :\mbfh \epi \st$$ where $\widehat{\pi}_n:\widehat{\mathcal{B}}_n \epi \mathfrak{S}_n$ is as in the previous section. The kernel of $s_{43}$ is what is called the Klein four group $$V_4 \ \defeq \ \ker(s_{43}) \ = \ \{id, (12)(34), (13)(24), (14)(23)\}
\subset\frakS_4.$$ Denote by $p_{43}: \mpfh \to \mpth$ the restriction of $b_{43}:\mbfh\to \mbth$ and write $\pzfh:=\ker(p_{43})$. We note that $p_{43}$ is not the same as the usual homomorphism obtained by forgetting one strand of pure 4-braids. These maps fit into the following commutative diagram of horizontal and vertical exact sequences: $$\label{BigDiagram}
\vcenter{
\xymatrix{
{} & 1\ar[d] & 1\ar[d] & 1\ar[d] & {} \\
1 \ar[r] & \pzfh \ar[d] \ar[r] & \ker(b_{43}) \ar[d] \ar[r] & V_4 \ar[d] \ar[r] & 1 \\
1 \ar[r] & \mpfh \ar[r] \ar[d] ^{p_{43}}&\mbfh \ar[rd]_{\mathfrak{P}} \ar[r]^{\mpifh} \ar[d] _{b_{43}}
& \Sf \ar[r] \ar[d]^{s_{43}} &1 \\
1 \ar[r] &\mpth \ar[r] \ar[d] &\mbth \ar[r]_{\mpith} \ar[d] &\St \ar[d] \ar[r] &1 \\
{} & 1 & 1 & 1 & {}.
}}$$
Concerning the two sequences of subgroups $\mbfh\supset\ker(b_{43})\supset\pzfh$ and $\mbfh\supset\ker(\mathfrak{P})\supset\mpfh
$, we shall prove
\[charaKer\] (i) $\pzfh$ is a characteristic subgroup of $\ker(b_{43})$.\
(ii) $\ker(\mathfrak{P})$ is a characteristic subgroup of $\mbfh$.\
(iii) $\ker(b_{43})$ is a characteristic subgroup of $\mbfh$.\
(iv) $\pzfh$ is a characteristic subgroup of $\mbfh$.\
(v) $\mpfh$ is a characteristic subgroup of $\mbfh$.
Proposition \[charaP4B4\] is obtained as (v) of the above Proposition. Here is a simple immediate consequence of it:
$\widehat{P}_n$ is a charactersitic subgroup of $\widehat{B}_n$ for every $n\ge 3$.
Proposition \[char\] and Proposition \[charaP4B4\] show that $\mpnh$ is a characteristic subgroup of $\mbnh$ for every $n\ge 3$. Assertion follows from this and the fact that $\widehat{P}_n$ is the inverse image of $\mpnh$ by the projection $\widehat{B}_n\epi\mbnh$ whose kernel is the center $\cnh$ of $\widehat{B}_n$.
For the proof of Proposition \[charaKer\], note first that (iv) follows from (i) and (iii). We will apply (iv) for the proof of (v). Assertion (ii) will be used to prove (iii). In fact, (ii) follows from a stronger assertion that every epimorphism $\mbfh \epi \st$ has the same kernel as $\ker(\mathfrak{P})$. In fact, it is not difficult to see that every (discrete group) homomorphism $B_4\epi \st$ is conjugate to the standard one $B_4\epi B_3\epi \frakS_3$ (cf.e.g., [@lin Theorem 3.19 (a)]). Since $\mbf$ is residually finite, the profinite version follows from Lemma \[ker\]. To complete the proof of Proposition \[charaKer\], it remains to prove (i), (iii) and (v).
[*Proof of Proposition*]{} \[charaKer\] (i): Let us begin with geometric interpretation of $\pzfh\subset \ker(b_{43})$ which has been well studied by topologists (see, e.g., [@aswy §2.1], [@KS §3]). One may regard the standard lift $\beta_{43}:\widehat B_4\epi \widehat B_3$ of $b_{43}:\widehat\cB_4\to\widehat\cB_3$ (given by $\sigma_1,\sigma_2,\sigma_3\mapsto \sigma_1,\sigma_2,\sigma_1$ respectively) as the $\pi_1^\et$-transform of the “Cardano-Ferrari mapping $\mathfrak{F}_0: (\mathbb{A}^4\setminus D)_0
\to (\mathbb{A}^3\setminus D)_0$” assigning to a monic quartic (with no multiple zeros) its cubic resolvent (in the notations of [@N13 §5.4]). The kernel of $\beta_{43}$ is isomorphic to the free profinite group $\widehat F_2$ of rank 2. In fact, after Mordell transformation, the homomorphism $\beta_{43}=\pi_1^\et(\mathfrak{F}_0)$ turns to interpret the monodromy of the universal family of the (affine part of) elliptic curves $$\xymatrix{
{} & \mathcal{E}\setminus\{O\}= \{Y^2=4X^3-g_2X-g_3\} \ar[d] \\
{} & \mathfrak{M}_{1,1}^{\,\omega} =\{(g_2,g_3)\mid \Delta\ne 0\}.
}$$ Let $\sqrt{\zeta_4}:=(\sigma_1\sigma_2\sigma_3)^2$ so that $\beta_{43}(\sqrt{\zeta_4})=\zeta_3\in B_3$. Then, the reduced sequence $$\begin{CD}
1 @>>> F_2 @>>> B_4/\la\sqrt{\zeta_4}\ra @>>> B_3/\la\zeta_3\ra=\mathrm{PSL}_2(\Z) @>>> 1
\end{CD}$$ fits in the orbifold quotient of the complex model of elliptic curve family over the upper half plane. Taking into account that $\sqrt{\zeta_4}\pmod{\la\zeta_4\ra}$ acts on each elliptic curve $E:Y^2=4X^3-g_2X-g_3$ by the switching $\pm Y$ involution, we see that $\ker(b_{43})$ can be regarded as the fundamental group of an orbicurve $\mathbb{P}^1_{\infty,2,2,2}$ obtained as the $X$-line from $(E\setminus\{O\})/\{\pm 1\}$; it turns out to be isomorphic to the profinite free product of three copies of $\Z/2\Z$: $$\begin{aligned}
\ker(b_{43})&=\pi_1(\mathbb{P}^1_{\infty,2,2,2}/\C) \\
&=(\Z/2\Z)\freeprod(\Z/2\Z)\freeprod(\Z/2\Z)
\notag\end{aligned}$$ which may also be regarded as the profinite completion of discrete free product $(\Z/2\Z)\ast(\Z/2\Z)\ast(\Z/2\Z)$ ([@rz §9.1]). The normal subgroup $\pzfh$ of $\ker(b_{43})$ corresponds to the fundamental group of the Galois cover of $\mathbb{P}^1_{\infty,2,2,2}$ with group $V_4$ given in the Lattés cover diagram: $$\vcenter{
\xymatrix{
E\setminus E[2] \ar[d] \ar[r] & \mathbb{P}^1
-\{e_0,e_1,e_2,e_3\} \ar[d] \\
E\setminus \{O\} \ar[r] &\mathbb{P}^1_{\infty,2,2,2}
}}$$ where the left vertical arrow is the isogeny of punctured elliptic curves by multiplication by 2, and horizontal arrows correspond to the $\{\pm 1\}$-quotients. From this we obtain a cartesian diagram of profinite groups: $$\label{Pi04inkeralpha}
\vcenter{
\xymatrix{
\ker(b_{43})=(\Z/2\Z)\freeprod(\Z/2\Z)\freeprod(\Z/2\Z)
\ar @{->>}[r]
& (\Z/2\Z)\times(\Z/2\Z)\times(\Z/2\Z) \\
\widehat\Pi_{0,4} \ar @{_{(}->}[u]\ar @{->>}[r]
&(\Z/2\Z) \ar @{_{(}->}[u]_{{}_{diagonal \ map}},
}
}$$ where the upper horizontal arrow is the abelianization map. Moreover, according to Herfort-Ribes ([@HR Theorem 2 (i)]), the torsion elements of $(\Z/2\Z)\freeprod(\Z/2\Z)\freeprod (\Z/2\Z)$ form exactly the three conjugacy classes of order two which, therefore, must be preserved as a set under $\aut(\ker(b_{43}))$. This characterizes the diagonal image of $(\Z/2\Z)$ in the right hand side of (\[Pi04inkeralpha\]). Thus we conclude that $\widehat\Pi_{0,4}$ is characteristic in $\ker(b_{43})$ as the pull-back image of $(\Z/2\Z)\overset{diag.}{\hookrightarrow} (\Z/2\Z)^3$ along the abelianization of $\ker(b_{43})$.
[*Proof of Proposition*]{} \[charaKer\] (iii): To prove (iii), pick any $\phi \in \aut(\mbfh)$. We first show that $\phi(\ker(b_{43})) \subset \ker(b_{43})$. As $\ker(\mathfrak{P})$ is characteristic in $\widehat\cB_4$ as shown in (ii), it follows that $\phi(\ker(b_{43}))\subset \ker(\mathfrak{P})$. Hence $b_{43}$ maps $\phi(\ker(b_{43}))$ onto a subgroup of $\mpth(\isom\pgtoh\cong\widehat F_2$). But $\phi(\ker(b_{43}))$ is isomorphic to $\ker(b_{43})$ which is a topologically finitely generated closed normal subgroup of $\mbfh$. Since $\widehat F_2$ has no nontrivial non-free finitely generated normal subgroups ([@LvD Corollary 3.14]) and since $\ker(b_{43})\cong (\Z/2\Z)
\freeprod
(\Z/2\Z)\freeprod(\Z/2\Z)$ has finite abelianization $(\Z/2\Z)^3$, the image $\phi(\ker(b_{43}))$ must be annihilated by $b_{43}$, i.e., $ \phi(\ker(b_{43}))\subset \ker(b_{43})$. We can argue in the same way after replacing $\phi$ by $\phi^{-1}$ to obtain $ \phi^{-1}(\ker(b_{43}))\subset \ker(b_{43})$. Combining both inclusions implies $ \phi(\ker(b_{43}))=\ker(b_{43})$.
[*Proof of Proposition*]{} \[charaKer\] (v): Let us write $[\ast]^\ab$ for the abelianization of $[\ast]$. Since we already know ‘(iv): $\pzfh$ is characteristic in $\mbfh$’ from (i)-(iii), for proving $\mpfh$ characteristic in $\mbfh$, it suffices to show the assertion that $ \mpfh$ is the kernel of the conjugate representation $\rho:\mbfh \ {\rightarrow} \ \aut(\pzfh^{\mathrm{ab}})$. First we note that $\rho$ factors through $\bar\rho: \mbfh/\mpfh\cong\frakS_4\to
\aut(\pzfh^{\mathrm{ab}})$. This follows from the observation that $p_{43}^\ab$ injects $\pzfh^{\mathrm{ab}}$ into $\mpfh^\ab$: Indeed, writing $\{\bar x_{ij}\}$ for the image of the standard generator system $\{x_{ij}=\sigma_{j-1}\cdots\sigma_{i+1}\sigma_i^2
\sigma_{i+1}^{-1}\cdots\sigma_{j-1}^{-1}\mid
1\le i<j\le n\}$ of $\cP_n$, we find $$\label{p34ab}
p_{43}^\ab:\mpfh^\ab \to \mpth^\ab :\
\begin{cases}
\bar x_{12}, \bar x_{34} & \mapsto \bar x_{12}, \\
\bar x_{13}, \bar x_{24} & \mapsto \bar x_{13}, \\
\bar x_{14}, \bar x_{23} & \mapsto \bar x_{23}.
\end{cases}
$$ Taking into account the single relation $\bar x_{12}+ \bar x_{13} +\bar x_{14}
+\bar x_{23}+\bar x_{24}+\bar x_{34}=0$ for $\mpfh^\ab$ (respectively, $\bar x_{12}+ \bar x_{13} +\bar x_{14}=0$ for $\mpth^\ab$), we easily see from the description (\[p34ab\]) of $p_{43}^\ab:\zh^5\epi \zh^2$ that $\ker(p_{43}^\ab)$ is isomorphic to $\zh^3$ (torsion-free) into which $\pzfh^\ab$ must inject. Then, to complete proof of the assertion, it suffices to see faithfulness of $\bar\rho: \mbfh/\mpfh\cong\frakS_4\to
\aut(\pzfh^{\mathrm{ab}})$. This is easily seen from the general fact that the action of $\cB_n/\cP_n=\frakS_n$ on the $\bar x_{ij}\in\mathcal{P}^\ab_n$ is given by the natural action on indices, once declared $\bar x_{ij}=\bar x_{ji}$. The action of $\frak S_4$ on $\pzfh^\ab$ turns out to be the standard permutation representation $\zh^4$ modulo the diagonal line, which is faithful.
Proofs of Theorems A and B
==========================
By virtue of Propositions \[char\] and \[charaP4B4\], we know that $\mpnh $ is a characteristic subgroup of $\mbnh$ for $n\ge 3$. The following proposition follows immediately from this together with the well-known fact that $\out(\sn) = \{1\}$ in the case $n\ne 6$. However, the case $n=6$ requires a special care, since $\out(\mathfrak{S}_6) \cong \zz/2\zz$. Theorem \[hmm\] (Hoshi-Minamide-Mochizuki) allows us to give a uniform proof working for all $n\ge 4$.
\[out\] Regard $\sn$ as the quotient of $\mbnh$ and of $\widehat\Gamma_{0,[n]}$ by $\varpi_n:B_n\to\sn$ in §2.\
(i) Every automorphism of $\mbnh$ induces an inner automorphism of $\sn$ for $n\ge 3$.\
(ii) $\widehat\Gamma_{0,n}$ is a characteristic subgroup of $\widehat\Gamma_{0,[n]}$ in the profinite completion of (\[exactGn\]), and every automorphism of $\widehat\Gamma_{0,[n]}$ induces an inner automorphism of $\sn$ for $n\ge 5$.
\(i) As $\out(\St) = \{1\}$, the assertion is trivial when $n=3$. Suppose $n\ge4$ and pick any $\phi \in \aut(\mbnh)$. Then, it follows from Propositions \[char\] and \[charaP4B4\], that $\phi$ induces $(\phi_P,\phi_\frakS)
\in\aut(\mpnh)\times\aut(\frakS_n)$, Moreover $\phi_P$ induces $\phig\in\aut(\pgnoh)$ via the natural isomorphism $\mpnh \isom \pgnoh$ given by $\Phi_n$ of §2. Let $\bar\phi_\Gamma\in\out(\pgnoh)$ be the outer class of $\phi_\Gamma$, and let $(\phi_0,\phi_1)\in\gt\times\frakS_{n+1}$ be the image of $\bar\phi_\Gamma$ under the isomorphism $\out(\pgnoh)\isom\gt\times\frakS_{n+1}$ of Theorem \[hmm\]. Then we have the commutative diagram $$\label{diag41}
\vcenter{
\xymatrix{
\sn \ar[r]^-{\chi_n} \ar[d]_{\vertsim}^{\phis}
& \out(\pgnoh) \ar[r]^{\sim} \ar[d]^{\inn(\bar{\phi}_{\Gamma})}_{\vertsim}
& \gt\times\frakS_{n+1}
\ar[d]^{\inn(\phi_0,\phi_1)
}_{\vertsim}
\\
\sn \ar[r]^-{\chi_n} &\out(\pgnoh) \ar[r]^{\sim}
& \gt\times\frakS_{n+1}
}}$$ where $\chi_n : \sn\to \out(\mpnh)=\out(\pgnoh)$ is the natural isomorphism regarding the commutative diagram $(*)$ in Definition \[moduli\]. Since $\chi_n$ factors through $\iota_n : \sn \hra \sno$, the above (\[diag41\]) makes the diagram $$\begin{CD}
\sn @>\iota_n>> \sno \\
@V{\phis}V{\vertsim}V
@V{\vertsim}V\inn(\phi_1)V \\
\sn @>\iota_n>> \sno
\end{CD}$$ commutative, hence $\phi_1$ normalizes (hence lies in) the image of $\iota_n$. From this follows that $\phis$ is an inner automorphism of $\sn$.
(ii): Recall from §2 that there is a surjection sequence $\bnh \epi \mbnh \epi \widehat\Gamma_{0,[n]}\epi \frakS_n$. By Proposition \[char\], every epimorphism from $\mbnh$ to $\frakS_n$ has kernel $\mpnh$ for $n\ge 5$. This makes $\widehat\Gamma_{0,n}$ to be a characteristic subgroup of $\widehat\Gamma_{0,[n]}$ as the pull-back of $\mpnh \subset \mbnh$. For the rest, we can argue in exactly a similar (and simpler) way to the case (i) with employing $\chi_n':\sn\to\out(\widehat\Gamma_{0,n})\cong\gt\times\frakS_n$ for the role of $\chi_n$ in (i). We leave the rest of detail to the reader.
For the proof of Theorem B, we prepare a simple lemma of group theory. Let $$\begin{CD}
1 @>>> \Delta @>>> \Pi @>>> G @>>> 1
\end{CD}$$ be an exact sequence of profinite groups with $\rho : G \to \out(\Delta)$ the associated outer representation. Let $Z_{\out(\Delta)}(\im(\rho))$ denote the centralizer of the image $\rho(G)$ in $\out(\Delta)$. Assume that $\Delta$ and $G$ are topologically finitely generated so that $\aut(\Delta)$, $\aut(G)$ are profinite groups. Write $\aut_G(\Pi)$ (resp. $\inn_\Pi(\Delta)$) for the group of automorphisms of $\Pi$ which preserve $\Delta \sseq \Pi$ and induce the identity automorphism of $G$ (resp. for the group of inner automorphisms of $\Pi$ by the elements of $\Delta$). Then,
\[dpg\] Notations being as above, we have the following assertions.\
(i) Suppose $Z(\Delta)=\{1\}$. Then the restriction map $\aut_G(\Pi) \to \aut(\Delta)$ induces an isomorphism $$\aut_G(\Pi)/\inn_\Pi(\Delta) \ \isom \ Z_{\out(\Delta)}(\im(\rho)).$$ (ii) Suppose $Z(G)=\{1\}$ and that $\Delta$ is a characteristic subgroup of $\Pi$. Then we have an exact sequence of profinite groups $$\begin{CD}
1 @>>> \aut_G(\Pi)/\inn_\Pi(\Delta) @>\mathj>> \out(\Pi) @>\varpi>> \out(G).
\end{CD}$$
Assertion (i) follows immediately from [@n Corollary 1.5.7]. We consider (ii). First, observing $
\aut_G(\Pi) \ \cap \ \inn(\Pi) \ = \ \inn_\Pi(\Delta)$ under the assumption $Z(G)=\{1\}$, we obtain the monomorphism $$\mathj : \aut_G(\Pi)/\inn_\Pi(\Delta)\hra \ \aut(\Pi)/\inn(\Pi) \ = \ \out(\Pi)$$ from the natural injection $\aut_G(\Pi) \hra \aut(\Pi)$. Next, since $\Delta$ is a characteristic subgroup of $\Pi$, there exists a natural homomorphism $\varpi : \out(\Pi) \ \to \ \out(G)$ with $\varpi\circ\mathj=1$. Then, immediately from the surjectivity $\inn(\Pi)\epi \inn(G)$ follows that $\im(\mathj) = \Ker(\varpi)$, which completes the proof of (ii).
We now obtain Theorem B:
\[gtmbnh\] (i) Let $n\ge4$ be an integer. Then the composite $$\gt \ \to \ \out(\mbnh)$$ of the natural homomorphisms $\gt \to \out(\bnh)\to \out(\mbnh)$ is an isomorphism.\
(ii) Let $n\ge 5$. Then, the natural homomorphism $$\gt \ \to \ \out(\widehat\Gamma_{0,[n]})$$ induced from $\widehat{\Psi}_n : \widehat{B}_n \epi \widehat\Gamma_{0,[n]}$ (\[Psi\_n\]) is an isomorphism.
First, we note that $\sn$ and $\mpnh$ are center-free (\[various-centers\]), and that $\mpnh$ is a characteristic subgroup of $\mbnh$ (Propositions \[char\] and \[charaP4B4\]). Consider the upper exact sequence $$\begin{CD}
1 @>>> \mpnh @>>> \mbnh @>>> \sn @>>> 1 \\
\end{CD}$$ of $(*)$ in Definition \[moduli\], and write $\varphi_n:\sn \to \out(\mpnh)$ for the associated outer representation. Let us apply Lemma \[dpg\] to the above exact sequence. By virtue of Proposition \[out\] (i), the homomorphism $\varpi: \out(\mbnh)\to \out(\sn)$ of Lemma \[dpg\] (ii) turns out trivial, so $\mathj$ in loc.cit. together with Lemma \[dpg\] (i) gives an isomorphism $$Z_{\out(\mpnh)}(\varphi_n(\sn)) \ \isom \ \out(\mbnh).$$ Then observe that the natural isomorphism $\mpnh \isom \pgnoh$ in $(*)$ induces an isomorphism $$Z_{\out(\mpnh)}(\varphi_n(\sn)) \ \isom \ Z_{\out(\pgnoh)}(\chi_n(\sn)),$$ where $\chi_n: \sn \to \out(\pgnoh)$ is as in (\[diag41\]). But since $\iota_n(\sn)$ has trivial centralizer in $\frakS_{n+1}$, Theorem \[hmm\] implies $$\gt \ \isom \ Z_{\out(\pgnoh)}(\chi_n(\sn)).$$ It is easy to see that the composite of the above three displayed isomorphisms coincides with $\gt \to \out(\mbnh)$ of the assertion. This completes the proof of (i).
\(ii) Let $n\ge 5$. After Proposition \[out\] (ii), the argument goes in a similar (and simpler) way to the case (i) with applying Lemma \[dpg\] to the profinite completion of (\[exactGn\]): $$\begin{CD}
1 @>>> \widehat\Gamma_{0,n} @>>> \widehat\Gamma_{0,[n]} @>>> \sn @>>> 1.
\end{CD}$$ We leave the rest of detail to the reader.
Now, to prove Theorem A, let us follow an argument in [@dg] (Theorem 20) to look closely at the short exact sequence $$\label{Wells0}
\begin{CD}
1 @>>> \cnh @>>> \bnh @>>> \mbnh @>>> 1
\end{CD}$$ obtained as the profinite completion of (\[exactBn\]). Since $\cnh$ is characteristic in $\bnh$, this yields two natural homomorphisms $$\label{WellsPair}
\fp_0:\aut(\bnh)\to\aut(\cnh),
\quad
\fp_1:\aut(\bnh)\to\aut(\mbnh).$$ Recalling $\cnh=\la \zeta_n\ra\cong\zh$, we now [*canonically*]{} identify $\aut(\cnh)=\zh^\times$.
For $n> 1$, define the subgroup $Z_n\subset \zhat^\times$ by $$Z_n :=
\bigl(1+n(n-1)\zhat\bigr)^\times=
\ker\left(\zhat^{\times} \epi (\zhat/n(n-1)\zhat)^{\times}\right).$$
It is clear that each $\nu \in Z_n$ has a unique element $e \in \zhat$ such that $$\nu= 1 + n(n-1)e.$$ (But note that this form of $\nu$ is not always in $\zh^{\times}$ for arbitrary $e \in \zh$.)
The next key lemma enables us to identify $\ker(\fp_1)$ with $Z_n$:
\[lemma\_phi\_nu\] There is an isomorphism $$\phi: Z_n \isom \ker(\fp_1)\subset \aut(\bnh)$$ which assigns to every $\nu\in Z_n$ an automorphism $\phi_\nu\in \aut(\bnh)$ determined by $$\phi_\nu(\sigma_i)=\sigma_i\zeta_n^e
\qquad (\nu= 1 + n(n-1)e, \ i=1,\dots,n-1).$$
Given any $\nu \in Z_n$, let $e \in \zhat$ be the unique element with $\nu = 1 + n(n-1)e$. By using this $e \in \zh$, we define $\phi_{\nu} \in \Ker(\mathfrak{p}_1)$ as follows: First, set $\phi_\nu(\sigma_i):=\sigma_i\zeta_n^e$ for all $i=1,\dots,n-1$. Since $\zeta_n$ lies in the center of $\bnh$, it is easy to see that $\phi_\nu$ preserves the Artin’s braid relations. Therefore, $\phi_\nu$ extends to an endomorphism of $\bnh$ written by the same symbol $\phi_\nu$. One computes then $$\label{phinu_zeta_n}
\phi_\nu(\zeta_n)=
\phi_\nu((\sigma_1\cdots\sigma_{n-1})^n)
=(\sigma_1\cdots\sigma_{n-1})^n\cdot \zeta_n^{n(n-1)e}
=\zeta_n^{1+n(n-1)e}=\zeta_n^\nu.$$ From this we see that the image of $\phi_\nu$ contains $\zeta_n=(\zeta_n^\nu)^{\nu^{-1}}$ and hence does contain $\sigma_i=\phi_\nu(\sigma_i)\cdot\zeta_n^{-e}$ for $i=1,\dots,n-1$. This means the endomorphism $\phi_\nu:\bnh \to \bnh$ is a surjective homomorphism. As $\bnh$ is hopfian, we conclude $\phi_{\nu}\in \aut(\bnh)$ (cf.[@rz Proposition 2.5.2]). Write $\phi: Z_n \to \aut(\bnh)$ for the homomorphism defined by the above correspondence $\nu \mapsto \ \phi_{\nu}$. One verifies immediately that $\phi$ is injective and $Z_n \cong \mathrm{Im}(\phi) \subset \Ker(\mathfrak{p}_1)$. To see $\mathrm{Im}(\phi) = \Ker(\mathfrak{p}_1)$, pick any $\alpha\in\Ker(\fp_1)$ and set $\nu:=\fp_0(\alpha)\in\zh^\times$. Then, $\alpha(\zeta_n)=\zeta_n^\nu$ and there exist $e_i\in\zh$ ($i=1,\dots,n-1$) such that $\alpha(\sigma_i)=\sigma_i\, \zeta_n^{e_i}$. It is easy to see from the braid relation that all $e_i$ are the same constant $e \in \zh$. But then, since $\zeta_n=(\sigma_1\cdots\sigma_{n-1})^n$, we find $\nu=1+n(n-1)e$ which belongs to $Z_n$ and that $\alpha = \phi_{\nu}$.
Theorem A is obtained from Theorem \[gtmbnh\] (i) together with the last part of the following
\[ThA\_detail\] Let $n\ge 4$ be an integer.
1. There exists an exact sequence $$\label{aut-result}
\begin{CD}
1 @>>> Z_n @>\phi>> \aut(\bnh) @>\fp_1>> \aut(\mbnh) @>>> 1.
\end{CD}$$
2. $\inn(\bnh)\cap \phi(Z_n)=\{1\}$.
3. The exact sequence (i) provides a split central extension, i.e., $$\aut(\bnh) \cong \aut(\mbnh) \times Z_n,$$ and gives rise to $\out(\bnh) \cong \out(\mbnh) \times Z_n$.
\(i) It suffices to show $\fp_1$ is surjective. Note that $\inn(\bnh)$ is mapped onto $\inn(\mbnh)$. On the other hand, there is a well-known action $\iota_n:\gt \to \aut(\bnh)$ in the form $$\label{GTaction}
\begin{cases}
\sigma_1 & \mapsto \ \sigma_1^\lambda, \\
\sigma_i & \mapsto \ f(\sigma_i,\zeta_i)\sigma_i^\lambda f(\zeta_i,\sigma_i)
\quad (i=1,\dots, n-1)
\end{cases}$$ with $(\lambda,f)\in \zh^\times \times [\widehat F_2, \widehat F_2]$ the standard parameter for the elements of $\gt$ ([@Dr], [@I90], [@im]). Let $\bar\iota_n: \gt\to \aut(\mbnh)$ be the induced action. By virtue of Theorem \[gtmbnh\] (i), $\gt\cong \out(\mbnh)$, hence $\aut(\mbnh)=\bar\iota_n(\gt)\cdot \inn(\mbnh)$. From this follows that $\fp_1$ maps $\iota_n(\gt)\cdot \inn(\bnh) (\subset \aut(\bnh))$ onto $\aut(\mbnh)$.
\(ii) This is a consequence of Lemma \[dpg\] (ii) applied to (\[Wells0\]). Here is an alternative direct proof: Recall that the abelianization $\bnh^\ab$ of $\bnh$ is isomorphic to $\zh$. Each inner automorphism acts trivially on $\bnh^\ab$, while $\phi_\nu\in \phi(Z_n)$ ($\nu\in Z_n$) acts on it by $$(\sigma_i)^\ab
\mapsto (\sigma_i\cdot \zeta_n^e)^\ab=(\sigma_i^\ab)^{1+n(n-1)e}
\qquad(i=1,\dots,n-1)$$ which is nontrivial unless $e=0$. This concludes the assertion.
\(iii) It follows from (ii) that $\fp_1$ induces $\inn(\bnh)\isom \inn(\mbnh)$. Since $\iota_n(\gt)\isom \bar\iota_n(\gt)$, we find from Theorem \[gtmbnh\] (i) that $\fp_1$ restricts to the isomorphism $$\label{compl_factor}
\iota_n(\gt)\cdot\inn(\bnh)\isom
\aut(\mbnh) ,$$ i.e., $\iota_n(\gt)\cdot\inn(\bnh)$ gives a complementary factor of $\phi(Z_n)$ in $\aut(\bnh)$. To see that the exact sequence (i) gives a central extension, it suffices to show that both $\inn(\bnh)$ and $\iota_n(\gt)$ commutes with $\phi(Z_n)$. The commutativity of $\inn(\bnh)$ and $\phi(Z_n)$ follows immediately from the definition of $\phi_\nu$ ($\nu\in Z_n)$ in Lemma \[lemma\_phi\_nu\]. The commutativity of $\iota_n(\gt)$ and $\phi(Z_n)$ also follows from direct computation by using the above description of the $\gt$-action on $\bnh$. Indeed, given $(\lambda,f)\in\gt$, noting that $\zeta_n$ lies in the center of $\widehat{B}_n$, and $f$ lies in the commutator subgroup of $\widehat F_2$, we have $f(\sigma_i,\zeta_i)=f(\sigma_i\zeta_n^e,\zeta_i)$ ($i=1,\dots, n-1$). Since $(\lambda,f)\in\gt$ is known to act on $\zeta_n$ by $\zeta_n\mapsto \zeta_n^\lambda$ under the action (\[GTaction\]), one computes: $$\begin{aligned}
(\lambda,f)\circ\phi_\nu(\sigma_i) &=(\lambda,f)(\sigma_i\zeta_n^e)
=f(\sigma_i,\zeta_i)\sigma_i^\lambda f(\zeta_i,\sigma_i)
\zeta_n^{\lambda e},\\
&=f(\sigma_i,\zeta_i)(\sigma_i \zeta^e)^\lambda f(\zeta_i,\sigma_i)
=\phi_\nu(f(\sigma_i,\zeta_n)\sigma_i^\lambda f(\zeta_i,\sigma_i)) \\
&=\phi_\nu\circ(\lambda,f)(\sigma_i).\end{aligned}$$ for every $i=1,\dots, n-1$ (we understand $\zeta_1=1$ when $i=1$). Thus we settle the first assertion $\aut(\bnh) \cong \aut(\mbnh) \times Z_n$ after identifying $Z_n\cong \phi(Z_n)\subset \aut(\bnh)$ and $\aut(\mbnh)\cong \iota_n(\gt)\cdot\inn(\bnh)\subset \aut(\bnh)$ via (\[compl\_factor\]). The second assertion is then just a consequence of it.
In our above discussion for the proof of Theorem A, important roles have been played by the pair of two maps (\[WellsPair\]), which was motivated from the profinite Wells exact sequence (cf.[@n §1.5], [@JL]) associated to the short exact sequence (\[Wells0\]) in the form: $$\label{wells}
\begin{CD}
0 \longrightarrow \coz^1(\mbnh, \cnh) \longrightarrow \aut( \bnh, \cnh)
@>\fp>> \mathscr{C} @>q>> \co^2(\mbnh, \cnh).
\end{CD}$$ Since $\bnh$ in (\[Wells0\]) is a central extension and $\cB_n^\ab\cong\Z/n(n-1)\Z$, we easily see that $\coz^1(\mbnh, \cnh) = \{0\}$, $\aut(\bnh, \cnh) = \aut(\bnh)$, and find the group of “compatible pairs” $\mathscr{C}$ to be $\aut(\mbnh) \times \aut(\cnh).$ Thus, the exact sequence (\[wells\]) is reduced to $$\label{reduced-wells}
\begin{CD}
0 @>>> \aut(\bnh) @>\fp=(\fp_1,\fp_0)>> \aut(\mbnh) \times \aut(\cnh) @>q>> \co^2(\mbnh, \cnh),
\end{CD}$$ where $q$ is called the Wells pointed map (generally not a homomorphism).
The above sequence (\[reduced-wells\]) is simply useful, for example, to see that [*the exact sequence of*]{} Theorem \[ThA\_detail\] (i) [*provides a central extension*]{}, reproving the core part of Theorem \[ThA\_detail\] (iii) without use of the explicit $\gt$-action (\[GTaction\]): Indeed, according to (\[phinu\_zeta\_n\]), the image $\fp(\phi_\nu)=(\fp_1(\phi_\nu),\fp_0(\phi_\nu))=(id, \nu)$ for every $\nu\in Z_n$ is easily seen to lie in the center of $\aut(\mbnh) \times \aut(\cnh)$. Besides this simple observation, it is a natural question to measure the size of the image of $\aut(\bnh)$ by the injection $\fp=(\fp_1,\fp_0)$ into $\aut(\mbnh) \times \aut(\cnh)$. Now, recalling $\gt\subset\{(\lambda,f)\in\zhat^\times \times \widehat F_2\}$, $\aut(\mbnh) = \gt \cdot \inn(\cB_n)$ and $\aut(\cnh) = \zh^\times$, we define two characters $$\lambda:\aut(\mbnh)\to \zh^\times
\quad\text{and}\quad
\nu:\aut(\cnh)\to \zh^\times
$$ in the obvious way. One finds:
Notations being as above, we have $$\im(\fp)=\{(\alpha,\beta)\in \mathscr{C}\mid
\lambda(\alpha)\equiv \nu(\beta) \mod n(n-1)
\}.$$ In particular, $\mathscr{C}/\im(\fp)\cong (\Z/n(n-1)\Z)^\times$.
Let $Z_n\subset\aut(\cnh)=\zh^\times$ be as above, and define $A_n\subset\aut(\cB_n)$ to be $\lambda^{-1}(Z_n)$. It is not difficult to see $A_n\times Z_n\subset\im(\fp)$. The assertion is derived from the observation that the image of $\im(\fp)$ in the quotient group $\mathscr{C}/(A_n\times Z_n)\cong
(\Z/n(n-1)\Z)^\times\times(\Z/n(n-1)\Z)^\times$ forms the diagonal subgroup. This follows from the well-known fact that the restriction of the action of $(\lambda,f)\in\gt$ on $\bnh$ to $\cnh = \la\zeta_n\ra$ is given by $\zeta_n\mapsto \zeta_n^\lambda$, which completes the proof.
Before closing the paper, let us add some remark on the Wells map $q:\mathscr{C} \to \co^2(\mbnh, \cnh)$. Let $[\mu] \in \co^2(\mbnh, \cnh)$ be the class of factor sets associated to the central extension (\[Wells0\]). For each pair $(\alpha,\nu)\in\mathscr{C}=
\aut(\mbnh) \times \aut(\cnh)$, we denote by $[\mu]^{(\alpha,\nu)}\in \co^2(\mbnh, \cnh)$ the class of a central extension obtained by twisting (\[Wells0\]) by $(\alpha, \nu)$. Then, one finds: $$q(\alpha,\nu)=[\mu]-[\mu]^{(\alpha,\nu)}.$$ This means that $\im(\fp)\subset\mathscr{C}$ can be characterized as the stabilizer of the twisting action of $\mathscr{C}$ on $[\mu]$. Concerning the precise position and size of $[\mu] \in \co^2(\mbnh, \cnh)$, we remark the following
Let $n\ge 4$. The cohomology group $\co^2(\mbnh, \cnh) $ is isomorphic to $\Z/n(n-1)\Z$, and is generated by the class $[\mu]$.
According to V.Arnold [@A68], $H^2(B_n,\Z)=\{0\}$ and $H^3(B_n,\Z)=\Z/2\Z$. Applying this to the long exact sequence associated with $0\to\Z\to\Z\to\Z/r\Z\to 0$ ($r\in \N$), we obtain $H^2(B_n, \Z/r\Z)\cong \{0\}$, $\cong\Z/2\Z$ according to whether $r$ is odd or even respectively. For a positive integer $N$, (part of) the five term exact sequence for the central extension $1\to C_n\to B_n\to \mathcal{B}_n\to 1$ reads $$\begin{aligned}
\label{5term0}
&H^1(B_n,\Z/N\Z)
\overset{\mathrm{res}_N}{\longrightarrow}
H^1(C_n,\Z/N\Z)
\overset{\mathrm{tg}_N}{\longrightarrow}
H^2(\mathcal{B}_n,\Z/N\Z)
\\
&\overset{\mathrm{inf}_N}{\longrightarrow}
H^2(B_n, \Z/N\Z),
\notag
$$ where $\mathrm{res}_N$, $\mathrm{tg}_N$ and $\mathrm{inf}_N$ are respectively the restriction, transgression and inflation maps. Suppose first $N$ is a positive integer divisible by $n(n-1)$. Then, (\[5term0\]) yields the exact sequence $$\label{5term}
\begin{CD}
0 \to \Z/n(n-1)\Z
@>{{\overline{\mathrm{tg}}_N}}>>
H^2(\mathcal{B}_n,\Z/N\Z)
@>{\mathrm{inf}_N}>>
H^2(B_n, \Z/N\Z) \ (\cong \Z/2\Z),
\end{CD}$$ where $\Z/n(n-1)\Z$ is regarded as the cokernel of the restriction $\mathrm{res}_N: \mathrm{Hom}(B_n,\Z/N\Z)\to \mathrm{Hom}(C_n,\Z/N\Z)$ followed by the factorization ${\overline{\mathrm{tg}}_N}$ of transgression ${\mathrm{tg}}_N$. Let us vary $N$ multiplicatively. The goodness of $\mathcal{B}_n$ (in the sense of Serre) together with [@NSW Corollary 2.7.6] allows us to interpret $\co^2(\mbnh, \cnh)
=\varprojlim_N H^2(\mathcal{B}_n,\Z/N\Z)$ after identification $C_n=\zeta_n^\Z\cong \Z$ with trivial (conjugate) action of $\mathcal{B}_n$. The term $\mathrm{coker}(\mathrm{res}_N)\cong \Z/n(n-1)\Z$ in (\[5term\]) is constant in the projective system along $N\in \N$ divisible by $n(n-1)$. On the other hand, we have (\#): $\varprojlim_N H^2(B_n,\Z/N\Z)=\{0\}.$ In fact, since $B_n^\ab\cong\Z$, in the long exact sequence associated with $0\to \Z/2\Z\to \Z/2N\Z\to \Z/N\Z\to 0$, we find that $H^1(B_n,\Z/2N\Z)\to H^1(B_n,\Z/N\Z)$ is surjective, hence that the former arrow in $H^2(B_n,\Z/2\Z)\to H^2(B_n,\Z/2N\Z)\to H^2(B_n,\Z/N\Z)$ gives an isomorphism between groups of order two so that the latter arrow is 0-map. This settles (\#) which concludes the first assertion $\co^2(\mbnh, \cnh) \cong\Z/n(n-1)\Z$.
It remains to show that the class $[\mu]$ has order $n(n-1)$ in $\co^2(\mbnh, \cnh) $. For an integer $d>0$, let $[\mu_d]\in H^2(\mathcal{B}_n,\Z/d\Z)$ be the class of factor sets corresponding to the central extension $$\label{extension_d}
1\to C_n/C_n^d(\cong \Z/d\Z)\to B_n/C_n^d\to \mathcal{B}_n\to 1.$$ It is known that $[\mu_d]$ is the transgression image of the projection $\mathrm{pr}_d:C_n\to C_n/C_n^d$ regarded as an element of $H^1(C_n,C_n/C_n^d)$, i.e., $$\label{suzuki-id}
[\mu_d]=\mathrm{tg}_d(\mathrm{pr}_d)\in
\im(\mathrm{tg}_d)\subset H^2(\mathcal{B}_n,\Z/d\Z),$$ where $C_n/C_n^d\isom \Z/d\Z$ is given by $\zeta_n\mapsto 1$ (cf.e.g.,[@Sz Chap.2 §9(9.4)]). Let us observe that the extension (\[extension\_d\]) splits if and only if $n(n-1)\in (\Z/d\Z)^\times$. In fact, a system of lifts of the generators $\bar\sigma_i\in \mathcal{B}_n$ ($i=1,\dots,n-1$) can be written in the form of images of $\sigma_i \zeta_i^{a_i} \in B_n$ in $B_n/C_n^d$ ($a_i\in\Z$). It is easy to see that they satisfy the braid relations modulo $C_n^d$ if and only if $a_1\equiv \cdots\equiv a_{n-1}$ and $1+n \sum_i a_i\equiv 0$ in $\Z/d\Z$ (cf.(\[phinu\_zeta\_n\])). This condition to be held by a collection $\{a_i\}_i$ is equivalent to $n(n-1)\in (\Z/d\Z)^\times$ as desired. Let $p$ be a prime dividing $n(n-1)$ and consider $[\mu_p]\in H^2(\mathcal{B}_n,\Z/p\Z)$. It follows from the above observation that $[\mu_p]\ne 0$. Since the restriction map $\mathrm{res}_p: H^1(B_n,\Z/p\Z)
\to H^1(C_n,\Z/p\Z)$ is trivial under the assumption $p \mid n(n-1)$, the transgression $\mathrm{tg_p}$ injects $H^1(C_n,\Z/p\Z)\cong\Z/p\Z$ into $H^2(\mathcal{B}_n,\Z/p\Z)$ whose image is generated by $[\mu_p]\ne 0$. But for any multiple $N$ of $n(n-1)$, the class $[\mu_N]\in \im(\mathrm{tg}_N) (\cong \Z/n(n-1)\Z)
\subset H^2(\mathcal{B}_n,C_n/C_n^N)$ is mapped to $[\mu_p]\in H^2(\mathcal{B}_n,C_n/C_n^p)$ via the reduction of central extensions induced from the surjective homomorphism $B_n/C_n^N\epi B_n/C_n^p$ in virtue of (\[suzuki-id\]). In particular, the reduction map $\im(\mathrm{tg}_N)\to \im(\mathrm{tg}_p)$ is given simply by the mod $p$ surjection between the cyclic groups: $$\label{diag-suzuki}
\begin{matrix}
\quad [\mu_N]& \in\im(\mathrm{tg}_N) & (\cong\Z/n(n-1)\Z)
&\subset H^2(\mathcal{B}_n,C_n/C_n^N) &\qquad
\\
\quad\downarrow & & \qquad\downarrow \mathrm{ mod}\ p \quad &
\downarrow &
\\
0\ne [\mu_p] &\in\im(\mathrm{tg}_p) & (\cong\Z/p\Z) \quad
&\subset H^2(\mathcal{B}_n,C_n/C_n^p). &
\end{matrix}$$ Since the class $[\mu]\in \co^2(\mbnh, \cnh)$ is the common limit of those $[\mu_N]$, it follows that $[\mu]$ generates the $p$-primary component of the cyclic group $\co^2(\mbnh, \cnh)
\cong \Z/n(n-1)\Z$ for every prime $p\mid n(n-1)$, hence gives a generator of it.
[ZZ99]{}
H. Akiyoshi, M. Sakuma, M. Wada, and Y. Yamashita, [*Punctured Torus Groups and 2-Bridge Knot Groups (I)*]{}, [*Lecture Notes in Mathematics [**1909**]{}*]{}, Springer (2007).
V. Arnold, [*Braids of algebraic functions and the cohomology of swallowtails*]{}, Translation of Usp. Mat. Nauk [**23**]{} (1968), 247–248;. [*English Translation*]{} in Collected Works, Volume II (2014), pp.171–173.
E. Artin, [*Braids and permutations*]{}, Ann. Math. [**48**]{} (1947), pp. 643-–649.
B. Collas, [*Action of the Grothendieck-Teichmüller group on torsion elements of full Teichmüller group in genus zero*]{}, J. Théorie Nombres Bordeaux, [**24**]{} (2012), 605–622.
V. G. Drinfeld, [*On quasitriangular quasi-Hopf algebras and on a group that is closely connected with $\Gal(\bQ/\Q)$*]{}, Algebra i Analiz [**2**]{} (1990); English translation Leningrad Math. J. [**2**]{} (1991), 829–860
J. L. Dyer and E. K. Grossman, The Automorphism Groups of the Braid Groups, [*Amer. J. Math*]{}. [**103**]{} (1981), pp. 1151–1169.
B. Farb and D. Margalit, [*A primer on mapping class groups*]{}, [**49**]{} [*Princeton Mathematical Series.*]{} Princeton University Press (2012).
D. Harbater and L. Schneps, [*Fundamental Groups of Moduli and the Grothendieck-Teichmüller Group*]{}, Transactions of the American Mathematical Society, [**352**]{} (2000), 3117–3148.
W. Herfort and L. Ribes, [*Torsion elements and centralizers in free products of profinite groups*]{}, J. reine angew. math. [**358**]{} (1985), 155-161.
Y. Hoshi, A. Minamide, and S. Mochizuki, [*Group-theoreticity of numerical invariants and distinguished subgroups of configuration space groups*]{}, RIMS Preprint [**1870**]{} (March 2017).
Y. Ihara, [*Braids, Galois groups, and some arithmetic functions*]{}, Proc. Intern. Congress of Math. Kyoto 1990, 99–120.
Y. Ihara, [*Some details on the $\gt$-action on $\hat B_n$*]{}, Appendix to : Y. Ihara and M. Matsumoto, [*On Galois Actions on Profinite Completions of Braid Groups*]{}, in [“Recent Developments in the Inverse Galois Problem”]{}, [*Contemp. Math*]{}. [**186**]{}, AMS (1995), pp.173–200.
P. Jin and H. Liu, [*The Wells exact sequence for the automorphism group of a group extension*]{}, J. Algebra, [**324**]{} (2010), 1219–1228.
M. Korkmaz, [*Automorphisms of complexes of curves on punctured spheres and on punctured tori*]{}, Topology Appl., [**95**]{} (1999), 85–111.
C. Kassel and V. Turaev, [*Braid Groups*]{}, [*Graduate Texts in Mathematics*]{} [**247**]{}, Springer-Verlag (2008).
Y. Komori and T. Sugawa, [*Bers embedding of the Teichmüller space of a once-punctured torus*]{}, Conformal geometry and dynamics, [**8**]{} (2004), 115–142.
A. Lubotzky and L. van den Dries, [*Subgroups of free profinite groups and large subfields of $\tilde{\Q}$*]{}, Israel J. Math., [**39**]{} (1981), 25–45.
V. Lin, [*Braids and Permutations*]{}, [arXiv:math/0404528 \[math.GR\]]{}, Arxiv e-prints (April 2004).
P. Lochak and L. Schneps, [*The Grothendieck-Teichmüller group and automorphisms of braid groups*]{}, London Math. Soc. Lecture Note Ser., [**200**]{} (1994), 323–358.
P. Lochak and L. Schneps, [*Open problems in Grothendieck-Teichmüller theory*]{}, in “Problems on mapping class groups and related topics”, Proc. Sympos. Pure Math., [**74**]{}, 165–186.
H. Nakamura, Galois rigidity of pure sphere braid groups and profinite calculus, [*J. Math. Sci. Univ. Tokyo*]{} [**1**]{} (1994), pp. 71–136.
H. Nakamura, [*On arithmetic monodromy representations of Eisenstein type in fundamental groups of once punctured elliptic curves*]{}, Publ. RIMS, Kyoto University, [**49**]{} (2013), 413–496.
J.Neukirch, A.Schmidt, K.Wingberg, [*Cohomology of Number Fields*]{}, 2nd Edition, Springer-Verlag 2013.
L. Ribes and P. Zalesskii, [*Profinite Groups (Second Edition)*]{}, [*Ergebnisse der Mathematik und ihrer Grenzgebiete 3. Folge, A Series of Modern Surveys in Mathematics*]{} [**40**]{}, Springer-Verlag 2010.
M. Suzuki, [*Group Theory I*]{}, Springer-Verlag 1982.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The general constraints on the parameter space of soft-breaking terms, in order to avoid dangerous charge and color breaking minima, are applied to the four-dimensional string scenario where the dilaton is the source of supersymmetry breaking (dilaton-dominated limit). The results indicate that the whole parameter space is excluded on these grounds after imposing the present experimental data on the top mass. The inclusion of a non-vanishing cosmological constant does not improve essentially the prospects. Possible way-outs to this situation are briefly discussed.'
author:
- |
[**J.A. Casas[^1] ${}^{ {\footnotesize,\S}}$**]{}, [**A. Lleyda${}^{ {\footnotesize\P}}$** ]{} and [**C. Muñoz${}^{\footnotesize\P}$**]{}\
\
${}^{\footnotesize\S}$ [Santa Cruz Institute for Particle Physics]{}\
[University of California, Santa Cruz, CA 95064, USA]{}\
[[email protected]]{}\
\
${}^{\footnotesize\P}$ [Departamento de Física Teórica C–XI]{}\
[Universidad Autónoma de Madrid, 28049 Madrid, Spain]{}\
[[email protected] $$ [email protected]]{}
title: '[**Problems for Supersymmetry Breaking by the Dilaton in Strings from Charge and Color Breaking**]{} [^2] '
---
4.5mm =-0.5cm 4.5mm
‘@=11 stequation ‘@=12
-23.cm
Introduction
============
The presence of scalar fields with color and electric charge in supersymmetric (SUSY) theories induces the possible existence of dangerous charge and color breaking minima, which would make the standard vacuum unstable \[1-8\]. This is not necessarily a shortcoming since many SUSY models can be discarded on these grounds, thus improving the predictive power of the theory. A complete analysis of all the potentially dangerous directions in the field space of the minimal supersymmetric standard model (MSSM) was carried out in ref.[@CCB]. It was shown there that the corresponding constraints on the soft parameter space ($m$, $M$, $A$, $B$) are very strong. As a matter of fact, there are extensive regions of this space that [*become forbidden*]{} producing important bounds, not only on the value of the trilinear scalar term ($A$), but also on the values of the bilinear scalar term ($B$) and the scalar and gaugino masses ($m, M$ respectively).
On the other hand, in four-dimensional strings, working at the perturbative level, it is possible to obtain information about the structure of soft SUSY-breaking terms \[9-14\]. The basic idea is to identify some chiral fields whose auxiliary components could break SUSY by acquiring a vacuum expectation value (VEV). This is the case of the dilaton and the moduli fields. The important point in this assumption of locating the seed of SUSY breaking in the dilaton/moduli sector, is that it leads to some interesting relationships among different soft terms which could perhaps be experimentally tested. This general analysis was applied in particular to the gaugino condensation scenario in ref.[@gaugino], whereas in refs.\[10-14\] no special assumption was made about the possible origin of SUSY breaking.
The dilaton-dominated limit [@Kaplunovsky; @Brignole], where only the dilaton field, $S$, contributes to SUSY breaking[^3] is specially interesting. The dilaton field, whose VEV determines the tree-level gauge coupling, is present in any four-dimensional string and couples at tree-level in a universal manner to all particles. Therefore, this limit is [*model independent*]{} and, as a consequence, the soft terms are independent of the four-dimensional string considered. Besides, their expressions are quite simple since they are universal and essentially depend on only two parameters, $m_{3/2}$ and $B$, where $m_{3/2}$ is the gravitino mass. Actually, universality is a desirable property, not only to reduce the number of independent parameters in the MSSM, but also for phenomenological reasons, particularly to avoid flavour changing neutral currents. Because of the simplicity of this scenario, the corresponding low-energy predictions are quite precise [@Barbieri; @Brignole; @Vissani]. For example, the first and second generation squarks are almost degenerate with the gluino and are much heavier than sleptons[^4].
From all the above reasons it is clearly of the utmost importance to study the consistency of the dilaton-dominated scenario with the possible existence of dangerous charge and color breaking minima. This is the aim of this paper. In fact, we will show that charge and color breaking constraints are so important that the whole parameter space is [*forbidden*]{} and, as a consequence, the dilaton-dominated limit is excluded on these grounds.
Basic Ingredients
=================
Let us briefly review the basic ingredients required for this analysis. First we will concentrate on the form of soft breaking terms. The general form of the soft SUSY-breaking Lagrangian in the context of the MSSM is given by \[lsoft\] [L]{}\_[soft]{}= \_[a=1]{}\^[3]{} M\_a \_a \_a - \_[i]{} m\_i\^2 |\_i|\^2 -(A\_[ijk]{} W\_[ijk]{} + B H\_1 H\_2 + [h.c.]{}), where $W_{ijk}$ are the usual terms of the Yukawa superpotential of the MSSM with $i$=$Q_L, u_L^c, d_L^c, L_L, e_L^c, H_1, H_2,$ and $\phi_i,\lambda_a$ are the canonically normalized scalar and gaugino fields respectively. In the dilaton-dominated scenario [@Kaplunovsky; @Brignole], neglecting string loop corrections, one obtains the following results for the above scalar masses, gaugino masses and soft trilinear terms \[softtermsV\] m\_i\^2 &=& m\_[3/2]{}\^2 +V\_0 ,\
M\_a &=& e\^[-i ]{} ,\
A\_[ijk]{} &=& - e\^[-i ]{} , where, for the sake of completeness, we have included the VEV of the scalar potential (i.e. the cosmological constant) $V_0$, and a possible phase $\alpha$ of the dilaton F-term [@Brignole]. Notice that we are using the standard supergravity mass units where $M_{Planck}/{\sqrt{8 \pi}}$=$1$.
The value of the bilinear term $B$ is more model dependent and deserves some additional comments. Indeed, $B$ depends not only on the dilaton-dominance assumption but also on the particular mechanism which could generate the associated (electroweak size) $\mu$ term [@Munoz1]. For example, the interesting possibilities of generating it through the Kähler potential [@Giudice; @Casas; @Kaplunovsky; @Lopes; @Antoniadis] or the superpotential [@Casas; @Antoniadis] give rise, in the dilaton-dominated limit, to the following value of $B$ [@Brignole; @Kaplunovsky; @Munoz2; @Munoz1; @Brignole2]: \[BV\] B = 2 m\_[3/2]{} + . The previous expressions for the soft terms can be simplified taking into account several experimental restrictions. From the limits on the electric dipole moment of the neutron it seems reasonable to impose in what follows $\alpha$=$0$ mod $\pi$. On the other hand, experimental constraints in present cosmology allow us to assume vanishing cosmological constant $V_0$=$0$ (we will see later on that our conclusions will not be modified if we give up this assumption). Then \[softterms\] m\_i\^2 &=& m\_[3/2]{}\^2 ,\
M\_a &=& m\_[3/2]{} ,\
A\_[ijk]{} &=& - M\_a , and the $B$ term associated with the mechanisms explained above in order to solve the $\mu$ problem is[^5] \[B\] B = 2 m\_[3/2]{} . Expressions (\[softtermsV\]–\[B\]) are to be understood at the string scale $M_{string}\simeq 0.5\times g_{string}\times{10}^{18}$ GeV [@Kaplu], where $g_{string}=(Re\; S)^{-1/2}\simeq 0.7$. In the following we will assume that the discrepancy between the unification scale of the gauge couplings, $M_X$, and the string unification scale, $M_{string}$, can be explained by the effect of string threshold corrections [@Ross].
In the present paper we have taken the expressions of the soft terms given by eqs.(\[softterms\], \[B\]) as our starting point to work, considering the value of $B$ given by eq.(\[B\]) as a guiding example. Then, by varying the value of $B$ (and also $V_0$) we will eventually obtain the most general results. Concerning the value of the $\mu$ parameter, we will fix it as usual from the requirement of correct electroweak breaking[^6].
The second basic ingredient of our analysis concerns the constraints associated with the existence of dangerous directions in the field space. As was mentioned in the introduction, a complete analysis of this issue, including in a proper way the radiative corrections to the scalar potential, was carried out in ref.[@CCB]. The most relevant results obtained there for our present task are the following.
There are two types of constraints: the ones arising from directions in the field-space along which the (tree-level) potential can become unbounded from below (UFB), and those arising from the existence of charge and color breaking (CCB) minima in the potential deeper than the standard minimum.
Concerning the UFB directions (and corresponding constraints), there are three of them, labelled as UFB-1, UFB-2, UFB-3 in [@CCB]. It is worth mentioning here that in general the unboundedness is only true at tree-level since radiative corrections eventually raise the potential for large enough values of the fields, but still these minima can be deeper than the realistic one (i.e. the SUSY standard model vacuum) and thus dangerous. The UFB-3 direction, which involves the fields $\{H_2,\nu_{L_i},e_{L_j},e_{R_j}\}$ with $i \neq j$ and thus leads also to electric charge breaking, yields the [*strongest*]{} bound among [*all*]{} the UFB and CCB constraints. The explicit form of this bound is as follows. By simple analytical minimization it is possible to write the value of all the relevant fields along the UFB-3 direction in terms of the $H_2$ one. Then, for any value of $|H_2|<M_{string}$ satisfying \[SU6\] |H\_2| > - , the value of the potential along the UFB-3 direction is simply given by \[SU8\] V\_[UFB-3]{}=(m\_2\^2 -\^2+ m\_[L\_i]{}\^2 )|H\_2|\^2 + ( m\_[L\_j]{}\^2+m\_[e\_j]{}\^2+m\_[L\_i]{}\^2 ) |H\_2| - . Otherwise \[SU9\] V\_[UFB-3]{}= (m\_2\^2 -\^2 ) |H\_2|\^2 + ( m\_[L\_j]{}\^2+m\_[e\_j]{}\^2 ) |H\_2| + (g’\^2+g\_2\^2)\^2 . In eqs.(\[SU8\],\[SU9\]) $\lambda_{e_j}$ is the leptonic Yukawa coupling of the $j-$generation and $m_2^2$ is the sum of the $H_2$ squared soft mass, $m_{H_2}^2$, plus $\mu^2$. Then, the UFB-3 condition reads \[SU7\] V\_[UFB-3]{}(Q=Q) > V\_[real min]{} , where $V_{\rm real \; min}=-\frac{1}{8}\left(g'^2 + g_2^2\right)
\left(v_2^2-v_1^2\right)^2$, with $v_{1,2}$ the VEVs of the Higgses $H_{1,2}$, is the realistic minimum evaluated at $M_S$ (see below) and the $\hat Q$ scale is given by $\hat Q\sim {\rm Max}(g_2 |e|, \lambda_{top} |H_2|,
g_2 |H_2|, g_2 |L_i|, M_S)$ with $|e|$=$\sqrt{\frac{|\mu|}{\lambda_{e_j}}|H_2|}$ and $|L_i|^2$=$-\frac{4m_{L_i}^2}{g'^2+g_2^2}$ +($|H_2|^2$+$|e|^2$). Finally, $M_S$ is the typical scale of SUSY masses (normally a good choice for $M_S$ is an average of the stop masses, for more details see refs.[@Gamberini; @bea; @CCB]). Notice from (\[SU8\]-\[SU9\]) that the negative contribution to $V_{UFB-3}$ is essentially given by the $m_2^2-\mu^2$ term, which can be very sizeable in many instances. On the other hand, the positive contribution is dominated by the term $\propto 1/\lambda_{e_j}$, thus the larger $\lambda_{e_j}$ the more restrictive the constraint becomes. Consequently, the optimum choice of the $e$–type slepton is the third generation one, i.e. ${e_j}=$ stau.
Concerning the CCB constraints, let us mention that the “traditional” CCB bounds [@Frere], when correctly evaluated (i.e. including the radiative corrections in a proper way), turn out to be extremely weak. However, the “improved" set of analytic constraints obtained in ref.[@CCB], which represent the necessary and sufficient conditions to avoid dangerous CCB minima, is much stronger. It is not possible to give here an account of the explicit form of the CCB constraints used in the present paper. This can be found in section 5 of ref.[@CCB], to which we refer the interested reader.
Results
=======
In Fig.1 we have presented in detail the interesting case (and guiding example) $B$=$2m_{3/2}$ for the two possible values of gaugino masses, $M \equiv M_a$=$\; \pm \sqrt{3} m_{3/2}$, see eqs.(\[softterms\], \[B\]). In the plots of the figure we have shown how the parameter space defined by $m_{3/2}$ (the only soft parameter left independent) and $M_{top}$ (which we let vary for completeness) is totally excluded by the different constraints in the game. Some comments are in order here.
First, we have taken $m_{3/2}\leq 500$ GeV since larger values of $m_{3/2}$ would induce too large SUSY-mass spectra; e.g. $m_{3/2}$=500 GeV implies gluino and first and second generation squark masses of order 2.5 TeV, conflicting the absence-of-fine-tuning requirements [@bar; @bea]. On the other hand, rather than assuming a particular value of the top mass, we have preferred to vary the physical (pole) top mass, $M_{top}^{phys}$, between 150 and 200 GeV. Actually, it is not always possible to choose the boundary condition of the top Yukawa coupling $\lambda_{top}$ so that the physical (pole) mass is reproduced because the renormalization group (RG) infrared fixed point of $\lambda_{top}$ puts an upper bound on the running top mass $M_{top}$, namely $M_{top}\simlt 197 sin\beta$ GeV [@Inoue], where $tan \beta$=$ v_2/v_1$. The corresponding restriction in the parameter space (black regions in Fig.1) is certainly substantial in the case $M$=$\sqrt{3} m_{3/2}$, yielding $M_{top}^{phys}<167$ GeV, which is by itself a remarkable result[^7]. The region excluded by the CCB bounds is denoted by circles in the figure. The sign of the trilinear soft term $A_{ijk}$ is important in these type of constraints as can be seen in the figure (recall that $A\equiv A_{ijk}$=$-M$). Whereas the case $M$=$-\sqrt{3}m_{3/2}$ is not constrained at all, in the case $M$=$\sqrt{3}m_{3/2}$ the whole parameter space left allowed by the previous “top-fixed-point constraint” is excluded by the CCB bounds.
Anyway, it is apparent from Fig.1 that the the restrictions coming from the UFB constraints (small filled squares) are very strong in both cases. Most of the parameter space is in fact excluded by the UFB-3 constraint, which has been explained in the previous section.
Finally, we have also plotted in Fig.1 the region excluded by the experimental bounds on SUSY particle masses (filled diamonds). Conservatively enough, we have imposed \[experimentalb\] & &M\_[g]{} 120 [GeV]{} , M\_[\^]{}45 [GeV]{} , M\_[\^o]{} 18 [GeV]{} ,\
& &M\_[q]{}100 [GeV]{} , M\_[t]{} 45 [GeV]{}, M\_[l]{} 45 [GeV]{} , in an obvious notation. The corresponding forbidden area comes mainly from too small masses for neutralinos and charginos in the case $M$=$\sqrt{3} m_{3/2}$, and for the sleptons when $M$=$-\sqrt{3} m_{3/2}$. The ants indicate regions which are excluded by negative squared mass eigenvalues, in this case the sneutrinos.
Notice from Fig.1 that there are areas that are simultaneously constrained by different types of bounds.
At the end of the day, the allowed region (white) left is very small. Only in the case $M$=$-\sqrt{3}m_{3/2}$ and for $m_{3/2}$ larger than 320 GeV, which on the other hand corresponds to gluino and first and second generation squark masses heavier than 1.5 TeV, the dangerous minima are not present. However, this occurs for $M_{top}^{phys} <$ 157 GeV. Thus, using the present experimental data on the top mass, $M_{top}^{exp}$=$180 \pm 12$ GeV [@PDG], we conclude that the dilaton dominance limit for SUSY breaking in strings with $B$=$2m_{3/2}$ is excluded on charge and color breaking (CCB and UFB) grounds. It is worth mentioning here that, even if we relax the previous fine-tuning requirement (i.e. $m_{3/2}\leq 500$ GeV) by admitting higher values of $m_{3/2}$, we have checked that the growing of the allowed region is anyway remarkably slow, so one has to go to extremely high values of $m_{3/2}$ to get acceptable top masses.
Let us generalize now the previous analysis by varying the value of $B$. In Fig.2 we have shown the representative examples $B$=$0, 3m_{3/2}$ with $M$=$-\sqrt{3}m_{3/2}$. Whereas for large values of $B$, $B\geq 3m_{3/2}$, the whole parameter space is excluded, for $B$=$0$ there is still a very small allowed region. However, this region is in fact excluded by the value of $M_{top}^{exp}$ in a similar fashion as it happened in the $B=2m_{3/2}$ case (see above). Intermediate values of $B$ do not improve the situation. For negative values of $B$ the corresponding figures are the same, since they are invariant under the transformation $B$, $A$, $M\rightarrow$ -$B$, -$A$, -$M$. The same conclusion is obtained for the case $M$=$\sqrt{3}m_{3/2}$. Examples of this scenario are given in Fig.1 for $B$=$2m_{3/2}$ and in Fig.2 for $B$=$0$ (notice that from the previous invariance the $B$=$0$ figure is valid for $M$=$-\sqrt{3}m_{3/2}$ as well as for $M$=$\sqrt{3}m_{3/2}$).
From the various figures it is clear that the CCB and UFB constraints do not allow the possibility of SUSY breaking in the dilaton-dominance limit in strings.
Let us now consider for the sake of completeness the possibility of a non-vanishing cosmological constant $V_0$. This may correspond to different attitudes concerning this problem. For example, one might think that the experimentally constrained cosmological term in present cosmology is not directly connected to the particle physics vacuum energy $V_0$. Another possibility is to admit a non-vanishing tree-level cosmological constant, requiring a vanishing fully renormalized one [@Choi]. Anyway, whatever the assumption is, it is worth relaxing the condition $V_0$=$0$ since, from eq.(\[softtermsV\]), this might be the only possibility to avoid the previous dramatical conclusions for the dilaton-dominated limit. Notice, however, that $V_0$ is constrained to be $V_0\ge -m_{3/2}^2$ in order to avoid negative squared scalar masses (see eq.(\[softtermsV\])) and, on the other hand, it should not be too large in order to keep the SUSY spectrum in the range of 1 TeV. In addition, to simplify the analysis we have initially taken the (theoretically well-motivated) value of $B$ given in eq.(\[BV\]).
In Fig.3 we have presented three representative examples of a non-vanishing cosmological constant ($V_0$=$-0.5 m_{3/2}^{2},\; 3 m_{3/2}^{2},
\;10 m_{3/2}^{2}$) with $M$=$-\sqrt{3}m_{3/2}$. We see that for $V_0$=$-0.5m_{3/2}^2$ the whole parameter space is excluded. Larger values of $V_0$ do improve the situation, the best case being $V_0$=$3m_{3/2}^2$. However, once more, the present value of $M_{top}^{exp}$ essentially excludes this possibility for any reasonable SUSY spectrum. More precisely, notice that for $m_{3/2}\geq 300$ GeV we are already in the range of 2 TeV for the gluino and first and second generation squark masses, which are far too large from fine-tuning considerations. Finally, for the case $M$=$\sqrt{3}m_{3/2}$ (not represented in the figures) the constraints are even stronger, e.g. for $V_0$=$3m_{3/2}^2$ the whole parameter space is excluded.
Even if we let $B$ vary as an independent parameter (at the same time as $V_0$) the results do not improve much. More precisely, the best case is found for $V_0$=$3m_{3/2}^2$, $B$=$3m_{3/2}$. But, even in this extreme possibility, if one demands a reasonable SUSY spectrum (e.g. masses of order 1 TeV, which in this case requires $m_{3/2}\sim $ 150 GeV) the top mass becomes too small ($M_{top}^{phys}\leq$ 172 GeV), almost inconsistent with the present experimental lower bound ($M_{top}^{exp}\geq$ 168 GeV).
To summarize the results, the dilaton-dominated limit is essentially excluded on charge and color breaking (CCB and UFB) grounds. Even in the few extreme cases consistent with the charge and color breaking constraints, the spectrum is inviable since either the top mass is too small or the SUSY spectrum is far too heavy from any fine-tuning consideration. The addition of a non-vanishing cosmological constant in the game does not improve this situation.
Concluding Remarks
==================
The dilaton-dominated limit in strings, defined as the scenario where the only source of SUSY breaking is the dilaton, is highly interesting since at the string tree-level approximation its formulation is model independent (i.e. holds for any four-dimensional string) and yields quite precise low-energy predictions (SUSY spectrum). However, we have seen along this paper that, after imposing the present experimental data on the top mass, the whole parameter space of this scenario ($m_{3/2}, B$) is excluded on charge and color breaking grounds [@CCB], i.e. by the existence of charge and color breaking minima deeper than the standard vacuum. Even allowing a non-vanishing cosmological constant does not improve essentially the situation. Due to the attractiveness of the dilaton-dominated limit, let us briefly discuss some possible way-outs to the previous dramatical conclusions.
One possibility is to accept that we live in a metastable vacuum, provided that its lifetime is longer than the present age of the universe [@Claudson; @Riotto], thus rescuing some points in the parameter space. This possibility, however, poses the cosmological problem of why our universe does not correspond to the global minimum of the potential (without invoking an anthropic principle). Even if a solution to that problem is found we would still have to face the rather bizarre situation of a future cosmological catastrophe, which does not seem very attractive. In addition, it is hard to understand how the cosmological constant is vanishing precisely in such “interim” vacuum. Anyway, despite the previous shortcomings, this is still a possible scenario which would be worth analyzing [@Nosotros].
A different possibility is to assume that also the moduli fields contribute to SUSY breaking. Then the soft terms are modified and possibly some regions in the parameter space would be allowed. This more general situation deserves further analysis [@Nosotros]. Of course, this amounts to a departure of the pure dilaton-dominated scenario. On the other hand, it is interesting to note that explicit possible scenarios of SUSY breaking by gaugino condensation in strings, when analyzed at the one-loop level, lead to the mandatory inclusion of the moduli in the game (in fact the moduli are the main source of SUSY breaking in these cases) [@gaugino; @gaugino2].
Finally, one may think that the perturbative and non-perturbative corrections to the “standard” tree-level dilaton-dominated scenario are important and can modify the previous conclusions. Due to analyticity and non-renormalization theorems these contributions are likely to affect in a substantial way only the Kähler potential [@Banksdine] (in fact, the gauge kinetic function does also receive perturbative corrections at one loop level, but not beyond). Actually, the one-loop string corrections to the Kähler potential (and the gauge kinetic function) have been calculated for orbifold models [@Kaplder] and they are rather small for sensible values of the moduli. Thus, it is reasonable to expect that further perturbative corrections will be even smaller. However, this is not the case for the string non-perturbative corrections, whose size could be much larger (see e.g. ref.[@Shenker]). These corrections could be crucial to understand both the SUSY-breaking mechanism and the vanishing of the cosmological constant [@Banksdine]; actually, it is possible to show [@Progress] that a tree-level dilaton-dominated scenario cannot have a global minimum of the dilaton potential at vanishing cosmological constant. Unfortunately, the form of this type of corrections is very poorly known, which introduces additional sources of uncertainty in the analysis [@Banksdine; @Progress].
Acknowledgements {#acknowledgements .unnumbered}
================
We thank L.E. Ibáñez for useful comments.
\#1 \#2 \#3 [[*Mod. Phys. Lett.*]{} [**\#1**]{} (\#2) \#3 ]{} \#1 \#2 \#3 [[*Nucl. Phys.*]{} [**B\#1**]{} (\#2) \#3 ]{} \#1 \#2 \#3 [[*Phys. Lett.*]{} [**B\#1**]{} (\#2) \#3 ]{} \#1 \#2 \#3 [[*Phys. Rep.*]{} [**\#1**]{} (\#2) \#3 ]{} \#1 \#2 \#3 [[*Phys. Rev.*]{} [**D\#1**]{} (\#2) \#3 ]{} \#1 \#2 \#3 [[*Phys. Rev. Lett.*]{} [**\#1**]{} (\#2) \#3 ]{} \#1 \#2 \#3 [[*Prog. Theor. Phys.*]{} [**\#1**]{} (\#2) \#3 ]{} \#1 \#2 \#3 [[*Rev. Mod. Phys.*]{} [**\#1**]{} (\#2) \#3 ]{} \#1 \#2 \#3 [[*Z. Phys.*]{} [**C\#1**]{} (\#2) \#3 ]{}
[99]{} J.M. Frere, D.R.T. Jones and S. Raby, 222 1983 11;\
L. Alvarez-Gaumé, J. Polchinski and M. Wise, 221 1983 495;\
J.P. Derendinger and C.A. Savoy, 237 1984 307;\
C. Kounnas, A.B. Lahanas, D.V. Nanopoulos and M. Quirós, 236 1984 438. M. Claudson, L.J. Hall and I. Hinchliffe, 228 1983 501. M. Drees, M. Glück and K. Grassie, 157 1985 164;\
J.F. Gunion, H.E. Haber and M. Sher, 306 1988 1;\
H. Komatsu, 215 1988 323. G. Gamberini, G. Ridolfi and F. Zwirner, 331 1990 331. P. Langacker and N. Polonsky, 50 1994 2199;\
A. Bordner, [*KUNS-1351*]{}, [*hep-ph/9506409*]{}. J.A. Casas, A. Lleyda and C. Muñoz, [*FTUAM 95/11*]{}, [*hep-th/9507294*]{}, to be published in [*Nuclear Physics B*]{}. T. Falk, K. Olive, L. Roszkowski and M. Srednicki, [*UMN-TH-1411/95, hep-ph/9510308*]{}. A. Riotto and E. Roulet, [*SISSA-163/95/EP, hep-ph/9512401*]{}. A. Font, L.E. Ibáñez, D. Lüst and F. Quevedo, 245 1990 401;\
M. Cvetic, A. Font, L.E. Ibáñez, D. Lüst and F. Quevedo, 361 1991 194;\
B. de Carlos, J.A. Casas and C. Muñoz, 299 1993 234; 399 1993 623;\
A. de la Macorra and G.G. Ross, 325 1994 85. L.E. Ibáñez and D. Lüst, 382 1992 305. V.S. Kaplunovsky and J. Louis, 306 1993 269. A. Brignole, L.E. Ibáñez and C. Muñoz, 422 1994 125 \[Erratum: [**B436**]{} (1995) 747\]. S. Ferrara, C. Kounnas and F. Zwirner, 429 1994 589 \[Erratum: [**B433**]{} (1995) 255\]. A. Brignole, L.E. Ibáñez, C. Muñoz and C. Scheich, [*FTUAM 95/26*]{}, [*hep-ph/9508258*]{}. A. de la Macorra and G.G. Ross, 404 1993 321;\
V. Halyo and E. Halyo, [*SU-ITP-96-4, hep-ph/9601328*]{}. R. Barbieri, J. Louis and M. Moretti, 312 1993 451 \[Erratum: [**B316**]{} (1993) 632\]. S. Khalil, A. Masiero and F. Vissani, [*SHEP-95-30*]{}, [*hep-ph/9511284*]{}. J.L. Lopez, D.V. Nanopoulos and A. Zichichi, 319 1993 451. For a recent review, see: C. Muñoz, “Soft supersymmetry-breaking terms and the $\mu$ problem”, [*FTUAM 95/20*]{}, [*hep-th/9507108*]{}. G.F. Giudice and A. Masiero, 206 1988 480. J.A. Casas and C. Muñoz, 306 1993 288. G. Lopes-Cardoso, D. Lüst and T. Mohaupt, 432 1994 68. I. Antoniadis, E. Gava, K.S. Narain and T.R. Taylor, 432 1994 187. C. Muñoz, Proceedings of the International Workshop “Beyond the Standard Model IV”, [*World Scientific*]{} (1995) 200, [*hep-ph/9503314*]{}. B. de Carlos and J.A. Casas, 349 1995 300 \[Erratum: [**B351**]{} (1995) 604\];\
M.A. Diaz and S.F. King, [*SHEP-95-30*]{}, [*hep-ph/9601230*]{}. V. Kaplunovsky, 307 1988 145 \[Erratum: [**B382**]{} (1992) 436\]. K. Choi, 37 1988 1564;\
L.E. Ibáñez, D. Lüst and G.G. Ross, 272 1991 251;\
L.E. Ibáñez and D. Lüst, 382 1992 305;\
H.P. Nilles and S. Stieberger, [*TUM-HEP-225/95*]{}, [*hep-th/9510009*]{}. C. Muñoz, Proceedings of the 5th Hellenic School and Workshops on Elementary Particle Physics, Corfu (1995), [*FTUAM 96/04*]{}, [*hep-ph/9601325*]{}. B. de Carlos and J.A. Casas, 309 1993 320. R. Barbieri and G.F. Giudice, 306 1988 63. K. Inoue, A. Kakuto, H. Komatsu and S. Takeshita, 67 1982 1889;\
L.E. Ibáñez and C. López, 126 1983 54;\
L. Alvarez-Gaumé, J. Polchinski and M. Wise, in ref.\[1\]. L.E. Ibáñez and C. López, 233 1984 511;\
L.E. Ibáñez, C. López and C. Muñoz, 256 1985 218. L. Montanet et al., 50 1994 1173 and 1995 off-year partial update for the 1996 edition available on PDG WWW pages (URL:http://pdg.lbl.gov/). K. Choi, J.E. Kim and H.P. Nilles, 73 1994 1758. J.A. Casas, A. Lleyda and C. Muñoz, in preparation. S. Ferrara, N. Magnoli, T.R. Taylor and G. Veneziano, 245 1990 409. T. Banks and M. Dine, 50 1994 7454. L.J. Dixon, V. Kaplunovsky and J. Louis, 355 1991 649;\
J.P. Derendinger, S. Ferrara, C. Kounnas and F. Zwirner, 372 1992 145. S.H. Shenker, Proceedings of the Cargese Workshop on Random Surfaces, Quantum Gravity and Strings, Cargese (France) 1990. Work in progress.
Figure Captions {#figure-captions .unnumbered}
===============
Fig.1
: Excluded regions in the parameter space of the MSSM assuming SUSY breaking by the dilaton, with $B$=$2m_{3/2}$. The black region is excluded because it is not possible to reproduce the experimental mass of the top. The small filled squares indicate regions excluded by Unbounded From Below constraints. The circles indicate regions excluded by Charge and Color Breaking constraints. The filled diamonds correspond to regions excluded by the experimental lower bounds on SUSY-particle masses. The ants indicate regions excluded by negative scalar squared mass eigenvalues.
Fig.2
: The same as Fig.1 but with $B$=$0,\;3m_{3/2}$ and $M$=$-\sqrt{3} m_{3/2}$.
Fig.3
: The same as Fig.1 but with $M$=$-\sqrt{3} m_{3/2}\;$ and $\;V_0$=$-0.5 m_{3/2}^{2},\;3 m_{3/2}^{2},\;10 m_{3/2}^{2}$.
[^1]: On leave of absence from Instituto de Estructura de la Materia CSIC, Serrano 123, 28006 Madrid, Spain.
[^2]: Research supported in part by: the CICYT, under contracts AEN95-0195 (JAC) and AEN93-0673 (ALL, CM); the European Union, under contracts CHRX-CT92-0004 (JAC), CHRX-CT93-0132 (CM) and SC1-CT92-0792 (CM); the Ministerio de Educación y Ciencia, under FPI grant (ALL).
[^3]: For possible explicit SUSY-breaking mechanisms where this limit might be obtained see ref.[@Halyo].
[^4]: The phenomenology of SUSY breaking by the dilaton in the context of a flipped SU(5) model was also studied in ref.[@Lopez].
[^5]: Phenomenological aspects of the $B$=$2m_{3/2}$ scenario have been studied in refs.[@Barbieri; @Brignole; @MAD].
[^6]: The value of $\mu$ can also be fixed once we choose a particular mechanism for solving the $\mu$ problem, e.g. if $\mu$ is generated through the Kähler potential in dilaton-dominated orbifold models, then $\mu$=$m_{3/2}$ [@Brignole2; @Munoz3]. However, we prefer to eliminate $\mu$ in terms of the other parameters by imposing appropriate electroweak breaking, as mentioned above, the reason being that this provides the most general analysis.
[^7]: However, in the case $M$=$-\sqrt{3} m_{3/2}$ the restriction is small allowing in principle large top masses. This possibility was not taken into account in ref.[@Barbieri] where an upper bound of 180 GeV on the top mass was obtained for $M$=$\sqrt{3}m_{3/2}$, $A$=$-\sqrt{3}m_{3/2}$. The discrepancy between that upper bound and the one that we obtain here for the same case, 167 GeV, is due to the following: the RGEs [@Ibanez] used in ref.[@Barbieri] correspond to the soft Lagrangian of eq.(\[lsoft\]) with a minus sign in front of the gaugino masses and therefore the associated boundary conditions that should have been used in order to get the correct result are $M$=$-\sqrt{3}m_{3/2}$, $A$=$-\sqrt{3}m_{3/2}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A fast Discrete Cosine Transform (DCT) algorithm is introduced that can be of particular interest in image processing. The main features of the algorithm are regularity of the graph and very low arithmetic complexity. The 16-point version of the algorithm requires only 32 multiplications and 81 additions. The computational core of the algorithm consists of only 17 nontrivial multiplications, the rest 15 are scaling factors that can be compensated in the post-processing. The derivation of the algorithm is based on the algebraic signal processing theory (ASP).'
author:
-
-
bibliography:
- 'IEEEabrv.bib'
- 'pcs\_ref.bib'
title: 'A low multiplicative complexity fast recursive DCT-2 algorithm'
---
Introduction
============
The Discrete Cosine Transform (DCT) has found many applications in image processing, data compression and other fields due to its decorrelation property [@Rao1990]. Despite the fact that a number of fast DCT algorithms has been proposed already, designing new efficient schemes is still of great interest [@Parf2006; @Liang01; @Vash09]. Majority of proposed fast DCT algorithms have been obtained using graph transformation, equivalence relation or sophisticated manipulation of the transform coefficients. Recently an algebraic approach to derivation of fast DCT has been presented [@Pusch3]. The approach uses [*polynomial algebra*]{} associated with DCT to obtain fast algorithms. Subsequently this theory has been called [*algebraic signal processing theory*]{} (ASP) [@Pusch8d]. The theory provides consistent algebraic interpretation of fast DCT algorithms.
The paper presents derivation of fast DCT-2 $n$-point algorithm ($n$ is a power of two) based on ASP. The algorithm is recursive and has a regular graph. Another feature of the algorithm is very low arithmetic complexity: 16-point DCT requires only 32 multiplications and 81 additions (that is only one multiplication greater than [@Loeff89]), but the computational core of algorithm contains only 17 multiplication while other 15 are scaling factors that can be compensated in the post-processing. Because of the mentioned properties the algorithm is a very attractive choice for hardware DCT implementations.
Algebraic approach to DCT
=========================
In this section the fundamentals of algebraic signal processing theory [@Pusch8d] are considered that are used further for derivation of the fast DCT-2 algorithm.
Background: polynomial algebras
-------------------------------
A [*polynomial algebra*]{} is a vector space over the field $\mathbb{F}$ denoted as $$\mathcal{A}_{\mathbb{F}} = \mathbb{F}[x]/p(x). \label{eq:pol_alg}$$ The elements of algebra is the set of all polynomials in $x$ over $\mathbb{F}$ of degree smaller than $\deg(p)=n$. $\mathcal{A}_{\mathbb{F}}$ is equipped with the operations of usual polynomial addition and multiplication modulo the polynomial $p(x)$.
Using the Chinese remainder theorem (CRT) a polynomial algebra (\[eq:pol\_alg\]) can be decomposed into a direct sum of one-dimensional subalgebras $$\mathcal{F} \; \colon \;\; \mathbb{F}[x]/p(x)\rightarrow \bigoplus_{0\leq k < n} \mathbb{F}_e[x]/(x-\alpha_k),
\label{eq:CRT}$$ provided that zeros $\alpha=(\alpha_0,\ldots,\alpha_{n-1})$ of $p(x)$ are pairwise distinct and $\alpha_k\in\mathbb{F}$. The mapping $\mathcal{F}$ is represented in matrix form $$\mathcal{F} = \mathcal{P}_{b,\alpha}=[p_\ell(\alpha_k)]_{0\leq k,\ell <n},
\label{eq:pol_trans}$$ if a basis $b=(p_0,\ldots,p_{n-1})$ is set in $\mathbb{F}[x]/p(x)$ and unit bases $(x^0)=(1)$ is chosen in each $\mathbb{F}[x]/(x-\alpha_k)$. $\mathcal{P}_{b,\alpha}$ is referred to as [*polynomial transform*]{} for $\mathcal{A}_{\mathbb{F}}$ with basis $b$ [@Pusch8d]. A [*scaled polynomial transform*]{} is obtained for a different basis $\beta_k$ in each $\mathbb{F}[x]/(x-\alpha_k)$: $$\mathcal{F} ={\mathop{\mathrm{diag}}}(1/\beta_1,\ldots,1/\beta_{n-1}) \cdot \mathcal{P}_{b,\alpha}.
\label{eq:sc_pol_trans}$$
Derivation of fast transform algorithms in ASP
----------------------------------------------
In ASP transforms is represented as matrix-vector products $$\mathbf{y} = T \mathbf{x}, \text{ where } T=[t_{k,\ell}]_{0\leq k,\ell<n}.$$ The fast transform algorithm is viewed as factorization of $T$ into a product of sparse structured matrices. This approach has advantages from an algorithmic point of view. It reveals the algorithm structure and simplifies manipulation with it to derive a new variants.
In the paper the following basic matrices are used: $$I_n = \begin{bmatrix}
1 & & \\
& \ddots & \\
& & 1
\end{bmatrix} \quad
J_n = \begin{bmatrix}
& & 1 \\
& \iddots & \\
1 & &
\end{bmatrix}.$$ Permutation matrices that has exactly one entry 1 in row $i$ at position $f(i)$ and each column and 0 elsewhere is defined as: $$P\colon i\mapsto f(i),\quad 0\leq i < n.$$ One important is the $n \times n$ stride permutation matrix defined for $m|n$ as $$L_m^n \colon i_2\frac{n}{m}+i_1 \mapsto i_1 m + i_2$$ for $0 \leq i_1< \frac{n}{m}$, $0 \leq i_2 < m$.
ASP states that every DCT corresponds to some polynomial algebra $\mathbb{F}[x]/p(x)$ with basis $b$. In this case DCT is given by the CRT (\[eq:CRT\]) and its matrix takes the form of polynomial transform (\[eq:pol\_trans\]) or a scaled polynomial transform (\[eq:sc\_pol\_trans\]). From (\[eq:CRT\]) it can be seen that $\mathcal{F}$ decomposes $\mathbb{F}[x]/p(x)$ into one-dimensional polynomial algebras. Fast algorithm is obtained by complying this decomposition [*in step*]{} using an intermediate subalgebras.
One possible way to perform decomposition of $\mathbb{F}[x]/p(x)$ in step is to use factorization $p(x)=q(x)\cdot r(x)$. If $\deg(q)=k$ and $\deg(r)=m$ then $$\begin{aligned}
&& \mathbb{F}[x]/p(x) \notag \\
& \rightarrow & \mathbb{F}[x]/q(x)\oplus \mathbb{F}[x]/r(x) \label{eq6}\\
& \rightarrow & \bigoplus\limits_{0\leq i<k}\mathbb{F}[x]/(x-\beta_i) \oplus \bigoplus\limits_{0\leq j<m} \mathbb{F}[x]/(x-\gamma_j) \label{eq7}\\
& \rightarrow & \bigoplus\limits_{0\leq i<n}\mathbb{F}[x]/(x-\alpha_i) \label{eq8}\end{aligned}$$ where $\beta_i$ and $\gamma_j$ are the zeros of $q(x)$ and $r(x)$ correspondingly. If $c$ and $d$ are the bases of $\mathbb{F}[x]/q(x)$ and $\mathbb{F}[x]/r(x)$, respectively, then (\[eq6\])-(\[eq8\]) are expressed in the following matrix form [@Pusch8d]: $$\label{th1}
\mathcal{P}_{b,\alpha} = P(\mathcal{P}_{c,\beta}\oplus \mathcal{P}_{d,\gamma})B,$$ where $A\oplus B=[\begin{smallmatrix} A & \\ & B\end{smallmatrix}]$ denotes the direct sum of matrices. Step (\[eq7\]) uses the CRT to decompose $\mathbb{F}[x]/q(x)$ and $\mathbb{F}[x]/r(x)$. This step corresponds to the direct sum of matrices $\mathcal{P}_{c,\beta}$ and $\mathcal{P}_{d,\gamma}$. Finally permutation matrix $P$ maps the concatenation $(\beta,\gamma)$ to the ordered list of zeros $\alpha$ in (\[eq8\]). Given that $B$ is sparse (\[th1\]) leads to a fast algorithm.
Polynomial algebras for $\mathrm{DCT}$-$2$ and $\mathrm{DCT}$-$4$
-----------------------------------------------------------------
This subsection introduces polynomial algebras which is connected with $\mathrm{DCT}\text{-}4$ and $\mathrm{DCT}\text{-}2$. Let us first consider the polynomial algebra associated with $\mathrm{DCT}\text{-}4_n$ $$\mathcal{A}_\mathbb{F} = \mathbb{F}[x]/2T_n(x),\quad b=(V_0,\ldots,V_{n-1}), \label{eq:dct4_pol_albr}$$ where $T$ and $V$ are Chebyshev polynomials of the first and third kind, respectively. This Chebyshev polynomials have the following closed form expressions ($\cos\theta=x$) $$\textstyle
T_n(x) = \cos(n\theta),\,\,\, V_n(x) = \frac{\cos(n+\frac12)\theta}{\cos\frac12\theta}.$$ $\alpha_k=\cos(k+\frac12)\frac{\pi}{n}$, $0\leq k<n$ are zeros of $2T_n(x)$. in accordance with (\[eq:pol\_trans\]) polynomial transform for algebra (\[eq:dct4\_pol\_albr\]) is defined as $$\label{eq4}
\mathcal{P}_{\alpha,b} = [ V_\ell(\alpha_k)]_{0\leq k,\ell<n} = \left[ \frac{\textstyle \cos(k+\frac{1}{2})(\ell+\frac{1}{2})\frac{\pi}{n}} {\textstyle\cos(k+\frac{1}{2})\frac{\pi}{2n}} \right].$$ In order to get the matrix of $\mathrm{DCT}$-$4_n$ (\[eq4\]) is multiplied from the left by scaling diagonal matrix $$\label{eq:sc_mtx}\textstyle
D_n^{(C4)} = {\mathop{\mathrm{diag}}}\nolimits_{0\leq k< n} \left({\cos(k+\frac12)\frac{\pi}{2n}}\right)$$ that yields $$\label{eq5}\textstyle
\mathrm{DCT}\text{-}4_n =\left[\cos(k+\frac12)(\ell+\frac12)\frac{\pi}{n} \right]_{0\leq k,\ell<n}.$$ Eq. (\[eq4\])–(\[eq5\]) show that $\mathrm{DCT}$-$4$ is a scaled polynomial transform of the form (\[eq:sc\_pol\_trans\]) for the specified polynomial algebra (\[eq:dct4\_pol\_albr\]).
$\mathrm{DCT}\text{-}2_n$ is arisen from polynomial algebra $$\mathcal{A}_\mathbb{F} = \mathbb{F}[x]/(x-1)U_{n-1}(x),\quad b=(V_0,\ldots,V_{n-1}), \label{eq:dct2_pol_albr}$$ where $U$ is Chebyshev polynomial of the second kind that can be written as ($\cos\theta=x$): $$\textstyle
U_n(x) = \frac{\sin(n+1)\theta}{\sin\theta}.$$ Since zeros of $U_n(x)$ is given by $\alpha_k=\cos\frac{(k+1)\pi}{n+1}$, $0\leq k<n$ polynomial transform for (\[eq:dct2\_pol\_albr\]) takes the form $$\label{eq:dct2_pol_trans}
\mathcal{P}_{\alpha,b} = [ V_\ell(\alpha_k)]_{0\leq k,\ell<n} = \left[ \frac{\textstyle \cos k(\ell+\frac{1}{2})\frac{\pi}{n}} {\textstyle\cos\frac{k\pi}{2n}} \right].$$ To obtain $\mathrm{DCT}\text{-}2$ matrix (\[eq:dct2\_pol\_trans\]) need to be multiplied from the left by the scaling diagonal $$\label{eq:sc_mtx_c2}\textstyle
D_n^{(C2)} = {\mathop{\mathrm{diag}}}\nolimits_{0\leq k< n} \left({\cos\frac{k\pi}{2n}}\right).$$ Polynomial transform corresponding to discrete trigonometric transform ($\mathrm{DTT}$) is denoted as $\overline{\mathrm{DTT}}$, for instance $\overline{\mathrm{DCT}\text{-}4}_n$ stands for the matrix in (\[eq4\]).
In what follows we need skew $\mathrm{DCT}\text{-}4(r)$. In [@Pusch8d] this transform was introduced since it appears to be important building blocks of Cooley-Tukey type of algorithms for $\mathrm{DCT}$. Skew $\mathrm{DCT}\text{-}4(r)$ associates with polynomial algebra $$\mathcal{A}_{\mathbb{F}}=\mathbb{F}[x]/(2T_n(x)-2\cos r\pi)$$ with the basis $b=(V_0,\ldots,V_{n-1})$, where $0<r<1$. The conventional $\mathrm{DCT}\text{-}4_n$ is the special case of skew $\mathrm{DCT}\text{-}4_n(r)$ for $r=1/2$.
Derivation of fast $\mathrm{DCT}\text{-}2_{2^k}$ algorithm
==========================================================
In this section the procedure of algebraic derivation of fast $\mathrm{DCT}\text{-}2_{2^k}$ algorithm is given in detail. According to (\[eq:dct2\_pol\_albr\]) the polynomial algebra corresponding to $\mathrm{DCT}\text{-}2_{2^k}$ is given by $$\mathcal{A}_\mathbb{F} = \mathbb{F}[x]/(x-1)U_{2^k-1}(x),\quad b=(V_0,\ldots,V_{2^k-1}).$$ Important issue is to choose the base field $\mathbb{F}$. Since Chebyshev polynomials $V$ and $U$ which is included in definition (\[eq:dct2\_pol\_albr\]) have integer coefficients (for example $V_2(x)=4x^2-2x-1$), the base field $\mathbb{F}$ is set to the field of rational numbers $\mathbb{Q}$. The filed is extended during factorization of polynomial $U_{2^k-1}(x)$, since the polynomial is not factored over $\mathbb{Q}$.
It is well known [@Rao1990] that fast $\mathrm{DCT}\text{-}2_{2n}$ algorithm can be reduced to fast $\mathrm{DCT}\text{-}2_{n}$ and $\mathrm{DCT}\text{-}4_{n}$ algorithms. Using factorization for the Chebyshev polynomial of the second kind $$U_{2n-1}(x) = U_{n-1}(x)\cdot 2T_n(x),$$ the algebra $\mathbb{Q}[x]/(x-1)U_{2n-1}(x)$ with basis $b=(V_0,\ldots $ $V_{2n-1})$ can be decomposed as $$\begin{aligned}
&&\mathbb{Q}[x]/(x-1)U_{2n-1}(x) \notag \\
&\rightarrow&\mathbb{Q}[x]/(x-1)U_{n-1}(x)\oplus \mathbb{Q}[x]/2T_{n}(x), \label{eq:dct2_2n_decomp}\end{aligned}$$ that according to (\[eq6\])–(\[eq8\]) leads to the following fast algorithm [@Pusch8d] $$\label{eq:fast_dct2}
\overline{\mathrm{DCT}\text{-}2}_{2n} = L_n^{2n}(\overline{\mathrm{DCT}\text{-}2}_{n}\oplus \overline{\mathrm{DCT}\text{-}4}_{n})B_{2n},$$ where $L_n^{2n}$ is the stride permutation matrix and $B_{2n}$ is change of basis matrix. $B_{2n}$ maps basis $b$ to the concatenation $(c,d)$, where $c=d=(V_0,\ldots V_{n-1})$ are the basis for subalgebras in the right-hand side of (\[eq:dct2\_2n\_decomp\]). The first $n$ columns of $B_{2n}$ are $$B_{2n}=\begin{bmatrix}
I_n & * \\
I_n & *
\end{bmatrix},$$ since the elements $V_\ell \in b$ for $0\leq \ell < n$ are already contained in $c$ and $d$. The rest entries are determined by the following expressions $$\begin{aligned}
V_{n+\ell} \equiv& V_{n-\ell-1} &\mod{(x-1)U_{n}} \label{eq:new_dct2_basis:a}\\
V_{n+\ell} \equiv& -V_{n-\ell-1} &\mod{2T_{n}}, \label{eq:new_dct2_basis:b}\end{aligned}$$ which yields $$B_{2n}=\begin{bmatrix}
I_n & J_n \\
I_n & -J_n
\end{bmatrix}.$$ (\[eq:new\_dct2\_basis:a\])–(\[eq:new\_dct2\_basis:b\]) can be induced using the following relation $2T_n = V_n + V_{n-1}$, $(x-1)U_{n-1} = V_n - V_{n-1}$ and $V_n = 2xV_{n-1}-V_{n-2}$. Note that decomposition (\[eq:dct2\_2n\_decomp\]) does not require extension of based field $\mathbb{Q}$. This leads to multiplication-free change of basis matrix $B_{2n}$.
When the size of $\mathrm{DCT}\text{-}2$ is power of two (\[eq:fast\_dct2\]) can be applied recursively to obtain fast algorithm. Thus, the problem of derivation of fast $\mathrm{DCT}\text{-}2_{2^k}$ algorithm reduces to derivation of fast $\mathrm{DCT}\text{-}4_{2^{k-1}}$ algorithm. From the ASP point of view the question is how to factor polynomial $2T_n$ (when $n$ is power of 2) in step. We propose to use the following general recursive formula $$\begin{gathered}
\textstyle 2T_{2n}(x) -2\cos r\pi = \mathstrut\left(2T_{n}(x)-2\cos\frac{r\pi}{2}\right) \\ \textstyle
\times\left(2T_{n}(x)-2\cos \pi(1-\frac{r}{2})\right), \label{eq:gen_fact}\end{gathered}$$ that can be proved using the closed form of $T_{2n}$, parameter $r\in(0,\;1)$. The special case of (\[eq:gen\_fact\]) for $r=1/2$ specify factorization of $2T_{2n}$. Using (\[eq:gen\_fact\]) polynomial algebra related to $\mathrm{DCT}\text{-}4_{2n}(r)$ is decomposed as[^1] $$\begin{aligned}
&& \mathbb{Q}_{\cos r\pi}[x]/(2T_{2n}(x)-2\cos r\pi) \notag \\
& \rightarrow & \textstyle \mathbb{Q}_{\cos\frac{r\pi}{2}}[x]/(2T_{n}(x)-2\cos\frac{r\pi}{2}) \oplus \notag\\
&& \textstyle \mathbb{Q}_{\cos\frac{r\pi}{2}}[x]/(2T_{n}(x)-2\cos \pi(1-\frac{r}{2})). \label{eq:dct4_2n_decomp}\end{aligned}$$ The decomposition leads to the following fast algorithm $$\label{my_alg}
\begin{split}
\textstyle
\overline{\mathrm{DCT}\text{-}4}_{2n}(r) = P\cdot(& \textstyle\overline{\mathrm{DCT}\text{-}4}_{n}(\frac{r}{2}) \\ \oplus&\textstyle\overline{\mathrm{DCT}\text{-}4}_{n}(1-\frac{r}{2}) )\cdot B_{2n}^{(C4)}(r),
\end{split}$$ where $P$ is a permutation matrix of the form $$P= \left[
\begin{smallmatrix} 1 & & & & & & \\
& & & & I_2 & & \\
& I_2& & & & & \\
& & & & &\ddots & \\
& &\ddots & & & & \\
& & & & & & I_2\\
& & & 1& & &
\end{smallmatrix}\right],$$ and $B_{2n}^{(C4)}(r)$ is the change of basis matrix $$\begin{split}
B_{2n}^{(C4)}(r)=&
\begin{bmatrix}
I_m & (2\cos\frac{r\pi}{2}I_m-J_m) \\
I_m & (-2\cos\frac{r\pi}{2}I_m-J_m)
\end{bmatrix} \\
=&\begin{bmatrix}
I_m & I_m \\
I_m & -I_m
\end{bmatrix}\cdot
\begin{bmatrix}
I_m & -J_m \\
& 2\cos\frac{r\pi}{2}I_m
\end{bmatrix},
\end{split}\label{eq:BB_matrix}$$ which is determined by $$\begin{split}\textstyle
V_{n+\ell} \equiv\textstyle - V_{n-\ell-1} + & \textstyle 2\cos\frac{r\pi}{2}V_\ell
\textstyle \mod{2T_{n} - 2\cos\frac{r\pi}{2}}
\\ \textstyle
V_{n+\ell} \equiv\textstyle - V_{n-\ell-1} - & \textstyle 2\cos\frac{r\pi}{2}V_\ell
\textstyle\mod{2T_{n} - 2\cos \pi(1-\frac{r}{2})}.
\end{split}$$
Decomposition (\[eq:dct4\_2n\_decomp\]) requires extension of the based field $\mathbb{Q}_{\cos r\pi}$ to $\mathbb{Q}_{\cos\frac{r\pi}{2}}$. New elements of the field appears in matrix $B_{2n}^{(C4)}(r)$.
Joint use of factorizations (\[eq:fast\_dct2\]) and (\[my\_alg\]) leads to the new fast $\mathrm{DCT}\text{-}2_{2^k}$ recursive algorithm. The basic operation of the algorithm is multiplication by the matrix $B_{2n}^{(C4)}(r)$. All nontrivial multiplication concentrate in it that is very similar to butterfly operation in FFT algorithm.
{width="160mm"}
![Building block of the fast DCT algorithm[]{data-label="fig:2"}](building_block){width="72mm"}
Fast $\mathrm{DCT}\text{-}2_{16}$ algorithm
===========================================
In this section the proposed approach is applied to derivation of fast $\mathrm{DCT}\text{-}2_{16}$ algorithm. At first the transform expressed as a product $$\mathrm{DCT}\text{-}2_{16} = D_{16}^{(C2)} \cdot \overline{\mathrm{DCT}\text{-}2}_{16}.$$ Then factorization (\[eq:fast\_dct2\]) and (\[my\_alg\]) is applied recursively to obtain fast transform algorithm. Flow graph of this algorithm is shown in Fig. \[fig:1\] (for simplicity scaling of the output is omitted). Fig. \[fig:2\] explains the basic building block (BB) of the algorithm that performs the multiplication by matrix (\[eq:BB\_matrix\]). All operations inside the BB are implemented on the input $m$ components vectors. Evaluation of one BB requires $3m$ addition and $m$ multiplication.
The presented 16-point $\mathrm{DCT}\text{-}2$ algorithm uses 32 multiplication and 81 addition. However only 17 multiplication constitute the core of algorithm while other 15 is scaling factors that can be compensate in the post-processing. Also the Fig. \[fig:1\] shows that algorithm include computation of 8-point $\mathrm{DCT}\text{-}2$ that requires only 5 multiplication. It is the same result as in [@Arai88]. In fact, proposed algorithm can be considered as generalization of Arai’s DCT algorithm since the resulting computational scheme has very low multiplicative complexity and scaling outputs.
Conclusion
==========
A fast $2^k$-point algorithm of DCT-2 based on ASP is presented. The key features of the algorithm are regularity of the graph ($\mathrm{DCT}\text{-}2_{n/2}$ available inside of a $\mathrm{DCT}\text{-}2_n$) and very low arithmetic complexity (computational core of the $\mathrm{DCT}\text{-}2_{2^k}$ algorithm contains only $\sum_{p=1}^{k-1}2^pp$ multiplications). Regular graph of proposed algorithm is well suited for development of new parallel-pipeline architecture of DCT processor. Also, the algorithm extends existing space of alternative fast algorithms of the DCT. It can be used by automatic code generation programs that search alternative implementations for the same transform to find the one that is best tuned to desired platform [@Wanh1999; @Pusch:05].
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported by the Belarusian Fundamental Research Fund (F11MS-037).
[^1]: Here $\mathbb{Q}_{\cos r\pi}$ is used as a short notation for field extension $\mathbb{Q}[\cos r\pi]$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Piero Ullio and Marc Kamionkowski\
Mail Code 130-33, California Institute of Technology, Pasadena, CA 91125\
E-mail: ,
title: 'Velocity distributions and annual-modulation signatures of weakly-interacting massive particles'
---
Introduction
============
Weakly interacting massive particles (WIMPs) are among the leading candidates for dark matter in galactic halos. Such particles arise naturally in extensions to the standard model (SM) of particle physics; an example is the neutralino, plausibly the lightest superpartner in supersymmetric versions of the SM. Massive particles whose coupling with lighter SM particles have interactions of electroweak strength have a cosmological abundance of order the critical density of the Universe. Hence, WIMPs appear naturally as dark-matter candidates. The possibility to link these two apparently separate problems (electroweak symmetry breaking and dark matter) was realized a couple of decades ago, and since then the search for WIMPs in the Milky Way halo has been a major endeavor both theoretically and experimentally (for a comprehensive review see Ref. [@jkg]).
Numerous complementary techniques have been developed in order to detect relic WIMPs. Currently, the most promising method is probably direct detection through observation, in a low-background laboratory detector, of nuclear recoils due to WIMP-nucleus elastic scattering [@dirdet; @annmod1]. The chance for a given WIMP to interact in the detector is very low and the energy released in case of interaction is expected to be tiny (in the keV range). Nevertheless, this detection method has already had a few successes. It has been exploited to exclude as the main component of the dark Galactic halo WIMP candidates such as a fourth-generation heavy neutrino and the sneutrino [@direxcl]. Detectors have now reached the sensitivity to start probing the region of parameter space of interest if a neutralino is the dark matter (see e.g. Ref. [@dirnow]). Recently, the [Dama]{} [@DAMA] and [Cdms]{} [@CDMS] collaborations, while probing roughly the same region of parameter space, have presented apparently contradictory results, a possible WIMP signal in the first case and a null result in the second. It is probably premature to derive any conclusion from these results, but, with further data and even more sensitive detectors being developed, the next years promise to be very exciting for the field.
To claim a positive detection, an experiment must be able to discriminate the signal from backgrounds. In principle, the shape of the recoil spectrum can be used, since the recoil spectra from WIMPs and background should generally differ. However, the shape of the recoil-energy spectrum for WIMP-nucleus scattering cannot be predicted with enough precision to separate it from the background, the spectrum of which is generally not understood in detail. A possible way out is to look for a slight annual modulation in the event rate (see Refs. [@annmod1; @annmod2]; among more recent works see, e.g., Refs. [@Brhlik; @Belli]). Such an effect is expected for the WIMP signal, but not for the background. This is the signature exploited in the data analysis by the [Dama]{} collaboration to claim detection of WIMPs. The underlying idea is quite simple. Like all other stars in the rotationally-supported disk, the Sun is moving around the Galactic center on a roughly circular orbit, passing through the dark halo which is believed on the other hand to be static and not rotationally supported. The Earth, and detectors on it, contain this velocity component plus an additional component due to the orbital motion around the Sun. The azimuthal velocity of the Sun and the projection of the velocity of the Earth on the galactic plane are most closely aligned near June 2 and most anti-aligned six months later. The WIMP-nucleus interaction rate in a detector depends on the velocities of the incident WIMPs. Hence, a yearly modulation of the signal is expected.
In prior analyses of the modulation effect, the local dark-matter velocity distribution was assumed to be a Maxwell-Boltzmann distribution (which is, of course, isotropic), as would arise if the Galactic halo is isothermal. Velocity distributions for halos with some bulk rotation have also been considered [@Belli]. Although these velocity distributions are consistent with current data on the Milky Way, there are other plausible, consistent, and possibly even better-motivated alternatives. For example, results from N-body simulations of hierarchical clustering favor density profiles which are steeper at large galactocentric distances than the $r^{-2}$ decline in the isothermal sphere and which are cuspy in the Galactic center, rather than cored, and the velocity distribution corresponding to a cuspy halo should differ from the Maxwell-Boltzmann distribution that corresponds to an isothermal halo.
Moreover, it is plausible that the velocity distribution may be anisotropic rather than isotropic as usually assumed. In fact, most of the visible populations in the Galactic halo show some degree of anisotropy (e.g., the stars in the local neighborhood and globular clusters). Furthermore, the inefficiency of phase mixing that results in a cuspy profile (rather than an isothermal sphere) should leave some degree of anisotropy in the velocity-dispersion tensor. Some evidence for a global preference for predominantly radial velocities is already seen in the simulations [@VDB], as well as in globular clusters [@OBGT]. Even if the global velocity distribution is isotropic, clumping in velocity space, which may also arise if phase mixing is not perfectly efficient during gravitational collapse, may yield a locally anisotropic velocity dispersion.
Prior work has shown that the direct-detection rate should not depend sensitively on the details of the velocity distribution [@Kink]. However, this work considered only the [*total*]{} detection rate, integrated over all nuclear-recoil energies. The modulation signal in [Dama]{} depends on details of the [*differential*]{} energy distribution. The purpose of this paper will be to show that the amplitude of the modulation can thus depend quite sensitively on the precise form of the velocity distribution. The inferred WIMP cross sections and masses could thus be altered.
The outline of the paper is the following. In the next Section we discuss a procedure to relate the velocity distribution to the Galactic density distribution. In Section 3 we review WIMP direct detection rates and the annual modulation effect. The main results are given in Section 4. In Section 5 we summarize and make some concluding remarks.
Dark-matter distribution functions {#sec:df}
==================================
We suppose that the dark-matter halo of the Milky Way is roughly spherical, and among the general family of profiles, $$\rho_{\rm{dm}}(r) = \rho_0 \left(\frac{R_0}{r}\right)^{\gamma}
\left[\frac{1+(R_0/a)^{\alpha}}{1+(r/a)^{\alpha}}\right]^
{(\beta-\gamma)/\alpha},
\label{eq:dmprof}$$ we focus on functional forms suggested by N-body simulations (in the equation above $\rho_0$ and $R_0$ are respectively the local dark-matter density and the Sun galactocentric distance). We restrict ourselves mainly to the Navarro, Frenk, and White profile [@NFW], which has $(\alpha,\beta,\gamma)=(1,3,1)$ (hereafter the NFW profile). We will show also that the behavior of the profile towards the Galactic center is not critical in our analysis by considering the more cuspy Moore et al. profile [@Moore], $(\alpha,\beta,\gamma)=(1.5,3,1.5)$, and the less singular profile of Kravtsov et al. [@Kravtsov], $(\alpha,\beta,\gamma)=(2,3,0.4)$. The value of the scale radius $a$ which appears in Eq. (\[eq:dmprof\]) is determined in the N-body simulations as well, depending on the mass of the simulated halo. We infer its approximate value for the NFW and Moore et al. profiles in case of an $\Omega_M = 0.3$ cosmology from Refs. [@Moore; @Bullock]. The approach we follow to fix the remaining unknown parameters, both in the dark-halo profile and in the functions that describe the luminous components of the Milky Way, is to perform a combined best fit of available observational data, taking into account the kinematics of local stars, the rotation curve of the Galaxy, the dynamics of the satellites, and more (details are given in Ref. [@Ullio1]). Sample values for the subset of parameters relevant in the present analysis are specified in Table \[tab:rho\] (mass decompositions for the Milky Way are highly degenerate, so slightly different values are compatible as well). Thus, we have a family of spherically-symmetric radial profiles that are all theoretically plausible and consistent with all known observational constraints.
We will now find the velocity distributions that correspond to these halo profiles. The density distribution does not determine the velocity distribution uniquely. To sample the possibilities, we will therefore first find velocity distributions that have isotropic velocity distributions, and then find some distributions that have preferentially radial velocities.
If we assume an isotropic velocity distribution, then there is a one-to-one correspondence between the spherically symmetric density profile $\rho(r)$ and its distribution function given by Eddington’s formula [@BT], $$F(\eps) = \frac{1}{\sqrt{8} \pi^2}
\left[ \int_0^{\eps} \frac{d^2\rho}{d\Psi^2}
\frac{d\Psi}{\sqrt{\eps-\Psi}} +
\frac{1}{\sqrt{\eps}} \left(\frac{d\rho}{d\Psi}\right)_{\Psi=0}
\right]
\label{eq:edd}$$ where $\Psi(r) = - \Phi(r) + \Phi (r=\infty)$, with $\Phi$ the potential of the system, $\eps = - E + \Phi (r=\infty) = - E_{\rm{kin}} + \Psi(r)$, and $E$ and $E_{\rm{kin}}$, respectively, the total and kinetic energy. Eq. (\[eq:edd\]) works for a single isolated self-gravitating system. However, the Milky Way has a complex structure containing a bulge elongated into a bar, a flattened disk, and maybe a triaxial dark halo. For the present purpose, however, it is sufficient to consider a toy model in which all components are assumed to be spherical. Even the awkward approximation of a “spherical” disk will have little influence on our conclusions. In such a toy model, we can alter Eq. (\[eq:edd\]) to provide the dark-matter distribution function by replacing $\Psi$ and $\rho$ (appropriate for an isolated system) by $\Psi_{\rm{tot}}$ and $\rho_{\rm{dm}}$, respectively, the gravitational potential due to all components and the dark matter density profile. Actually, it is easier from the numerical point of view to implement Eq. (\[eq:edd\]) by changing the integration variable from $\Psi_{\rm{tot}}$ to the radius of the spherical system $r$. Then Eq. (\[eq:edd\]), in case of the dark-matter halo distribution function, becomes, $$\begin{aligned}
F_{\rm{dm}}(\eps) & = & \frac{1}{\sqrt{8} \pi^2}
\int_{\Psi_{\rm{tot}}^{-1}(\eps)}^{\infty}
\frac{dr}{\sqrt{\eps-\Psi_{\rm{tot}}(r)}}
\left[\frac{d\rho_{\rm{dm}}}{dr} \frac{d^2\Psi_{\rm{tot}}}{dr^2}
\left(\frac{d\Psi_{\rm{tot}}}{dr}\right)^{-2} \right.
\nonumber \\
&& \left. -\frac{d^2\rho_{\rm{dm}}}{dr^2}
\left(\frac{d\Psi_{\rm{tot}}}{dr}\right)^{-1} \right]\; .
\label{eq:edd2}\end{aligned}$$
If we relax the hypothesis of isotropy of the velocity dispersion tensor, the most general distribution function corresponding to a spherical density profile is a function of $\eps$ and $L$, the magnitude of the angular-momentum vector. In such systems the velocity dispersion in the radial direction is different from that in the azimuthal direction (which is equal to the one in the other tangential direction) [@BT]. For a given radial density profile, the distribution function is not unique. We investigate a special class of models, the Osipkov-Merritt models [@Osipkov; @Merritt], in which $F$ is a function of $\eps$ and $L$ only through the variable $\cal Q$: $${\cal Q} \equiv \eps - \frac{L^2}{2 r_a^2}\;.
\label{eq:q}$$ Here $r_a$ is called the anisotropy radius, as in the Osipkov-Merritt models the anisotropy parameter is [@Merritt]: $$\beta(r) \equiv 1 - \frac{\overline{v_{\phi}^2}}{\overline{v_{r}^2}}
= \frac{r^2}{r^2+r_a^2}\; .$$ Therefore, $r_a$ is the radius within which the dispersion velocity is nearly isotropic. As already mentioned, in analogy with other observed populations, we will entertain the possibility that for the dark-matter halo $\overline{v_{r}^2} > \overline{v_{\phi}^2}$ and thus that $\beta > 0$. The distribution function for this class of models is again easy to derive. It is sufficient to replace in Eq. (\[eq:edd2\]) $\eps$ with $\cal Q$ and $\rho_{\rm{dm}}(r)$ with $\rho_{\rm{dm}}^{\cal Q}(r) = (1 + r^2/r_a^2)\, \rho_{\rm{dm}}(r)$.
Direct-detection rates and annual-modulation effect
===================================================
The differential direct-detection rate for dark-matter WIMPs in a given material (per unit detector mass) is [@jkg], $$\frac{dR}{dQ} = \frac{\rho_0}{M_{\chi}}
\int_{|\vec{v^{\prime}}| \geq v_{\rm{min}}} d^3\vec{v^{\prime}}\;
f(\vec{v^{\prime}}) \,|\vec{v^{\prime}}|\, \frac{d\sigma}{dQ} \;,
\label{eq:ddrate}$$ where $Q$ is the energy deposited in the detector and $d\sigma/dQ$ is the differential cross section for WIMP elastic scattering with the target nucleus. We assumed here that WIMPs of mass $M_{\chi}$ account for the local dark matter density $\rho_0$ and have a local distribution in velocity space (in the rest frame of the detector) $f = F_{\rm{dm}}(r=R_0)/\rho_0$. The lower limit of integration $v_{\rm{min}}$ is the minimum velocity required for a WIMP to deposit the energy $Q$.
Assuming that scalar interactions dominate (as is probably the case for neutralino elastic scattering with Ge and NaI, the materials used respectively by the [Cdms]{} and [Dama]{} experiments) and that the couplings with protons and neutrons are roughly the same, Eq. (\[eq:ddrate\]) can be rewritten as, $$\frac{dR}{dQ} = \left(\frac{\rho_0\,
\sigma_{p}^{\rm{scalar}}}{2}\right)
\left[ A_N^2 \frac{M_N}{{M_{\chi}}^3}
\left(1+\frac{M_{\chi}}{M_p}\right)^2 {\cal F}^2(Q)\right]
\int_{|\vec{v^{\prime}}| \geq v_{\rm{min}}} d^3\vec{v^{\prime}}\;
\frac{f(\vec{v^{\prime}})}{|\vec{v^{\prime}}|} \; ,
\label{eq:ddrate2}$$ where $\sigma_{p}^{\rm{scalar}}$ is the WIMP-proton cross section at zero momentum transfer, $A_N$ and $M_N$ are the detector nucleus atomic number and mass, while ${\cal F}(Q)$ is the nuclear form factor. In the equation above, the terms in the round bracket are energy and detector independent; we will not consider them in what follows. The terms in the square brackets depend on the nucleus chosen for the detector, as well as on the energy and WIMP mass. When considering annual modulation, they play a weighting effect for those detectors, like NaI, which are not monatomic (the generalization of Eq. (\[eq:ddrate2\]) to this case is straightforward). The last term, $$\begin{aligned}
T(Q,M_N,M_{\chi},t) &=&
\int_{|\vec{v^{\prime}}| \geq v_{\rm{min}}} d^3\vec{v^{\prime}}
\; \frac{f(\vec{v^{\prime}})}{|\vec{v^{\prime}}|} =
\int_{v_{\rm{min}}}^{\infty} dv^{\prime} \;
v^{\prime} \int d\Omega^{\prime} \; f(v^{\prime},\Omega^{\prime})
\nonumber \\
&\equiv &
\int_{v_{\rm{min}}}^{\infty}
dv^{\prime} \;g(v^{\prime}) \; ,
\label{eq:g}\end{aligned}$$ depends on $Q$, $M_N$, and $M_{\chi}$ through $v_{\rm{min}}= [(Q M_N)/(2 M_r^2)]^{1/2}$, where $M_r$ is the WIMP-nucleus reduced mass. It is time dependent and gives rise to the annual-modulation effect. This might not be clear at first sight, as we wrote implicitly the temporal dependence in the change of variables between the detector rest frame and the galactic frame. In polar coordinates, the change of variable to the detector frame (primed system in our notation) is simply: $v_r = v_r^{\prime}$, $v_{\theta} = v_{\theta}^{\prime}$, and $v_{\phi} = v_{\phi}^{\prime} + v_{\oplus}$. The azimuthal shift $v_{\oplus}$ varies during the year; in June it is roughly $v_{\oplus}= \Theta_0 + v_E$, while in December it is $v_{\oplus}= \Theta_0 - v_E$, where $v_E \simeq 15 \,\rm{km}\, \rm{s}^{-1}$ is the projection of the earth orbital velocity on the galactic plane, while $\Theta_0$ is the galactic circular velocity at the Sun’s position. The latter is given in terms of Oort’s constants and the galactocentric distance by: $$\Theta_0 = (A - B) R_0 = (27.2 \pm 0.9) \, R_0
\; \rm{km} \,\rm{s}^{-1} \,\rm{kpc}^{-1}\;,
\label{eq:theta}$$ where the numerical value of $A - B$ comes from the determination from Cepheid proper motions measured by the Hipparcos satellite [@FW].
The amplitude of the annual modulation (keeping track of whether the signal is greater in June or December) for a monatomic detector is then, $${\cal{A}}(Q,M_N,M_{\chi}) =
\frac{T(Q,M_N,M_{\chi},{\rm June})-T(Q,M_N,M_{\chi},{\rm December})}
{T(Q,M_N,M_{\chi},{\rm June})+T(Q,M_N,M_{\chi},{\rm December})}\;.$$ As mentioned, the formula for NaI has instead a weighting factor for each of the two nuclei. For a given detector and distribution function the value of $\cal{A}$ follows. To compute $g$ as defined in Eq. (\[eq:g\]) the appropriate choice of integration variable are, in the isotropic case, $$\begin{aligned}
g(v^{\prime}) & = & 2 \pi\,v^{\prime}
\int_0^{\pi} d\alpha \; \sin\alpha\,
\frac{F_{\rm{dm}}(\eps)}{\rho_0} \nonumber \\
\eps & = &\Psi(R_0) - \frac{1}{2}
\left( {v^{\prime}}^2 + 2 \cos\alpha \, v^{\prime} v_{\oplus}
+ v_{\oplus}^2 \right) \;,\end{aligned}$$ while in the anisotropic case, $$\begin{aligned}
g(v^{\prime}) & = & 2\,v^{\prime}
\int_0^{2 \pi} d\psi \int_0^{\pi} d\eta
\; \sin\eta\,
\frac{F_{\rm{dm}}(\cal{Q})}{\rho_0} \nonumber \\
\cal{Q} & = &\Psi(R_0) - \frac{1}{2}
\left( {v^{\prime}}^2 + 2 \sin\psi\, \sin\eta \,v^{\prime} v_{\oplus}
+ v_{\oplus}^2 \right) \nonumber \\
&& - \frac{R_0^2}{2 r_a^2} \left( {v^{\prime}}^2\, \sin^2\eta
+ 2 \sin\psi\, \sin\eta \,v^{\prime} v_{\oplus} + v_{\oplus}^2 \right)\;.\end{aligned}$$
Results
=======
Isotropic Velocity Distributions
--------------------------------
We first consider distribution functions with isotropic velocity dispersions. In Fig. \[fig:1\] we plot with a solid line the function $g(v^{\prime})$ defined above in case of a NFW profile, assuming the galactocentric distance to be $R_0 = 8\; {\rm kpc}$ and $\Theta_0$ as derived from Eq. (\[eq:theta\]). There are two solid lines in the figure; the one which is higher at the peak refers to the function $g$ in December, while the second one is appropriate for June. As shown in the previous Section, the amplitude ${\cal{A}}$ of the annual modulation is proportional to the difference between June and December of the integral of $g$ above the value $v_{\rm{min}}$, which in turn depends on the energy deposited in the detector and WIMP and nucleus masses. As a visual aid to identify which are the relevant portions of the curves in each case, we plot in the figure the value of $v_{\rm{min}}$ for a Germanium detector and a few values of $Q$ and $M_{\chi}$ (e.g. $v_{\rm{min}}(Q=30\;{\rm keV}, M_{\chi}= 60 \;{\rm GeV})$ is given by the abscissa of the point at the intersection between the horizontal dotted line labeled $Q=30\;{\rm keV}$ and the vertical dotted line labeled $M_{\chi}= 60 \;{\rm GeV}$). Analogous plots for Na and I are given in Figs. \[fig:3\] and \[fig:4\].
In Fig. \[fig:2\] we plot the predicted annual-modulation amplitude as a function of $Q$, for this NFW profile, for a Germanium detector and for four sample values for the WIMP mass. As known from previous analyses, the modulation amplitude changes sign going to higher values of the deposited energy. At least for low-mass WIMPs, the largest values of ${\cal{A}}$ correspond to the largest displayed value of $Q$. Note however that at such large $Q$s the differential rate is almost negligible (being suppressed by the form factor ${\cal F}$).
To compare with the case previous analyses focussed on, we display in Fig. \[fig:1\] the functions $g$ expected for an isothermal distribution function. The value for the velocity dispersion $\sigma$ is assumed accordingly to the naive (in the sense that it does not correspond to a self-consistent solution) prescription $\sigma = \sqrt{3/2} \;\Theta(R=\infty)$ and $\Theta(R=\infty) = \Theta_0$. We find a fairly good agreement with the NFW case and hence a consistency as well in the values of the modulation amplitude in Fig. \[fig:2\].
We have checked that the effect we are trying to address does not depend sensitively on the steepness of the profile towards the Galactic center. The Moore et al. profile in Table \[tab:rho\] gives curves for $g$ barely distinguishable from the NFW curves in Fig. \[fig:1\]. This is because these two profiles have similar amounts of dark matter inside $R_0$ and analogous ratios of dark to luminous matter. The Kravtsov et al. profile in Table \[tab:rho\] has best-fit values for the local halo density $\rho_0$ and for length scale $a$ respectively higher and lower than in the previous cases; the dark matter happens then to be appreciably more concentrated toward the inner part of the Galaxy. In this case the velocity dispersion gets larger and hence values for the modulation amplitudes are reduced. We go in the direction of a slightly larger velocity dispersion also by dropping the hypothesis of having a “spherical disk”. We claim this in analogy to the isothermal case where it is relatively easy to construct a self-consistent solution for a thin disk and a flattened dark halo; this system has a velocity dispersion somewhat larger than in the purely spherical case.
Anisotropic Velocity Distributions
----------------------------------
We sketch now what happens in case of anisotropy in the velocity dispersion tensor. As mentioned in Section \[sec:df\], different approaches are possible; we consider the Osipkov-Merritt models applied to the NFW density profile introduced above. This will turn out to be sufficient to address the main qualitative effects. We suppose that the distribution function favors radial velocities. To illustrate the effects of anisotropy in the velocity distribution, we consider values for the anisotropy parameter $\beta(R_0)$ in the range (0, 0.48). The upper value is close to the value of $\beta(R_0)$ above which, with our particular choice of potential and dark-matter-density profile, the Osipkov-Merritt scheme breaks down.
In Fig. \[fig:3\] we plot the forms for the function $g$ corresponding to $\beta(R_0) = 0.2$ and 0.4, as well as $\beta(R_0) = 0$. In Fig. \[fig:4\] we plot instead the $\beta(R_0) = 0.3$ and 0.48 cases. It is evident that detector rest-frame values of the WIMP kinetic energy are, in anisotropic models, significantly redistributed (even though the spherically symmetric density profile remains unaltered). The enhancement at large $v^{\prime}$ is due to the fact that in these models there is a higher probability to have a contribution to the signal from particles on very elongated and nearly radial orbits (i.e. particles with $\cal{Q}$ close to zero; distribution functions analogous to the case considered here are given in Fig. 5 of Ref. [@Widrow]). Obviously this implies that the recoil energy spectra changes to some extent. However, without knowing the WIMP mass it may be hard, in case of a detection, to tell one spectrum induced by an anisotropic distribution from another due to an isotropic population for a different WIMP mass.
Regarding instead the annual-modulation signature, the effect can be quite dramatic. In Fig \[fig:5\] we show the modulation amplitude for a NaI detector, plotting on the horizontal axis the electron equivalent energy $Q_{ee}$, rather than $Q$, in the range interesting for the [Dama]{} experiment (to derive this plot we assumed quenching factors and Woods-Saxon form factors as suggested by the [Dama]{} Collaboration [@DAMA]). As can be seen, for $\beta(R_0) = 0.4$, which corresponds to a local velocity ellipsoid for dark-matter particles that is still relatively close to spherical (the axis ratio is equal to 0.7), the modulation amplitudes in a NaI detector are severely damped. For intermediate neutralino masses and values of the recoil energy, the amplitude even changes sign. The $\beta(R_0) = 0.2$ case lies between the isotropic and 0.4 models. Still, the influence on the annual modulation is still rather large. Results for a Ge detector and in case of $\beta(R_0) = 0.48$ and 0.3 are shown in Fig \[fig:6\] and are analogous to the NaI case.
Conclusions
===========
We have investigated the possible effects on an annual-modulation signal of additional structure in the WIMP velocity distribution beyond the canonical Maxwell-Boltzmann distribution. To do so, we have considered isotropic distribution functions that correspond to density profiles other than the isothermal profile that goes with a Maxwell-Boltzmann velocity distribution, as well as some simple but plausible anisotropic velocity distribution functions.
The measured local rotation curve of the Milky Way fixes the velocity dispersion of any consistent dark-matter phase-space distribution. Thus, the [*total*]{} detection rate, integrated over all recoil energies, is expected to be independent of the detailed form of the velocity distribution. However, uncertainties in the velocity distribution can lead to larger uncertainties in the predicted detection rate if a signal is dominated primarily by events in a small recoil-energy bin, as occurs, for example, in the [Dama]{} annual-modulation signal. Moreover, the [*sign*]{}, as well as the magnitude, of the annual modulation can be changed. Thus, the constraints to the WIMP mass that are inferred from the sign of the modulation in [Dama]{} may be loosened if we allow for some structure in the phase-space distribution.
The anisotropic velocity distributions we used were chosen as they provide simple deviations from a Maxwell-Boltzmann distribution that illustrate our point. There are other possibilities for the phase-space distribution that are more complicated, although perhaps better motivated. For example, the existence of an NFW profile—rather than an isothermal profile—in numerical simulations suggests that phase mixing via violent relaxation is not fully efficient in gravitational collapse. If so, then some of the pre-collapse phase-space structure (recall that the pre-collapse phase-space structure of cold dark matter is very highly peaked around zero velocity) should be preserved. It is thus reasonable to expect some clumping in velocity space, even if the halo is smooth in physical space. Thus, for example, the [*local*]{} velocity distribution might be highly anisotropic even if the velocity distribution averaged over a larger volume of the Galaxy is isotropic. Although numerical simulations will be required to quantify this further, it is important to note the possible implications for WIMP-detection rates with the simple models we have considered.
We would like to thank George Lake for discussions. This work was supported in part by NSF AST-0096023, NASA NAG5-8506, and DoE DE-FG03-92-ER40701.
[99]{}
G. Jungman, M. Kamionkowski, and K. Griest, \[\].
M. W. Goodman and E. Witten, ;\
I. Wasserman, ;\
K. Griest, .
A. Drukier, K. Freese, and D. Spergel, ;\
K. Freese, J. Frieman, and A. Gould, .
K. Griest and J. Silk, [*Nature* [**343**]{} (1990) 26]{};\
L.M. Krauss, ;\
T. Falk, K.A. Olive, and M. Srenicki, .
L. Bergstrom and P. Gondolo, [*Astropart. Phys.* [**5**]{} (1996) 263]{} \[\];\
A. Bottino, F. Donato, N. Fornengo, and S. Scopel, \[\].
R. Bernabei et al. ([Dama]{} Collaboration), preprint ROM2F/2000/01 and INFN/AE-00/01.
R. Abusaidi et al. ([Cdms]{} Collaboration), \[\].
K. Griest, .
M. Brhlik and L. Roszkowski, \[\].
P. Belli et al., \[\].
F.C. van den Bosch, G.F. Lewis, G. Lake, and J. Stadel, [*ApJ* [**515**]{} (1999) 56]{} \[\].
M. Odenkirchen, P. Brosche, M. Geffert, and H.J. Tucholke, [*New A* [**2**]{} (1997) 477]{}.
M. Kamionkowski and A. Kinkhabwala, \[\].
J.F. Navarro, C.S. Frenk, and S.D.M. White, [*ApJ* [**462**]{} (1996) 563]{} \[\].
B. Moore et al., [*MNRAS* [**310**]{} (1999) 1147]{} \[\].
A.V. Kravtsov et al., [*ApJ* [**502**]{} (1998) 48]{} \[\].
J.S. Bullock et al., \[\].
P. Ullio, in preparation.
J. Binney and S. Tremaine, *Galactic Dynamics*, Princeton University Press, 1987.
L.P. Osipkov, [*Pis’ma Astron.* [**55**]{} (1979) 77]{}.
D. Merritt, [*AJ* [**90**]{} (1985) 1027]{}.
M. Feast and P. Whiterlock, [*MNRAS* [**291**]{} (1997) 683]{} \[\].
L.M. Widrow, \[\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The MGT fluid model has been used extensively to guide designs of AQM schemes aiming to alleviate adverse effects of Internet congestion. In this paper, we provide a new analysis of a TCP/AQM system that aims to improve the accuracy of the MGT fluid model especially in heavy traffic conditions. The analysis is based on the consideration of two extreme congestion scenarios that leads to the derivation of upper and lower bounds for the queue length and marking probability dynamics and showing that they approach each other in steady state. Both discrete and continuous time models are provided. Simulation results demonstrate that the new model achieves a significantly higher level of accuracy than a simplified version of the MGT fluid model.'
author:
- 'Qin Xu, Fan Li, Jinsheng Sun, and Moshe Zukerman, [^1][^2]'
bibliography:
- 'xu\_fan.bib'
title: 'A New TCP/AQM System Analysis'
---
MGT fluid model, control system, networking, congestion control, TCP/AQM
INTRODUCTION {#sec:Intro}
============
For over three decades researchers have been seeking efficient and distributed means to control Internet traffic congestion. The common approach has been based on congestion indications generated by links and sent to their sources which then adapt their send rate. The scheme implemented at the links which decides on how and when to indicate congestion is called Active Queue Management (AQM) and the protocol that dictates how to adapt sources rate is mainly TCP (herein TCP means TCP Reno). The combined TCP/AQM system aims to achieve efficient resource utilization, acceptable packet loss and a stable and robust operation.
Congestion indication can be done by either dropping or marking incoming packets. If Explicit Congestion Notifications (ECN) [@Floyd:1994:ECN] is enabled, the packets will be marked instead of being discarded, which increases TCP goodput. Given this advantage of ECN, we focus in this paper on the AQM schemes with ECN enabled unless otherwise mentioned.
Many AQM schemes have been proposed. One way to classify AQM schemes is according to their congestion detection method into queue-based, rate-based and queue and rate-based combined [@2004:Survey]. RED [@1993:Floyd:RED] and PI [@Hollot2001PI] are queue-based. Green [@2002:Wydrowski:Green], BLUE [@Feng:2002:BLUE] and AVQ [@Kunniyur:2001:AVQ] belong to the rate-based class. The combined queue and rate-based category includes REM [@Athuraliya:2001:REM], Yellow [@2005:Long:Yellow], and RaQ [@Sun2007RaQ]. An alternative classification into event- and time-driven AQM schemes is based on the method of updating marking probability [@Suzer:2012:event]. RED is a typical event-driven AQM scheme. The marking probability of RED is reset at packet arrival events. AVQ and E-AVQ [@Wang2012adaptiveAVQ] are also event-driven where resets take place at packet arrivals. BLUE [@Feng:2002:BLUE] is event-driven as it updates marking probability when the buffer is empty or full. However, many AQM schemes are time-driven, including PI, PID [@Fan:2003:PID], REM, RaQ. AQM schemes of this type update the marking probability when a certain timeout expires. Table \[table:classificatioin\] provides classifications of various AQM schemes.
[|c|p[2cm]{}|p[2cm]{}|]{} & Event-Driven & Time-Driven\
Queue Base & RED [@1993:Floyd:RED] FPID [@2012:chen:FPID]& PI [@Hollot2001PI] PID [@Fan:2003:PID] MPAQM [@2011:Wang:MPAQM]\
Rate Based & Green [@2002:Wydrowski:Green] BLUE [@Feng:2002:BLUE] AVQ [@Kunniyur:2001:AVQ] &\
----------------------
Queue and Rate Based
----------------------
: The Classification of AQM Schemes[]{data-label="table:classificatioin"}
& Yellow [@2005:Long:Yellow] & REM [@Athuraliya:2001:REM] RaQ [@Sun2007RaQ]\
Designing efficient, stable and robust TCP/AQM systems requires the understanding of their dynamics. To achieve such understanding, various models have been proposed [@Mathis:1997:model; @Kelly:1998:Model; @Low:2003:model]. However, the model of [@Mathis:1997:model] is based on certain simplified assumptions which introduce significant inaccuracies, the models proposed in [@Kelly:1998:Model] and [@Low:2003:model] are not scalable. Misra, Gong and Towsley [@Misra:2000:model] proposed an analytical model for TCP/AQM systems which is called in [@Ajmone2005] the [*MGT fluid model*]{} after the authors’ initials. A simplified version of the MGT fluid model was provided in [@Hollot:2001:linearizemodel] and has been widely used. It provided theoretical foundations for numerous analyzes and syntheses of TCP/AQM systems and gave rise to many newly proposed AQM schemes, e.g. [@Yu:2011:robust; @Wang2008ImcPID].
In this paper, we use the model of [@Hollot:2001:linearizemodel] as a benchmark and we refer to it as the [*Simplified MGT*]{} model. It considers TCP/AQM as a feedback system where the number of packets allowed to be sent without acknowledgement by the source (sender), designated as the congestion window, is adapted based on congestion indications generated by the AQM (either as lost or marked packets). However, it introduces certain inaccuracy because it assumes independence between the congestion window and the number of lost or marked packets. The inaccuracy tends to increase with increased load. Furthermore, the MGT model of [@Misra:2000:model] does not consider practical effects, such as the effect of ECN and modes of TCP operation, (e.g. slow start and congestion avoidance).
In our modeling of a TCP/AQM system we overcome the weaknesses of the Simplified MGT model by considering the above mentioned practical effects. An important distinction between our model and the Simplified MGT model is associated with the roles that TCP and AQM play in the feedback TCP/AQM system. The Simplified MGT model adopts the traditional approach, where TCP is the plant and the AQM is the controller that provides feedback to the TCP. In our model their roles are reversed where the TCP/AQM system is viewed as a system where AQM is the plant and TCP is the controller that provides feedback to the AQM.
We analyze the system dynamics by considering two congestion scenarios which give rise to upper and lower bounds of the queue length and marking probability processes, and show that these bounds are close to each other. Furthermore, we derive a continuous-time model based on differential equations that describe the TCP policy and the link queueing process. We also provide an equivalent discrete-time model for ease of implementation. For a given AQM scheme, our model can provide the transient behavior which include both queue length and marking probability. Both the Simplified MGT model and our model are directly applicable to time-driven AQM schemes and could be applicable to an event-driven AQM if it is approximated by a time-driven system. This paper focuses on time-driven AQM schemes and provides numerical results on PI, REM, RaQ which are all time driven.
The remainder of this paper is organized as follows. In Section \[sec:MGT\], the Simplified MGT model is described. In Section \[sec:Analysis\], we provide an extensive analysis that leads to a new model of a TCP/AQM system which yields bounds for queue length and marking probability dynamics. Section \[sec:Numerical\] presents simulation results over a wide range of scenarios and parameter values that validate our new model, compare its accuracy to that of the Simplified MGT model, and illustrate the closeness of the its bounds. Finally, we present our conclusions in Section \[sec:Conclusions\].
THE SIMPLIFIED MGT MODEL {#sec:MGT}
========================
In a TCP/AQM system, TCP adjusts the send rate to avoid congestion according to the congestion indication received from the AQM. Fig. \[fig:bidirection\] shows the working mechanism of a TCP/AQM system. In the figure, the variable $w$ is the congestion window at the TCP end, and $p$ is the marking probability calculated by the AQM.
![Bidirectional relationship of a TCP/AQM system.[]{data-label="fig:bidirection"}](TCPAQM.pdf){width="8cm"}
The original MGT fluid model [@Misra:2000:model] assumed that packet losses in a flow follow a non-homogeneous Poisson process. It considers the individual behaviors TCP sessions and their timeout mechanisms. In the Simplified MGT model [@Hollot:2001:linearizemodel], the authors ignored the TCP timeout mechanism and provided a set of differential equations that describe the dynamics of the TCP window at the sources and the queueing at the link for this case. Then, this Simplified MGT model is written as:
$$\begin{cases}
\ \dot{W}(t)=\dfrac{1}{R(t)} - \dfrac{W(t-R(t))W(t)}{2 R(t)}p(t-R(t))\\
\ \rule{0ex}{2.5em} \dot{q}(t)= -C + N\dfrac{W(t)}{R(t)} ,
\end{cases}
\label{eq:modelw}$$
where $N$ is the number of TCP sessions, $W(t)$ is the expected congestion window size for a TCP session at time $t\geq0$, $R(t)$ is the expected Round Trip Time (RTT) at time $t\geq0$, $p(t)$ is the marking probability at time $t\geq0$, $C$ is the link capacity, and $q(t)$ is the expected link queue length at time $t\geq0$. The operating point $(W_0, q_0, p_0)$ is defined by $\dot{W}(t) = 0$ and $\dot{q}(t) = 0$, where $q_0$ denotes the queue length at the operating point and $$p_0 = \dfrac{2N^2}{R_0^2 C^2},
\label{eq:MGT_p0}$$ $$W_0 = \dfrac{R_0 C}{N},
\label{eq:MGT_W0}$$ where $R_0$ is the RTT at the operating point.
The Simplified MGT model is known to be accurate for light traffic, but it is inaccurate under heavy traffic. This is partly because the model may yield $p_0 > 1$ in heavy traffic, and then if $p_0$ is truncated to 1, the model may not be able to converge to the target queue size, and if it is not truncated, there the marking probability is erroneous and the queue dynamic may exhibit much slower convergence than the real system (see Section \[sec:Numerical\]). Our modeling approach overcomes the $p_0>1$ weakness as discussed below. For brevity, the Simplified MGT model hereafter designates the truncated version of the model. We will also discuss the untruncated version of the Simplified MGT model in Section \[sec:Numerical\].
A NEW TCP/AQM MODEL {#sec:Analysis}
===================
We begin our study of the TCP/AQM system by modeling and analysis of the AQM marking. As mentioned, a time-driven AQM updates its marking probability every sampling interval. This enables the consideration of the sources reactions on the queue length in one sampling period. Next, two scenarios of TCP associated with two extreme congestion levels are considered, which lead to upper and lower bounds for the queue and marking probability dynamics. Both discrete- and continuous-time models are provided. Afterwards, we discuss the effect of ECN on the queue dynamics, followed by an analysis of the operating point where the two scenarios converge at the steady state under the assumption that the entire system is stable. To be able to test the accuracy of the model for a wide range of cases, three non-overlapping and exhaustive congestion levels are then defined and corresponding working scenario(s) are presented. This follows by a discussion on the closeness of the bounds, the system stability and on the setting of a key parameter.
Analysis of an AQM Marking Process {#sub:AQM}
----------------------------------
In most AQM schemes, packets will be marked or dropped according to marking probability and traffic conditions during a sampling interval. In our analysis, we assume that ECN is always ON and the buffers are sufficiently large so that packet dropping is excluded. Then for the marking process during a sample interval, we make the following assumptions:
1. For each packet, there are two mutually exclusive outcomes: marked or unmarked.
2. During each sampling internal, the marking of each packet is independent of the marking of other packets.
3. The marking probability remains constant within the duration of a sampling interval.
Accordingly, given the marking probability and the number of packets that arrive during a sampling interval, the number of marked packets within this interval is a Binomial random variable.
Let $m(t)$ be the number of packets that arrive at the link during the time period from $t$ to $t+ \tau$, where $\tau>0$ is an infinitesimally small time period. Let $p(t)$ be the marking probability. As the number of marked packets is Binomially distributed, its expectation is $m(t)p(t)$, and the expected number of unmarked packets is $m(t)(1-p(t))$. Let the sum of all TCP congestion windows $W_s(t)$ increases by $a(t)$ for every unmarked packet, and decreases by $b(t)$ for every marked packet, at time $t\geq0$. Let $\Delta W_s(t+R(t))$ be defined by: $$\Delta W_s(t+R(t)) \doteq W_s(t+R(t)+\tau)-W_s(t+R(t)) .
\label{eq:delta_W_s}$$ Then, we obtain $$\Delta W_s(t+R(t)) = a(t)m(t)(1-p(t)) + b(t) m(t)p(t).
\label{eq:binomial}$$ The function $R(t)$ here is equivalent to the $R(t)$ of the Simplified MGT model used in (\[eq:modelw\]). It represents the average over all TCP sessions, that are active during the time period from $t$ to $t+\tau$, of their time delay from the moment a packet arrives at a link until its acknowledgement arrives at the sender. This delay is the RTT minus the time it takes the packet to reach the link from the moment it leaves the sender. However, for tractability, $R(t)$ is approximated as the average RTT which is given by $$\label{eq:RTT}
R(t) = T_p + \dfrac{q(t)}{C},$$ where $T_p$ is the average propagation time, $C$ is the link capacity, $q(t)$ is the average queue length during the time period from $t$ to $t+\tau$, and $q(t)/C$ is the mean queueing delay of a packet arriving in that time period. By the approximation we made in (\[eq:RTT\]), our $R(t)$ is the same as that of the MGT Fluid model which is also based on this approximation.
We assume that $m(t)$ takes the form: $$m(t) = \dfrac{W_s(t)}{R(t)} \tau ,
\label{eq:m(t)}$$ where ${W_s(t)}/{R(t)}$ represents the packet arrival rate.
Marked packets may belong to different TCP sessions and therefore have different effects on TCP window sizes, for simplicity, we assume that the processes $a(t)$ and $b(t)$ are the same for all sessions. Appropriate values for them will be their expected values over all sessions which will be obtained through the analysis of TCP dynamics in the next subsection.
Two Working Scenarios of TCP Data Transmission {#sub:TCP}
----------------------------------------------
TCP is composed of four algorithms [@1997RFC2001; @1999RFC2581; @2009RFC5681] which give rise to four data transfer phases: slow start, congestion avoidance, fast retransmit and fast recovery. There are two important variables for these algorithms: congestion window (*cwnd*) that limits the amount of data that TCP can send, and slow start threshold size (*ssthresh*) that determines whether the slow start or congestion avoidance algorithm is used. Slow start (involving exponential growth of *cwnd*) continues until *cwnd* reaches the threshold *ssthresh* which has a lower bound of two packets. Then, congestion avoidance (involving linear *cwnd* growth) is used. When congestion occurs, fast retransmit and fast recovery are used. A typical single-source single-link TCP cwnd process is illustrated in Fig. \[fig:four phases\].
![A TCP cwnd process.[]{data-label="fig:four phases"}](Lightload.pdf){width="50.00000%"}
To introduce our new model, it is convenient to consider the following two scenarios:
- Scenario A: In all the TCP sessions *cwnd* $<$ *ssthresh*, so they are either in slow start, or in fast recovery.
- Scenario B: In all the TCP sessions *cwnd* $\geq$ *ssthresh*, so they are either in congestion avoidance, or in fast recovery.
Scenarios A and B represent two extreme cases where all the TCP sessions behave in a similar way. For tractability, these scenarios exclude cases where in some sessions *cwnd* $<$ *ssthresh* and in others *cwnd* $\geq$ *ssthresh*. Nevertheless, we demonstrate by simulations in Section \[sec:Numerical\] that our analysis of Scenario A and B is sufficient to lead to accurate evaluations for a wide range of system states even if the conditions of neither Scenario A nor B are satisfied.
In the example presented in Fig. \[fig:four phases\], Scenario A takes place at time 0, and then Scenario B takes over at time 8. This example represents a light traffic case involving a single TCP session. As the number of sessions increases, and more sessions need to share a limited capacity, their *cwnds* may be less than 2 packets and therefore less than *ssthresh*, then Scenario A dominates.
To derive the mathematical model of the two scenarios, consider a system where there is a single congested link with capacity $C$. Let $N$ be the number of TCP flows that use this link; they are labeled by $i = 1, 2, \ldots, N$. Then let $W_i(t)$ and $R_i(t)$ denote the window size and the RTT of flow TCP$_i$ ($i = 1, 2, \ldots, N$) at time $t>0$, respectively. In the slow start phase, when an acknowledgement is received by the sender of an unmarked packet, the congestion window will increase by one packet. We obtain: $$a(t) = \sum_{i = 1}^{N} {\dfrac{W_i(t)}{W_s(t )}} = 1 , \qquad t>0,$$ where the ratio ${W_i(t)}/{W_s(t)}$ represents the proportion of total data packets that generated by TCP$_i$.
In the congestion avoidance phase, the congestion window will increase by one *packet* in every RTT, then the increment of the congestion window can be expressed as: $$\label{aca}
a(t) = \sum_{i = 1}^{N} {\frac{W_i(t)}{W_s(t)} \frac{1}{W_i (t+R(t))}} \approx \frac{N}{W_s(t+R(t))} , \qquad t>0.$$ Notice that in (\[aca\]) we make the following approximation: $$\sum_{i = 1}^{N} \frac{W_i(t)}{W_s(t)} \frac{1}{W_i(t+R(t))} \approx \sum_{i = 1}^{N} \frac{W_i(t+R(t))}{W_s(t+R(t))} \frac{1}{W_i(t+R(t))} .$$ This approximation is made to simplify the derivation of the window increment $a(t)$, which has negligible affect as packets marking happens repeatedly and $R(t)$ is relatively small. This approximation is also applicable to the derivation of the window decrement $b(t)$.
In the fast recovery process, the congestion window will decrease by half its size when the sender receives a notification of a marked packet. We have $$b(t) = -\sum_{i = 1}^{N} {\frac{W_i(t)}{W_s(t)} \frac{W_i(t+R(t))}{2}} \approx -\sum_{i = 1}^{N} {\frac{W_i^2(t+R(t))}{2 W_s(t+R(t))}} ,$$ where $$\begin{aligned}
\sum_{i=1}^{N} {\frac{W_i^2(t+R(t))}{W_s(t+R(t))}}\geq \frac{(\sum_{i=1}^{N} {W_i(t+R(t))})^2}{W_s(t+R(t))N}, \nonumber \\
\frac{(\sum_{i=1}^{N} {W_i(t+R(t))})^2}{W_s(t+R(t))N}=\frac{W_s(t+R(t))}{N} .
\label{eq:ro}\end{aligned}$$ Then, we write $$b(t) = -\frac{\rho W_s(t+R(t))}{2N} ,$$ where $\rho$ is a function of $N$ and the $cwnd$ values of all the sessions, and it satisfies $1 \leq \rho \leq N$. For large $N$, the extreme cases of $\rho = 1$ and $\rho = N$ are very rare. Having $\rho=1$ corresponds to the case where the congestion windows $W_i (t+R(t))$s of all TCP sessions are the same and nonzero, and $\rho=N$ corresponds to the case where there is always exactly one TCP session with $\emph{cwnd}>0$ and in all other sessions $cwnd = 0$. In practice, both $N$ and the $cwnd$ values of each session vary, so $\rho$ also varies in time. Nevertheless, we will show by simulations in Section \[sec:Numerical\] that if $N$ is fixed and if a fixed value for $\rho$ is correctly chosen, the analysis based on a fixed $\rho$ leads to accurate results despite the fact that $cwnd$ values vary. We also show that to model a case where $N$ varies, if we can select the right $\rho$ value for each value of $N$, we still obtain accurate results. Obtaining the parameter $\rho$ is a key issue in our modeling, and although we do not have an analytical way to derive it, we argue and demonstrate in the sequel that setting $\rho = 1$ improves on the accuracy of the Simplified MGT model.
Modeling Window Dynamics for Scenarios A and B {#subsec:new_model}
----------------------------------------------
Based on (\[eq:binomial\]), (\[eq:m(t)\]) and the above results for $a(t)$ and $b(t)$, we have the following equation for Scenario A: $$\begin{aligned}
\Delta W_s(t+R(t))=&\dfrac{W_s(t)}{R(t)}(1-p(t))\tau \nonumber \\
&-\dfrac{\rho W_s(t+R(t))W_s(t)}{2N R(t) }p(t) \tau .
\label{eq:SceA}\end{aligned}$$ Similarly, for Scenario B, we obtain: $$\begin{aligned}
\Delta W_s(t+R(t)) =&\dfrac{N W_s(t)}{R(t) W_s(t+R(t))}(1-p(t))\tau \nonumber \\
&-\dfrac{\rho W_s(t+R(t))W_s(t)}{2N R(t) }p(t) \tau .
\label{eq:SceB}\end{aligned}$$
Based on above two equations which give the sum of all TCP congestion windows for the two Scenarios, we can present our new models in both discrete- and continuous-time as described in the following.
### Discrete-time Model
Let time be divided into consecutive fixed-length time intervals each of size $\Delta t$. Let $k$ be an index for these intervals so that the $k$th interval represents the time between $(k-1)\Delta t$ and $k\Delta t$. Recalling our definition of $R(t)$ associated with the TCP sessions that are active between time $t$ and time $t+\Delta t$, let $R(k)$ be the expected RTT of the TCP sessions that are active during the $k$th time interval. Let $n(k)$ be defined as $n(k)=\lfloor R(k)/\Delta t \rfloor$, and $\Delta W_s(k+n(k))$ be defined as $\Delta W_s(k+n(k)) \doteq W_s(k+n(k)+1)-W_s(k+n(k))$. We then obtain the following for Scenario A. $$\begin{aligned}
\Delta W_s(k+n(k)) =& \frac{W_s(k)}{R(k)}(1-p(k)) \Delta t \nonumber \\
&- \frac{\rho W_s(k+n(k))W_s(k)}{2N R(k) }p(k) \Delta t .\end{aligned}$$ For Scenario B, we have: $$\begin{aligned}
\Delta W_s(k+n(k)) =& \frac{N W_s(k)}{R(k) W_s(k+n(k))}(1-p(k)) \Delta t \nonumber \\
&- \frac{\rho W_s(k+n(k))W_s(k)}{2NR(k) }p(k) \Delta t .\end{aligned}$$ In the above two equations, the variables $W_s(k)$ and $p(k)$ are the expected values of the sum of all TCP congestion windows and marking probability at the $k$th time interval, respectively.
### Continuous-time Model
According to the definition of time-derivative, we have: $$\dot{W_s}(t+R(t)) = \lim_{dt\rightarrow 0} \frac{W_s(t+R(t)+dt) - W_s(t+R(t))}{dt} .$$ Recall the (\[eq:SceA\]) and (\[eq:SceB\]), we obtain the following form for Scenario A: $$\dot{W_s}(t+R(t)) = \frac{W_s(t)}{R(t)}(1-p(t)) - \frac{\rho W_s(t+R(t))W_s(t)}{2N R(t)}p(t) .
\label{eq:SceAcon}$$ For the Scenario B, we have: $$\begin{aligned}
\dot{W_s}(t+R(t)) =& \frac{N W_s(t)}{R(t) W_s(t+R(t))}(1-p(t)) \nonumber \\
&- \frac{\rho W_s(t+R(t))W_s(t)}{2N R(t)}p(t) .
\label{eq:SceBcon}\end{aligned}$$
The marking probability $p(k)$ or $p(t)$ depends on the specific AQM scheme, namely, different AQM schemes will update such marking probability based on different measures, for example, PI uses queue length error while REM uses both queue length error and sending rate error. Later in Section \[sec:Numerical\], we demonstrate the effect of marking probability of different AQM schemes on the dynamics of queue length and marking probability. As in our model, Scenarios A and B are used as upper and lower bounds, respectively, for the predicted values of queue lengths and marking probability, it is important to have these bounds close to each other. In Subsection \[sub:Closeness\] we will provide intuitive arguments and a rigorous proof for the closeness of the bounds, and in Section \[sec:Numerical\] we will provide a wide range of numerical results to provide further evidence this closeness.
Effect of ECN on the Queue Dynamics {#sub:ECN}
-----------------------------------
ECN is enabled in some TCP sessions and it is not enabled in others. We will discuss now the different queue dynamics considering operation alternatives when the ECN is ON and OFF. The queue dynamics depends on the packet arrival rate which equals the sum of the send rate of the sources and the packet departure rate which is the link capacity $C$. In the discrete-time model, the queue dynamics is described as: $$\Delta q(k) = (W_s(k)/R(k) - C) \Delta t ,
\label{eq:DeltaQ_Dis_ECN}$$ where $\Delta q(k) \doteq q(k+1)-q(k)$, and $W_s(k)/R(k)$ represents the packet arrival rate.
Equivalently, for the continuous-time model, we obtain, $$\dot{q}(t) = W_s(t)/R(t) - C ,
\label{eq:DeltaQ_Con_ECN}$$ where $\dot{q}(t)$ is the time-derivative of $q(t)$. The first and second term represent the packet arrival and departure rates, respectively.
When the ECN is off, the method of informing the sources of congestion is by dropping packets, instead of marking packets. The queue dynamics, which considers the dropped packets by disabling the ECN, can be expressed in both the discrete- and continuous-time models as: $$\Delta q(k) = (W_s(k)/R(k) - C - p(k)W_s(k)/R(k)) \Delta t ,
\label{eq:DeltaQ_Dis_NECN}$$ and $$\dot{q}(t) = W_s(t)/R(t) - C - p(t)W_s(t)/R(t) ,
\label{eq:DeltaQ_Con_NECN}$$ where the third term corresponds to the dropped packets, which is often ignored in the literature. However, it is important since the operating point can be affected by the dropped packets. This will be discussed next.
The Operating Point {#sub:Opera}
-------------------
The operating point ($W_{s0}$, $q_0$, $p_0$) is defined by $\dot{W_s}(t)=0$ and $\dot{q}(t)=0$ according to the continuous-time model, so that $$\label{eq:w0}
\dot{q}(t)=0 \Rightarrow W_{s0} = \left\{
\begin{array}{rcl}
R_0C , & & \text{if ECN is ON}\\
\rule{0ex}{2em} \dfrac{R_0C}{1-p_0} , & & \text{if ECN is OFF}
\end{array} \right.$$ $$\label{eq:RTT_op}
R_0 = T_{p} + \dfrac{q_0}{C} ,$$ where $W_{s0}$, $q_0$ and $p_0$ are the sum of the congestion windows, the queue length, and the marking (dropping) probability when TCP/AQM system is in steady state, respectively. The value of $W_{s0}$ as shown in (\[eq:w0\]) is dependent on whether ECN is ON or OFF.
In order to obtain $p_0$ for Scenario A, we write $$\dot{W_s}(t)=0 \Rightarrow p_0 = \dfrac{1}{1+\frac{\rho W_{s0}}{2N}} = \dfrac{2N}{2N + \rho W_{s0}} .
\label{eq:p0a}$$ For Scenario B, we obtain $$\dot{W_s}(t)=0 \Rightarrow p_0 = \dfrac{1}{1+\frac{\rho W_{s0}^2}{2N^2}} = \dfrac{2N^2}{2N^2 + \rho W_{s0}^{2}}.
\label{eq:p0b}$$ Note that the value of $p_0$ in (\[eq:p0a\]) and (\[eq:p0b\]) cannot exceed 1, which is an important property that helps improve accuracy over the Simplified MGT model where $p_0$ in (\[eq:MGT\_p0\]) can take values higher than 1. Specifically, equation (\[eq:p0a\]) and (\[eq:p0b\]) correct such error by adding the value $\rho W_{s0}$ or $\rho W_{s0}^2$ to the denominator. Comparing the denominators of (\[eq:MGT\_p0\]) and (\[eq:p0b\]), and recalling that $\rho$ must satisfy $1 \leq \rho \leq N$, we find that $(2N^2 + \rho W_{s0}^2) > W_{s0}^2$, so $p_0$ of Scenario B will be always smaller than $p_0$ of the Simplified MGT model. Furthermore, empirically, we have noticed that we examined that the real $p_0$ was always lower than $p_0$ based on the Scenario B model with $\rho=1$ in all the cases. If we take this empirical result as always true, it leads to an important conclusion that conservatively setting $\rho=1$, Scenario B always yields a better approximation than the Simplified MGT model.
Note that $W_{s0}$, $q_0$ and $p_0$, as well as $\rho$ are steady state parameters, so their values will be the same for a set of AQM schemes which means that the corresponding TCP/AQM systems stabilize at the same operating point, while the transient states will distinguish between the dynamics of the different AQM schemes before they converge to steady-state. This is consistent with the results shown in Section \[sec:Numerical\].
Different Congestion Levels {#sub:ConLev}
---------------------------
As discussed, our model is based on Scenarios A and B which represent two extreme cases. These two scenarios are definitely not exhaustive as there are many cases where in some TCP sessions *cwnd* $<$ *ssthresh* and in others *cwnd* $\geq$ *ssthresh*, and such cases are not covered by scenarios A and B. To demonstrate that our model is applicable in general, we will need to consider cases that are not included in Scenarios A and B, preferably cases that are very different from Scenarios A and B.
When the congestion level is mild, most TCP sessions are in the congestion avoidance phase, where Scenario B represents the system behavior. When the congestion level increases, the sizes of more and more congestion windows become less than two packets, which means that the *ssthresh* must be equal to 2 (since 2 is the lower bound of *ssthresh*). When most sessions are in the slow start phase, the congestion level is considered to be severe, where Scenario A represents the system behavior. Cases where neither Scenario A nor B can entirely represents the system behavior are characterized by many TCP sessions working in slow start phase and many other sessions working in the congestion avoidance phase.
Having discussed the relationship between Scenarios A and B and the congestion level, it is convenient to introduce a measure of the congestion level, denoted $\overline{w}$, defined by $\overline{w}=W_{s0}/N$. This measure is the average congestion window of all sessions at the operating point.
When $\overline{w}<1$, most of the TCP sessions have small *cwnd* which is less than 2 packets. In such case, the *cwnd* is even lower than the lower bound of the *ssthresh* which indicate that the majority of TCP sessions work in slow start phase. The link is said to be *severely* congested in this condition. The condition that *cwnd* is greater than the 2 packet lower bound is the necessary condition for a TCP session to work in congestion avoidance phase. The larger $\overline{w}$ is, the more TCP sessions work in congestion avoidance phase. Therefore, we use the condition $\overline{w}>2$ to indicate a region where the link is said to be *mildly* congested where Scenario B accurately represents the system behavior. Having defined the congestion regions of mild and severe, we define the condition $1 \leq \overline{w} \leq 2$ for the region where the link is [*moderately*]{} congested. Having the two thresholds of $\overline{w}$, i.e, 1 and 2, we now have three non-overlapping and exhaustive regions for the congestion level. Table \[table:congestion level\] shows the partition of the three congestion level regions of $\overline{w}$ and Scenarios. As our Scenarios A and B do not cover the moderate congestion region where $1 \leq \overline{w} \leq 2$, we will use Scenarios A and B as the lower and upper bounds to approximate the system behavior.
Mild Moderate Severe
---------------- ------------------ ---------------------------- ------------------
$\overline{w}$ $\overline{w}>2$ $1\leq\overline{w} \leq 2$ $\overline{w}<1$
Scenario B Both A and B A
: The Classification of Congestion Level[]{data-label="table:congestion level"}
Closeness of the Bounds {#sub:Closeness}
-----------------------
As discussed, Scenarios A and B are applicable to the severe and mild congestion levels, respectively. However, for all the cases in-between these two extreme congestion levels, Scenario A and B models are used as bounds. It is therefore important to discuss their closeness to each other. Here we explain that their closeness can be achieved if we can choose an appropriate $\rho$ parameter for each of the two scenarios and if the system is stable. Then, we show stability for a specific linearized PI AQM system.
Although Scenario A and B use different ways to increase *cwnd*, the queue length and marking probability of the two models converge to the operating point under steady-state conditions if the system is stable. As mentioned, the operating point is defined by $W_{s0}$, $q_0$ and $p_0$. Firstly, the value of $q_0$ depends on the AQM scheme. Hence, the queue length of our model can be stabilized at $q_0$ if the system is stable. Secondly, for a given set of network parameters, the values of $W_{s0}$ of both scenarios are the same according to (\[eq:w0\]). Let $\rho_A$ and $\rho_B$ be the $\rho$ values for Scenario A and B, respectively. Then, if the settings satisfy the following equation: $$\rho_A =\frac{W_{s0}}{N} \rho_B,$$ the values of $p_0$ of both scenarios will be the same according to (\[eq:p0a\]) and (\[eq:p0b\]). Hence, for a given set of network parameters, Scenario A and B can have the same operating point if a suitable value of $\rho$ is set for each scenario. The models of the two scenarios will converge to the same value under the assumption that the system is stable.
Although it is difficult to prove stability of every TCP/AQM system, we consider a particular system based on PI AQM under the settings used in Section \[sec:Numerical\] and in Appendix \[sec:app1\] we provide stability analysis and numerical verification for this TCP/AQM system for the case where the system is linearized at the operating point. Routh-Hurwitz stability criterion is then used to determine whether this TCP/AQM system is stable.
The Parameter $\rho$ {#sub:rho}
--------------------
We have already discussed the importance of setting $\rho$ to achieve accuracy and closeness of the bounds. As discussed, we do not have an analytical method to obtain it. Nevertheless, we know that it is bounded within $1 \leq \rho \leq N$ and that simply and conservatively setting $\rho = 1$ gives improved results over the Simplified MGT model, which will be demonstrated through a large sample of cases over a wide range of parameters in the next section. To appreciate the difficulty in obtaining $\rho$, notice that $\rho$ could have been obtained by (\[eq:w0\]), (\[eq:RTT\_op\]), (\[eq:p0a\]), and (\[eq:p0b\]) if $p_0$ is available analytically. However, obtaining $p_0$ analytically is a well known open problem. By the same equations, if $\rho$ is analytically available then $p_0$ can be obtained. This explain that obtaining $\rho$ is equivalent to the known difficult problem of finding $p_0$.
One way is to obtain $p_0$ (and $\rho$) is by simulation. Such values will be used in many simulations in the next section to demonstrate how accurate our model could be if we were able to obtain $\rho$ analytically. As discussed, the results for the cases of $\rho=1$ will also be discussed and compared with. As mentioned, $\rho$ is a steady-state parameter, therefore if the parameter $\rho$ is correct for one AQM scheme, it can be reused for the other AQM schemes who has the same operating point. This property will be illustrated in the next section.
NUMERICAL RESULTS {#sec:Numerical}
=================
In this section, we show the improvement of the proposed model compared to the Simplified MGT model for two settings of the parameter $\rho$: one is the conservative case of $\rho=1$ and the other is the case where $\rho$ is obtained based on the NS2 simulation results.
We first introduce the network topology used for the NS2 simulation, followed by information on the setting of the default parameters of three AQM schemes: PI, REM and RaQ. Then, we conduct a series of simulations to evaluate the new model by comparing the analytical results with simulation results. All the time-dependent analytical results are based on the discrete-time model of Scenarios A and B and a discrete time version of the Simplified MGT model. In all the figures in this section that show comparative simulation results, for clarity, we use the label Simplified MGT to represent the results obtained for the Simplified MGT model, we use the labels Scenario A, Scenario B, with $\rho = 1$ to represent the analytical results obtained for Scenarios A and B setting $\rho = 1$, and equivalently we use the labels Scenario A, Scenario B together with the specific $\rho$ value obtained from the NS2 simulations to represent an ideal analytical results for Scenarios A and B indicating the limit of possible improvement based on our model. We observe in all the numerical results that all Scenario A and B curves (including the cases of $\rho=1$) converge to the target queue length. This is consistent with our discussions in Subsections \[sub:Closeness\].
Experiment Topology and Default Parameters {#sub:topology}
------------------------------------------
In our simulations, we use a dumb-bell network topology shown in Fig. \[fig:topology\]. The link between router B and router C is the bottleneck link. Unless mentioned otherwise, the following parameters are set as default: the mean packet size is 1000 bytes, the propagation time is 100 ms, the bottleneck link capacity is 45 Mb/s, the buffer size is 1125 packets, the target queue length is 500 packets, and the number of TCP sessions is 500. TCP/Reno is adopted as the TCP plant and ECN is set ON unless otherwise stated.
![The topology of simulation network.[]{data-label="fig:topology"}](topology.pdf){width="\linewidth"}
The default parameters of the AQM schemes are set as follows: For PI, $a = 0.00001822$ and $b = 0.00001816$. For REM, $\gamma = 0.001$ and $\Phi=1.001$. For RaQ, $q_{kp} = 0.0077$, $q_{ki} = 0.0005$, and $r_{kp} = 0.0095$. The above PI [@Hollot2001PI] and REM [@Athuraliya:2001:REM] parameters are the NS2 defaults and the RaQ parameters are as in [@Sun2007RaQ].
The default sampling period $\Delta t$ used for the analytical results for Scenarios A and B and the Simplified MGT model is 0.0005 s, and the default sampling period $T$ used for the three AQM schemes PI, REM and RaQ is 0.005 s. Note that $\Delta t$ should be smaller than or equal to $T$. Further discussion of these two sampling periods will be given in Subsection \[subsec:step time\] below.
Performance Under Different Congestion Levels {#sub:Perfo_ConLev}
---------------------------------------------
As discussed, our analytical model that are based on Scenario A and B should be tested for a wide range of congestion levels. To this end, we consider five cases, three of which represent the three congestion levels: mild, severe and moderate, and the other two represent borderline cases *mild/moderate* and *moderate/severe* that approximately correspond to the cases $\overline{w}=2$ and $\overline{w}=1$, respectively. These five cases are described in Table \[table:simulation cases\]. The number of TCP sessions $N$ is adjusted to achieve different congestion levels. In particular, we use the $N$ values of 200, 500, 800, 1100 and 2000. For example, in the first case, of $N = 200$, $\overline{w} = CR_0/N=5625*0.1889/200 = 5.3128$, which is more than 2 packets and therefore, this case is within the mild congestion level region.
$N$ $\overline{w}$ Congestion level
------ ---------------- ------------------
200 5.3128 Mild
500 2.1251 Mild/Moderate
800 1.3282 Moderate
1100 0.9660 Moderate/Severe
2000 0.5313 Severe
: Congestion Levels and the Corresponding $N$[]{data-label="table:simulation cases"}
### Mild Congestion
In this case, we consider PI as the AQM scheme and $N=200$. As this case falls within the mild congestion region, it is well modeled by Scenario B. Based on NS2 simulations, we obtain $p_0 = 0.0442$, which gives by (\[eq:SceB\]), $\rho = 1.5318$. By (\[eq:MGT\_p0\]) and (\[eq:p0b\]), the $p_0$ values of the Simplified MGT model and the Scenario B model (with $\rho=1$) are given in Table \[table:p0\_Mild\]. We reuse the value $\rho=1.5318$ in the other two AQM schemes REM and RaQ. Results for all three AQM schemes are presented in Fig. \[fig:PI200\], Fig. \[fig:REM200\] and Fig. \[fig:RaQ200\], for PI, REM and RaQ, respectively. We observe that the results of all analytical models for all three AQM schemes are fairly accurate in both in steady state and during the transient periods before steady state. For example, all analytical results for all three AQM schemes exhibit convergence to the queue length target of 500 packets. In addition, we notice for all three AQM schemes that the marking probability dynamics based on Scenario B with $\rho=1$ is somewhat more accurate than the results obtained by the Simplified MGT model which is consistent with the steady state results shown in Table \[table:p0\_Mild\].
[|c|c|c|c|]{} $N$ &
------------
Simplified
MGT
------------
: Values of $p_0$ for $N = 200$.[]{data-label="table:p0_Mild"}
&
------------
Scenario B
($\rho=1$)
------------
: Values of $p_0$ for $N = 200$.[]{data-label="table:p0_Mild"}
& NS2\
200 & 0.0708 & 0.0662 & 0.0442\
![Queue length and marking probability versus time for PI AQM with $N = 200$.[]{data-label="fig:PI200"}](PI200.pdf){width="50.00000%"}
![Queue length and marking probability versus time for REM AQM with $N = 200$.[]{data-label="fig:REM200"}](REM200.pdf){width="50.00000%"}
![Queue length and marking probability versus time for RaQ AQM with $N = 200$.[]{data-label="fig:RaQ200"}](DEM200.pdf){width="50.00000%"}
### Severe Congestion
In this case, we again first consider PI as the AQM scheme and we set $N=2000$. This case falls within the severe congestion region, so it is modeled by Scenario A. Based on NS2 simulations, we obtain $p_0 = 0.4879$, which gives by (\[eq:SceA\]), $\rho = 3.9516$. By (\[eq:MGT\_p0\]) and (\[eq:p0a\]), the $p_0$ values of the Simplified MGT model and Scenario A (with $\rho=1$) are given in Table \[table:p0\_Severe\]. We notice that the $p_0=0.7901$ predicted by Scenario A (with $\rho=1$) improves the $p_0=7.0827$ of the Simplified MGT model, while still it is not close to the NS2 $p_0 = 0.4879$. We reuse the value $\rho=3.9516$ in the other two AQM schemes REM and RaQ. Results for all three AQM schemes are presented in Fig. \[fig:PI2000\], Fig. \[fig:REM2000\] and Fig. \[fig:RaQ2000\], for PI, REM and RaQ, respectively.
We observe that the results of Scenario A (with $\rho=1$) for all three AQM schemes improve the accuracy both in steady state and during the transient periods. We observe in Fig. \[fig:PI2000\] (a), that although the curve of Scenario A ($\rho=1$) does not very accurately predict the transient period, it converges to the target queue length of 500 packets, despite the fact that the $p_0$ is inaccurate. By comparison, the Simplified MGT model predicts that the queue length saturates the buffer. As for the marking probability dynamics, the curve of Scenario A (with $\rho=1$) converges to the value 0.7901 after 100 second, while the Simplified MGT model converges to 1 only because it is truncated to 1. In the other two figures associated with REM and RaQ AQM, Scenario A (with $\rho=1$) again converges to the target queue length and provides more accurate marking probability than the Simplified MGT model. The results also demonstrate that Scenario A (with $\rho=3.9516$) provides curves closest to the curves of the NS2 simulations by comparison to the Simplified MGT model and Scenario A (with $\rho=1$) for both queue length dynamics and marking probability.
We also presented the result of MGT model without truncation (for both marking probability and queue length) in Fig. \[fig:untruncated\]. It is shown that the queue length for RaQ or PI can converge to the target queue length of 500 packets, while the marking probability is stabilized to more than 7. However, the marking probability of REM does not exceed 1 because its calculation is based on the function $1- \Phi^{-pl}$, but the consequence is that the queue length cannot converge to the target queue length. We also present the queue length dynamics of the Simplified MGT model without truncation in Fig. \[fig:RaQ2000\]. We observe that although some improvement in the modeling of queue length dynamic is achieved without truncation, relative to the truncated counterpart, the accuracy under severe congestion is not satisfactory. Note that the convergence of the untruncated Simplified MGT for PI takes 2000 seconds for the queue size and 4000 for the marking probability, so the results are not included in Fig. \[fig:PI2000\]. We also exclude the untruncated MGT results from Fig. \[fig:REM2000\] because they do not improved on their truncated counterparts.
[|c|c|c|c|]{} $N$ &
------------
Simplified
MGT
------------
: Values of $p_0$ for $N=2000$[]{data-label="table:p0_Severe"}
&
------------
Scenario A
($\rho=1$)
------------
: Values of $p_0$ for $N=2000$[]{data-label="table:p0_Severe"}
& NS2\
2000 & 7.0827 & 0.7901 & 0.4879\
![Queue length and marking probability of untruncated MGT model for PI, REM and RaQ schemes with $N=2000$.[]{data-label="fig:untruncated"}](untruncated.pdf){width="50.00000%"}
![Queue length and marking probability versus time for PI AQM with $N = 2000$.[]{data-label="fig:PI2000"}](PI2000.pdf){width="50.00000%"}
![Queue length and marking probability versus time for REM AQM with $N = 2000$.[]{data-label="fig:REM2000"}](REM2000.pdf){width="50.00000%"}
![Queue length and marking probability versus time for RaQ AQM with $N = 2000$.[]{data-label="fig:RaQ2000"}](DEM2000new.pdf){width="50.00000%"}
### Mild/Moderate, Moderate and Moderate/Severe Congestion
We now consider the congestion levels associate with the cases $N=500, 800, 1100$. For each case, we obtain $p_0$ from the NS2 simulations of PI AQM, and derive the $\rho$ values from (\[eq:p0a\]) and (\[eq:p0b\]) provided in Table \[table:ro\_moderate\]. We also obtain $p_0$ of the Simplified MGT model and Scenarios A and B for $\rho=1$ by (\[eq:MGT\_p0\]), (\[eq:p0a\]) and (\[eq:p0b\]), respectively. The values are presented in Table \[table:p0\_moderate\]. Fig. \[fig:PI500\], Fig. \[fig:PI800\] and Fig. \[fig:PI1100\] show the numerical results of the NS2 simulation and of the analytical models. Comparing to the results of NS2 simulations, it can be found that the convergence time of queue length and marking probability at the steady state increase when $N$ increases. For the cases that $\rho$ are set to the values based on NS2 simulation (Table \[table:ro\_moderate\]), the results show that, as expected, Scenario B performs better than Scenario A near the bound of Mild/Moderate, while Scenario A performs better near the bound of Moderate/Severe. In the moderate congestion range, the curves of NS2 simulations are between the Scenario B and Scenario A curves. Therefore, these results confirm that the milder the congestion level is, the better Scenario B model performs, and vice versa. This conclusion also applies to the cases where $\rho=1$. In addition, Scenario A and B curves are approaching each other in the cases where the $\rho$ values are set based on the NS2 simulations. This is consistent with the discussion in Subsection \[sub:Closeness\]. Also Scenario A and B curves with $\rho=1$ are reasonable close to each other for queue dynamic results of the three AQM schemes.
[|c|c|c|c|c|]{} $N$ &
------------
Simplified
MGT
------------
: Values of $p_0$ for $N = 500, 800, 1100$[]{data-label="table:p0_moderate"}
&
------------
Scenario B
($\rho=1$)
------------
: Values of $p_0$ for $N = 500, 800, 1100$[]{data-label="table:p0_moderate"}
&
------------
Scenario A
($\rho=1$)
------------
: Values of $p_0$ for $N = 500, 800, 1100$[]{data-label="table:p0_moderate"}
& NS2\
500 & 0.4429 & 0.3069 & 0.4848 & 0.2004\
800 & 1.1337 & 0.5313 & 0.6009 & 0.3504\
1100 & 2.1434 & 0.6819 & 0.6743 & 0.4212\
$N$ $\rho$ for Scenario B $\rho$ for Scenario A
------ ----------------------- -----------------------
500 1.7670 3.7551
800 2.1022 2.7921
1100 2.9450 2.8448
: Parameter $\rho$ Of New Model For $N = 500, 800, 1100$ Based On NS2 Simulations[]{data-label="table:ro_moderate"}
![Queue length and marking probability versus time for PI AQM with $N = 500$.[]{data-label="fig:PI500"}](PI500.pdf){width="50.00000%"}
![Queue length and marking probability versus time for PI AQM with $N = 800$.[]{data-label="fig:PI800"}](PI800.pdf){width="50.00000%"}
![Queue length and marking probability versus time for PI AQM with $N = 1100$.[]{data-label="fig:PI1100"}](PI1100.pdf){width="50.00000%"}
Turning OFF ECN
---------------
In the case where ECN is turned off, by NS2 simulations, we obtain $p_0 = 0.1416$. The congestion level index is $\overline{w} = CR_0/(N-Np_0) = 2.4756$. By Table \[table:congestion level\], such congestion level falls into the mildly congested region, so the model of Scenario B is adopted. Then from (\[eq:w0\]) and (\[eq:p0a\]), we obtain $$\rho = \frac{(1-p_0)^3 * 2N^2}{C^2 R_0^2 p_0} = 1.9789.$$
Since the Simplified MGT model does not specifically consider ECN, the previous case of $N=500$, which yields $p_0 (\rm MGT) = 0.4429$ is used here. While the $p_0$ of Scenario B model ($\rho=1$) when turning off ECN is obtained by (\[eq:w0\]) and (\[eq:p0b\]):
$$\label{eq:p_0b_ECN_OFF}
p_0 (Scenario B, ECN\_OFF)= \dfrac{2N^2}{2N^2 + \dfrac{R_{0}^{2}C^{2}}{(1-p_0)^2}} .$$
Isolating $p_0$ and solving this cubic equation, we obtain $p_0 (Scenario B, \rho=1)= 0.2146$. The simulation and analytical results are shown in Fig. \[fig:PI500drop\], Fig. \[fig:REM500drop\] and Fig. \[fig:DEM500drop\].
![Queue length and marking probability versus time for PI AQM with $N = 500$ when ECN is OFF.[]{data-label="fig:PI500drop"}](PI500drop.pdf){width="50.00000%"}
![Queue length and marking probability versus time for REM AQM with $N = 500$ when ECN is OFF.[]{data-label="fig:REM500drop"}](REM500drop.pdf){width="50.00000%"}
![Queue length and marking probability versus time for RaQ AQM with $N = 500$ when ECN is OFF.[]{data-label="fig:DEM500drop"}](DEM500drop.pdf){width="50.00000%"}
The results exhibit that the curves of Scenario B (with $\rho=1$) are closer to the curves of NS2 simulations in both Fig. \[fig:PI500drop\] and Fig. \[fig:REM500drop\] compared to the Simplified MGT model¡¯s results. The marking probabilities are consistent with the theoretical value calculated by (\[eq:p\_0b\_ECN\_OFF\]).
Note that the target queue length is set as 500 packets in NS2 simulations. In Fig. \[fig:REM500drop\], where REM AQM is adopted, the stable queue length of the NS2 simulation is around 450 packets, which is lower than the target queue length (500 packets). In this case, Scenario B ($\rho=1.9789$) can still captures this discrepancy, while the queue length of the Simplified MGT model converges to 500 packets. This fact demonstrates that the correct parameter $\rho$ is reusable under the stability assumption, regardless of which AQM scheme is used. For the other two AQM schemes, Scenario B model ($\rho=1.9789$) provide accurate match for both the queue length and dropping probability dynamic.
Varying Number of TCP Sessions
------------------------------
This set of simulations tests the ability of tracking a link where the number of TCP sessions is varied. Specifically, 300 TCP sessions start at the time 0, 200 additional sessions join at 65 s, and 200 TCP sessions stop at 130 s. According to (\[eq:w0\]), (\[eq:RTT\_op\]) and (\[eq:p0b\]) and based on NS2 simulation, we have the values of $\rho = 1.6575$ and $\rho = 1.7670$ when $N = 300$ and $N = 500$, respectively. By (\[eq:MGT\_p0\]) and (\[eq:p0b\]), the theoretical $p_0$ of Scenario B ($\rho=1$) and the Simplified MGT model for $N=300$ and $N=500$ are given in Table \[table:p0\_varying\].
[|c|c|c|c|]{} $N$ &
------------
Simplified
MGT
------------
: Values of $p_0$ for $N$ Changes[]{data-label="table:p0_varying"}
&
------------
Scenario B
($\rho=1$)
------------
: Values of $p_0$ for $N$ Changes[]{data-label="table:p0_varying"}
& NS2\
300 & 0.1594 & 0.1375 & 0.0877\
500 & 0.4429 & 0.3069 & 0.2004\
![Queue length and marking probability versus time for PI AQM when $N$ varies.[]{data-label="fig:PI300200"}](PI300200.pdf){width="50.00000%"}
![Queue length and marking probability versus time for REM AQM when $N$ varies.[]{data-label="fig:REM300200"}](REM300200.pdf){width="50.00000%"}
![Queue length and marking probability versus time for RaQ AQM when $N$ varies.[]{data-label="fig:RaQ300200"}](DEM300200.pdf){width="50.00000%"}
The results are shown in Fig. \[fig:PI300200\], Fig. \[fig:REM300200\] and Fig. \[fig:RaQ300200\]. In Fig. \[fig:PI300200\] and Fig. \[fig:REM300200\], the marking probability of PI and REM AQM schemes have not reach the theoretical values presented in the above table. This is because the TCP/AQM system has not reached steady state when $N$ changes at times 65 s and 130 s. Nevertheless, the results in Fig. \[fig:RaQ300200\] confirm the theoretical values due to the fast convergence of the RaQ scheme. As expected, in such Mild and Mild/Moderate congestion region when $N$ changes, the curves of Scenario B ($\rho=1$) are closer to the NS2 simulation curves in both queue dynamic and marking probability compared to the curves of the Simplified MGT model.
Different RTT
-------------
Now simulation and analytical results of two different RTTs are presented. The propagation time are set as 0.05 s and 0.15 s, then the corresponding RTTs (\[eq:RTT\_op\]) are $0.05+500/5625=0.1389$ and $0.15+500/5625=0.2389$, respectively. The marking probabilities are 0.2894 and 0.1402 from NS2 simulation, then the corresponding $\rho=2.0107$ and $\rho=1.6984$ can be derived. By (\[eq:MGT\_p0\]) and (\[eq:p0b\]), the $p_0$ of Simplified MGT model and of Scenario B ($\rho=1$) under different RTTs are provided in Table \[table:p0\_RTT\].
[|c|c|c|c|]{} RTTs (s) &
------------
Simplified
MGT
------------
: Values of $p_0$ for Different RTTs[]{data-label="table:p0_RTT"}
&
------------
Scenario B
($\rho=1$)
------------
: Values of $p_0$ for Different RTTs[]{data-label="table:p0_RTT"}
& NS2\
0.1389 & 0.8191 & 0.4503 & 0.2894\
0.2389 & 0.2769 & 0.2168 & 0.1402\
![Queue length and marking probability versus time for PI AQM with propagation time is 0.05 s.[]{data-label="fig:PIR005"}](PIR005.pdf){width="50.00000%"}
![Queue length and marking probability versus time for PI AQM with propagation time is 0.15 s.[]{data-label="fig:PIR015"}](PIR015.pdf){width="50.00000%"}
The simulation results are presented in Fig. \[fig:PIR005\] and Fig. \[fig:PIR015\], it is shown that Scenario B ($\rho=1$) improves the accuracy of queue length and marking probability in both cases relative to the Simplified MGT model. Note that Scenario B ($\rho$=1.6984) in Fig.\[fig:PIR015\] matches NS2 simulations better than Scenario B ($\rho=2.0107$) in Fig.\[fig:PIR005\]. This can be explained by the congestion level. According to (\[eq:w0\]) and $\overline{w} = W_{s0}/N$, we obtain $\overline{w}=1.5626 $ and $\overline{w}=2.6876$ for $RTT=0.1389$ and $RTT=0.2389$, respectively. As shown in Table \[table:congestion level\], the second case falls into the mildly congested region while the first one falls into the moderately congested region, hence the Scenario B model performs better in the second case.
Different Link Capacities
-------------------------
We now consider two cases where the link capacities are 15Mb/s and 95 Mb/s. By (\[eq:MGT\_p0\]) and (\[eq:p0b\]), the $p_0$ of Simplified MGT and of Scenario B model ($\rho=1$) for the two cases are given in Table \[table:p0\_Link\]. From the NS2 simulations, the marking probabilities for the cases of $C=15$ Mb/s and $C=95$ Mb/s are 0.3426 and 0.0973, respectively. Then the corresponding $\rho$ values are 2.0297 and 1.6286, respectively. The results are shown in Fig. \[fig:PIC15M\] and Fig. \[fig:PIC95M\]. In the two figures, the Scenario B curves with the $\rho$ values of 2.0297 and 1.6286 match the NS2 simulation resuls very well, and the curves generated by Scenario B ($\rho=1$) are closer to the NS2 simulation results than the curves of the Simplified MGT model.
[|c|c|c|c|]{}
-----------------
Link
Capacities (Mb)
-----------------
: Values of $p_0$ for Different Link Capacities[]{data-label="table:p0_Link"}
&
------------
Simplified
MGT
------------
: Values of $p_0$ for Different Link Capacities[]{data-label="table:p0_Link"}
&
------------
Scenario B
($\rho=1$)
------------
: Values of $p_0$ for Different Link Capacities[]{data-label="table:p0_Link"}
& NS2\
15 & 1.0577 & 0.5140 & 0.3426\
95 & 0.1756 & 0.1494 & 0.0973\
![Queue length and marking probability versus time for PI AQM with link capacity C is 15 Mb/s.[]{data-label="fig:PIC15M"}](PIC15M.pdf){width="50.00000%"}
![Queue length and marking probability versus time for PI AQM with link capacity C is 95 Mb/s.[]{data-label="fig:PIC95M"}](PIC95M.pdf){width="50.00000%"}
Time Interval $\Delta t$ {#subsec:step time}
------------------------
In practice the marking probability may change during the time interval $\Delta t$, while we assume that it stays constant in our model (and also the discrete-time implementation of the Simplified MGT model). Therefore, the longer $\Delta t$ is, the larger the error it introduces. It is required that $\Delta t$ is bounded above by the AQM sampling period denoted by $T$, to allow sufficient time to update the marking probability before the next sampling period. Here we increase the value of $\Delta t$ from its previous value of 0.0005 s to 0.2 s and the AQM sampling period from 0.005 s to 0.2 s, and we examine the effect on the accuracy of our model and of the Simplified MGT model. Notice that the $\Delta t$ and AQM sampling period are now longer that the RTT which is 0.1889 s by (\[eq:RTT\_op\]). Because of RaQ¡¯s short convergence time, we only present results here for RaQ AQM. The parameter $\rho = 1.7670$ by (\[eq:p0b\]) is used, and Scenario B is chosen of this case. The result is shown in Fig. \[fig:RaQ02\], which demonstrates that even if $\Delta t$ is greater than the RTT, the Scenario B model (with $\rho=1.7670$) still has the capability to track the simulation results well. The performance of the Simplified MGT model is also quite good under the same settings, while it is not as good as that of Scenario B (with $\rho=1$) for both the queue length and marking probability dynamics.
![Queue length and marking probability versus time for RaQ AQM with step time $\Delta t$ is 0.2 s.[]{data-label="fig:RaQ02"}](DEM02.pdf){width="50.00000%"}
CONCLUSIONS {#sec:Conclusions}
===========
In this paper, we have proposed a new model for a time-driven TCP/AQM system. This new model rigorously considers the four algorithms embedded in TCP, and it contributes to performance studies of AQM schemes. We have provided both discrete and continuous-time models. Our proposed model is based on bounds associated with heavy and light congestion, but we have shown that the bounds are tight in steady state and the model is applicable to a wide range of congestion levels. In addition, the model has been validated using extensive simulation results for a wide range of parameters and various AQM schemes and it was demonstrated that it is more accurate than the Simplified MGT model. In fact, the accuracy of our model was consistently demonstrated in all cases studied.
\[sec:app1\] Recalling Scenario A described by (\[eq:SceAcon\]), we define: $$\begin{aligned}
f(W_s, W_{sR}, p_R, q_R)\doteq & \dot{W_s}(t) = \frac{W_{sR}}{T_p + \frac{q_R}{C}} (1-p_R) \nonumber \\
&-\frac{\rho W_s W_{sR} }{2N(T_p + \frac{q_R}{C})} p_R ,\end{aligned}$$
where $W_{sR}(t) \doteq W_s(t-R)$, $p_R(t) \doteq p(t-R)$, $q_R(t) \doteq q(t-R)$.
Taking partial derivatives at the operating point ($W_{s0}, q_0, p_0$) of this model yields: $$\frac{\partial f}{\partial W_s}=-\frac{\rho W_{s0} p_0}{2NR_0}$$
$$\frac{\partial f}{\partial W_{sR}} = \frac{1-p_0}{R_0} - \frac{\rho W_{s0} p_0}{2NR_0}$$
$$\frac{\partial f}{\partial p_R} = -\frac{W_{s0}}{R_0} - \frac{\rho W_{s0}^2}{2NR_0}$$
$$\frac{\partial f}{\partial q_R} = -\frac{W_{s0}(1-p_0)}{R_0^2 C} + \frac{\rho W_{s0}^2 p_0}{2NR_0^2 C} .$$
Hence Scenario A linearized form is expressed as: $$\begin{aligned}
\delta\dot{W_s}(t)= &\frac{\partial f}{\partial W_s} \delta W_s(t) + \frac{\partial f}{\partial W_{sR}} \delta W_s(t-R) \nonumber \\
& + \frac{\partial f}{\partial p_R} \delta p(t-R) + \frac{\partial f}{\partial q_R} \delta q(t-R) ,
\label{eq:linearized model}\end{aligned}$$ where $$\begin{cases}
\ \delta W_s(t) \doteq W_s(t) - W_{s0}\\
\ \delta q(t) \doteq q(t) - q_0\\
\ \delta p(t) \doteq p(t) - p_0.
\end{cases}$$
We linearize Scenario B model in a similar way to (\[eq:linearized model\]), where the partial derivatives are replaced by: $$\begin{cases}
\dfrac{\partial f}{\partial W_s} = -\dfrac{N(1-p_0)}{R_0W_{s0}} - \dfrac{\rho W_{s0} p_0}{2NR_0}\\
\dfrac{\partial f}{\partial W_{sR}} = \dfrac{N(1-p_0)}{R_0W_{s0}} - \dfrac{\rho W_{s0} p_0}{2NR_0}\\
\dfrac{\partial f}{\partial p_R} = -\dfrac{N}{R_0} - \dfrac{\rho W_{s0}^2}{2NR_0}\\
\dfrac{\partial f}{\partial q_R} = -\dfrac{N(1-p_0)}{R_0^2 C} + \dfrac{\rho W_{s0}^2 p_0}{2NR_0^2 C} .
\end{cases}
\label{eq:parameters of linearized B}$$
Linearizing (\[eq:DeltaQ\_Con\_ECN\]) when ENC is enabled gives: $$\delta\dot{q}(t)=\frac{1}{R_0} \delta W_s(t) - \frac{W_{s0}}{R_0^2 C}\delta q(t) .$$
We obtain the block diagram presented in Fig. \[fig:block1\]. We choose PI as the AQM algorithm and we obtain a closed-loop system. In Laplace area, the PI controller has the form $k_p + \frac{k_i}{s}$, where $k_p$ and $k_i$ are the proportion and integral coefficient, respectively. Then the diagram of Fig. \[fig:block1\] is transformed to the diagram of Fig. \[fig:block2\]. For simplicity, we approximate the time delay as the first-order lag [@2006:Franklin] $e^{sR_0} \approx \frac{1}{1+sR_0}$. Thus, we obtain the characteristic equation: $$s^4 + \alpha_1 s^3 + \alpha_2 s^2 + \alpha_3 s + \alpha_4=0 ,$$ where $$\begin{cases}
\ \alpha_1 \doteq \frac{1}{R_0} - \dfrac{\partial f}{\partial W_{s}} + \dfrac{W_{s0}}{R_0^2 C}\\
\ \alpha_2 \doteq -\dfrac{\dfrac{\partial f}{\partial W_{s}}+\dfrac{\partial f}{\partial W_{sR}}}{R_0} + \dfrac{W_{s0}(1-\dfrac{\partial f}{\partial W_{s}} R_0)}{R_0^3 C}\\
\ \alpha_3 \doteq -\dfrac{W_{s0}(\dfrac{\partial f}{\partial W_{s}}+\dfrac{\partial f}{\partial W_{sR}})}{R_0^3 C} - \dfrac{\dfrac{\partial f}{\partial q_R} + \dfrac{\partial f}{\partial p_R} k_p}{R_0^2}\\
\ \alpha_4 \doteq -\dfrac{\dfrac{\partial f}{\partial p_R} k_i}{R_0^2} .
\end{cases}$$
![The block diagram of TCP/AQM system.[]{data-label="fig:block1"}](Block1.pdf){width="50.00000%"}
![Simplified block diagram of TCP/AQM system.[]{data-label="fig:block2"}](Block2.pdf){width="50.00000%"}
According to the Routh stability criterion, the system is stable if and only if $$\begin{cases}
\ \alpha_1 >0\\
\ \beta_1 \doteq \alpha_1 \alpha_2 - \alpha_3 >0\\
\ \beta_2 \doteq \alpha_3(\alpha_1 \alpha_2 - \alpha_3) - \alpha_1^2 \alpha_4>0\\
\ \alpha_4 >0 .
\end{cases}$$
As an example we demonstrate that the settings of the Subsection \[sub:Perfo\_ConLev\] leads to a stable system. Computation based on these settings are given in Table \[tab:stability A\] and \[tab:stability B\], where PI is chosen as the AQM scheme. The values shown in the two tables demonstrate that they all satisfy the conditions, so the system is stable.
$\alpha_1$ $\beta_1$ $\beta_2$ $\alpha_4$
------ ------------ ----------- ----------- ------------
500 14.8205 946.6351 1.2581e+3 0.0472
800 14.0721 799.4057 8.3581e+4 0.0270
1100 13.6513 732.6899 6.7882e+4 0.0225
2000 13.2988 672.6120 5.5029e+4 0.0194
: The Values for Scenario A []{data-label="tab:stability A"}
$\alpha_1$ $\beta_1$ $\beta_2$ $\alpha_4$
------ ------------ ----------- ----------- ------------
200 12.4921 536.3968 3.5160e+4 0.0403
500 45.8829 2.6924e+4 2.6975e+7 0.0422
800 15.7664 1.1551e+3 1.7474e+5 0.0203
1100 16.9312 1.4268e+3 2.6368e+5 0.0232
: The Values for Scenario B []{data-label="tab:stability B"}
ACKNOWLEDGEMENT {#acknowledgement .unnumbered}
===============
This paper was supported by a grant from City University of Hong Kong (Project No. 9380044), and two grants from National Natural Science Foundation of China (Project No. 60974129) and (Project No. 70931002).
[Qin Xu]{} is a PhD candidate in the School of Automation, Nanjing University of Science and Technology, Nanjing, P. R. China. She received her B. Eng degree in electronic information engineering from the same school. Her research interest is network congestion control.
[Fan Li]{} is a research assistant with the department of Electronic Engineering, City University of Hong Kong. He received his MSc. degree from the same department in 2009, and the B.Eng. degree in electronic information engineering from Nanjing University of Science and Technology, Nanjing, P. R. China. His research interests include teletraffic theory, network transmission control, and optical network dimensioning.
[Jinsheng Sun]{}received the B.S., M.S. and Ph.D. degrees in Control Science from Nanjing University of Science and Technology in 1990, 1992 and 1995, respectively. Since 1995, he has been with the Department of Automation, Nanjing University of Science and Technology. Currently, he is a full professor. In 2006 and 2007, he took up a Research Fellow position at the Department of Electrical and Electronic Engineering, The University of Melbourne, Victoria, Australia. His research interests include congestion control and fault-tolerant control.
[Moshe Zukerman]{} (M’87-SM’91-F’07) received his B.Sc. and M.Sc. degrees from the Technion, and his Ph.D. degree from UCLA in 1985. During 1986-1997, he was with the Telstra Research Laboratories, first as a Research Engineer and, in 1988-1997, as a Project Leader. During 1997-2008, he was with The University of Melbourne, Victoria, Australia. In 2008 he joined City University of Hong Kong as a Chair Professor of Information Engineering, and a Team Leader. He has over 250 publications in scientific journals and conference proceedings. He has served on various editorial boards and technical program committees.
[^1]: Q. Xu and J. Sun are with the School of Automation, Nanjing University of Science and Technology, Nanjing, JiangSu Province, P. R. China, 210094 e-mail: [email protected]
[^2]: F. Li and M. Zukerman are with Department of Electronic Engineering, City University of Hong Kong, Hong Kong SAR.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We cast the problem of combinatorial auction design in a Bayesian framework in order to incorporate prior information into the auction process and minimize the number of rounds to convergence. We first develop a generative model of agent valuations and market prices such that clearing prices become maximum a posteriori estimates given observed agent valuations. This generative model then forms the basis of an auction process which alternates between refining estimates of agent valuations and computing candidate clearing prices. We provide an implementation of the auction using assumed density filtering to estimate valuations and expectation maximization to compute prices. An empirical evaluation over a range of valuation domains demonstrates that our Bayesian auction mechanism is highly competitive against the combinatorial clock auction in terms of rounds to convergence, even under the most favorable choices of price increment for this baseline.'
author:
- |
Gianluca Brero\
University of Zurich\
[[email protected]]{} Sébastien Lahaie\
Google Research\
[[email protected]]{}
bibliography:
- 'aaai18.bib'
title: A Bayesian Clearing Mechanism for Combinatorial Auctions
---
Introduction
============
Combinatorial auctions address the problem of allocating multiple distinct items among agents who may view the items as complements or substitutes. In such auctions, agents can place bids on entire packages of items in order to express complex preferences, leading to higher allocative efficiency. Nevertheless, bidding in a combinatorial auction places a substantial cognitive burden on agents, because the process of valuing even a single bundle can be a costly exercise [@kwasnica2005new; @parkes2006mit]. There is therefore great interest in developing *iterative* combinatorial auctions, which help to guide the bidding process using price feedback, and in devising techniques to limit the number of rounds needed to reach convergence (ideally in the dozens rather than hundreds) [@petrakis2012ascending; @bichler2017coalition].
In this work, we propose to incorporate *prior information* on agent valuations into the auction procedure in a principled manner, thereby achieving a low number of rounds in practice. We cast the problem of combinatorial auction design in a Bayesian framework by developing a joint generative model of agent valuations and market prices. Our generative model defines a likelihood function for clearing prices given agent valuations. If these valuations are observed, the maximum a posteriori (MAP) estimate for prices corresponds to market clearing prices. If they remain latent, valuations can be marginalized away, weighed by their own likelihood according to observed bids. This forms the basis for an auction scheme to solve the more general clearing problem where valuations are unknown.
We consider settings where several indivisible items are up for sale, and agents have super-additive valuation functions over bundles of items (i.e., the items are pure complements). We provide an auction implementation using item prices consisting of two components. In the *knowledge update* component, we maintain a Gaussian posterior over agent valuations, which is updated as new bids are placed using assumed density filtering [@opper1998bayesian]. Prior information can be incorporated into the auction by suitably initializing this component. The knowledge update step presumes that agents follow *myopic best-response* strategies and bid on utility-maximizing bundles at each round. Accordingly, we discuss an extension to our auction scheme using multiple price trajectories that incentivizes this behavior in *ex post* Nash equilibrium. In the *price update* component, we obtain an analytical expression for the clearing price objective, based on the Gaussian model of valuations that the auction maintains. We establish that the form of the objective is suitable for optimization using expectation maximization. By alternating the two components, we obtain an intuitive and tractable auction scheme where agents place bids, knowledge over latent valuations is updated given bids, and prices are updated given current knowledge of valuations.
For evaluation purposes, we first illustrate our auction on a stylized instance to gain insight into the auction’s behavior under both unbiased and biased prior information. We then conduct simulation experiments to compare our auction implementation against a combinatorial clock auction that updates prices according to excess demand, which is the standard price update scheme used in practice [@ausubel2014practical]. The prior information in our Bayesian auction is obtained by fitting a Gaussian process prior on a training sample of valuations. The baseline clock auction is parametrized by a step size, or price increment. We find in our experiments that our Bayesian auction is competitive against the strongest possible version of the baseline auction, where the price increment is chosen separately for each instance to lead to the fewest possible rounds. In particular, the Bayesian auction almost matches the strongest possible version of baseline auction in terms of number of instances cleared, and uses fewer rounds on average when it is able to clear. 0 ============================ Motivation:
- The way standard auction designs look for clearing prices is not well principled. Usually candidate prices are updated on the basis of demand/supply observations that do not provide a principled way to establish the intensity of the update.
- Standard auction designs do not incorporate a priori knowledge of agents’ values in a principled way. This may lead to great inefficiencies in the elicitation process, since approximate knowledge of agents’ values may already be sufficient to determine clearing prices.
Papers addressing the problem of high number or rounds in Combinatorial Auctions:
- [@schneider2010robustness]: comparison between different combinatorial auction designs in terms of number of rounds. Non-linear personalized price auctions take lot of rounds (hundreds of them) while linear price auctions take a reasonably small amount of rounds. In the light of their results, they advocate for linear prices. Domains are ad hoc generated and bidders are multi-minded.
- [@petrakis2012ascending]: they introduce pricing rules for combinatorial auctions with non-linear personalized prices based on deadness levels (lower bound to bids which still can become winning in the course of the auction) and winning levels (lowest possible bids which would win if no other bids are submitted). The goal is to decrease the number of rounds.
- [@bichler2017coalition]: they introduce combinatorial auction designs suggesting prices for members of losing coalitions which would make the coalition winning. The goal is again to decrease the number of rounds that, according to them, if too high leads to inefficiencies.
Preliminaries {#sec:preliminaries}
=============
We consider a setting with $m$ distinct and indivisible items, held by a single seller. The items are to be allocated among $n$ agents (i.e., buyers). We will use the notation $[n] = \{1,\dots,n\}$, so that $[n]$ and $[m]$ denote the index sets of agents and items, respectively. There is unit supply of each item. A *bundle* is a subset of the set of items. We associate each bundle with its indicator vector, and denote the set of bundles as ${\mathcal{X}}=\{0,1\}^m$. The component-wise inequality $x \leq x'$ therefore means that bundle $x$ is contained in bundle $x'$. The empty bundle is denoted by $\emptyset$.
Each agent $i$ is *single-minded* so that its valuation can be encoded via a pair $(x_i, w_i)$ where $x_i \in {\mathcal{X}}$ is a bundle and $w_i \in {\mathbb{R}}_+$ is a non-negative value (i.e., willingness to pay) for the bundle. The agent’s valuation function $v_i: {\mathcal{X}}\rightarrow {\mathbb{R}}_+$ is defined as $v_i(x)=w_i$ if $x \geq x_i$, and $v_i(x)=0$ otherwise. In words, the agent only derives positive value if it acquires all the items in $x_i$ (which are therefore complements), and any further item is superfluous. Our auction and results all extend to agents with OR valuations, which are concise representations of super-additive valuations [@nisan2000bidding].[^1] This is due to the fact that an agent with an OR valuation will behave and bid in our auction exactly like a set of single-minded agents, under myopic best-response [@parkes1999bundle]. Under super-additive valuations, items are pure complements, and complementarities are a key motivation for using package bidding. For the sake of simplicity, however, we limit the exposition to single-minded agents.
An *allocation* is represented as a vector of bundles ${\bm{y}}= (y_1,\dots,y_n)$, listing the bundle that each agent obtains (possibly $\emptyset$). An allocation is *feasible* if the listed bundles are pairwise disjoint (i.e., each item is allocated to at most one agent). We denote the set of feasible allocations by ${\mathcal{F}}$. The purpose of running a combinatorial auction is to find an *efficient* allocation of the items to the agents, meaning an allocation that maximizes the total value to the agents.[^2] More formally, a feasible allocation ${\bm{y}}\in {\mathcal{F}}$ is efficient if $
{\bm{y}}\in \operatorname*{arg\,max}_{{\bm{y}}' \in {\mathcal{F}}} \sum_{i \in [n]} v_i(y'_i).
$ However, an iterative auction proceeds via a price adjustment process, so prices will be our central object of study, rather than allocations. The allocation in an iterative auction is adjusted according to agents’ responses to prices.
#### Clearing Prices
In the context of a combinatorial auction, we encode prices as a non-negative function $\theta: {\mathcal{X}}\rightarrow {\mathbb{R}}_+$ over the bundles. We assume that prices are normalized and monotone: $\theta(\emptyset) = 0$, and $\theta(x) \le \theta(x')$ if $x\le x'$. An iterative auction adjusts prices to balance demand and supply. To formalize this notion, we need several additional concepts. We assume that agents have quasi-linear utility, so that the utility to agent $i$ of obtaining bundle $x$ at prices $\theta$ is $v_i(x) - \theta(x)$. The *indirect utility* function provides the maximum utility that agent $i$ can achieve, when faced with prices $\theta$, by choosing among bundles from ${\mathcal{X}}$: $$\label{eq:indirect-buyer}
V_i(\theta) = \max \left\{\, v_i(x) - \theta(x) : x \in {\mathcal{X}}\,\right\}.$$ Note that for single-minded agents, the indirect utility reduces to $V_i(\theta; x_i, w_i) = {\left( w_i - \theta(x_i) \right)_+}$, where the notation ${\left( a \right)_+} = \max\{a, 0\}$ refers to the positive part of the argument. It will sometimes be useful to make explicit the parametrization of the indirect utility on the agent’s type $(x_i, w_i)$, as we have just done. The *demand set* of agent $i$ is defined as $D_i(\theta) = \left\{x \in {\mathcal{X}}: v_i(x) - \theta(x) = V_i(\theta)
\right\}$. Similarly, the *indirect revenue* function provides the maximum revenue that the seller can achieve, when faced with prices $\theta$, by selecting among feasible allocations: $$\label{eq:indirect-seller}
R(\theta) = \max \left\{\, \sum_{i \in [n]} \theta(y_i) : {\bm{y}}\in {\mathcal{F}}\,\right\}.$$ The seller’s *supply set* consists of the feasible allocations that maximize revenue: $S(\theta) = \{ {\bm{y}}\in {\mathcal{F}}: \sum_{i \in [n]} \theta(y_i) = R(\theta) \}.$
We say that prices $\theta$ are *clearing prices* if there is a feasible allocation ${\bm{y}}$ such that, at prices $\theta$, the seller’s revenue is maximized, and each agent’s utility is maximized. Formally, we require the following conditions: ${\bm{y}}\in S(\theta)$ and $y_i \in D_i(\theta)$ for all $i \in [n]$. We say that the clearing prices $\theta$ $\emph{support}$ allocation ${\bm{y}}$.
It is a standard result that the set of allocations supported by any given clearing prices $\theta$ coincides with the set of efficient allocations. (This is a special case of the Fundamental Theorems of Welfare Economics [@mas1995microeconomic 16.C–D].) Moreover, @bikhchandani2002package have shown that clearing prices exist and coincide with the minimizers of the following objective function, which corresponds to the linear programming dual of the problem of allocating the items efficiently: $$\label{eq:clearing-pot}
\sum_{i \in [n]} V_i(\theta) + R(\theta) .$$ This is a piece-wise linear, convex function of the price function $\theta$. Importantly, this result is guaranteed to hold only if the prices are an unrestricted function over the bundles (except for non-negativity and normalization). In practice, it is common to use certain parametrizations for the prices. For instance, taking $\theta(x) = p \cdot x$ for some vector $p \in {\mathbb{R}}^m_+$ corresponds to using linear prices (i.e., item prices). These parametrizations may not achieve the unrestricted minimum in ; in particular, linear clearing prices may not exist. We will use unrestricted prices in the development of our auction, and postpone the question of price parametrization until needed to achieve a practical implementation.
It is useful to view as a potential function that quantifies how close prices $\theta$ are to supporting an efficient allocation. Indeed, if some prices achieve a value of that differs from the optimum by an additive error of $\delta$, then the agents (and seller) can be induced to accept an efficient trade using transfers totaling $\delta$. 0
Let $\theta$ be prices that minimize the potential (\[eq:clearing-pot\]) to within an additive error of $\delta$, and let $x^*$ be an efficient allocation. Then there exist transfers totaling no more than $\delta$ such that each $x_i^*$ maximizes agent $i$’s utility, and $x^*$ maximizes the seller’s profit.
In the following, we will therefore refer to the function $$\label{eq:clearing-potential}
U(\theta; \bm v) = \exp{\left[-\sum_{i \in [n]} V_i(\theta) - R(\theta)\right]}$$ as the *clearing potential* for the valuation profile $ \bm v = (v_1,...,v_n)$, which will capture, in a formal sense, how likely a price function $\theta$ is to clearing valuation profile $\bm v$.
#### Iterative Auction and Incentives
The goal of our paper is to design an iterative auction that exploits the auctioneer’s prior knowledge over agent valuations in order to speed up the clearing process. The auction proceeds over rounds. Agents report their demand at the current prices and, if the market is not cleared, the information provided by the agents is used to update the knowledge about their valuations. Candidate clearing prices are computed based on the updated knowledge, and the procedure iterates. A schematic representation of the auction process is presented in Figure \[fig:BayesianClearingMechanism\]. The knowledge update and price update components constitute the core of the auction that must be implemented.
![Bayesian iterative auction.[]{data-label="fig:BayesianClearingMechanism"}](ClearingMechanism){width="45.00000%"}
The correctness of our auction relies on the agents following a strategy of *myopic best-response* bidding, meaning that each agent bids on a utility-maximizing bundle at each round. There is evidence that myopic bidding may be a reasonable assumption in practice. For instance, in the FCC broadband spectrum auction, jump bids were the exception [@cramton1997fcc]. Nonetheless, a robust auction design should incentivize agents to follow the appropriate strategies. For this purpose, we can use an extension of our auction that maintains $n+1$ price trajectories in order to compute clearing prices when all agents are present, and when each agent is removed in turn. This allows one to compute final VCG payments and bring myopic best-response bidding into an *ex post* Nash equilibrium [@gul2000english; @bikhchandani2006ascending]. The technique of using multiple trajectories was previously used by @ausubel2006efficient and @mishra2007ascending among others. We will provide a more precise treatment of incentives in the formal description of our auction mechanism.
0 rely on a proxy agent architecture which has been commonly used in the literature [@parkes2000preventing; @ito2005new]. Agents submit their valuations to a proxy (e.g., using the OR bidding language), and the proxy software then bids myopically on the agent’s behalf. If upon termination the auction charges the agents VCG payments, then submitting *truthful* valuation to the proxy becomes a dominant strategy for each agent [@gul2000english; @bikhchandani2006ascending]. We will show how our auction can be adapted to compute VCG payments by maintaining $n+1$ price trajectories, an approach previously used by @ausubel2006efficient and @mishra2007ascending among others, and relying on the proxies to finalize the payments.
Generative Model {#sec:generative-model}
================
The purpose of this section is to define a probabilistic relationship between prices and valuations that will allow us to use the auctioneer’s prior knowledge over valuations to make inferences over clearing prices. We write ${\bm{w}}= (w_1,\dots,w_n)$ and ${\bm{x}}= (x_1,\dots,x_n)$ for the vectors of agents’ values and bundles, and denote the probabilistic model as $P({\bm{x}},{\bm{w}},\theta)$. Below, our convention is that $Q$ refers to distributions—possibly unnormalized—that form the building blocks of the generative model, whereas $P$ refers to the normalized distribution resulting from the generative model. We represent the prior knowledge of the auctioneer over agent valuations via the probability density function $Q(\bm w) = \prod_{i\in[n]}Q(w_i)$.
The structure of our probability model is inspired by the work of @sollich2002bayesian, who provides a Bayesian interpretation of the support vector machine (SVM) objective. To establish a proper relationship between prices and valuations, the key is to require that $$\label{eq:generative-model}
P({\bm{x}}, {\bm{w}}, \theta) \propto U(\theta; {\bm{x}}, {\bm{w}})Q({\bm{w}}),$$ where $U(\theta; {\bm{x}}, {\bm{w}})$ is the clearing potential introduced in , adapted to single-minded valuations. Under this joint probability model, we have that the posterior probability of prices $\theta$ takes the form $$\label{eq:posterior}
P(\theta {\:|\:}{\bm{x}}, {\bm{w}}) \propto U(\theta; {\bm{x}}, {\bm{w}}).$$ Therefore, the maximum a posteriori (MAP) estimate maximizes $U(\theta; {\bm{x}}, {\bm{w}})$, or equivalently minimizes (\[eq:clearing-pot\]), and corresponds to clearing prices.
To establish that a probability model of the form (\[eq:generative-model\]) is possible—namely, that it can indeed be normalized—we will derive it as the result of a generative model. This process may be of independent interest as a means of generating agents together with market prices. 0 In this section we develop a generative model that jointly generates the $n$ single-minded agents together with prices $\theta$. Under this generative model, the maximum a posteriori (MAP) estimate for the prices will correspond to clearing prices for the set of agents. The model will provide the basis for a principled and systematic derivation of an auction procedure that computes clearing prices. The clearing price objective resembles the objective of a support vector machine (SVM): the $V_i(\theta)$ terms are similar to hinge losses, and the $R(\theta)$ term corresponds to a convex regularizer. Our generative model is inspired by the one constructed by @sollich2002bayesian, which recovers the SVM solution as a MAP estimate. The model is as follows:
1. Draw prices $\theta$ according to $Q(\theta) \propto \exp\left[ -R(\theta) \right].$
2. For each agent $i \in [n]$:
- Draw $w_i \in {\mathbb{R}}_+$ from $Q(w_i)$.
- Draw $x_i \in {\mathcal{X}}$ from $$Q(x_i {\:|\:}w_i, \theta) = \frac{1}{2^m}
\exp \left [ -V_i(\theta; x_i, w_i) \right].$$
- With probability $1 - \nu(w_i, \theta)$, restart from step 1, where $$\nu(w, \theta) = \sum_{x \in {\mathcal{X}}} Q(x {\:|\:}w, \theta).$$
Above, we must ensure that the $Q(\theta)$ prior normalizes; this is the case under our assumption that the domain of $\theta$ falls within the positive orthant. The prior distribution $Q(w_i)$ on value $w_i$ is left free in the model, so that it may correspond to the auctioneer’s prior in practice. Note that the bundle likelihood $Q(x_i {\:|\:}w_i, \theta)$ is *not* normalized; because $V_i(\theta; x_i, w_i) \geq 0$, summing over the set of bundles leads to the aggregate probability mass $\nu(w_i, \theta) \leq 1$. Rather than normalizing by this quantity, we use the “remaining probability” $1 - \nu(w_i, \theta)$ of not drawing any bundle to restart the process. Because of the possible restart, the agent types (bundle-value pairs) and clearing prices are not independent in the overall generative distribution. In particular, the number of agents $n$ in the economy affects the distribution of prices.
0 We write ${\bm{w}}= (w_1,\dots,w_n)$ and ${\bm{x}}= (x_1,\dots,x_n)$ for the vectors of agents’ values and bundles. We will use $P$ to refer to the generative probability model implied by the process given above. Our convention is that $Q$ refers to distributions—possibly unnormalized—that form the building blocks of the generative model, whereas $P$ refers to the normalized distribution resulting from the generative model. The following result is fundamental to the development of our auction scheme. The following proposition confirms that our model satisfies . All proofs are deferred to the appendix.
\[prop:gen-model\] The generative model of agent types and prices takes the form $$\label{eq:gen-model}
P({\bm{x}}, {\bm{w}}, \theta) \propto U(\theta; {\bm{x}}, {\bm{w}})Q({\bm{w}}).$$
0 As a corollary, we have that the posterior probability of prices $\theta$ also takes the form $$\label{eq:posterior}
P(\theta {\:|\:}{\bm{x}}, {\bm{w}}) \propto Q({\bm{x}}{\:|\:}{\bm{w}}, \theta) Q(\theta).$$ However, unlike in , the omitted normalization constant in depends on $({\bm{x}}, {\bm{w}})$. We can assume that bundle $x_i$ is observed for each agent $i$, since the agent will bid on the bundle in the very first round of the auction when prices are set to zero. We can therefore ignore dependence on ${\bm{x}}$ in the generative model since there is no uncertainty around bundles. Instead, it will be important to track any dependence on ${\bm{w}}$ in our auction, because these values remain latent. When values are observed, equation implies that the MAP estimate for $\theta$ is the value that minimizes $- \log P(\theta {\:|\:}{\bm{x}}, {\bm{w}})$, which up to an additive constant is $$\label{eq:posterior-clearing}
-\log Q({\bm{x}}{\:|\:}{\bm{w}}, \theta) -\log Q(\theta) = \sum_{i \in [n]} V_i(\theta) + R(\theta).$$ Recalling the objective , we see that the MAP estimate for prices $\theta$, when agent types $({\bm{x}}, {\bm{w}})$ are fully observed, corresponds to clearing prices for these agents. This justifies using our generative model as the basis of an auction that estimates clearing prices based on values inferred from observed bidding behavior.
The generative process defines a probability distribution over prices once valuations are *observed*, but during the auction the valuations remain latent, and must be inferred based on observed bids placed across rounds. Under appropriate incentives, the auctioneer can infer valuations assuming that the agents follow myopic best-response bidding. However, if there are any bidding errors or corruption in communication, assuming exact best-response can cause singularities in the inference process (e.g., there may be no valuation consistent with all observed bids). To guard against this, our mechanism will integrate bids as if they were generated from the following stochastic model: Let $b_i \in \{-1,+1\}$ be an indicator variable to denote whether the agent bids on bundle $x_i$ ($b_i = +1$) or not ($b_i = -1$); the latter is equivalent to bidding on $\emptyset$. If the cost of bundle $x_i$ is $c_i$, then the choice of bid follows the probability distribution $$\label{eq:bid-behavior}
Q(b_i = +1 {\:|\:}c_i, w_i) \propto \Phi(\beta(w_i - c_i)),$$ where $\Phi$ is the cumulative distribution function of the standard normal, and $\beta > 0$ is a scalar parameter. This is known as the *probit* variant of approximate best-response, which arises from random utility models [@train2009discrete]. As $\beta \rightarrow \infty$, we obtain exact best-response: the agent bids on $x_i$ if and only if this bundle yields positive utility under bundle cost $c_i$. Using a large (but finite) $\beta$ allows the auctioneer to model agents as essentially following a best-response strategy, but occasionally allowing for bidding errors or inconsistencies.
Auction Description {#sec:AuctionDescription}
===================
Our auction proceeds over rounds; we use $k$ to denote the current round, and $\ell$ to index the rounds up to $k$. At each round, prices are updated, which imputes a cost to each agent’s bundle. Let $c_i^\ell$ be the cost of agent $i$’s desired bundle $x_i$ in round $\ell$ according to the current prices. The prices at each round should not be confused with the latent clearing prices $\theta$, which we are trying to compute as a MAP estimate of the generative model. Given its value $w_i$ and the bundle cost $c_i^\ell$, agent $i$ places bid $b_i^\ell \in \{-1,+1\}$ in round $\ell$. We write ${\bm{c}}^\ell = (c_1^\ell, \dots, c_n^\ell)$ to denote the vector of bundle costs in round $\ell$, and ${\bm{c}}^{(\ell)} = ({\bm{c}}^1, \dots, {\bm{c}}^\ell)$ to denote the vector of costs *up to* round $\ell$. For brevity we also write ${\bm{c}}= {\bm{c}}^{(k)}$ to denote the vector of all costs up to the current round. We use the notation ${\bm{b}}^\ell$, ${\bm{b}}^{(\ell)}$, and ${\bm{b}}$ to denote the analogous vectors of bids. The bundle costs and agent bids in a round depend on the current prices, which themselves depend on the bids placed in all earlier rounds. Assuming that the first round prices are zero, we have the following intuitive posterior over bids and costs.
\[lem:bids-cost-posterior\] The posterior distribution over bids and costs placed during the auction, given the generated prices and agent types, is given by $$P({\bm{b}}, {\bm{c}}{\:|\:}{\bm{x}}, {\bm{w}}, \theta) = Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}})
\prod_{\ell=1}^k P({\bm{c}}^\ell {\:|\:}{\bm{b}}^{(\ell-1)}, {\bm{c}}^{(\ell-1)}, {\bm{x}}),$$ where ${\bm{b}}^\ell$ and ${\bm{c}}^\ell$ are the vectors of agent bids and costs at round $\ell$, and ${\bm{b}}^{(\ell)}$ and ${\bm{c}}^{(\ell)}$ are the vectors of agent bids and costs *up to* round $\ell$.
We see that the posterior over bids and costs does not depend on the underlying clearing prices $\theta$, conditional on agent types $({\bm{x}}, {\bm{w}})$, because the initial prices and agent valuations fully determine how the auction proceeds. More specifically, the posterior decomposes into the likelihood of the observed bids $Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}})$ under stochastic model , times the likelihood of the observed sequence of costs. The latter does not involve ${\bm{w}}$, because current round prices are fully determined by the bids and costs of previous rounds. Our auction is based on the following characterization of the overall posterior over prices and agent values.
\[prop:full-posterior\] The posterior distribution of latent variables $({\bm{w}}, \theta)$ given observed variables $({\bm{b}}, {\bm{c}}, {\bm{x}})$ takes the form $$P({\bm{w}}, \theta {\:|\:}{\bm{b}}, {\bm{c}}, {\bm{x}}) \propto Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}})Q({\bm{w}})
\cdot Q({\bm{x}}{\:|\:}{\bm{w}}, \theta) Q(\theta),$$ where the proportionality constant depends solely on $({\bm{b}}, {\bm{c}},
{\bm{x}})$.
The posterior factors into two terms, which motivates our auction procedure. The first term, $Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}})Q({\bm{w}})$, can be construed as a posterior over agent values given bids and costs, since $Q({\bm{w}})$ corresponds to a prior and $Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}})$ corresponds to a likelihood. We will maintain an approximation ${\hat{P}}({\bm{w}})$ to this posterior over agent values and update it as new bids are placed in response to bundle costs. This is the *knowledge update* component.
Recalling , the second term $Q({\bm{x}}{\:|\:}{\bm{w}}, \theta) Q(\theta) \propto U(\theta; {\bm{x}}, {\bm{w}})$ in the posterior corresponds (up to a constant factor) to the price posterior given knowledge of agent types. This leads to an approximation to the price posterior when values remain latent: $$\label{eq:price-map}
P(\theta {\:|\:}{\bm{b}}, {\bm{c}}, {\bm{x}}) \approx \int \mathrm{d}{\bm{w}}\, {\hat{P}}({\bm{w}})
\cdot Q({\bm{x}}{\:|\:}{\bm{w}}, \theta) Q(\theta).$$ Here we have simply integrated the full posterior as given by Proposition \[prop:full-posterior\], and made use of our approximation to the value posterior. (We have also omitted the normalization constant.) In the context of an auction, we quote a specific price function to the agents, rather than a distribution over prices. Therefore, in the *price update* component, we will compute and quote the MAP estimate of prices by maximizing . Note that if we have exact knowledge of agent values (i.e., ${\hat{P}}$ is a point mass), computing the MAP estimate is equivalent to minimizing (\[eq:clearing-pot\]) and to computing clearing prices, as one would expect.
0 As far we as we are aware, iterative auction designs to date lack a ‘knowledge update’ component, and directly update prices based on any imbalance between demand and supply.[^3] Our auction also needs a termination criterion to check whether the current round prices clear the market (perhaps approximately). There are several sensible options for the criterion, and we defer this aspect to the experimental evaluation.
#### Knowledge Update
We observe that the value posterior $Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}})Q({\bm{w}})$ consists of a separate factor for each agent $i$, taking the form $
Q(w_i) \prod_{\ell=1}^k Q(b_i^\ell {\:|\:}c_i^\ell, w_i),
$ where $k$ is the current round. This represents a posterior on agent $i$’s individual value $w_i$. To obtain an approximation to this posterior, we use an online scheme known as *assumed density filtering*, which is a special case of expectation propagation [@cowell1996comparison; @minka2001family; @opper1998bayesian]. Under this approach, a Gaussian distribution ${\hat{P}}(w_i; m_i, \sigma^2_i)$ is used to approximate the posterior; its mean $m_i$ and variance $\sigma^2_i$ are updated at each round given the bidding observations. The Gaussian is initially set to approximate the prior $Q(w_i)$ via moment matching: $m_i$ and $\sigma^2_i$ are set to the mean and variance of this prior $Q$. In each later round $\ell = 1,\dots,k$ the posterior is again updated by matching the moments of $
Q(b_i^\ell {\:|\:}c_i^\ell, w_i) {\hat{P}}(w_i; m_i, \sigma^2_i),
$ which is an online update. Using moment matching as an approximation is justified by the fact that it corresponds to minimizing the Kullback-Leibler divergence $\mathrm{KL}(Q \| {\hat{P}})$ under the constraint that ${\hat{P}}$ is Gaussian.\
Due to the form of the likelihood and the fact that ${\hat{P}}$ is Gaussian, the update has a closed-form solution [see @williams2006gaussian p. 74]: $$\begin{aligned}
m_i & \leftarrow & m_i + \frac{b_i^\ell\, \sigma_i^2\, \beta
{\mathcal{N}}(z_i)}{\sqrt{1+\sigma_i^2\beta^2} \, \Phi(z_i) } \\
\sigma^2_i & \leftarrow & \sigma^2_i - \frac{\sigma_i^4\, \beta^2 {\mathcal{N}}(z_i)}{1+\sigma_i^2\beta^2 \Phi(z_i) }\left(z_i+\frac{{\mathcal{N}}(z_i)}{\Phi(z_i)}\right),
\end{aligned}$$ where ${\mathcal{N}}$ and $\Phi$ are the probability density and cumulative distribution functions of the standard normal, respectively, and where $z_i = b_i^\ell \,\beta(m_i - c_i^\ell) / \sqrt{1+\sigma_i^2\,\beta^2}$. Recall that $\beta$ is a positive parameter characterizing the extent to which the auctioneer assumes that agents make mistakes in placing best-response bids. Since $\beta$ is positive, the mean $m_i$ is updated in the direction of the bid $b_i^{\ell}$. On the other hand, the variance $\sigma^2_i$ is strictly decreasing, thus ensuring that the beliefs over bidder values converge to a point mass in the limit as the rounds progress, and that the auction converges to a final vector of prices.
#### Price Update
To implement the price update component we need an algorithm to maximize the approximate posterior . This posterior factors into $Q(\theta)$ and a term for each agent $i$, which we denote as $${\mathcal{L}}_i(\theta) = \int \mathrm{d}w_i\, {\hat{P}}_i(w_i; m_i, \sigma_i^2) Q(x_i {\:|\:}w_i, \theta).$$ Because $Q_i(x_i {\:|\:}w_i, \theta)$ has an exponential form, and ${\hat{P}}_i$ is a Gaussian, this integral has a closed form solution (see appendix). Let $q_i \in \{0,1\}$ be a binary auxiliary variable. We have $
{\mathcal{L}}_i(\theta)
= \sum_{q_i = 0}^1 {\mathcal{L}}_i(\theta, q_i),
$ where we define $$\begin{aligned}
{\mathcal{L}}_i(\theta, 1) &\!\!\!\!\! = & \!\!\!\!\! \Phi\left( \frac{m_i - \theta(x_i)}{\sigma_i} - \sigma_i \right) \exp\left[ \theta(x_i) - m_i + \frac{\sigma_i^2}{2} \right] ,\\
{\mathcal{L}}_i(\theta, 0) &\!\!\!\!\! = & \!\!\!\!\! \Phi\left( \frac{\theta(x_i) - m_i}{\sigma_i} \right).
\end{aligned}$$ Here $\Phi$ is again the cumulative distribution function of the standard normal. To summarize, taking the log of , the objective we seek to maximize with respect to $\theta$ is $$\label{eq:price-objective}
\log Q(\theta) + \sum_{i \in [n]} \log \sum_{q_i = 0}^1 {\mathcal{L}}_i(\theta, q_i).$$ Now, because $\Phi$ is log-concave, both ${\mathcal{L}}_i(\theta, 0)$ and ${\mathcal{L}}_i(\theta, 1)$ are log-concave in $\theta$. Ignoring the first term for an instant, we see that the objective consists of a sum of mixtures ${\mathcal{L}}_i(\theta)$ of log-concave functions for each agent. This kind of objective is well-suited to optimization using the expectation-maximization (EM) algorithm [@dempster1977maximum]. The $q_i$ amount to “latent” variables and the “marginal” likelihood appears within the objective . (However, we do not claim any intuitive interpretation for the latent $q_i$—they are simply used to fit the objective into the mold of EM.)
The remaining term is $\log Q(\theta)$, which is $-R(\theta)$ up to an additive constant. Recalling the definition of the seller’s indirect utility , we see that this term is very complex for unrestricted $\theta$, because the set of feasible allocations ${\mathcal{F}}$ has a very complicated structure. To address this we will impose a linear structure on prices: $\theta(x) = p \cdot x$ where $p \in {\mathbb{R}}^m_+$ denote item prices. With this parametrization, we have $R(\theta) = \sum_{j=1}^m p_j$, because any allocation that allocates all the items maximizes revenue under linear prices. The $\log Q(\theta)$ term therefore becomes a linear term in $p$, which is straightforward to incorporate within the EM algorithm.
#### Incentive Compatibility
Our auction converges to an efficient allocation and clearing prices under myopic best-response bidding, but to ensure that agents follow such a strategy, they must be incentivized to do so. The standard technique used to achieve this in the literature on iterative auctions is to charge VCG payments upon completion [@gul2000english; @bikhchandani2006ascending]. But whereas VCG payments (together with an efficient allocation) induce truthful bidding in dominant strategies for single-shot auctions, weaker results hold for iterative auctions.
A strategy profile constitutes an *ex post* Nash equilibrium if no agent would like to deviate from its strategy, holding the others’ strategies fixed, even with knowledge of the private valuations of the other agents. @gul2000english prove the following result:
Truthful myopic best-response bidding is an *ex post* Nash equilibrium in an iterative auction that myopically-implements the VCG outcome.
Above, the VCG outcome refers to an efficient allocation along with VCG payments, and an auction myopically-implements this outcome if the auction converges to it under myopic best-response bidding. The reason that truthfulness only holds in *ex post* Nash equilibrium, rather than dominant strategies, is that profitable deviations may exist if another agent bids in a manner inconsistent with any valuation.
Our auction already computes the efficient allocation under these conditions by virtue of converging to clearing prices. To compute VCG payments, we can simply extend or auction drawing on the idea of multiple price trajectories: the usual trajectory traced by our auction, and the trajectories that would result if each agent were removed in turn. This technique was previously used by @ausubel2006efficient and @mishra2007ascending. In this extended design, at each round, agents place bids against $n+1$ different price vectors. Upon completion, the agents place last-and-final bids for their allocated bundles, thereby communicating their value for the allocations; importantly, agents do not need to communicate values for any bundles they did not win. This information is precisely what is needed to compute VCG payments [see, e.g., @parkes2000preventing].
Empirical Evaluation {#sec:empirical-evaluation}
====================
In this section we evaluate our Bayesian auction design with two different kinds of experiments: a small experiment to illustrate the behavior of our auction under biased and unbiased prior information, and a larger-scale experiment to compare our auction against a competitive baseline.
Our simulations are conducted in Matlab. In all our experiments, we assume that agents best respond to the proposed prices (i.e., they always bid on their most profitable bundle), and that the auctioneer considers their bids as if they were generated from the response model presented in (\[eq:bid-behavior\]) with $\beta = 10$. However, simulations where real bids follows (\[eq:bid-behavior\]) with $\beta = 4$ provide results similar to the ones presented. For the price update, the objective (\[eq:price-objective\]) is maximized using the “active-set” algorithm in Matlab. To avoid numerical singularities we place a lower bound of 0.01 on the variance of valuation estimates.
#### LLG Experiments
We consider an instance of the *Local-Local-Global* (LLG) domain [@ausubel2006lovely], which has been considered several times in the combinatorial auctions literature. There are two items and three single-minded agents. Two of the agents are *local*, meaning that they are interested in just one item, respectively the first and second item. The last agent is *global* in the sense that it is interested in both items.
The two local agents have a value of 4 for their respective items, and the global one has a value of 10 for both. The items are efficiently allocated when they are both assigned to the global agent, and linear prices are expressive enough to clear the market (e.g., we can use a price of 4 for each item).
We assume that the auctioneer has accurate knowledge of the local agents’ values: $Q(w_1) = \mathcal N(w_1; m_1=4, \sigma_1^2=0.01)$ and $Q(w_2) = \mathcal N(w_2; m_2=4, \sigma_2^2=0.01)$. Note the very low variance, reflecting certainty. We test how different kinds of prior knowledge over the global agent’s value affect the number of rounds that the Bayesian auction takes to clear the market. In the first case we assume unbiased prior knowledge: $m_3=10$. In the second case we assume that it is biased below: $m_3=4$. Here, the auctioneer tends to allocate to the local agents instead of the global one.
Figure \[figure:LLG\] plots the number of rounds that our Bayesian auction takes to clear the market against the variance $\sigma^2_3$ of the prior over the global agent’s value. We see that, in the unbiased scenario, the number of rounds monotonically increases as the variance grows. This can be easily explained since increasing the variance only adds noise to the exact prior estimate. In the biased scenario, we have an optimal range of variances between 8 and 16. If the variance is too low, the auction needs many observations to correct the biased prior. If it is too high, the low confidence leads to many rounds because the auctioneer needs to refine its estimate of the value regardless of the bias.
![Auction rounds in LLG.[]{data-label="figure:LLG"}](LLG){width="40.00000%"}
#### CATS Experiments
For our second set of experiments, we generate instances using the Combinatorial Auction Test Suite (CATS), which offers four generator distributions: [`paths`]{}, [`regions`]{}, [`arbitrary`]{}, and [`scheduling`]{} [@leyton2000towards]. These are meant to model realistic domains such as truck routes, real estate, and pollution rights. We generate 1000 instances from each distribution, each with 12 items and 10 single-minded agents.
The instances are generated as follows. First, 100 input files with 1000 bids each are generated. Each input file is partitioned into a “training set” and “test set”, each with 500 bids. From the test set, 10 bids (representing 10 single-minded agents) are sampled uniformly at random. The training set is used to fit the prior knowledge of our Bayesian auction. Specifically, we fit a linear regression model of bundle value according to items contained, using a Gaussian process with a linear covariance function, leading to Gaussian prior knowledge. The fit is performed using the publicly available GPML Matlab code [@williams2006gaussian].
As a baseline we implemented a standard linear-price auction scheme closely related to the combinatorial clock auction [@ausubel2014practical]. The scheme is parametrized by a positive step size $\tau$. At each round $\ell$, the price of an item is incremented by its *excess demand*, scaled by $\tau/\sqrt{\ell}$. The excess demand of an item is the number of bidded bundles that contain it, minus the number of units of the item offered by the seller at current prices. This can be viewed as a subgradient descent method for computing clearing prices, for which a step size proportional to $1/\sqrt{\ell}$ yields the optimal worst-case convergence rate [@bertsekas2015convex Chap. 3].
Both the Bayesian auction and the baseline subgradient auction use linear prices, but these may not be expressive enough to support an efficient allocation in instances generated by CATS. We therefore set a limit of 100 rounds for each auction run, and record reaching this limit as a failure to clear the market.
![Cleared instances in CATS.[]{data-label="figure:CfrNumOfClearing"}](CATSPercOfClearings){width="50.00000%"}
On each instance we run a single Bayesian auction, and 100 subgradient auctions with the step size uniformly spanning the interval from zero to the maximum agent value. This leads to several baseline results. The *standard instance optimized* (SIO) results refer to the performance of the baseline when using the optimal step size for each instance. The *standard average clearing-optimized* (SAOc) results refer to the performance of the baseline under the fixed step size that leads to the best clearing performance on average, for each valuation domain. Analogously, the *standard average round-optimized* (SAOr) results refer to baseline performance under the step size leading to lowest average number of rounds. [For each instance, the step size that leads to the lowest number of rounds naturally leads to the best clearing rate. But the fixed step sizes that optimize these two criteria on average may be different.]{} Note that SIO is an extremely competitive baseline, since it is optimized for each instance; a priori, we hoped to be competitive against it, but did not expect to beat it. The SAOc and SAOr baselines reflect more realistic performance that could be achieved in practice.
We first consider clearing performance. The results are reported in Figure \[figure:CfrNumOfClearing\]. We find that the Bayesian auction is competitive with SIO on all four domains, and that it always outperforms SAOc. In fact, our auction even outperforms SIO on the [`arbitrary`]{} domain. This means that it was able to clear some instances that the subgradient auction could not clear within 100 rounds at any step size. In general, there is good agreement between our Bayesian auction and the baselines on which instances can be cleared or not according to the 100-round criterion. This indicates that failure to clear is typically a property of the instance rather than the algorithm.
Figure \[figure:CfrAuctionRounds\] summarizes the distributions of rounds needed to achieve clearing using box plots. To enable fair comparisons, for this plot we only considered instances that were cleared by all auction types: the Bayesian auction, SAOr, and SIO. This yields 770 valid instances for [`paths`]{}, 910 for [`regions`]{}, 624 for [`arbitrary`]{} and 955 for [`scheduling`]{}. The mean rounds for the Bayesian auction, SAOr, and SIO are always statistically different at the 0.01 level. We see from the plot that, in terms of the median number of rounds, the Bayesian auction clearly outperforms SAOr, but also remarkably outperforms SIO. Furthermore, the distribution of rounds for the Bayesian auction has a much lower spread than the baselines. It is able to clear almost all instances in less than 10 rounds.
![Auction rounds in CATS.[]{data-label="figure:CfrAuctionRounds"}](CATSBoxPlots){width="40.00000%"}
Conclusion {#sec:conclusion}
==========
In this work we developed a Bayesian clearing mechanism for combinatorial auctions that allows one to incorporate prior information into the auction process in a principled manner. Our auction mechanism is based on a joint generative model of valuations and prices such that clearing prices are the MAP estimate given observed valuations. Our empirical evaluation confirmed that our Bayesian mechanism performs remarkably well against a conventional clock auction scheme, in terms of rounds to convergence. Our auction’s performance simply relies on reasonable priors for valuations, rather than careful tuning of price increments.
0 Our analysis was limited to single-minded agents, and a natural generalization of our work would be to consider a setting with multi-minded agents. While the knowledge-update component of our auction can be straightforwardly generalized via Gaussian process classification [@williams2006gaussian], the price-update component presents some challenges because the price posterior (\[eq:price-map\]) no longer has an analytical form. Nevertheless, the literature on Bayesian methods offers several potential approaches for approximating this integral, ranging from variational approximations to sampling-based methods [@mackay2003information], which represent promising avenues for generalizing our auction scheme.
We believe that the Bayesian perspective on auction design developed in this paper could be leveraged to improve other aspects beyond rounds to convergence. For instance, the Bayesian paradigm offers a principled way to select hyperparameters [@mackay1992bayesian]; in our context, this could be used to choose the right structure of prices (linear, nonlinear) to clear the market, a priori. The knowledge update component could also form the basis of more refined activity rules; for instance, one could reject bids that are highly unlikely, given the valuation posterior based on previous bids. We intend to pursue these directions in future work.
#### Proof of Proposition \[prop:gen-model\]
Let $\nu(\bm w, \theta)=\prod_{i} \nu(w_i, \theta) $ and $Q(\bm w) = \prod_{i} Q(w_i)$.\
We denote the probability of a restart as $$r = 1- \int \int \mathrm{d}\bm w \,\mathrm{d}\theta \, \nu(\bm w, \theta) Q(\bm w) Q(\theta),$$ where $\mathrm{d}\theta$ is a shorthand for $\prod_{x\in \mathcal X} \mathrm{d}\theta(x)$. (Note that any price function $\theta$ can be seen as a $2^m$-dimensional vector of positive real numbers.) The probability that $(\bm x, \bm w, \theta)$ is drawn after $\ell$ restarts will then be $$P(\bm x, \bm w, \theta , \ell ) = Q( \bm x {\:|\:}\bm w, \theta) Q(\bm w) Q(\theta) \cdot r^{\ell}.$$ Thus, $$\begin{array}{lll}
P(\bm x, \bm w, \theta) &= & \displaystyle \sum_{\ell=0}^{\infty} Q( \bm x {\:|\:}\bm w, \theta) Q(\bm w) Q(\theta) \cdot r^{\ell} \\ \\
&= & {\displaystyle}\frac{ Q(\bm x {\:|\:}\bm w, \theta) Q(\bm w) Q(\theta)}{1-r}.
\end{array}$$ Recalling equation , we have that $$U(\theta; {\bm{x}}, {\bm{w}}) \propto Q({\bm{x}}{\:|\:}{\bm{w}}, \theta) Q(\theta) .$$ We can conclude that $$P(\bm x, \bm w, \theta) \propto U(\theta; {\bm{x}}, {\bm{w}}) Q(\bm w) .$$
#### Proof of Lemma \[lem:bids-cost-posterior\]
$$\begin{aligned}
P({\bm{b}}, {\bm{c}}{\:|\:}{\bm{x}}, {\bm{w}}, \theta) & = & \prod_{\ell=1}^k P({\bm{b}}^\ell,
{\bm{c}}^\ell {\:|\:}{\bm{b}}^{(\ell-1)},
{\bm{c}}^{(\ell-1)},{\bm{x}}, {\bm{w}}, \theta) \\
&\hspace{-1.3cm} = &\hspace{-0.65cm} \prod_{\ell=1}^k
Q({\bm{b}}^\ell {\:|\:}{\bm{c}}^\ell, {\bm{w}})
P({\bm{c}}^\ell {\:|\:}{\bm{b}}^{(\ell-1)},
{\bm{c}}^{(\ell-1)}, {\bm{x}})\\
&\hspace{-1.3cm} = & \hspace{-0.65cm}
Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}}) \prod_{\ell=1}^k
P({\bm{c}}^\ell {\:|\:}{\bm{b}}^{(\ell-1)},
{\bm{c}}^{(\ell-1)}, {\bm{x}})
\end{aligned}$$
The third line follows because bids only depend on bundle costs and values according to , and costs are independent of values and clearing prices given previous round bids and costs, which fully determine the current round prices.
#### Proof of Proposition \[prop:full-posterior\]
We have that $$\begin{aligned}
P({\bm{w}}, \theta {\:|\:}{\bm{b}}, {\bm{c}}, {\bm{x}}) & \propto & P({\bm{b}}, {\bm{c}}{\:|\:}{\bm{x}}, {\bm{w}}, \theta) P({\bm{x}}, {\bm{w}}, \theta) \\
& \propto & Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}}) P({\bm{x}}, {\bm{w}}, \theta) \\
& \propto & Q({\bm{b}}{\:|\:}{\bm{c}}, {\bm{w}}) Q({\bm{x}}{\:|\:}{\bm{w}}, \theta) Q({\bm{w}}) Q(\theta).
\end{aligned}$$ The second line follows from Lemma \[lem:bids-cost-posterior\], and the last line from Proposition \[prop:gen-model\].
#### Price Likelihood Derivation\
\
We aim to solve the integral $$\begin{aligned}
\int \mathrm{d}w_i\, {\hat{P}}_i(w_i; m_i, \sigma_i^2) Q(x_i {\:|\:}w_i, \theta) = \\
& & \hspace{-4.8cm} \frac{1}{2^m}\, \int \mathrm{d}w_i\, \mathcal N(w_i| m_i, \sigma_i^2) \exp{[-(w_i-\theta(x_i))_+]},
\end{aligned}$$ where $\mathcal N(w_i| m_i, \sigma_i^2)$ is the probability density function of the normal distribution with mean $m_i$ and variance $\sigma_i^2$. We can rewrite the integral as $$\begin{aligned}
\frac{1}{2^m}\,
\int_{\theta(x_i)}^{+\infty} \mathrm{d}w_i\, \mathcal N(w_i| m_i, \sigma_i^2) \exp{[-(w_i-\theta(x_i))]}
+ \\
& & \hspace{-4.5cm} \displaystyle\frac{1}{2^m}\, \int_{-\infty}^{\theta(x_i)} \mathrm{d}w_i\, \mathcal N(w_i| m_i, \sigma_i^2).
\end{aligned}$$ Now, since $$\begin{aligned}
\mathcal N(w_i| m_i, \sigma_i^2) \exp{[-(w_i-\theta(x_i))]}
=\\
& & \hspace{-4.8cm} \mathcal N(w_i| m_i-\sigma_i^2, \sigma_i^2) \exp{[-(m_i-\sigma^2_i/2-\theta(x_i))]},
\end{aligned}$$ and given that $\Phi(-z)=1-\Phi(z)$, we can rewrite our integral as $$\begin{aligned}
\frac{1}{2^m}\,
\Phi \left( \frac{m_i-\theta(x_i)}{\sigma_i} - \sigma_i\right)\, \exp{\left [\theta(x_i) - m_i + \frac{\sigma_i^2}{2}\right ]} + \\
& & \hspace{-3.6cm}
\displaystyle\frac{1}{2^m}\, \Phi \left( \frac{\theta(x_i)-m_i}{\sigma_i}
\right).
\end{aligned}$$
[^1]: More formally, an OR valuation takes the form $v(x) = \max\{v_1(x_1) + v_2(x_2) : x_1+x_2=x,\: x_1 \wedge x_2
= 0 \}$, where $v_1$ and $v_2$ are themselves OR valuations or single-minded.
[^2]: This is in contrast to the goal of maximizing *revenue*. In auction design, one typically begins with an efficient auction, which is then modified (e.g., using reserve prices) to achieve optimal revenue [@myerson1981optimal]. We therefore consider the problem of designing an efficient auction as more fundamental.
[^3]: One might argue that activity rules are a form of knowledge update, but activity rules only constrain bidding [@ausubel2011activity; @ausubel2014practical]. In our case, knowledge update informs price update.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The paper concerns multiobjective linear optimization problems in $\mathbb{R}^{n}$ that are parameterized with respect to the right-hand side perturbations of inequality constraints. Our focus is on measuring the variation of the feasible set and the Pareto front mappings around a nominal element while paying attention to some specific directions. This idea is formalized by means of the so-called epigraphical multifunction, which is defined by adding a fixed cone to the images of the original mapping. Through the epigraphical feasible and Pareto front mappings we describe the corresponding vector subdifferentials and employ them to verifying Lipschitzian stability of the perturbed mappings with computing the associated Lipschitz moduli. The particular case of ordinary linear programs is analyzed, where we show that the subdifferentials of both multifunctions are proportional subsets. We also provide a method for computing the optimal value of linear programs without knowing any optimal solution. Some illustrative examples are also given in the paper.
**Key words.** Epigraphical set-valued mappings, feasible set mappings, Lipschitz moduli, linear programming, optimal value functions, multiobjective optimization.
**AMS Subject Classification:*** *49J53, 90C31, 15A39, 90C05, 90C29.
author:
- 'M. J. Cánovas[^1]'
- 'M. A. López[^2]'
- 'B. S. Mordukhovich[^3]'
- 'J. Parra'
title: |
Subdifferentials and Stability Analysis\
of Feasible Set and Pareto Front Mappings\
in Linear Multiobjective Optimization[^4]
---
Introduction and Overview {#intro}
=========================
The original motivation for this paper comes from analyzing Lipschitzian behavior of the so-called Pareto front mapping associated with the *multiobjective linear programming* (MLP) problem given by $$\begin{tabular}{ccc}
$MLP\left( b\right) :$ & \emph{\textrm{minimize}} & $\left( \left\langle
c_{1},x\right\rangle ,...,\left\langle c_{q},x\right\rangle \right) $ \\
& \emph{subject to} & $x\in \mathcal{F}\left( b\right) $,\end{tabular}
\label{eq_MOP}$$where $x\in \mathbb{R}^{n}$ is the decision variable, where $c_{1},...,c_{q}\in \mathbb{R}^{n}$ are fixed, and where $\mathcal{F}\left(
b\right) $ is the *feasible set* of the linear inequality system in $\mathbb{R}^{n}$ parameterized by its right-hand side (RHS) as $$\sigma \left( b\right) :=\big\{\left\langle a_{t},x\right\rangle \leq
b_{t},\,\,\,t\in T:=\left\{ 1,...,m\right\} \big\} \label{eq_sigma}$$with the coefficients $a_{t}\in \mathbb{R}^{n}$ fixed for each $t\in T$ and the perturbation parameter $b=\left( b_{t}\right) _{t\in T}\in \mathbb{R}^{T} $ in the RHS of .
For each $b\in \mathbb{R}^{T}$ denote by $\mathcal{S}\left( b\right) $ the set of *nondominated solutions* to $MLP\left( b\right) $, i.e., $\mathcal{S}\left( b\right) $ is formed by all $x\in \mathcal{F}\left(
b\right) $ such that there does not exist any other feasible point $y\in
\mathcal{F}\left( b\right) $ for which $\left\langle c_{i},y\right\rangle
\leq \left\langle c_{i},x\right\rangle $ whenever $i=1,...,q$ and $\left\langle c_{i_{0}},y\right\rangle <\left\langle c_{i_{0}},x\right\rangle
$ for some $i_{0}\in \{1,...,q\}$. Alternatively it can be reformulated as follows: considering the mapping $\mathcal{C}\colon \mathbb{R}^{n}\rightarrow \mathbb{R}^{q}$ defined by $\mathcal{C}(x):=\left(
\left\langle c_{1},x\right\rangle ,...,\left\langle c_{q},x\right\rangle
\right) $, we have the equivalence $$x\in \mathcal{S}\left( b\right) \ \Leftrightarrow \ \big(\mathcal{C}(\mathcal{F}\left( b\right) -x)\big)\cap \left( -\mathbb{R}_{+}^{q}\right)
=\{0_{q}\}.$$
Associated with the parameterized problem (\[eq\_MOP\]), we define the* Pareto front mapping* $\mathcal{P}\colon\mathbb{R}^{T}\mathbb{\rightrightarrows R}^{q}$ by $$\label{eq_pareto}
\mathcal{P}\left(b\right):=\left\{\left( \left\langle
c_{1},x\right\rangle,...,\left\langle c_{q},x\right\rangle\right),\text{ }
x\in\mathcal{S}\left(b\right)\right\}=\mathcal{C}(\mathcal{S}\left(b\right)).$$ Observe that in the case of ordinary/scalar linear programming (LP) problem, i.e., ** **when $q=1,$ the Pareto front mapping $\mathcal{P}$ reduces to the real-valued *optimal value function* known also as the ‘marginal function’ in variational analysis.
Appropriate tools of *variational analysis* and *generalized differentiation* are our primary machinery to study the major (robust) *Lipschitzian stability* notion for the feasible set and Pareto front mappings. To proceed, we need to compute the *subdifferential* of these set-valued mappings/multifunctions, which is defined via the *coderivative* of the corresponding epigraphical multifunctions; see Section 2. At this moment we advance that a natural definition of the *epigraphical Pareto front mapping* $\mathcal{E}_{\mathcal{P}}\colon\mathbb{R}^{T}\mathbb{\rightrightarrows R}^{q}$ is given by $$\label{eq_epi_Pareto_front}
\mathcal{E}_{\mathcal{P}}\left(b\right):=\mathcal{P}\left(b\right)+\mathbb{R}_{+}^{q},$$ where $\mathbb{R}_{+}^{q}$ is formed by the elements of $\mathbb{R}^{q}$ with nonnegative components.
Roughly speaking, while analyzing optimality in MLP we are interested only in that region of the feasible set where optimal/nondominated solutions may be located. A possible idea to skip the noninteresting regions is to consider a certain epigraphical mapping associated with the feasible set mapping. In this way we define the *epigraphical feasible set mapping* $\mathcal{E}_{\mathcal{F}}\colon\mathbb{R}^{T}\mathbb{\rightrightarrows R}^{n} $ by $$\label{eq_epi mapp}
\mathcal{E}_{\mathcal{F}}\left(b\right):=\mathcal{F}\left(b\right)+\left\{c_{1},...,c_{q}\right\}^{\circ },$$ where $\Omega^{\circ}:=\left\{y\in\mathbb{R}^{n}|\left\langle
y,x\right\rangle\ge 0\;\text{ for all }\;x\in\mathbb{R}^{n}\right\}$ stands for the (positive) polar cone of the set $\Omega\subset\mathbb{R}^{n}$.
The *main contributions* of our paper are precise calculations of the subdifferentials of the set-valued mappings $\mathcal{F}$ and $\mathcal{P}$ with the subsequent usage of them to verify Lipschitzian stability of these mappings and computing the corresponding Lipschitz moduli by invoking the powerful machinery of variational analysis. We show below that the subdifferentials of these multifunctions and their Lipschitz moduli are closely related as seen in Theorems \[Cor\_lip\_EP\] and \[Prop\_subdiffF\], and the established relationships are particularly clear in the case of ordinary (single-objective) linear programs; see Proposition \[Prop\_sect5\] and Theorem \[Th lipV\].
Given a mapping $\mathcal{M}\colon Y\rightrightarrows Z$ between metric spaces $Y$ and $Z$ with the graph $$\mathrm{gph}\mathcal{M}:=\big\{(y,z)\in Y\times Z\big|\;z\in \mathcal{M}(y\big\}$$and with the same notation $d$ for the metrics on $Y$ and $Z$, its Lipschitzian behavior is analyzed locally around a fixed point $\left(
\overline{y},\overline{z}\right) \in \mathrm{gph}\mathcal{M}$ while reflecting the rate of variation of its images with respect to the variation of the corresponding preimages. Here we focus on the most natural graphical extension of the classical local Lipschitz continuity to set-valued mappings that is spread in variational analysis as the Lipschitz-like/pseudo-Lipschitz/Aubin property. For definiteness let us say that $\mathcal{M}$ is *Lipschitz-like* around $\left( \overline{y},\overline{z}\right) \in \mathrm{gph}\mathcal{M}$ if there exist neighborhoods $U\subset Y$ and $V\subset Z\,\ $of $\overline{y}$ and $\overline{z}$, respectively, and a constant $\ell \geq 0$ such that we have the linear estimate $$d\big(z,\mathcal{M}\left( y^{\prime }\right) \big)\leq \ell \,d\left(
y,y^{\prime }\right) \;\text{ for all }\;y,y^{\prime }\in U\;\text{ and all }\;z\in V\cap \mathcal{M}\left( y\right) . \label{eq_Aubin_prop}$$Each constant $\ell $ ensuring (\[eq\_Aubin\_prop\]) for associated neighborhoods $U$ and $V\,\ $is called a Lipschitz constant and the infimum of such Lipschitz constants is called the *Lipschitz modulus*, or the *exact Lipschitz bound* of $\mathcal{M}$ around $\left( \overline{y},\overline{z}\right) $, and is denoted by $\mathrm{lip}\mathcal{M}\left(
\overline{y},\overline{z}\right) $. We can easily check that $$\mathrm{lip}\mathcal{M}\left( \overline{y},\overline{z}\right) =\underset{\QTATOP{y,y^{\prime }\rightarrow \overline{y}}{z\rightarrow \overline{z},z\in \mathcal{M}\left( y\right) }}{\lim \sup }\frac{d\big(z,\mathcal{M}\left( y^{\prime }\right) \big)}{d\left( y,y^{\prime }\right) } \label{mod}$$under the convention that $0/0:=0$. It has been well recognized in variational analysis that the Lipschitz-like property and its inverse mapping equivalences known as *metric regularity* and *linear openness/covering* play a fundamental role in many aspects of optimization, equilibrium, systems control, and applications; see the monographs [@bz05; @DoRo; @Ioffe17; @KlKu02; @Mor06i; @Mor18; @rw] and the references therein.
Using the modulus representation , we can rephrase that the main contribution of this paper is to *explicitly compute* the quantities $\mathrm{lip}\mathcal{E}_{\mathcal{F}}\left( \overline{b},\overline{x}\right)
$ and $\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left( \overline{b},\overline{p}\right) ,$ with $\left( \overline{b},\overline{x}\right) \in \mathrm{gph}\mathcal{E}_{\mathcal{F}}$ and $\left( \overline{b},\overline{p}\right) \in
\mathrm{gph}\mathcal{E}_{\mathcal{P}}$ respectively, entirely in terms of the given data of and . Furthermore, we advance here that the number $\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left(
\overline{b},\overline{p}\right) $ provides a lower estimate of $\mathrm{lip}\mathcal{P}\left( \overline{b},\overline{p}\right) $, and that both Lipschitz moduli *agree* for ordinary *linear programs* as shown in Section 5. Having the precise formulas for computing the moduli $\mathrm{lip}\mathcal{E}_{\mathcal{F}}\left( \overline{b},\overline{x}\right) $ and $\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left( \overline{b},\overline{p}\right)
$, the *necessary and sufficient conditions* for *Lipschitzian stability* of the mappings (\[eq\_epi\_Pareto\_front\]) and (\[eq\_epi mapp\])—in the sense of the validity of the Lipschitz-like property for these mappings around the reference points—are formulated now as, respectively, $$\mathrm{lip}\mathcal{E}_{\mathcal{F}}\left( \overline{b},\overline{x}\right)
<\infty \;\mbox{ and }\;\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left(
\overline{b},\overline{p}\right) <\infty .$$
These achievements are largely based on the *subdifferential* notion for *set-valued mappings* with *ordered values* introduced in [BaMo07]{} (see also [@BaMo10; @Mor18]) via the coderivatives concept for mappings and on the *coderivative criterion* for the Lipschitz-like property of multifunctions established in [@Mo93]. The passage from coderivatives to subdifferentials of ordered mappings was accomplished in [@BaMo07] via the usage of *epigraphical multifunctions*: the pattern well understood in variational analysis for the subdifferential-coderivative relationship concerning scalar (extended-real-valued) functions; see, e.g., [@Mor06i Vol. 1, p. 84].
It is worth mentioning that some coderivative analysis of frontier and efficient solution mapping was provided in [@huy-mor-yao08] for problems of vector optimization with respect to the so-called *generalized order optimality* (including Pareto efficiency) in infinite-dimensional spaces. However, neither precise coderivative formulas, nor subdifferential analysis, nor computations of Lipschitz moduli were obtained in the general setting of [@huy-mor-yao08] in contrast to what is done in this paper.
Furthermore, while confining to the case of ordinary/scalar linear programs where $\mathcal{P}$ is the optimal value function, the reader is addressed to [@GCPT19] for different formulas concerning Lipschitz moduli in various parametric frameworks. Note also that Lipschitzian behavior of the ‘ordinary’ feasible set mapping $\mathcal{F}$ and the computation of its modulus were derived for more general models of semi-infinite and infinite programming in [@CDLP05] and [@clmp09]. Other stability properties of the feasible set mapping of linear semi-infinite systems were analyzed in [@GoLo98 Chapter 6].** **Lipschitzian behavior of the optimal set, again in the context of linear programming problems (in fact, in a continuous convex semi-infinite setting allowing also perturbations of the objective function) was studied in [@CKLP07], whereas the associated Lipschitz modulus was computed in [@CGP08].
The rest of the paper is organized as follows. In Section \[prelim\] we present the necessary notation, definitions, and results about coderivatives, subdifferentials, and Lipschitz moduli that are needed later on. Section \[feas\] is devoted to subdifferential analysis and Lipschitzian stability of the epigraphical feasible set mappings $\mathcal{E}_{\mathcal{F}}$ from . Specifically, we provide explicit descriptions of the subdifferential of $\mathcal{F}$ and the Lipschitz modulus of $\mathcal{E}_{\mathcal{F}}$ at a given point of its graph. In the subsequent Section \[scal\] we develop a constructive procedure for deriving the representation of such a mapping as the feasible set mapping associated with new parameterized systems of linear programming. Section [pareto]{} is focussed on the precise computations of subdifferential and Lipschitz modulus of the epigraphical Pareto front mapping $\mathcal{E}_{\mathcal{P}}$ from . In Section \[lp\] we consider the case of ordinary linear programs (with only one objective function) and show that even in this case our results are new. In particular, we establish exact relationships between the subdifferentials and Lipschitz moduli of the set-valued mappings $\mathcal{E}_{\mathcal{F}}$ and $\mathcal{E}_{\mathcal{P}}$ under consideration. Both Sections [pareto]{} and \[lp\] contain illustrative examples of their own interest. The final Section \[conc\] summarizes the obtained results and discusses some directions of future research.
Throughout the paper we use the standard notion in variational analysis and optimization. Recall that $\mathrm{conv\,}\Omega $, $\mathrm{cone\,}\Omega $, and $\mathrm{span\,}\Omega $ stand, respectively, for the *convex hull*, the *conic convex hull*, and the *linear subspace* generated by the set $\Omega \subset \mathbb{R}^{n}$ under the convention that $\mathrm{cone\,}\emptyset =\{0_{n}\}$, where $0_{n}$ is the origin of $\mathbb{R}^{n}$. If $\Omega $ is convex, by $O^{+}(\Omega )$ we represent the *recession cone* of $\Omega $. The space of decision variables $\mathbb{R}^{n} $ is endowed with an arbitrary norm $\left\Vert \cdot \right\Vert $, while the space of parameters $\mathbb{R}^{T}$ is equipped with the supremum norm $$\left\Vert b\right\Vert _{\infty }:=\sup_{t\in T}\left\vert b_{t}\right\vert
. \label{sup-norm}$$
Preliminaries and First Results {#prelim}
===============================
In this section, unless otherwise stated, $\mathcal{M}\colon
Y\rightrightarrows Z$ is a set-valued mapping between Banach spaces $Y$ and $Z$ which topological duals are denoted by $Y^{\ast }$ and $Z^{\ast }$, respectively. The *coderivative* of $\mathcal{M}$ at $\left( \overline{y},\overline{z}\right) \in \mathrm{gph}\mathcal{M}$ is a positively homogeneous multifunction $D^{\ast }\mathcal{M}\left( \overline{y},\overline{z}\right) \colon Z^{\ast }\rightrightarrows Y^{\ast }$ defined by $$y^{\ast }\in D^{\ast }\mathcal{M}\left( \overline{y},\overline{z}\right)
\left( z^{\ast }\right) \Longleftrightarrow \left( y^{\ast },-z^{\ast
}\right) \in N\big(\left( \overline{y},\overline{z}\right) ;\mathrm{gph}\mathcal{M}\big), \label{cod}$$where $N\left( \left( \overline{y},\overline{z}\right) ;\mathrm{gph}\mathcal{M}\right) $ is the (basic, limiting, Mordukhovich) *normal cone* to $\mathrm{gph}\mathcal{M}$ at $\left( \overline{y},\overline{z}\right) $; see, e.g., [@Mor06i] and [@rw]. For simplicity, $\left\Vert \cdot
\right\Vert $ stands for the norm in any Banach space $X$, and $\left\Vert
\cdot \right\Vert _{\ast }$ is the corresponding dual norm, i.e., $$\left\Vert x^{\ast }\right\Vert _{\ast }=\sup \big\{\left\langle x^{\ast
},x\right\rangle \big|\;\left\Vert x\right\Vert \leq 1,\text{ }x\in X\big\},\text{ }x^{\ast }\in X^{\ast },$$where $\left\langle .,.\right\rangle $ denotes the canonical pairing between $X$ and $X^{\ast }$. If no confusion arises, from now on we skip the subscript ‘$_{\ast }$’ in the dual norm notation.
When both spaces $Y$ and $Z$ are finite-dimensional and the graph of $\mathcal{M}$ is locally closed around $(\overline{y},\overline{z})\in
\mathrm{gph}\mathcal{M}$, there is the following *precise formula* for the *computing the Lipschitz modulus* of $\mathcal{M}\left( \overline{y},\overline{z}\right) $: $$\mathrm{lip}\mathcal{M}\left( \overline{y},\overline{z}\right) =\left\Vert
D^{\ast }\mathcal{M}\left( \overline{y},\overline{z}\right) \right\Vert
:=\sup \big\{\left\Vert y^{\ast }\right\Vert _{\ast }\big|y^{\ast }\in
D^{\ast }\mathcal{M}\left( \overline{y},\overline{z}\right) (z^{\ast
}),\Vert z^{\ast }\Vert _{\ast }=1\big\}, \label{eq_lipM}$$which was obtained in [@Mo93]. We also refer the reader to [@rw Theorem 9.40] for another proof of this result, which was labeled therein as the Mordukhovich criterion. An infinite-dimensional extension of was derived in [@Mor06i Theorem 4.10]. It is more involved and is not used in this paper dealing with finite-dimensional multiobjective optimization problems of type . A simplified proof of in finite dimensions was given in [Mor18]{}.
If the graph of $\mathcal{M}$ is *convex*, the normal cone in reduces to the normal convex of convex analysis, and hence $y^{\ast}\in D^{\ast}\mathcal{M}\left(\overline{y},\overline{z}\right)\left(z^{\ast}\right)$ if and only if $$\left\langle\left(y^{\ast},-z^{\ast}\right),\left(y^{\prime}-\overline{y},z^{\prime}-\overline{z}\right)\right\rangle\le 0\;\text{ for all }\;\left(y^{\prime},z^{\prime}\right)\in\mathrm{gph}\mathcal{M},$$ which is equivalent to the description $$\label{eq_conderiv_convex}
\left\langle y^{\ast},y^{\prime}-\overline{y}\right\rangle\le\left\langle
z^{\ast},z^{\prime}-\overline{z}\right\rangle\;\text{ for all }\;\left(y^{\prime},z^{\prime }\right)\in\mathrm{gph}\mathcal{M}.$$
Given further a closed and convex *ordering cone* $\Theta\subset Z$, the *epigraphical multifunction *$\mathcal{E}_{\mathcal{M}}\colon X\rightrightarrows Z$ associated with $\mathcal{M}$ and the cone $\Theta$ is that which graph $\mathrm{gph}\mathcal{E}_{\mathcal{M}}$ coincides with the *epigraph* of $\mathcal{M}$ with respect to $\Theta$. In other words, we have $\mathrm{\mathcal{E}_{\mathcal{M}}}\left(y\right):=\mathcal{M}\left(y\right)+\Theta$ and $$\mathrm{epi\mathcal{M}}:=\mathrm{gph\mathcal{E}_{\mathcal{M}}}=\big\{\left(y,z\right)\big|\;z\in\mathcal{M}\left(y\right)+\Theta\big\},$$ where we skip indicating $\Theta$ in the epigraphical notation.
In accordance with [@BaMo07], we present the following definition of the subdifferential of $\mathcal{M}$ at the reference point of its epigraph with respect to $\Theta$.
\[def fre subdiff\] Let $\left( \overline{y},\overline{z}\right) \in
\mathrm{epi\mathcal{M}}$ be given. The <span style="font-variant:small-caps;">subdifferential</span> of $\mathcal{M}$ at $\left( \overline{y},\overline{z}\right) $ denoted as $\partial
\mathcal{M}\left( \overline{y},\overline{z}\right) $ is a subset of $Y^{\ast
}$ defined by $$\partial \mathcal{M}\left( \overline{y},\overline{z}\right) :=\big\{y^{\ast
}\in D^{\ast }\mathcal{E}_{\mathcal{M}}\left( \overline{y},\overline{z}\right) \left( z^{\ast }\right) \big|-z^{\ast }\in N\left( 0;\Theta \right)
,\;\Vert z^{\ast }\Vert _{\ast }=1\big\}, \label{eq_def_subdif_gen}$$where $N\left( 0;\Theta \right) \subset Z^{\ast }$ is the convex normal cone to the set $\Theta $ at the origin of $Z$.
Note that if $\mathcal{M}\colon Y\rightarrow\mathbb{R}$ is a proper convex function with** **$\Theta=\mathbb{R}_{+}$, then $\mathrm{gph\mathcal{E}_{\mathcal{M}}}$ is its standard epigraph, and for any $\overline{y}\in\limfunc{dom}\mathcal{M}$ the set $\partial \mathcal{M}\left(\overline{y},\mathcal{M}(\overline{z})\right)$ is the classical subdifferential of** **$\mathcal{M}$ at $\overline{y}$ in the sense of convex analysis.
Observe also that the set $-N\left(0;\Theta\right) $ is nothing else but the polar cone $\Theta^{\circ }$, and thus we have the following representation of the coderivative of $\mathrm{\mathcal{E}_{\mathcal{M}}}$ in terms of the graph $\mathrm{gph}\mathcal{M}$ instead of the epigraph $\mathrm{epi}\mathcal{M}$.
\[coder-epi\] Assume that $\mathrm{epi\mathcal{M}}$ is a convex set, and let $\left( \overline{y},\overline{z}\right)\in\mathrm{epi\mathcal{M}}$. Then for any $z^{\ast}\in Z^{\ast}$ we have the representation $$\begin{aligned}
&D^{\ast}\mathcal{E}_{\mathcal{M}}\left(\overline{y},\overline{z}\right)
\left(z^{\ast}\right)\medskip \\
&=\left\{\begin{array}{l}
\{y^{\ast}\in Y^{\ast}\mid\left\langle y^{\ast},y^{\prime}-\overline{y}\right\rangle\le\left\langle z^{\ast},z^{\prime}-\overline{z} \right\rangle\text{ }\forall\left(y^{\prime},z^{\prime}\right)\in\mathrm{gph}\mathcal{M\}}\;\text{ if }\;z^{\ast}\in\Theta^{\circ},\medskip \\
\emptyset\;\text{ if }\;z^{\ast}\notin\Theta^{\circ}.\end{array}
\right.\end{aligned}$$
Take $z^{\ast }\in \Theta ^{\circ }$. By using the definitions of the coderivative and of the epigraphical multifunction $\mathcal{E}_{\mathcal{M}}$, we get due to the convexity of $\mathrm{epi\mathcal{M}}$ ($=\mathrm{gph}\mathcal{E}_{\mathcal{M}}$) that $$D^{\ast }\mathcal{E}_{\mathcal{M}}\left( \overline{y},\overline{z}\right)
\left( z^{\ast }\right) =\big\{y^{\ast }\mid \left\langle y^{\ast
},y^{\prime }-\overline{y}\right\rangle \leq \left\langle z^{\ast
},z^{\prime }-\overline{z}\right\rangle \;\forall \left( y^{\prime
},z^{\prime }\right) \in \mathrm{gph}\mathcal{E}_{\mathcal{M}}\big\}.
\label{eq_EM_M}$$Let us show that $\mathrm{gph}\mathcal{E}_{\mathcal{M}}$ can be equivalently replaced by $\mathrm{gph}\mathcal{M}$ in . Indeed, take any $y^{\ast }\in Y^{\ast }$ satisfying . Pick further any $\left( \widetilde{y},\widetilde{z}\right) \in \mathrm{gph}\mathcal{E}_{\mathcal{M}}$ and write $\widetilde{z}=z^{\prime }+u\,\ $with $z^{\prime
}\in \mathcal{M}\left( \widetilde{y}\right) $ and $u\in \Theta $. Then we obtain the inequalities $$\left\langle y^{\ast },\widetilde{y}-\overline{y}\right\rangle \leq
\left\langle z^{\ast },z^{\prime }-\overline{z}\right\rangle \leq
\left\langle z^{\ast },\widetilde{z}-\overline{z}\right\rangle$$due to $\left\langle z^{\ast },u\right\rangle \geq 0$. It gives us the claimed coderivative formula for $z^{\ast }\in \Theta ^{\circ }$.
Suppose now that $z^{\ast }\notin \Theta ^{\circ }$ and find $u\in \Theta $ such that $\left\langle z^{\ast },u\right\rangle <0$. Arguing by contradiction, assume that there is $y^{\ast }\in D^{\ast }\mathcal{E}_{\mathcal{M}}\left( \overline{y},\overline{z}\right) \left( z^{\ast }\right) $, i.e., by we have $$\left\langle y^{\ast },y^{\prime }-\overline{y}\right\rangle \leq
\left\langle z^{\ast },z^{\prime }-\overline{z}\right\rangle \;\text{ for
all }\;\left( y^{\prime },z^{\prime }\right) \in \mathrm{gph}\mathcal{E}_{\mathcal{M}}.$$Since $\mathcal{E}_{\mathcal{M}}\left( \overline{y}\right) +\Theta =\mathcal{E}_{\mathcal{M}}\left( \overline{y}\right) $, it follows that $\left(
\overline{y},\overline{z}+u\right) \in \mathrm{gph}\mathcal{E}_{\mathcal{M}}$, and therefore $$0=\left\langle y^{\ast },\overline{y}-\overline{y}\right\rangle \leq
\left\langle z^{\ast },u\right\rangle ,$$which is a contradiction that completes the proof of the proposition.
Employing Proposition \[coder-epi\] leads us to deriving effective representations of the subdifferential of $\mathcal{M}$ and the Lipschitz modulus of $\mathcal{E}_{\mathcal{M}}$ as well as to a relation between the latter and the Lipschitz modulus of $\mathcal{M}$ at the reference point.
\[Cor\_lip\_subdiff\] Let the epigraphical set $\mathrm{epi\mathcal{M}}$ be convex, and let $\left( \overline{y},\overline{z}\right) \in \mathrm{epi\mathcal{M}}$. Then we have the subdifferential representation $$\partial \mathcal{M}\left( \overline{y},\overline{z}\right) =\bigcup_{\QATOP{\scriptstyle{z^{\ast }\in \Theta ^{\circ }}}{\scriptstyle{\Vert z^{\ast
}\Vert }_{\ast }{=1}}}\big\{y^{\ast }\big|\;\left\langle y^{\ast },y^{\prime
}-\overline{y}\right\rangle \leq \left\langle z^{\ast },z^{\prime }-\overline{z}\right\rangle \forall \left( y^{\prime },z^{\prime }\right) \in
\mathrm{gph}\mathcal{M}\big\}. \label{eq_subdif_convex}$$If in addition $Y$ and $Z$ are finite-dimensional and if the set $\mathrm{epi\mathcal{M}}$ is locally closed around $\left( \overline{y},\overline{z}\right) $, then the Lipschitz modulus of $\mathcal{E}_{\mathcal{M}}$ at $\left( \overline{y},\overline{z}\right) $ is computed by $$\mathrm{lip}\mathcal{E}_{\mathcal{M}}\left( \overline{y},\overline{z}\right)
=\sup \big\{\Vert y^{\ast }\Vert _{\ast }\;\big|\;y^{\ast }\in \partial
\mathcal{M}\left( \overline{y},\overline{z}\right) \big\}. \label{eq_lip_EM}$$Assuming furthermore that $\left( \overline{y},\overline{z}\right) \in
\mathrm{gph}\mathcal{M}$ and that the set $\mathrm{gph}\mathcal{M}$ is locally closed around this point, we conclude that $$\mathrm{lip}\mathcal{E}_{\mathcal{M}}\left( \overline{y},\overline{z}\right)
\leq \mathrm{lip}\mathcal{M}\left( \overline{y},\overline{z}\right) .
\label{eq_lipEmenorlip}$$
Representation follows directly from definition of the subdifferential $\partial\mathcal{M}\left(\overline{y},\overline{z}\right)$ combined with Proposition \[coder-epi\].
Assuming now that the spaces $Y$ and $Z$ are finite-dimensional, applying the Lipschitz modulus formula to the epigraphical mapping $\mathcal{E}_{\mathcal{M}}$, and appealing again to Proposition [coder-epi]{} tell us that $$\begin{aligned}
\mathrm{lip}\mathcal{E}_{\mathcal{M}}\left( \overline{y},\overline{z}\right)
&=&\sup \big\{\Vert y^{\ast }\Vert _{\ast }\;\big|\;y^{\ast }\in D^{\ast }\mathcal{E}_{\mathcal{M}}\left( \overline{y},\overline{z}\right) (z^{\ast
}),\;\Vert z^{\ast }\Vert _{\ast }=1\big\} \\
&=&\sup \big\{\Vert y^{\ast }\Vert _{\ast }\;\big|\;y^{\ast }\in D^{\ast }\mathcal{E}_{\mathcal{M}}\left( \overline{y},\overline{z}\right) (z^{\ast
}),\;z^{\ast }\in \Theta ^{\circ },\;\Vert z^{\ast }\Vert _{\ast }=1\big\}.\end{aligned}$$Thus the claimed formula follows from the definition of $\partial \mathcal{M}\left( \overline{y},\overline{z}\right) $.
To verify finally the inequality , denote by $\widehat{N}\left(\left(\overline{y},\overline{z}\right);\Omega\right)$ the prenormal/regular normal cone to $\Omega\subset Y\times Z$ at $\left(\overline{y},\overline{z}\right)$ (see, e.g., [@Mor06i; @rw]) and using the convexity of $\mathrm{epi\mathcal{M}}$, we get $$N\left(\left(\overline{y},\overline{z}\right);\mathrm{epi}\mathcal{M}\right)=\widehat{N}\left(\left(\overline{y},\overline{z}\right);\mathrm{epi}\mathcal{M}\right) \subset\widehat{N}\left(\left(\overline{y}, \overline{z}\right);\mathrm{gph}\mathcal{M}\right)=N\left(\left(\overline{y},\overline{z}\right);\mathrm{gph}\mathcal{M}\right),$$ where the inclusion comes from [@Mor06i Proposition 1.5]. This gives us $$D^{\ast}\mathcal{E}_{\mathcal{M}}\left(\overline{y},\overline{z}\right)\left(z\right)\subset D^{\ast}\mathcal{M}\left(\overline{y},\overline{z}\right)\left(z\right)\;\text{ for all }\;z\in Z$$ and thus deduces from the basic coderivative formula .
\[Rem\_lip\_EM\_M\]
Stability Analysis of Epigraphical Feasible Sets {#feas}
================================================
The underlying goal of this section is explicit computing the Lipschitz modulus of epigraphical feasible set mapping associated with the parameterized MLP problem . As we know from Sections [intro]{} and \[prelim\], our approach reduces this computation to deriving a verifiable formula to calculate the subdifferential in the sense of Definition \[def fre subdiff\] of the perturbed feasible set $\mathcal{F}$ in terms of its given data. Proceeding in this way, we concentrate here on obtaining the representations of the subdifferential and Lipschitz modulus with involving the graph of the nondominated solution mapping $\mathcal{S}$.
Let us begin with two lemmas. The first one is a well-known result that gives a characterization of nondominated solutions to MLP$\left(b\right)$ via optimal solutions to a scalarized linear program. We formulate it without a proof. The second lemma is a new result, which plays a key role throughout the paper.
\[Lem\_optimality\_conditions\] Let $x_{0}\in\mathcal{F}\left(b\right)$ for some $b\in\mathbb{R}^{T}$. Then the following are equivalent:
**(i)** $x_{0}\in\mathcal{S}\left(b\right)$.
**(ii)** There exist numbers $\lambda _{i}>0$ for $i=1,...,q,$ such that $$x_{0}\in\arg\min \left\{\dsum_{i=1}^{q}\lambda _{i}\left\langle
c_{i},x\right\rangle\Big|\;x\in\mathcal{F}\left(b\right)\right\}.$$
To formulate the second lemma, recall that $$\mathrm{dom}\mathcal{S}:=\big\{b\in \mathbb{R}^{T}\big|\;\mathcal{S}\left(b\right)\ne\emptyset\big\}.$$
\[Lem\_F\_notS\] Let $b\in\mathrm{dom}\mathcal{S}$. Then for any $x_{0}\in\mathcal{F}\left( b\right)\setminus \mathcal{S}\left(b\right)$ there exists $\widetilde{x}_{0}\in\mathcal{S}\left(b\right)$ such that $\left\langle c_{i},\widetilde{x}_{0}\right\rangle\le\left\langle c_{i},x_{0}\right\rangle$ whenever $i=1,...,q$.
Fix $x_{0}\in\mathcal{F}\left(b\right)\setminus\mathcal{S}\left(b\right)$ and proceed step-by-step as follows:
*Step* 1. Let us proof the existence of solutions to the linear program: $$x_{1}\in \arg \min \big\{\left\langle c_{1},x\right\rangle \big|\;x\in
\mathcal{F}\left( b\right) ,\text{ }\left\langle c_{i},x\right\rangle \leq
\left\langle c_{i},x_{0}\right\rangle ,\text{ }i=1,...,q\big\}.
\label{eq_001}$$Arguing by contradiction, suppose that has no optimal solutions. Since $x_{0}$ is a feasible solution to , our assumption is equivalent to the unboundedness of the set of feasible solutions to the linear program . Thus there exists a sequence $\{w_{r}\}_{r\in \mathbb{N}}\subset \mathbb{R}^{n}$ such that $$w_{r}\in \mathcal{F}\left( b\right) ,\text{ }\left\langle
c_{i},w_{r}\right\rangle \leq \left\langle c_{i},x_{0}\right\rangle ,\text{ }i=1,...,q,\text{ for all }r\in \mathbb{N},$$while we have the infinite limit $$\lim_{r\rightarrow \infty }\left\langle c_{1},w_{r}\right\rangle =-\infty .$$Remembering that $\mathcal{S}\left( b\right) \neq \emptyset $, pick any $\widetilde{x}\in \mathcal{S}\left( b\right) $ and find by Lemma [Lem\_optimality\_conditions]{} numbers $\lambda _{i}>0$ with $i=1,...,q$ such that $$\widetilde{x}\in \arg \min \left\{ \dsum_{i=1}^{q}\lambda _{i}\left\langle
c_{i},x\right\rangle \Big|\;x\in \mathcal{F}\left( b\right) \right\} .$$This readily brings us to the contradiction: $$\dsum_{i=1}^{q}\lambda _{i}\left\langle c_{i},\widetilde{x}\right\rangle
\leq \dsum_{i=1}^{q}\lambda _{i}\left\langle c_{i},w_{r}\right\rangle \leq
\lambda _{1}\left\langle c_{1},w_{r}\right\rangle +\dsum_{i=2}^{q}\lambda
_{i}\left\langle c_{i},x_{0}\right\rangle \underset{r\rightarrow \infty }{\longrightarrow }-\infty ,$$which therefore verifies the existence of the solution $x_{1}$ to . Note furthermore that if $x_{1}$ satisfies $x_{1}\in
\mathcal{S}\left( b\right) $, then the proof of the lemma is complete. Otherwise we go to the next step as follows.
*Step* 2. Suppose that $x_{1}\in\mathcal{F}\left(b\right)\setminus\mathcal{S}\left(b\right)$. Then arguing as in Step 1 ensures the existence of a vector $x_{2}\in{\mathbb{R}}^n$ satisfying $$\label{eq_002}
x_{2}\in \arg\min\big\{\left\langle c_{2},x\right\rangle\big|\;x\in\mathcal{F}\left(b\right),\text{ }\left\langle c_{i},x\right\rangle\le\left\langle
c_{i},x_{1}\right\rangle,\text{ }i=1,...,q\big\}.$$ As before, the proof of the lemma is finished if $x_{2}\in\mathcal{S}\left(b\right)$. Otherwise we go to *Step* 3 and proceed similarly.
Reaching in this way *Step* $j$ with some $j<q$, we either finish the proof, or arrive at *Step* $q$ that is described below.
*Step* $q$. Suppose that $x_{q-1}\in\mathcal{F}\left(b\right)\setminus\mathcal{S}\left(b\right)$. Again we get $$x_{q}\in\arg\min\big\{\left\langle c_{q},x\right\rangle\big|\;x\in\mathcal{F}\left(b\right),\text{ }\left\langle c_{i},x\right\rangle\le\left\langle
c_{i},x_{q-1}\right\rangle,\text{ }i=1,...,q\big\}.$$
Let us show that now we do not have any choice but $x_{q}\in\mathcal{S}\left(b\right)$. Arguing by contradiction, assume that there exists $w\in\mathcal{F}\left(b\right)$ such that $$\left\{\begin{array}{l}
\left\langle c_{i},w\right\rangle\le\left\langle c_{i},x_{q}\right\rangle\text{ for all }\;i=1,...,q, \\
\left\langle c_{j},w\right\rangle<\left\langle c_{j},x_{q}\right\rangle\text{
for some }\;j\in\{1,...,q\}.\end{array}\right.$$ Then we arrive at a contradiction with the choice of $x_{j}$. Indeed, it follows that $$\begin{aligned}
\left\langle c_{i},w\right\rangle &\le&\left\langle
c_{i},x_{q}\right\rangle\le\left\langle c_{i},x_{q-1}\right\rangle\le
...\le\left\langle c_{i},x_{j-1}\right\rangle\text{ for all }i=1,...,q, \\
\left\langle c_{j},w\right\rangle&<&\left\langle
c_{j},x_{q}\right\rangle\le\left\langle
c_{j},x_{q-1}\right\rangle\le...\le\left\langle c_{j},x_{j}\right\rangle.\end{aligned}$$ This completes the proof of the lemma.
The next theorem provides a description of the subdifferential $\partial\mathcal{F}\left(\overline{b},\overline{x}\right)$ in terms of $\mathrm{gph}\mathcal{S}$ (instead of $\mathrm{gph}\mathcal{F}$ as in the definition), which eventually allows us to relate the subdifferential $\partial\mathcal{F}\left(\overline{b},\overline{x}\right)$ to the subdifferential of the Pareto front mapping . This leads us to new results even in the case of standard linear programs as shown in Section \[lp\].
\[Rem\_C\] *Using the notation of Section \[prelim\] gives us* $$\mathrm{\mathcal{E}_{\mathcal{F}}}\left(b\right)=\mathcal{F}\left(b\right)+\Theta\;\;\mbox{\rm for all }\;b\in\mathbb{R}^{T}\;\;\mbox{\rm with }\;\Theta:=\big\{c_{1},...,c_{q}\big\}^\circ.$$ *From now on we denote* $$C:=-N\left(0_{n};\Theta\right)=\Theta^{\circ}=\mathrm{cone}\big\{c_{1},...c_{q}\big\},$$ *where the last equality immediately follows from the classical Farkas Lemma.*
Here is the aforementioned theorem with the subdifferential calculation. In the paper, and despite $\mathbb{R}^{n}$ is self-dual, we are using $\left\Vert c\right\Vert _{\ast }$ and $\left\Vert a_{t}\right\Vert _{\ast }$ because $c$ and $a_{t}$ are regarded as linear functions ($x\mapsto
\left\langle c,x\right\rangle $ and $x\mapsto \left\langle
a_{t},x\right\rangle $, respectively).** **
\[Theo\_subdiffF\] Let $\left( \overline{b},\overline{x}\right) \in
\mathrm{gph\mathcal{E}_{\mathcal{F}}}$. Then we have the subdifferential formula $$\partial \mathcal{F}\left( \overline{b},\overline{x}\right) =\dbigcup_{\QATOP{\scriptstyle{c\in C}}{\scriptstyle{\Vert c\Vert }_{\ast }{=1}}}\big\{y\in \mathbb{R}^{T}\big|\left\langle y,b-\overline{b}\right\rangle \leq
\left\langle c,x-\overline{x}\right\rangle \;\text{for all}\;\left(
b,x\right) \in \mathrm{gph}\mathcal{S}\big\}. \label{eq_subdiffF}$$
By the convexity of the sets $\mathrm{gph}\mathcal{F}$ and $\mathrm{gph}\mathcal{E}_{\mathcal{F}}$ we get from that $$\partial \mathcal{F}\left( \overline{b},\overline{x}\right) =\dbigcup_{\QATOP{\scriptstyle{c\in C}}{\scriptstyle{\Vert c\Vert }_{\ast }{=1}}}\big\{y\in \mathbb{R}^{T}\big|\;\left\langle y,b-\overline{b}\right\rangle \leq
\left\langle c,x-\overline{x}\right\rangle \;\text{for all }\;\left(
b,x\right) \in \mathrm{gph}\mathcal{F}\big\}.$$Since $\mathrm{gph}\mathcal{S\subset }\mathrm{gph}\mathcal{F}$, we only need to verify the inclusion ‘$\supset $’ of .
To proceed, pick any $c\in C$ with ${\Vert c\Vert }_{\ast }=1$ and select $y\in \mathbb{R}^{T}$ such that $$\left\langle y,b-\overline{b}\right\rangle \leq \left\langle c,x-\overline{x}\right\rangle \;\text{for all }\left( b,x\right) \in \mathrm{gph}\mathcal{S}.
\label{eq_inequalities}$$Arguing by contradiction, suppose that there exists $\left(
b_{0},x_{0}\right) \in \mathrm{gph}\mathcal{F}$ with $$\left\langle y,b_{0}-\overline{b}\right\rangle >\left\langle c,x_{0}-\overline{x}\right\rangle ,$$which yields $\left( b_{0},x_{0}\right) \notin \mathrm{gph}\mathcal{S}$. Applying then Lemma \[Lem\_F\_notS\] to $\left( b_{0},x_{0}\right) $ ensures the existence of $\widetilde{x}_{0}\in \mathcal{S}\left( b_{0}\right) $ such that $\left\langle c_{i},\widetilde{x}_{0}\right\rangle \leq \left\langle
c_{i},x_{0}\right\rangle $ for all $i=1,...,q$. In particular, we get $\left\langle c,\widetilde{x}_{0}\right\rangle \leq \left\langle
c,x_{0}\right\rangle $. Therefore $$\left\langle y,b_{0}-\overline{b}\right\rangle >\left\langle c,x_{0}-\overline{x}\right\rangle \geq \left\langle c,\widetilde{x}_{0}-\overline{x}\right\rangle ,$$which contradicts and thus completes the proof of the theorem.
Now we are ready to establish a precise formula for computing the Lipschitz modulus at $\left(\overline{b},\overline{x}\right)\in\mathrm{gph\mathcal{E}_{\mathcal{F}}}$ of the epigraphical feasible set mapping from . In the next theorem we employ the $l_{1}$-norm $\left\Vert\cdot\right\Vert_{1}$ on $\mathbb{R}^{T}$, which is dual to the primal supremum norm used above.
\[Cor\_lip\_EF\] Let $\left( \overline{b},\overline{x}\right) \in \mathrm{gph\mathcal{E}_{\mathcal{F}}}$. Then we have $$\begin{aligned}
\mathrm{lip}\mathcal{E}_{\mathcal{F}}\left( \overline{b},\overline{x}\right)
&=&\sup \big\{\left\Vert y\right\Vert _{1}\big|\;y\in \partial \mathcal{F}\left( \overline{b},\overline{x}\right) \big\} \\
&=&\sup \big\{\dbigcup_{\QATOP{\scriptstyle{c\in C}}{\scriptstyle{\Vert
c\Vert }_{\ast }{=1}}}\big\{\left\Vert y\right\Vert _{1}\big|\;\left\langle
y,b-\overline{b}\right\rangle \leq \left\langle c,x-\overline{x}\right\rangle \text{ }\forall \left( b,x\right) \in \mathrm{gph}\mathcal{S}\big\},\end{aligned}$$and thus the multifunction $\mathcal{E}_{\mathcal{F}}$ is Lipschitz-like around $\left( \overline{b},\overline{x}\right) $ if and only if $$\sup \big\{\dbigcup_{\QATOP{\scriptstyle{c\in C}}{\scriptstyle{\Vert c\Vert }_{\ast }{=1}}}\Big\{\left\Vert y\right\Vert _{1}\Big|\;\left\langle y,b-\overline{b}\right\rangle \leq \left\langle c,x-\overline{x}\right\rangle
\text{ }\forall \;\left( b,x\right) \in \mathrm{gph}\mathcal{S}\Big\}<\infty
.$$
Observe that $\left( b,x\right) \in \mathrm{gph}\mathcal{F}\Longleftrightarrow a_{t}^{\prime }x-e_{t}^{\prime }b\leq 0$ for all $t=1,...,m$, where $e_{t}\in \mathbb{R}^{m}$** **is the $t$-th vector of the canonical basis of $\mathbb{R}^{m}$. Appealing to Remark \[Rem\_C\], we see that the set $\mathrm{gph}\mathcal{E}_{\mathcal{F}}$ is a polyhedral convex cone admitting the representation $$\mathrm{gph}\mathcal{E}_{\mathcal{F}}=\mathrm{gph}\mathcal{F+}\left(
\{0_{m}\}\times \big\{c_{1},...,c_{q}\big\}^{\circ }\right) ;$$so this set is closed and convex. Thus the claimed modulus formula follows from and Theorem \[Theo\_subdiffF\]. The last statement of this theorem follows directly from the definition of the Lipschitz modulus and the formula for its computation.
Examples \[Exa1\] and \[Exa2\] illustrate both Theorem [Theo\_subdiffF]{} and Theorem \[Cor\_lip\_EF\]. They are included in the next section for comparative purposes, specifically to point out the similarities between the subdifferentials $\partial\mathcal{F}$ and $\partial\mathcal{P}$.
Computation Formulas for Feasible Sets {#scal}
======================================
In this section we derive a precise formula for representing the epigraphical multifunction $\mathcal{E}_{\mathcal{F}}$ from via solutions of a new linear inequality system associated with $\mathcal{F}(b)$. More constructive representations are obtained for some specific forms of feasible solution sets that are especially important for applications. All of this constitutes, in particular, the basis for computations of the optimal value in linear programs, which is illustrated and further developed in Section \[lp\] in the framework of Example [Exa\_new]{}.
Let us start revealing the following relationship between the ‘multiobjective epigraphical feasible set mapping’ $\mathcal{E}_{\mathcal{F}} $ and its linear program counterpart $\mathcal{F}\left(b\right)+\left\{c\right\}^{\circ }$ coming from a certain scalarization technique.
\[Th\_MOP value\] For any $b\in\mathbb{R}^{T}$ we have the relationship $$\mathcal{E}_{\mathcal{F}}\left(b\right)=\bigcap_{c\in\mathrm{conv}\left\{c_{1},...,c_{q}\right\}}\left(\mathcal{F}\left(b\right)+\left\{c\right\}^{\circ}\right).$$
Confining ourselves to the nontrivial case where $\mathcal{F}\left(b\right)\ne\emptyset$, observe first that the inclusion ‘$\subset$’ follows from the obvious fact that $$\left\{c_{1},...,c_{q}\right\}^{\circ}=\bigcap_{c\in\mathrm{conv}\left\{c_{1},...,c_{q}\right\}}\left\{c\right\}^{\circ}.$$ To verify the opposite inclusion ‘$\supset$’, assume that $x\notin\mathcal{E}_{\mathcal{F}}\left(b\right)$ and then show that there exists $c\in\mathrm{conv}\left\{c_{1},...,c_{q}\right\}$ such that $x\notin\mathcal{F}\left(b\right)+\left\{c\right\}^{\circ }$. Denote by $\widehat{x}$ the Euclidean projection of $x$ onto $\mathcal{E}_{\mathcal{F}}\left(b\right)$. It is well known that $$\label{eq_ineq proy}
\left\langle\widehat{x}-x,y\right\rangle\ge\left\langle\widehat{x}-x,\widehat{x}\right\rangle \text{ for all }\;y\in\mathcal{E}_{\mathcal{F}}\left(b\right).$$ In particular, for any $y_{0}\in\mathcal{F}\left(b\right)$, all $u\in\left\{c_{1},...,c_{q}\right\}^{\circ}$, and all $\lambda>0$ we have $\left\langle\widehat{x}-x,y_{0}+\lambda u\right\rangle\ge\left\langle\widehat{x}-x,\widehat{x}\right\rangle$. Dividing both sides of the latter inequality by $\lambda>0$ and letting $\lambda\rightarrow\infty$ give us $\left\langle\widehat{x}-x,u\right\rangle\ge 0$, i.e., $$\widehat{x}-x\in\left\{c_{1},...,c_{q}\right\}^{\circ\circ}=\mathrm{cone}\left\{c_{1},...,c_{q}\right\}.$$ Thus we have $\widehat{x}-x=\mu c$ for some $c\in\mathrm{conv}\left\{c_{1},...,c_{q}\right\}$ and some $\mu>0$ by taking into account that $\widehat{x}-x\ne 0_{n}$. To verify now that $x\not\in\mathcal{F}\left(b\right)+\left\{c\right\}^{\circ}$, suppose the contrary and then deduce from the above that $x=y+u$ with some $y\in\mathcal{F}\left( b\right)$ and $u\in\left\{c\right\}^{\circ}$. It tells us that $$\left\langle c,x\right\rangle=\left\langle c,y\right\rangle+\left\langle
c,u\right\rangle\ge\left\langle c,y\right\rangle\ge\left\langle c,\widehat{x}\right\rangle>\left\langle c,x\right\rangle,$$ where the penultimate step comes from , while the last one follows from the projection inequality $$\left\langle c,\widehat{x}-x\right\rangle=\frac{1}{\mu}\left\Vert\widehat{x}-x\right\Vert_{2}^{2}$$ with $\left\Vert\cdot\right\Vert_{2}$ standing for the Euclidean norm. The obtained contradiction completes the proof of the theorem.
*Observe that in Theorem \[Th\_MOP value\] we cannot avoid the convex combination in the representation of $\mathcal{E}_{\mathcal{F}}\left(b\right) $, i.e., replace $\mathrm{conv}\left\{c_{1},...,c_{q}\right\} $ by $\left\{c_{1},...,c_{q}\right\}$.* *To illustrate it, consider the case where* $\mathbb{R}^{T}=\mathbb{R}^{2}$, $$\mathcal{F}\left( b\right)=\mathrm{conv}\left\{\left(1,0\right),\left(0,1\right)\right\},\;\emph{and }\;
c_{1}=\left(1,0\right),\;c_{2}=\left(0,1\right).$$ *However, the set* $\mathrm{conv}\left\{c_{1},...,c_{q}\right\}$ *can be replaced by any basis of the cone* $C$.
To establish efficient representations of the sets in the form $\mathcal{F}\left(b\right)+\left\{c\right\}^{\circ}$, and hence of $\mathcal{E}_{\mathcal{F}}\left(b\right)$ due to Theorem \[Th\_MOP value\], we focus now on multifunctions $\mathcal{F}$ defined as $$\mathcal{F}\left(\cdot\right)+\mathrm{cone}\left\{u\right\}\;\mbox{ and }\;\mathcal{F}\left(\cdot\right)+\mathrm{span}\left\{u\right\}.$$ This is done in the remainder of this section.
Given $u\in\mathbb{R}^{n}$, consider first the polyhedral set $\mathcal{F}\left(b\right)+\mathrm{cone\,}\left\{u\right\}$ and introduce the following partition $\left\{T_{1},T_{2}\right\}$ of $T=\left\{1,...,m\right\}$: $$\label{Eq_T12}
\begin{array}{cc}
T_{1}:=\big\{t\in T\big|\;\left\langle a_{t},u\right\rangle\le 0\big\}\;\mbox{ and }\; T_{2}:=\big\{t\in T\big|\;\left\langle a_{t},u\right\rangle>0\big\}. &
\end{array}$$ Then for each $t\in T_{1}$ we denote $$a_{\left(t,0\right)}:=a_{t}\text{ and }b_{\left(t,0\right)}:=b_{t}$$ and for each $\left(t,s\right)\in T_{1}\times T_{2}$ denote $$\label{eq_a_ts}
\begin{array}{cc}
a_{\left(t,s\right)}:=\left\langle a_{s},u\right\rangle a_{t}-\left\langle
a_{t},u\right\rangle a_{s}, & b_{\left(t,s\right)}:=\left\langle
a_{s},u\right\rangle b_{t}-\left\langle a_{t},u\right\rangle b_{s}.\end{array}$$ With $\widetilde{T}:=T_{1}\times\left(\{0\}\cup T_{2}\right)$ let us now define the linear inequality system $$\label{eq_sigmatilde}
\widetilde{\sigma}\left( b\right):=\left\{\left\langle
a_{\left(t,s\right)},x\right\rangle\le
b_{\left(t,s\right)},~\left(t,s\right)\in\widetilde{T}\right\}$$ and denote by $\widetilde{\mathcal{F}}\left(b\right)$ the set of feasible solutions to $\widetilde{\sigma}\left(b\right)$.
*If* $T_{1}=\emptyset $.* then* $\widetilde{T}=\emptyset $ * and* $\widetilde{\mathcal{F}}\left( b\right) =\mathbb{R}^{n}$. *Otherwise we have that* $\left( a_{\left( t,s\right) },b_{\left( t,s\right)
}\right) $* is a conic combination of* $\left( a_{t},b_{t}\right) $ *and* $\left( a_{s},b_{s}\right) $ *for all* $\left( t,s\right)
\in \widetilde{T}$. *It is clear then that* $\mathcal{F}\left( b\right)
\subset \widetilde{\mathcal{F}}\left( b\right) $. *Observe also that, in contrast to* $\sigma \left( b\right) $, *the new system* $\widetilde{\sigma }\left( b\right) $ *is no longer parameterized by its RHS.*
The next theorem represents $\mathcal{F}\left(b\right)+\mathrm{cone\,}\left\{u\right\}$ as the set of feasible solutions to the new linear inequality system .
\[Prop\_system\_epi\] In terms of the notation above, for any $b\in\mathbb{R}^{T}$ we have $$\label{eq_ef}
\widetilde{\mathcal{F}}\left(b\right)=\mathcal{F}\left(b\right)+\mathrm{cone\,}\left\{u\right\}.$$
Let us first verify the inclusion ‘$\supset$’ in . Taking $x\in\mathcal{F}\left(b\right)+\mathrm{cone\,}\left\{u\right\}$, we get the linear inequalities $$\label{eq_10}
\left\langle a_{t},x-\mu u\right\rangle\le b_{t}\text{ for all }\;t\in T
\text{ and some }\;\mu\ge 0.$$ There is nothing to prove if $T_1=\emptyset$. Otherwise we fix $t\in T_{1}$ and get $\left\langle a_{t},x\right\rangle\le b_{t}$. Taking further $s\in
T_{2}$, we distinguish the following two cases. If $\left\langle
a_{t},u\right\rangle=0$, then the aimed inequality $$\left\langle a_{\left(t,s\right)},x\right\rangle\le b_{\left(t,s\right)}$$ reduces to $\left\langle a_{t},x\right\rangle\le b_{t}$. In the case where $\left\langle a_{t},u\right\rangle<0$ we deduce from that $$\frac{\left\langle a_{s},x\right\rangle-b_{s}}{\left\langle
a_{s},u\right\rangle}\le\mu\le\frac{\left\langle a_{t},x\right\rangle-b_{t}}{\left\langle a_{t},u\right\rangle}.$$ In particular, it follows that $$\frac{\left\langle a_{s},x\right\rangle-b_{s}}{\left\langle
a_{s},u\right\rangle}\le\frac{\left\langle a_{t},x\right\rangle-b_{t}}{\left\langle a_{t},u\right\rangle},$$ which readily implies that $$\left\langle a_{\left(t,s\right)},x\right\rangle\le b_{\left(t,s\right)}$$ and thus verifies the inclusion ‘$\supset$’ in .
To prove now the opposite inequality ‘$\subset$’ in , pick any $x\in\widetilde{\mathcal{F}}\left(b\right)$ and let us verify the existence of $\mu\ge 0$ such that $$\left\langle a_{t},x-\mu u\right\rangle\le b_{t}\;\mbox{ for all }\;t\in T.$$ Indeed, when $T_{2}=\emptyset$ we get $x\in\mathcal{F}\left(b\right)$, which agrees in this case with $\widetilde{\mathcal{F}}\left(b\right)$. If $T_{2}\ne\emptyset$, it is sufficient to consider any $\mu\ge 0$ satisfying $$\max_{s\in T_{2}}\frac{\left\langle a_{s},x\right\rangle-b_{s}}{\left\langle
a_{s},u\right\rangle}\le\mu\le\inf_{\left\langle a_{t},u\right\rangle<0}\frac{\left\langle a_{t},x\right\rangle-b_{t}}{\left\langle
a_{t},u\right\rangle}$$ under the usual convention that $\inf\emptyset=\infty$. We complete the proof of the theorem by observing that such a number $\mu$ exists due to the choice of $x$.
Looking closely at the proof of Theorem \[Prop\_system\_epi\] tells us that the successive application of the procedure therein is instrumental to represent the more general feasible sets $\mathcal{F}\left( b\right) +\mathrm{cone\,}\left\{ u_{1},...,u_{p}\right\} $ via linear inequality systems. However, explicit forms of such representations may generally be rather complicated. In the next theorem we consider the important case where $$\mathcal{F}\left( b\right) +\mathrm{span\,}\big\{u\big\}$$for which we give a direct proof.
\[Prop\_F+Ru\] Given any $b\in \mathbb{R}^{T}$ and recalling the notation in , we have $$\mathcal{F}\left( b\right) +\mathrm{span}\big\{u\big\}=\left\{ x\in \mathbb{R}^{n}\left\vert
\begin{array}{ll}
\left\langle a_{t},x\right\rangle \leq b_{t}\text{ if }\left\langle
a_{t},u\right\rangle =0, & \\
\left\langle a_{\left( t,s\right) },x\right\rangle \leq b_{\left( t,s\right)
}\text{ if}\;\left\langle a_{t}u\right\rangle <0 & \\
\qquad \qquad \qquad \quad \;\;\mbox{and}\,\left\langle a_{s},u\right\rangle
>0 &
\end{array}\right. \right\} . \label{eq_F+span}$$
Let us introduce the index set$$T_{0}:=\{t\in T\mid \left\langle a_{t},u\right\rangle =0\}\text{(}\subset
T_{1}\text{).}$$Note that, in the case $T_{1}=\emptyset $ (or $T_{2}=T_{0}=\emptyset ),$ the system (\[eq\_F+span\]) has no inequality (i.e., its solution set is the whole space $\mathbb{R}^{n}$), but in such cases $\mathcal{F}\left( b\right)
+\mathrm{span\,}\left\{ u\right\} $ is also $\mathbb{R}^{n}$, as $-u\in
\limfunc{int}(O^{+}\mathcal{F}\left( b\right) )$ (or $u\in \limfunc{int}(O^{+}\mathcal{F}\left( b\right) )$, respectively), entailing that if $x_{0}\in \mathcal{F}\left( b\right) ,$ $$x_{0}+\mathrm{span\,}\left\{ u\right\} +\lambda \mathbb{B\subset \ }\mathcal{F}\left( b\right) +\mathrm{span\,}\left\{ u\right\} ,\text{ for all }\lambda
\geq 0.$$If $T_{1}\diagdown T_{0}$ and $T_{2}$ are both nonempty, the reasoning is the same followed in Proposition \[Prop\_system\_epi\], without taking into account the sign of $\mu .$
The reader will see in Example \[Exa\_new\] below a detailed illustration of both Theorems \[Prop\_system\_epi\] and \[Prop\_F+Ru\] together with additional comments on the relationship between the optimal value and the epigraphical mapping $\mathcal{E}_{\mathcal{F}}$ in linear programming.
Subdifferentials of Epigraphical Pareto Fronts {#pareto}
==============================================
This section concerns the epigraphical Pareto front multifunction $\mathcal{E}_{\mathcal{P}}\colon\mathbb{R}^{T}\mathbb{\rightrightarrows R}^{q}$ introduced in in the form $$\mathcal{E}_{\mathcal{P}}\left(b\right):=\mathcal{P}\left(b\right)+\mathbb{R}_{+}^{q},\quad b\in\mathbb{R}^{T},$$where the Pareto front mapping $\mathcal{P}$ is defined in . In contrast to $\mathrm{gph}\mathcal{F}$, the set $\mathrm{gph\mathcal{P}}$ is nonconvex in general, while the one of our interest $\mathrm{gph\mathrm{\mathcal{E}}_{\mathrm{\mathcal{P}}}}$ is *always convex*. This is shown in the next proposition.
\[Prop\_EP\_convex\] The set $\mathrm{gph\mathrm{\mathcal{E}}_{\mathrm{\mathcal{P}}}}$ is a closed and convex subset of $\mathbb{R}^{T}\times\mathbb{R}^{q}$.
First we observe that the set $\mathrm{gph\mathcal{P}}$ is a finite union of convex polyhedral cones as the KKT (or primal/dual) optimality conditions in linear programming allow us to express $\mathrm{gph\mathcal{P}}$ as the graph of a certain feasible set mapping of a linear system and we can apply then the classical result by Robinson [@Robinson]. Hence *a fortiori* the set $\mathrm{gph\mathrm{\mathcal{E}}_{\mathrm{\mathcal{P}}}}=\mathrm{gph\mathcal{P+}}\left( \left\{ 0_{m}\right\} \times \mathbb{R}_{+}^{q}\right) $ is also closed.
Let us now show that the set $\mathrm{gph\mathrm{\mathcal{E}}_{\mathrm{\mathcal{P}}}}$ is convex. Fix any two pairs $\left(b_{1},p_{1}+u_{1}\right)$, $\left(b_{2},p_{2}+u_{2}\right)\in\mathrm{gph\mathrm{\mathcal{E}}_{\mathrm{\mathcal{P}}}}$, i.e., such that $b_{i}\in\mathbb{R}^{T}$, $p_{i}\in\mathcal{P}\left(b_{i}\right)$, and $u_{i}\in\mathbb{R}_{+}^{q}$ as $i=1,2$. Then for every $\lambda\in\lbrack 0,1]$ we have $$p_{i}=\left(\left\langle c_{1},x_{i}\right\rangle,...,\left\langle
c_{q},x_{i}\right\rangle\right) \text{ with some }\;x_{i}\in\mathcal{S}\left(b_{i}\right),\text{ }i=1,2,$$ and so $\left(1-\lambda\right)x_{1}+\lambda x_{2}\in\mathcal{F}\left(\left(1-\lambda\right)b_{1}+\lambda b_{2}\right)$. In the nontrivial case where $$\left(1-\lambda\right)x_{1}+\lambda x_{2}\notin\mathcal{S}\left(\left(1-\lambda\right)b_{1}+\lambda b_{2}\right)$$ we apply Lemma \[Lem\_F\_notS\] to get the existence of $\widetilde{x}\in\mathcal{S}\left(\left(1-\lambda\right) b_{1}+\lambda b_{2}\right)$ with $\left\langle c_{i},\widetilde{x}\right\rangle\le\left\langle
c_{i},\left(1-\lambda\right)x_{1}+\lambda x_{2}\right\rangle$ for all $c_1,\ldots,c_q$. It implies that $$\left(1-\lambda\right)p_{1}+\lambda p_{2}\in\left(\left\langle c_{1},
\widetilde{x}\right\rangle ,...,\left\langle c_{q},\widetilde{x}\right\rangle\right)+\mathbb{R}_{+}^{q},$$ which can be equivalently written as $$\left(1-\lambda\right) p_{1}+\lambda p_{2}\in\mathcal{E}_{\mathcal{P}}\left(\left(1-\lambda\right)b_{1}+\lambda b_{2}\right).$$ Therefore, we arrived at the inclusion $$\left(1-\lambda\right)\left(p_{1}+u_{1}\right)+\lambda\left(p_{2}+u_{2}\right)\in\mathcal{E}_{\mathcal{P}}\left(\left(1-\lambda\right)b_{1}+\lambda
b_{2}\right),$$ which verifies the convexity of the set $\mathrm{gph\mathrm{\mathcal{E}}_{\mathrm{\mathcal{P}}}}$ .
Using the above proposition and employing the fundamental results of Theorem \[Cor\_lip\_subdiff\], we can now conduct a local stability analysis of the epigraphical Pareto front mapping similarly to that for the epigraphical feasible solution mapping developed in Section \[feas\].
\[Cor\_lip\_EP\] Let $\left( \overline{b},\overline{p}\right) \in \mathrm{gph\mathcal{P}}$. Then we have $$\partial \mathcal{P}\left( \overline{b},\overline{p}\right) =\dbigcup_{\QATOP{\scriptstyle{\alpha \in \mathbb{R}_{+}^{q}}}{\scriptstyle{\Vert
\alpha \Vert }_{\ast }{=1}}}\big\{y\in \mathbb{R}^{T}\big|\;\left\langle y,b-\overline{b}\right\rangle \leq \left\langle \alpha ,p-\overline{p}\right\rangle \text{ for all }\left( b,p\right) \in \mathrm{gph}\mathcal{P}\big\}.$$Furthermore, the Lipschitz modulus of the epigraphical Pareto front mapping $\mathcal{E}_{\mathcal{P}}$ at $\left( \overline{b},\overline{x}\right) $ is computed by the formula $$\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left( \overline{b},\overline{x}\right)
=\sup \left\{ \{\left\Vert y\right\Vert _{1}\big|\;y\in \partial \mathcal{P}\left( \overline{b},\overline{p}\right) \right\} ,$$which ensures that the mapping $\mathcal{E}_{\mathcal{P}}$ is Lipschitz-like around $\left( \overline{b},\overline{p}\right) $ if and only if $$\sup \Big\{\{\left\Vert y\right\Vert _{1}\Big|\;\dbigcup_{\QATOP{\scriptstyle{\alpha \in \mathbb{R}_{+}^{q}}}{\scriptstyle{\Vert \alpha \Vert }_{\ast }{=1}}}\big\{y\in \mathbb{R}^{T}\big|\;\left\langle y,b-\overline{b}\right\rangle \leq \left\langle \alpha ,p-\overline{p}\right\rangle \text{for all}\;\left( b,p\right) \in \mathrm{gph}\mathcal{P}\big\}\Big\}<\infty .$$
Having in hand Proposition \[Prop\_EP\_convex\], we can apply Theorem [Cor\_lip\_subdiff]{} and then proceed similarly to the proofs of Theorems [Cor\_lip\_subdiff]{} and \[Cor\_lip\_EF\].
The next result expresses the subdifferential $\partial\mathcal{F}\left(\overline{b},\overline{x}\right)$ in terms of $\mathrm{gph}\mathcal{P}$ instead of $\mathrm{gph}\mathcal{S}$. Observe that the difference between the expression for $\partial\mathcal{F}\left(\overline{b},\overline{x}\right) $ obtained below and the one for $\partial\mathcal{P}\left(\overline{b},\overline{p}\right)$ established in Theorem \[Cor\_lip\_EP\] is seen only in the sets where the vector $\alpha\in\mathbb{R}_{+}^{q}$ takes its values.
\[Prop\_subdiffF\] Let $\left( \overline{b},\overline{x}\right) \in
\mathrm{gph\mathcal{S}}$, and let $\overline{p}=\left( \left\langle c_{1},\overline{x}\right\rangle ,...,\left\langle c_{q},\overline{x}\right\rangle
\right) \in \mathrm{\mathcal{P}}\left( \overline{b}\right) $. Then the subdifferential of $\mathcal{F}$ at $\left( \overline{b},\overline{x}\right)
$ is computed by $$\partial \mathcal{F}\left( \overline{b},\overline{x}\right) =\bigcup_{\QATOP{\scriptstyle{\alpha \in \mathbb{R}_{+}^{q}}}{\scriptstyle{\left\Vert \dsum
\alpha _{i}c_{i}\right\Vert }_{\ast }{=1}}}\big\{y\in \mathbb{R}^{T}\big|\;\left\langle y,b-\overline{b}\right\rangle \leq \left\langle \alpha ,p-\overline{p}\right\rangle \;\text{ for all }\;\left( b,p\right) \in \mathrm{gph}\mathcal{P}\big\}.$$
Taking into account the previous considerations, we proceed similarly to the proof of Theorem \[Theo\_subdiffF\].
Let us now present a two-dimensional numerical example that illustrates how both Theorems \[Cor\_lip\_EP\] and \[Prop\_subdiffF\] can be applied in computation.
\[Exa1\] *Take any* $a>0$ *and consider the following multiobjective problem* *with $n=q=2$ and the Euclidean norms on both spaces*: $$\begin{tabular}{ccc}
$MLP\left( b\right) :$ & \textrm{minimize} & $\left( ax_{1},x_{2}\right) $
\\
& subject to & $x_{1}\geq b_{1}$, \\
& & $x_{2}\geq b_{2}$.\end{tabular}$$*Letting* $\overline{b}=0_{2}$, *we easily see that* $$\begin{aligned}
& \mathcal{F}\left( b\right) =\mathcal{E}_{\mathcal{F}}\left( b\right)
=\{\left( b_{1},b_{2}\right) \}+\mathbb{R}_{+}^{2},\;\mathcal{S}\left(
b\right) =\text{ }\{\left( b_{1},b_{2}\right) \},\text{ } \\
& \mathcal{P}\left( b\right) =\text{ }\{\left( ab_{1},b_{2}\right) \},\;\text{\textrm{and }}\mathcal{E}_{\mathcal{P}}\left( b\right) =\{\left(
ab_{1},b_{2}\right) \}+\mathbb{R}_{+}^{2}.\end{aligned}$$*Furthermore, using Theorems \[Theo\_subdiffF\] and \[Cor\_lip\_EP\] tells us, respectively, that* $$\begin{aligned}
\partial \mathcal{F}\left( 0_{2},0_{2}\right) &=&\bigcup_{\QATOP{\scriptstyle{\alpha ^{2}\alpha _{1}^{2}+\alpha _{2}^{2}=1}}{\scriptstyle{\alpha
_{1},\alpha _{2}\geq 0}}}\big\{y\in \mathbb{R}^{2}\big|\;b_{1}y_{1}+b_{2}y_{2}\leq a\alpha _{1}b_{1}+\alpha _{2}b_{2}\;\text{ }\forall \;b_{1},b_{2}\in \mathbb{R}\big\} \\
&=&\Big\{y\in \mathbb{R}^{2}\Big|\;y_{1}^{2}+y_{2}^{2}=1,\;y_{1},y_{2}\geq 0\Big\},\end{aligned}$$$$\begin{aligned}
\partial \mathcal{P}\left( 0_{2},0_{2}\right) &=&\bigcup_{\QATOP{\scriptstyle{\alpha _{1}^{2}+\alpha _{2}^{2}=1}}{\scriptstyle{\alpha _{1},\alpha
_{2}\geq 0}}}\Big\{y\in \mathbb{R}^{2}\Big|\;b_{1}y_{1}+b_{2}y_{2}\leq
a\alpha _{1}b_{1}+\alpha _{2}b_{2}\text{ }\forall \;b_{1},b_{2}\in \mathbb{R}\Big\} \\
&=&\Big\{\left( y_{1},y_{2}\right) \in \mathbb{R}^{2}\Big|\;\frac{y_{1}^{2}}{a^{2}}+y_{2}^{2}=1,\text{ }y_{1},y_{2}\geq 0\Big\}.\end{aligned}$$*Since $\mathcal{F}=\mathcal{E}_{\mathcal{F}}$ in this case,* *we can appeal to Theorem \[Cor\_lip\_EF\] (cf. also [CDLP05]{}) to compute the Lipschitz modulus:* $$\mathrm{lip}\mathcal{F}\left( 0_{2},0_{2}\right) =\mathrm{lip}\mathcal{E}_{\mathcal{F}}\left( 0_{2},0_{2}\right) =\sqrt{2}.$$*Considering now the mappings* $\mathcal{P}$* and* $\mathcal{E}_{\mathcal{P}}$, *we can treat them as the feasible set mappings for the equality and inequality systems with respect to the variables* $\left(
p_{1},p_{2}\right) $. *Namely, as $(1/a)p_{1}=b_{1},p_{2}=b_{2}$ and* $(1/a)p_{1}\geq b_{1},p_{2}\geq b_{2}$, *respectively. Appealing to Theorem \[Cor\_lip\_EP\] (cf. also [@CDLP05 Corollary 3.2]), we obtain* $$\mathrm{lip}\mathcal{P}\left( 0_{2},0_{2}\right) =\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left( 0_{2},0_{2}\right) =\sqrt{1+a^{2}}.$$
As follows from , we have $$\label{ex-ine}
\mathrm{lip}\mathcal{P}\left(\overline{b},\overline{p}\right)\ge\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left(\overline{b},\overline{p}\right), \text{ for
any }\left(\overline{b},\overline{p}\right)\in\mathrm{gph\mathcal{P}}.$$
The next example shows that the inequality in may be *strict*.
\[Exa2\] *Consider the multiobjective problem* *with $n=q=2$ and the Euclidean norms on both spaces*: $$\begin{tabular}{ccc}
$MLP\left( b\right) :$ & \textrm{minimize} & $\left( x_{1},x_{2}\right) $ \\
& subject to & $x_{1}\geq b_{1}$, \\
& & $x_{1}+x_{2}\geq b_{2}$.\end{tabular}$$*It is easy to check that* $\mathcal{F}\left( b\right) =\mathcal{E}_{\mathcal{F}}\left( b\right) $ for any $b\in \mathbb{R}^{2}$ *and that* $$\mathrm{lip}\mathcal{F}\left( 0_{2},0_{2}\right) =\mathrm{lip}\mathcal{E}_{\mathcal{F}}\left( 0_{2},0_{2}\right) =1.$$*On the other hand, we clearly have the expressions* $$\begin{aligned}
\mathcal{P}\left( b\right) & =\big\{p\in \mathbb{R}^{2}\big|\;p_{1}\geq
b_{1},p_{1}+p_{2}=b_{2}\},\text{ }b\in \mathbb{R}^{2}, \\
\mathcal{E}_{\mathcal{P}}\left( b\right) & =\{p\in \mathbb{R}^{2}\big|\;p_{1}\geq b_{1},p_{1}+p_{2}\geq b_{2}\},\text{ }b\in \mathbb{R}^{2},\end{aligned}$$*with the strict inequality* $$\mathrm{lip}\mathcal{P}\left( 0_{2},0_{2}\right) =\sqrt{5}>\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left( 0_{2},0_{2}\right) =1.$$
Observe that the situation of Example \[Exa2\] does not occur in the case of single-objective linear programs, where we always have $\mathrm{lip}\mathcal{P}\left(\overline{b},\overline{p}\right)=\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left(\overline{b},\overline{p}\right)$ with $\left(\overline{b},\overline{p}\right)\in\mathrm{gph\mathcal{P}}$. This is one of the main points of the next section.
Lipschitz Moduli in Linear Programming {#lp}
======================================
In this section we provide specifications and further developments of the results obtained above for the general linear multiobjective problem for the case of ordinary linear programs given by $$\label{eq_LP}
\begin{tabular}{ccc}
$LP\left(b\right):$ & \emph{\textrm{minimize}} & $\left\langle
c,x\right\rangle$ \\
& \emph{subject to} & $x\in\mathcal{F}\left(b\right)$\end{tabular}$$ where the vector $c\in\mathbb{R}^{n}$ is fixed. As shown below, the approach and results developed for linear multiobjective problems lead us to refined computation formulas for subdifferentials and Lipschitz moduli in standard problems of linear programming under parameter perturbations.
In what follows we assume that $-c\in \mathrm{cone}\{a_{t},t\in T\}$ (dual consistency); otherwise the problem $LP\left( b\right) $ is always unsolvable. Denote by $\vartheta \colon \mathbb{R}^{T}\rightarrow \mathbb{R}\cup \{+\infty \}$ the associated *optimal value function* defined by $$\vartheta \left( b\right) :=\inf \big\{\left\langle c,x\right\rangle \big|\;x\in \mathcal{F}\left( b\right) \big\}.$$We can easily see in this framework that $$\mathrm{dom}\mathcal{\vartheta }:=\big\{b\in \mathbb{R}^{T}\big|\;\vartheta
\left( b\right) <\infty )=\mathrm{dom}\mathcal{F}.$$Indeed, it is well known in linear programming that the boundedness of $LP\left( b\right) $ is equivalent to its solvability, which is in turn equivalent to the simultaneous fulfilment of primal and dual consistency.
Observe that in the setting of the multifunctions $\mathcal{S}$, $\mathcal{P}$, $\mathcal{E}_{\mathcal{P}}$, and $\mathcal{E}_{\mathcal{F}}$ admit the following specifications. For each $b\in\mathbb{R}^{T}$ we have that $\mathcal{S}\left(b\right)$ is the set of optimal solutions to $LP\left(b\right)$ while the mapping $\mathcal{P}\colon\mathbb{R}^{T}\rightrightarrows\mathbb{R}$ is actually single-valued given by $$\mathcal{P}\left(b\right)=\left\{\begin{array}{l}
\left\{\vartheta\left(b\right)\right\} \text{ if }\;b\in\mathrm{dom}\mathcal{S}, \\
\emptyset \text{ if }\;b\notin\mathrm{dom}\mathcal{S}.\end{array}\right.$$ Furthermore, we get the relationships $$\mathrm{dom}\mathcal{P}=\mathrm{dom}\mathcal{S}=\mathrm{dom}\mathcal{F}=\mathrm{dom}\mathcal{\vartheta}.$$ Taking into account that $\mathcal{P}\left(b\right)$ is a singleton for any $b\in\mathrm{dom}\mathcal{S}$, from now on we write $\partial\mathcal{P}\left(b\right)$ instead of $\partial\mathcal{P}\left(b,\vartheta\left(b\right)\right)$. Moreover, for each such $b$ it follows that $$\mathcal{E}_{\mathcal{P}}\left(b\right)=[\vartheta\left(b\right),\infty),$$ and it is easily to verify that $$\label{eq_F+co geq v}
\mathcal{E}_{\mathcal{F}}\left(b\right):=\mathcal{F}\left(
b\right)+\left\{c\right\}^{\circ }=\big\{x\in\mathbb{R}^{n}\big|\;\left\langle c,x\right\rangle\ge \vartheta\left(b\right)\big\}.$$ It allows us to show below that Lipschitzian behavior of $\mathcal{E}_{\mathcal{F}}$ is closely related to that of $\mathcal{P};$ see Theorem [Th lipV]{}. To proceed, we first present the following proposition.
\[Prop\_sect5\] For any $\left( \overline{b},\overline{x}\right) \in
\mathrm{gph}\mathcal{S}$ we have $$\partial \mathcal{P}\left( \overline{b}\right) =\left\Vert c\right\Vert
_{\ast }\partial \mathcal{F}\left( \overline{b},\overline{x}\right) .$$
It follows from Theorems \[Cor\_lip\_EP\] and \[Prop\_subdiffF\] that $$\begin{aligned}
\partial \mathcal{P}\left( \overline{b}\right) & \big\{y\in \mathbb{R}^{T}\big|\;\left\langle y,b-\overline{b}\right\rangle \leq p-\overline{p}\text{ }\forall \left( b,p\right) \in \mathrm{gph}\mathcal{P}\big\} \\
& =\left\Vert c\right\Vert _{\ast }\Big\{y\in \mathbb{R}^{T}\Big|\;\left\langle y,b-\overline{b}\right\rangle \leq \frac{1}{\left\Vert
c\right\Vert _{\ast }}\left( p-\overline{p}\right) \text{ }\forall \left(
b,p\right) \in \mathrm{gph}\mathcal{P}\Big\} \\
& =\left\Vert c\right\Vert _{\ast }\partial \mathcal{F}\left( \overline{b}\overline{x}\right) ,\end{aligned}$$which therefore verifies the claimed equality.
\[Rem\_neigh\] *Given $\overline{b}\in\mathrm{dom}\mathcal{P}$* *and remembering that* $\mathcal{P}\left(b\right)$* is a singleton for any* $b\in\mathrm{dom}\mathcal{P}$, *we can write* $$\partial\mathcal{P}\left(\overline{b}\right)=\big\{y\in\mathbb{R}^{T}\big|\;\left\langle y,b-\overline{b}\right\rangle\le\vartheta\left(b\right)-\vartheta\left(\overline{b}\right) \text{ }\forall\;b\in\mathrm{dom}\mathcal{P}\big\},$$ *which agrees with the classical subdifferential of $\mathcal{P}$ at $\overline{b}$ in the sense of convex analysis*. *It is actually not surprising since the convexity of the function* $\vartheta$ *can be clearly derived from Proposition \[Prop\_EP\_convex\]. Going a little further, observe that the set* $\mathrm{dom}\mathcal{F}$ *can be replaced by any intersection of the form* $\mathrm{dom}\mathcal{F\cap U}_{\overline{b}}$, *where* $\mathcal{U}_{\overline{b}}\in\mathbb{R}^{T}$ *is an arbitrary neighborhood of* $\overline{b}$.
Now we are ready to formulate and prove the last theorem of this paper.
\[Th lipV\] Let $\left( \overline{b},\overline{x}\right) \in \mathrm{gph}\mathcal{S}$. Then $$\mathrm{lip}\mathcal{P}\left( \overline{b}\right) =\mathrm{lip}\mathcal{E}_{\mathcal{P}}\left( \overline{b}\right) =\left\Vert c\right\Vert _{\ast }\mathrm{lip}\mathcal{E}_{\mathcal{F}}\left( \overline{b},\overline{x}\right)
.$$
The first equality is standard since $\mathcal{P}$ is single-valued on $\mathbb{R}$. The second equality follows from Theorem \[Cor\_lip\_EP\] and Proposition \[Prop\_sect5\] with taking into account the fact that the Lipschitz moduli under consideration agree with the suprema of the norms in the corresponding subdifferentials.
The next example shows how the obtained results are applied in the case of two-dimensional linear programming with multiply inequality constraints.
\[Exa\_new\] *Consider the following parameterized linear program in the space* $\mathbb{R}^{2}$* with the Euclidean norm on it*: $$\begin{tabular}{ccc}
$PL\left( b\right) :$ & \textrm{minimize} & $2x_{1}+x_{2}$ \\
& subject to & \multicolumn{1}{r}{$-x_{1}-x_{2}\leq b_{1}$} \\
& & \multicolumn{1}{r}{$-x_{1}+2x_{2}\leq b_{2}$} \\
& & \multicolumn{1}{r}{$-2x_{1}\leq b_{3}$} \\
& & \multicolumn{1}{r}{$3x_{1}+x_{2}\leq b_{4}$}\end{tabular}$$*around the nominal parameter* $\overline{b}=\left( -2,1,-2,7\right) $. *Since* $$\mathcal{F}\left( b\right) +\left\{ c\right\} ^{\circ }=\mathcal{F}\left(
b\right) +\mathrm{cone}\left\{ \left( 2,1\right) \right\} +\mathrm{span\,}\left\{ \left( -1,2\right) \right\} ,$$*we first apply Theorem \[Prop\_system\_epi\] with* $u=\left(
2,1\right) $. *Recalling the notation above tells us that* $T_{1}=\{1,2,3\},\;T_{2}=\{4\}$, *and* $$\widetilde{\sigma }\left( b\right) =\left\{
\begin{array}{l}
-x_{1}-x_{2}\leq b_{1},~-x_{1}+2x_{2}\leq b_{2},~-2x_{1}\leq b_{3}, \\
2x_{1}-4x_{2}\leq 7b_{1}+3b_{4},~-7x_{1}+14x_{2}\leq 7b_{2}, \\
-2x_{1}+4x_{2}\leq 7b_{3}+4b_{4}\end{array}\right\} .$$*Note that the fifth constraint is equivalent to the second one for all* $b\in \mathbb{R}^{\widetilde{T}}$, *while the sixth constraint is redundant at* $\overline{b}$ *but not at any* $b$. *Remark [Rem\_neigh]{} implies that the sixth inequality is irrelevant in a local analysis around* $\overline{b}$. *Anyway, let us remove just the fifth inequality and renumber the resulting* $\widetilde{T}$ *as* $\{1,...,5\}.$ *Then apply Theorem \[Prop\_F+Ru\] to the reduced and renumbered system* $\widetilde{\sigma }$ *with* $\widetilde{u}=\left(
-1,2\right) $ *to obtain* $\left\langle \widetilde{a}_{t},\widetilde{u}\right\rangle <0$ *for* $t\in \{1,4\}$ *and* $\left\langle
\widetilde{a}_{s},\widetilde{u}\right\rangle >0$ *for* $s\in \{2,3,5\}$. *It gives us* $$\widetilde{\widetilde{\sigma }}\left( b\right) =\left\{
\begin{array}{c}
-6x_{1}-3x_{2}\leq 5b_{1}+b_{2},~-4x_{1}-2x_{2}\leq 2b_{1}+b_{3}, \\
-12x_{1}-6x_{2}\leq 10b_{1}+7b_{3}+4b_{4},~0\leq 35b_{1}+10b_{2}+15b_{4}, \\
-16x_{1}-8x_{2}\leq 14b_{1}+10b_{3}+6b_{4},~0\leq 70b_{1}+70b_{3}+70b_{4}\end{array}\right\} .$$*Hence for any* $b\in \mathbb{R}^{4}$ *the system* $\widetilde{\widetilde{\sigma }}\left( b\right) $* is equivalent to the single inequality* $$2x_{1}+x_{2}\geq \max \left\{ -\frac{5b_{1}+b_{2}}{3},-\frac{2b_{1}+b_{3}}{2},-\frac{10b_{1}+7b_{3}+4b_{4}}{6},-\frac{7b_{1}+5b_{3}+3b_{4}}{4}\right\}$$*provided that* $\min \left\{
7b_{1}+2b_{2}+3b_{4},b_{1}+b_{3}+b_{4}\right\} \geq 0$, *while otherwise the system* $\widetilde{\widetilde{\sigma }}\left( b\right) $* is infeasible. Since the right-hand side in* $\widetilde{\widetilde{\sigma
}}\left( b\right) $ *is* $\left( -9,-6,-6,45,-6,210\right) $, *for any* $b$ *close to $\overline{b}$ we have* $\widetilde{\widetilde{\sigma }}(b)\equiv 2x_{1}+x_{2}\geq 3$ *and* $$\widetilde{\widetilde{\sigma }}\left( b\right) \equiv 2x_{1}+x_{2}\geq \max
\left\{ -\frac{5b_{1}+b_{2}}{3},-\frac{2b_{1}+b_{3}}{2}\right\} .$$*This readily implies for such $b$ that* $$\mathcal{P}\left( b\right) =\max \left\{ -\frac{5b_{1}+b_{2}}{3},-\frac{2b_{1}+b_{3}}{2}\right\} .$$*Employing now Remark \[Rem\_neigh\] and the classical formula of convex analysis for subdifferentiation of maximum functions gives us* $$\partial \mathcal{P}\left( \overline{b}\right) =\mathrm{conv}\big\{\left(
-5/3,-1/3,0,0\right) ,\left( -1,0,-1/2,0\right) \big\}.$$*Then we deduce from Theorem \[Cor\_lip\_EP\] that* $$\limfunc{lip}\mathcal{E}_{\mathcal{P}}\left( \overline{b}\right) =\left\Vert
\left( -5/3,-1/3,0,0\right) \right\Vert _{1}=2,$$*which ensures in turn by using Theorem \[Th lipV\] that* $$\limfunc{lip}\mathcal{P}\left( \overline{b}\right) =2\text{\emph{\ and} }\mathrm{lip}\mathcal{E}_{\mathcal{F}}\left( \overline{b},\overline{x}\right)
=2/\sqrt{5}$$*at any optimal solution* $\overline{x}$ *of* $PL\left( \overline{b}\right) $.
*Paper [@GCPT19] provides an alternative way to compute the Lipschitz modulus* $\limfunc{lip}\mathcal{P}\left( \overline{b}\right) $ *under the additional assumption that at least one optimal solution* $\overline{x}$ *of* $PL\left( \overline{b}\right) $ *is known. As we see, the procedure described in Example \[Exa\_new\] does not require such an* a priori* information.*
Concluding Remarks {#conc}
==================
This paper demonstrates that employing appropriate tools of variational analysis and generalized differentiation of set-valued mappings allows us to efficiently deal with major sensitivity characteristics of perturbed linear multiobjective optimization problems. Namely, in this way we explicitly computed the subdifferentials of the feasible set and Pareto front mappings in such problems together with the exact moduli of their Lipschitzian stability.
In future research we plan to extend the variational approach and results obtained in this paper to *convex* problems of multiobjective optimization by reducing them to linear systems with *block perturbations*. Observe that a similar procedure has been explored for feasibility mappings in *semi-infinite* programming with both decision and parameter variables living in Banach spaces.
[99]{} T. Q. BAO and B. S. MORDUKHOVICH, *Variational principles for set-valued mappings with applications to multiobjective optimization*, Control and Cybernetics **36** (2007), 531–562.
T. Q. BAO and B. S. MORDUKHOVICH, *Relative Pareto minimizers for multiobjective problems: existence and optimality conditions*, Math. Program. **122** (2010), 301–347.
J. M. BORWEIN and Q. J. ZHU, *Techniques of Variational Analysis*, Springer, New York, 2005.
M. J. CÁNOVAS, A. L. DONTCHEV, M. A. LÓPEZ and J. PARRA, *Metric regularity of semi-infinite constraint systems*, Math. Program. **104** (2005), 329–346.
M. J. CÁNOVAS, F. J. GÓMEZ-SENENT and J. PARRA, *On the Lipschitz modulus of the argmin mapping in linear semi-infinite optimization*, Set-Valued Anal. **16** (2008), 511–538.
M. J. CÁNOVAS, D. KLATTE, M. A. LÓPEZ and J. PARRA, *Metric regularity in convex semi-infinite optimization under canonical perturbations*, SIAM J. Optim **18** (2007), 717–732.
M. J. CÁNOVAS, M. A. LÓPEZ, B. S. MORDUKHOVICH and J. PARRA, *Variational analysis in semi-infinite and infinite programming, I: Stability of linear inequality systems of feasible solutions*, SIAM J. Optim. **20** (2009), 1504–1526.
M. J. CÁNOVAS, M. A. LÓPEZ, B. S. MORDUKHOVICH and J. PARRA, Quantitative stability of linear infinite inequality systems under block perturbations with applications to convex systems, *TOP* **20** (2012), 310–327.
A. L. DONTCHEV and R. T. ROCKAFELLAR, *Implicit Functions and Solution Mappings: A View from Variational Analysis*, 2nd edition, Springer, New York, 2014.
M. J. GISBERT, M. J. CÁNOVAS, J. PARRA and F. J. TOLEDO, *Lipschitz modulus of the optimal value in linear programming*, J. Optim. Theory Appl. **182** (2019), 133–152.
M. A. GOBERNA and M. A. LÓPEZ, *Linear Semi-Infinite Optimization*, John Wiley & Sons, Chichester, UK, 1998.
N. Q. HUY, B. S. MORDUKHOVICH and J. C. YAO, *Coderivatives of frontier and solution maps in parametric multiobjective optimization*, Taiwanese J. Math. **12** (2008), 2083–2111.
A. D. IOFFE, *Variational Analysis of Regular Mappings*, Springer, Cham, Switzerland, 2017.
D. KLATTE and B. KUMMER, *Nonsmooth Equations in Optimization: Regularity, Calculus, Methods and Applications*, Kluwer Academic, Dordrecht, The Netherlands, 2002.
B. S. MORDUKHOVICH, *Complete characterizations of openness, metric regularity, and Lipschitzian properties of multifunctions*, Trans. Amer. Math. Soc. **340** (1993), 1–35.
B. S. MORDUKHOVICH,*Variational Analysis and Generalized Differentiation, I: Basic Theory, II: Applications*, Springer, Berlin, 2006.
B. S. MORDUKHOVICH,*Variational Analysis and Applications*, Springer, Cham, Switzerland, 2018.
S. M. ROBINSON, *Some continuity properties of polyhedral multifunctions*, Mathematical Programming at Oberwolfach (Proc. Conf., Math. Forschungsinstitut, Oberwolfach, 1979). Math. Programming Stud. No. **14** (1981), 206–214.
R. T. ROCKAFELLAR and R. J-B. WETS, *Variational Analysis*, Springer, Berlin, 1998.
[^1]: Center of Operations Research, Miguel Hernández University of Elche, 03202 Elche (Alicante), Spain ([email protected], [email protected]).
[^2]: Department of Mathematics, University of Alicante, 03080 Alicante, Spain ([email protected]); CIAO, Federation University, Ballarat, Australia. Research of this author is also partially supported by the Australian Research Council (ARC) Discovery Grants Scheme (Project Grant \# DP180100602).
[^3]: Department of Mathematics, Wayne State University, Detroit, MI 48202, USA ([email protected]). Research of this author was partially supported by the USA National Science Foundation under grants DMS-1512846 and DMS-1808978, by the USA Air Force Office of Scientific Research grant \#15RT04, and by Australian Research Council under grant DP-190100555.
[^4]: This research has been partially supported by grants MTM2014-59179-C2-(1,2)-P and PGC2018-097960-B-C2(1,2).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Using new configuration spaces, we give an explicit construction that extends Kontsevich’s Lie-infinity quasi-isomorphism from polyvector fields to Hochschild cochains to a quasi-isomorphism of A-infinity algebras equipped with actions by homotopy derivations of the Lie algebra of polyvector fields. One may term this formality a formality of two-colored noncommutative Gerstenhaber homotopy algebras. In our result the action of polyvector fields by homotopy derivations of the wedge product on polyvector fields is not the adjoint action by the Schouten bracket, but a homotopy nontrivial and, in a sense, unique deformation of that action.
As an application we give an explicit Duflo-type construction for Lie-infinity algebras that generalizes the Duflo-Kontsevich isomorphism between the Chevalley-Eilenberg cohomology of the symmetric algebra on a Lie algebra and the Chevalley-Eilenberg cohomology of the universal enveloping algebra of the Lie algebra.
author:
- Johan Alm
bibliography:
- 'qocha.bib'
title: |
Two-colored noncommmutative Gerstenhaber formality\
and infinity Duflo isomorphism
---
**[Introduction]{}** {#introduction .unnumbered}
====================
Kontsevich’s Formality Map is best understood as a morphism of two-colored operads $$\mathcal{K}(\overline{C}(\mathbf{H}))\rightarrow\mathcal{E}nd(T_{poly},\mathcal{O}),$$ where $\mathcal{K}(\overline{C}(\mathbf{H}))$ is the operad of fundamental chains of a certain cellular operad $\overline{C}(\mathbf{H})$ of compactified configuration spaces of points in the closed upper half-plane and $\mathcal{E}nd(T_{poly},\mathcal{O})$ is the standard two-colored endomorphism operad on formal polyvector fields, $T_{poly}$, and formal smooth functions, $\mathcal{O}$, on some chosen graded vector space. The content of this map of operads is an $L_{\infty}$ map from $T_{poly}$ to the (differential) Hochschild cochain complex of $\mathcal{O}$. In this note we introduce a three-colored operad $\overline{CF}(\mathbf{H})$ of compactified configuration spaces of points in a closed upper half-plane equipped with a line parallel to the boundary, and, using the same techniques as Kontsevich, a representation $$\mathcal{K}(\overline{CF}(\mathbf{H}))\rightarrow\mathcal{E}nd(T_{poly},T_{poly},\mathcal{O})$$ of its fundamental chains. This representation implies
- Kontsevich’s $L_{\infty}$ map $T_{poly}\rightarrow C(\mathcal{O},\mathcal{O})$ to the Hochschild cochain complex of the associative algebra of functions,
- an $L_{\infty}$ map $T_{poly}\rightarrow C^{\geq 1}(T_{poly},T_{poly})$ to the Hochschild cochain complex of the associative algebra of polyvector fields, extending the canonical adjoint action of $T_{poly}$ on itself,
- and a morphism $T_{poly}\rightarrow C(\mathcal{O},\mathcal{O})$ of $A_{\infty}$ algebras equipped with actions of the Lie algebra $T_{poly}$ by homotopy derivations.
This data can be concisely encoded as a quasi-isomorphism of two-colored noncommutative $G_{\infty}$ algebras.
The three-colored operad $\overline{CF}(\mathbf{H})$ is closely related to the moduli spaces of quilted holomorphic disks introduced in the context of Floer homology by Mau and Woodward in [@MW]. The moduli spaces of quilted holomorphic disks form a two-colored operad that can be embedded as a suboperad of our three-colored operad.
As an application we give an explicit strong homotopy version of the Duflo isomorphism. This generalizes earlier work by Calaque, Kontsevich, Manchon, Pevzner, Rossi, Torossian and others; see [@MT; @PT; @CR; @Sh; @Kon03]. More specifically, we construct a universal and generically homotopy nontrivial $A_{\infty}$ deformation $C(\mathbf{g},S(\mathbf{g}))_{exotic}$ of the Chevalley-Eilenberg cochain algebra $C(\mathbf{g},S(\mathbf{g}))$ and an $A_{\infty}$ quasi-isomorphism $C(\mathbf{g},S(\mathbf{g}))_{exotic}\rightarrow C(\mathbf{g},U(\mathbf{g}))$ that on the cohomology level reproduces the Duflo-Kontsevich isomorphism of Chevalley-Eilenberg cohomologies. This implies that the Duflo-Kontsevich isomorphism can *not* be universally lifted to an $A_{\infty}$ quasi-isomorphism $C(\mathbf{g},S(\mathbf{g}))\rightarrow C(\mathbf{g},U(\mathbf{g}))$ of the Chevalley-Eilenberg cochain algebras.
It is a pleasure to thank Johan Gran[å]{}ker, Carlo A. Rossi, Bruno Vallette and Thomas Willwacher for discussions and helpful criticism. Special thanks to my supervisor, Sergei Merkulov.
Conventions and notation {#conventions-and-notation .unnumbered}
------------------------
Given a natural number $n$, $[n]$ denotes the set $\{1,2,\dots,n\}$. The cardinality of a finite set $A$ is written $\vert A\vert$, e.g. $\vert [n]\vert=n$. Given finite sets $A$ and $B$, we write $A+B$ for their disjoint union. We customarily write $0$ for the empty set. If $S$ is a subset of a finite set $A$, we customarily write $A-S$ for the complement of $S$ in $A$. We write $A/S$ for the set $A-S+\{S\}$. (So the cardinality of $A/S$ is $\vert A\vert -\vert S\vert +1$.) If $A$ is an ordered finite set, then we say $S\subset A$ is a connected subset, and write $S<A$, if $s,s''\in S$ and $s<s'<s''\in A$ implies also $s'\in S$. The group of permutations of a finite set $T$ is denoted $\Sigma_T$, and $\Sigma_{[n]}$ is denoted $\Sigma_n$.
The terms “chain complex” and “dg vector space” are used as synonyms and refer to (possibly unbounded) cohomologically graded chain complexes over the real numbers. The $n$-fold suspension of a chain complex $V$ is the chain complex $V[n]$ with $V[n]^d=V^{d+n}$. We employ the Koszul symmetry conventions for the category of chain complexes. Given a dg vector space $V$, $T(V):=\prod_{n\geq 0}V^{\otimes n}$ and $S(V):=\prod_{n\geq 0}(V^{\otimes n})_{\Sigma_n}$. The vector space $S(V)$ has a structure of commutative associative algebra; we denote $S(V)$ with this algebra structure $S^a(V)$. It can also be regarded as a cocommutative coassociative coalgebra, which we denote $S^c(V)$.
**[Configuration space models for [$NCG_{\infty}$]{} algebras and flag OCHAs]{}**
=================================================================================
In this section we define four different operads in the category of cellular compact smooth manifolds with corners. Two of the operads are our invention.
A configuration space model for [$L_{\infty}$]{}.
-------------------------------------------------
For an integer $\ell\geq 2$, let $\mathrm{Conf}_{\ell}(\mathbf{C})$ be the manifold of all injective maps of $[\ell]:=\{1,\dots,\ell\}$ into $\mathbf{C}$. The group of translations and positive dilations of the plane, $\mathbf{C}\rtimes\mathbf{R}_{>0}$, acts on the plane and hence (by postcomposition) on $\mathrm{Conf}_{\ell}(\mathbf{C})$. Define $C_{\ell}(\mathbf{C}):=\mathrm{Conf}_{\ell}(\mathbf{C})/\mathbf{C}\rtimes\mathbf{R}_{>0}$. Let $\overline{\mathrm{Conf}}_{\ell}(\mathbf{C})$ be the real Fulton-MacPherson compactification (in the literature also called the Axelrod-Singer compactification) of $\mathrm{Conf}_{\ell}(\mathbf{C})$, i.e. the real oriented blow-up of $\mathbf{C}^{\ell}$ along all diagonals. The action by translations and positive dilations is smooth; hence extends uniquely to a smooth action on $\overline{\mathrm{Conf}}_{\ell}(\mathbf{C})$. Define $\overline{C}_{\ell}(\mathbf{C})$ to be the quotient of $\overline{\mathrm{Conf}}_{\ell}(\mathbf{C})$ by this action. It is a smooth compact manifold with corners with codimension one boundary $$\bigsqcup_S C_{\ell-\vert S\vert+1}(\mathbf{C})\times C_S(\mathbf{C})$$ given by products labelled by subsets $S\subset[\ell]$ (of cardinality $2\leq\vert S\vert<\ell$). Moreover, the closure of $C_{\ell-\vert S\vert+1}(\mathbf{C})\times C_S(\mathbf{C})$ in $\overline{C}_{\ell}(\mathbf{C})$ is the product $\overline{C}_{\ell-\vert S\vert+1}(\mathbf{C})\times \overline{C}_S(\mathbf{C})$. This means that the family of spaces $\overline{C}(\mathbf{C})=\{\overline{C}_{\ell}(\mathbf{C})\}$ together with the inclusions of boundary components and permutation actions by permutation of points assemble into the structure of an operad. We promote it to an operad of oriented manifolds as follows. Let $C^{std}_{\ell}(\mathbf{C})$ be the submanifold of $\mathrm{Conf}_{\ell}(\mathbf{C})$ consisting of configurations $x$ satisfying $\sum_{i=1}^{\ell}x_i=0$ and $\sum_{i=1}^{\ell}\vert x_i\vert^2=1$. The manifolds $C_{\ell}(\mathbf{C})$ and $C^{std}_{\ell}(\mathbf{C})$ are diffeomorphic. The manifold $\mathrm{Conf}_{\ell}(\mathbf{C})$ is canonically oriented; hence so is $C^{std}_{\ell}(\mathbf{C})$. We orient $C_{\ell}(\mathbf{C})$ by pulling back the orientation on $C^{std}_{\ell}(\mathbf{C})$. Requiring Stokes’ formula (without a sign) to hold defines an orientation of the compactification $\overline{C}_{\ell}(\mathbf{C})$. It is easy to see that all permutations of $[\ell]$ preserve the orienation.
The boundary description describes a canonical stratification and the face complexes of the stratification of each component form an operad $\mathcal{K}(\overline{C}(\mathbf{C}))$ that this is freely generated as a graded operad by the set $\{[C_{\ell}(\mathbf{C})]\mid \ell\geq 2\}$ of “fundamental chains”. It is well-known that representations of $\mathcal{K}(\overline{C}(\mathbf{C}))$ in a dg vector space $V$ are in one-to-one correspondence with $L_{\infty}$ structures on the suspension $V[1]$ of $V$; see e.g. [@GJ].
A configuration space model for OCHA
------------------------------------
Set $\mathbf{H}:=\mathbf{R}\times\mathbf{R}_{\geq 0}$. For integers $m,n>0$, with $2m+n\geq2$, let $\mathrm{Conf}_{m,n}(\mathbf{H})$ be the manifold of injections of $[m]+[n]$ into $\mathbf{H}$ that map $[n]$ into the boundary $\mathbf{R}\times\{0\}$ of the half-plane and $[m]$ into the interior. The group of translations along the boundary and positive dilations, $\mathbf{R}\times\mathbf{R}_{>0}$, acts (by postcomposition) on $\mathrm{Conf}_{m,n}(\mathbf{H})$ and we let $C_{m,n}(\mathbf{H})$ be the quotient of this action. The embedding $$\mathrm{Conf}_{m,n}(\mathbf{H})\rightarrow\mathrm{Conf}_{2m+n}(\mathbf{C})$$ defined by sending a configuration in $[m]+[n]\hookrightarrow\mathbf{H}$ to its orbit under complex conjugation induces an embedding $$C_{m,n}(\mathbf{H})\rightarrow C_{2m+n}(\mathbf{C})\subset\overline{C}_{2m+n}(\mathbf{C}).$$ The compactification $\overline{C}_{m,n}(\mathbf{H})$ of $C_{m,n}(\mathbf{H})$ was in[@Kon03] defined as the closure under this embedding. It is a smooth manifold with corners with $n!$ connected components. Let $\overline{C}^+_{m,n}(\mathbf{H})$ be the connected component that has the boundary points “compatibly ordered”, by which we mean that if $i<j\in [n]=\{1<\dots<n\}$, then the point labelled by $i$ is before the point labelled by $j$ on the boundary for the orientation of the boundary induced by the orientation of the half-plane. This gives us a permutation-equivariant identification $\overline{C}_{m,n}(\mathbf{H})\cong \overline{C}^+_{m,n}(\mathbf{H})\times\Sigma_n$. The codimension one boundary of $\overline{C}^+_{m,n}(\mathbf{H})$ is $$\bigsqcup_I \bigl(C^+_{m-\vert I\vert+1,n}(\mathbf{H})\times C_I(\mathbf{C})\bigr)
\sqcup\bigsqcup_{S,T}\bigl(C^+_{m-\vert S\vert,n-\vert T\vert+1}(\mathbf{H})\times C^+_{S,T}(\mathbf{H})\bigr).$$ Here $C^+_{m-\vert I\vert+1,n}(\mathbf{H})$ is the interior of $\overline{C}^+_{m-\vert I\vert+1,n}(\mathbf{H})$, etc. The union is over all subsets $I\subset[m]$ and subsets $S\subset [m]$, $T<[n]$ such that all involved spaces are defined. This description of the boundary extends, via the identification $\overline{C}_{m,n}(\mathbf{H})\cong \overline{C}^+_{m,n}(\mathbf{H})\times\Sigma_n$, to boundary descriptions for all connected components, and defines the structure of a two-coloured operad on the collection $\overline{C}(\mathbf{H}):=\{\overline{C}_{\ell}(\mathbf{C}),\overline{C}_{m,n}(\mathbf{H})\}$, the points in the interior being inputs of one color and the points on the boundary being inputs of another color. The spaces $\overline{C}_{m,n}(\mathbf{H})$ are defined using embeddings into spaces fo the form $\overline{C}_{\ell}(\mathbf{C})$, for which we have chosen orientations. We orient the spaces $\overline{C}_{m,n}(\mathbf{C})$ by the pullback orientations of these embeddings.
The dg operad of face complexes of the stratification defined by the boundary decomposition is again generated by the fundamental chains. We denote this operad of fundamental chains $\mathcal{K}(\overline{C}(\mathbf{H}))$. A representation of it is referred to as an **open-closed homotopy algebra**, see [@Ho; @KS06], henceforth abbreviated as an OCHA. An OCHA consists of a pair of dg vector spaces $V$ and $W$, an $L_{\infty}$ structure on $V[1]$, an $A_{\infty}$ structure on $W$, and an $L_{\infty}$ morphism from $V$ to the Hochschild cochain complex of $W$.
We now define flag versions of the operads $\overline{C}(\mathbf{C})$ and $\overline{C}(\mathbf{H})$.
Flag version of $\overline{C}(\mathbf{C})$.
-------------------------------------------
Since the affine group preserves collinearity and parallel lines it makes sense to say that some points in a configuration $x\in C_{\ell}(\mathbf{C})$ are collinear on a line parallel to the real axis. For integers $p\geq 0$ and $q\geq 1$ with $p+q\geq 2$, define $CF_{p,q}(\mathbf{C})\subset C_{[p]+[q]}(\mathbf{C})$ to be the subset of configurations for which the points labelled by $[q]$ are collinear on a line parallel to the real axis. Define $\overline{CF}_{p,q}(\mathbf{C})$ to be its closure inside $\overline{C}_{p+q}(\mathbf{C})$. It has $q!$ connected components. Let $CF^+_{p,q}(\mathbf{C})$ denote the interior of the connected component that has the collinear points compatibly ordered, by which we mean that if $i<j\in [q]=\{1<\dots<q\}$, then the point labelled by $i$ is before the point labelled by $j$ on their common line for the orientation of the line induced by the orientation of the plane. Then $CF_{p,q}(\mathbf{C})\cong CF^+_{p,q}(\mathbf{C})\times\Sigma_q$. We deduce that the codimension one boundary of the corresponding compact connected component, $\overline{CF}^+_{p,q}(\mathbf{C})$, is $$\bigsqcup_I \bigl(CF^+_{p-\vert I\vert+1,q}(\mathbf{C})\times C_I(\mathbf{C})\bigr)
\sqcup\bigsqcup_{S,T}\bigl(CF^+_{p-\vert S\vert,q-\vert T\vert+1}(\mathbf{C})\times CF^+_{S,T}(\mathbf{C})\bigr).$$ The union is over all subsets $I\subset [p]$, $S\subset[p]$, $T<[q]$ for which all involved spaces are defined. One can use the inclusions of boundary components to define a two-colored operad structure on the collection $$\overline{CF}(\mathbf{C}):=\{\overline{C}_{\ell}(\mathbf{C}),\overline{CF}_{p,q}(\mathbf{C})\},$$ in a way completely analogous the previously discussed operadic structure on $\overline{C}(\mathbf{H})$.
We call $\overline{CF}(\mathbf{C})$ the operad of configurations on flags in the plane.
We orient the spaces of the form $\overline{CF}_{p,q}(\mathbf{C})$ by the pullback orientations of the defining embeddings into $\overline{C}_{p+q}(\mathbf{C})$. As before one then obtains a dg operad $\mathcal{K}(\overline{CF}(\mathbf{C}))$ of fundamental chains. It is almost identical to the operad $\mathcal{K}(\overline{C}(\mathbf{H}))$ of OCHAs: its representations also consist of an $L_{\infty}$ algebra $V[1]$, an $A_{\infty}$ algebra $W$ and an $L_{\infty}$ morphism from $V$ to the Hochschild cochain complex of $W$. The difference lies in that the latter operad contains chains $[C_{m,n}(\mathbf{H})]$ with $n=0$ while the former operad does not contain any chain of the form $[CF_{p,q}(\mathbf{C})]$ with $q=0$. This means that the $L_{\infty}$ map of an OCHA contains components $V^{\otimes p}\rightarrow W$, so called curvature terms, whilst the $L_{\infty}$ map of a $\mathcal{K}(\overline{C}(\mathbf{H}))$-representation can not, i.e. it maps into the truncated Hochschild cochain complex $C^{\geq 1}(W,W)$.
We call $\mathcal{K}(\overline{CF}(\mathbf{C}))$ the operad of two-colored noncommutative $G_{\infty}$ algebras.
Define a **two-colored noncommutative Gerstenhaber algebra** to be a pair $(L,A)$, where $L[1]$ is a dg Lie algebra and $A$ is a dg assciative algebra, together with a dg Lie algebra morphism $L[1]\rightarrow\mathrm{Der}(A)$. Such algebras are representations of an operad $\mathcal{NCG}$ and $\mathcal{K}(\overline{CF}(\mathbf{C}))$ is the cobar construction on the Koszul dual cooperad of $\mathcal{NCG}$. We prove in an appendix that $\mathcal{NCG}$ is Koszul. Thus $\mathcal{K}(\overline{CF}(\mathbf{C}))$ indeed deserves to be called the operad of two-colored noncommutative $G_{\infty}$ algebras.
We shall abbreviate “two-colored noncommutative $G_{\infty}$ algebra” as $NCG_{\infty}$ algebra.
Flag version of $\overline{C}(\mathbf{H})$.
-------------------------------------------
There is also a flag version of the operad $\overline{C}(\mathbf{H})$, defined as follows. Let $k,m,n\geq 0$ be integers with $2k+m+n\geq 1$ if $m\geq 1$ and $k+n\geq 2$ if $m=0$. Let $CF_{k,m,n}(\mathbf{H})$ be the subspace of $C_{k+m,n}(\mathbf{H})$ consisting of all configurations wherein the points labelled by $[m]$ are collinear on a line parallel to the boundary. Denote by $\overline{CF}_{k,m,n}(\mathbf{H})$ the closure inside $\overline{C}_{k+m,n}(\mathbf{H})$. Let $CF^+_{k,m,n}(\mathbf{H})$ denote connected component of $CF_{k,m,n}(\mathbf{H})$ that has both the collinear points and the boundary points compatibly ordered, i.e. if $i<j$ in $[m]$, then $x_i<x_j$ on their common line of collinearity, and if $r<s$ in $[n]$, then $x_r<x_s$ on the boundary. The codimension one boundary of its compactification, $\overline{CF}^+_{k,m,n}(\mathbf{H})$, has the form $$\bigsqcup_I \bigl(CF^+_{k-\vert I\vert+1,m,n}(\mathbf{H})\times C_I(\mathbf{C})\bigr)
\sqcup\bigsqcup_{P,Q}\bigl(CF^+_{k-\vert P\vert,m-\vert Q\vert+1,n}(\mathbf{H})\times CF^+_{P,Q}(\mathbf{C})\bigr)$$ $$\sqcup\bigsqcup_{S,T,U}\bigl(CF^+_{k-\vert S\vert,m-\vert T\vert,n-\vert U\vert+1}(\mathbf{H})\times CF^+_{S,T,U}(\mathbf{H})\bigr).$$ The union is over all subsets $I,P,S\subset[k]$, $Q,T<[m]$, $S<[n]$ for which all involved spaces are defined. These boundary factorizations define an operad structure, but now in three colors, on the collection $$\overline{CF}(\mathbf{H}):=\{\overline{C}_{\ell}(\mathbf{C}),\overline{CF}_{p,q}(\mathbf{C}),\overline{CF}_{k,m,n}(\mathbf{H})\}.$$
We call $\overline{CF}(\mathbf{H})$ the operad of configurations on flags in the half-plane.
Orient the spaces $\overline{CF}_{k,m,n}(\mathbf{H})$ by the pullback orientations of the embeddings into $\overline{C}_{k+m,n}(\mathbf{H})$. There is an associated operad $\mathcal{K}(\overline{CF}(\mathbf{H}))$ of fundamental chains.
We call $\mathcal{K}(\overline{CF}(\mathbf{H}))$ the operad of flag open-closed homotopy algebras, abbreviated as the operad of flag OCHAs.
A representation of the operad of flag open closed homotopy algebras in a triple $(L,A,B)$ of chain complexes is equivalent to
- an $NCG_{\infty}$ algebra structure on $(L,A)$;
- an OCHA structure on $(L,B)$;
- and a morphism from $A$ to $C(B,B)$ of $A_{\infty}$ algebras with $L_{\infty}$ actions of $L$ by homotopy derivations, where the Hochschild cochain complex of $B$ is considered with the $L$-action induced by the OCHA structure.
The first two listed items are obvious. Let $\mathcal{M}or_*(\mathcal{NCG})_{\infty}$ be the Koszul resolution of the operad, $\mathcal{M}or_*(\mathcal{NCG})$, whose representations are NCGAs $(L,A)$, $(L,A')$, with the same dg Lie algebra $L$ appearing in both pairs, and a morphism between the two dg associative algebras respecting the actions by $L$. See the appendix for some comments on why $\mathcal{M}or_*(\mathcal{NCG})$ is Koszul. The third item in the list is a $\mathcal{M}or_*(\mathcal{NCG})_{\infty}$-representation on $(L,A,C(B,B))$. The key to this correspondence is to change from the operadic perspective that the chains $[CF_{k,m,n}(\mathbf{H})]$ are represented as maps $L^{\otimes k}\otimes A^{\otimes m}\otimes B^{\otimes n}\rightarrow B$ to the perspective that they define maps $$L^{\otimes k}\otimes A^{\otimes m}\rightarrow\mathrm{Map}(B^{\otimes n},B).$$ (This hom-adjunction argument exactly parallels the argument used for interpreting an OCHA structure $\{[C_{p,q}(\mathbf{H})]:L^{\otimes p}\otimes B^{\otimes q}\rightarrow B\}$ as an $L_{\infty}$ morphism $L\rightarrow C(B,B)$, compare with [@KS06; @Ho].) After this reinterpretation of the chains the argument reduces to (i) recognizing the induced $NCG_{\infty}$ algebra structure on $(L,C(B,B))$ and (ii) comparing the differential on the chains to the differential on $\mathcal{M}or_*(\mathcal{NCG})_{\infty}$. The details are left to the reader. We work out some more explicit details in the subsequent sections.
Consider the two-colored suboperad of $\overline{CF}(\mathbf{H})$ on the components $$\{\overline{CF}_{0,q}(\mathbf{C}),\overline{CF}_{0,m,0}(\mathbf{H}),\overline{CF}_{0,0,n}(\mathbf{H})\}.$$ It is isomorphic as an operad of smooth manifolds with corners to the operad of quilted holomorphic disks introduced by Mau and Woodward in [@MW]. Its operad of cellular chains is the operad of morphisms of $A_{\infty}$ algebras.
**[(Co)operads of graphs]{}**
=============================
Kontsevich’s proof of his Formality Conjecture and construction of a universal deformation quantization formula can be regarded[@Mer10] as the construction of
- a map of cooperads $\omega:\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{H})}}\rightarrow\Omega(\overline{C}(\mathbf{H}))$, where $\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{H})}}$ is a cooperad of Feynman diagrams,
- and a map of operads $\Phi:\mathfrak{G}_{{\scriptscriptstyle \overline{C}(\mathbf{H})}}\rightarrow\mathcal{E}nd(T_{poly},\mathcal{O})$ from the dual operad of Feynman diagrams.
Dualizing the map of cooperads and composing, one gets a representation $$\Phi\circ\omega^*:\mathcal{K}(\overline{C}(\mathbf{H}))\rightarrow \mathfrak{G}_{{\scriptscriptstyle \overline{C}(\mathbf{H})}}
\rightarrow\mathcal{E}nd(T_{poly},\mathcal{O})$$ of the fundamental chains of half-plane configurations, i.e. an OCHA structure on $(T_{poly},\mathcal{O})$. We shall show that Kontsevich’s construction can be extended, essentially without any changes, to a representation $$\Phi\circ\omega^*:\mathcal{K}(\overline{CF}(\mathbf{H}))\rightarrow \mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}
\rightarrow\mathcal{E}nd(T_{poly},T_{poly},\mathcal{O})$$ of the operad of flag OCHAs. This is our $NCG_{\infty}$ Formality Theorem. The new data added by extending Kontsevich’s OCHA to a flag OCHA is a quasi-isomorphism $T_{poly}\rightarrow C(\mathcal{O},\mathcal{O})$ of $A_{\infty}$ algebras with homotopy actions by $T_{poly}$.
The first construction we need for our extension of the Kontsevich representation is a suitable operad $\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}$.
Directed graphs.
----------------
Choose a finite set $S$. Let $fdgra^d_S$ be the set of all injectve functions $\Gamma$ of the set $[d]$ into $(S\times S)-\Delta$, for $\Delta$ the diagonal of $S$. We refer to such a $\Gamma$ as a directed graph with $d$ edges on the set $S$ and introduce the following terminology:
- $E_{\Gamma}:={\textrm{im}}(\Gamma)$ is the set of edges of $\Gamma$. We consider it as ordered by the given isomorphism with $[d]$. The element $\Gamma(i)\in E_{\Gamma}$ is written $e_i$ and referred to as the $i$th edge.
- The function $s_{\Gamma}:E_{\Gamma}\subset S\times S\rightarrow S$ given by projection onto the first factor $S$ is called the source map of $\Gamma$. The projection $t_{\Gamma}:E_{\Gamma}\rightarrow S$ onto the second factor is called the target map of $\Gamma$. An edge $e$ is said to be directed from $s_{\Gamma}(e)$ to $t_{\Gamma}(e)$.
- The set $S$ is called the set of vertices of $\Gamma$.
- The valence of a vertex is the number of edges having that vertex as either source or target.
- A connected component of $\Gamma$ is a maximal (with respect to inclusions) subset $E\subset E_{\Gamma}$ with the property that $s_{\Gamma}(E)\cup t_{\Gamma}(E)$ and $s_{\Gamma}(E_{\Gamma}-E)\cup t_{\Gamma}(E_{\Gamma}-E)$ are disjoint. A graph with a single connected component is said to be connected.
Let $dgra^d_S$ be the subset of $fdgra^d_S$ of connected graphs. There is a natural action of the permutation groups $\Sigma_d$ and $\Sigma_S$ on $dgra^d_S$ by, respectively, reordering edges and permuting the vertices. Let $sgn_d$ be the one-dimensional sign representation of $\Sigma_d$. Define, for any finite set $I$, of cardinality at least $2$, the graded $\Sigma_I$-module $$\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(I):=\bigoplus_{j\geq 0} (\mathbf{R}\langle dgra^d_I \rangle \otimes_{\Sigma_d} sgn_d)[-d].$$
Elements of $\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(I)^d$ may be represented as (linear combinations of) connected graphs with $d$ directed edges ordered up to an even permutation, $\vert I\vert$ vertices labelled by $I$, without double edges and without tadpoles (edges that begin and end at the same vertex).
For a finite set $P$ and a nonempty finite set $Q$, with $\vert P\vert+\vert Q\vert\geq 2$, let $dgra^d_{P,Q}$ be a copy of the subset of $fdgra^d_{P+Q}$ consisting of those graphs which have no connected components $E\subset E_{\Gamma}$ with $s_{\Gamma}(E)\cup t_{\Gamma}(E)\subset P$, and put $$\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(P,Q):=\bigoplus_{d\geq 0}
(\mathbf{R}\langle dgra^d_{P,Q} \rangle \otimes_{\Sigma_d} sgn_d)[-d].$$ The vertices labelled by $P$ of a graph in $\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(P,Q)$ are called **free vertices** and the vertices labelled by $Q$ are called **collinear vertices**. Our restrictions informally say that there are no connected components with only free vertices.
Assume given a triple of finite sets $(K,M,N)$, with $2\vert K\vert+\vert M\vert+\vert N\vert\geq 1$ if $M$ is nonempty, and $2\vert K\vert+\vert N\vert\geq 2$ if $M$ is empty. Let $dgra^d_{K,M,N}$ be a copy of the subset of $dgra^d_{K,M+N}$ consisting of graphs $\Gamma$ having no edge with source a vertex labelled by $N$. Set $$\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}(K,M,N):=\bigoplus_{d\geq 0}
(\mathbf{R}\langle dgra^d_{K,M,N} \rangle \otimes_{\Sigma_d} sgn_d)[-d].$$ The vertices labelled by $K$ of a graph in $\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}(K,M,N)$ are called **free vertices**, the vertices labelled by $M$ are called **collinear vertices** and the vertices labelled by $N$ are called **boundary vertices**.
(Co)operad structures
---------------------
We shall now describe how the vector spaces of (equivalence classes of) graphs defined above assemble into cooperads.
Given $\Gamma_2\in dgra^{d_2}_{S_2}$ and $\Gamma\in dgra^{d}_{S}$, where $d_2\leq d$ and $S_2\subset S$, we define an **embedding of $\Gamma_2$ as a full subgraph of $\Gamma$** to be an order-preserving inclusion $f:[d_2]\hookrightarrow [d]$ which makes $$[d_2]\hookrightarrow[d]\stackrel{\Gamma}{\rightarrow}S\times S \;\;\mathrm{equal}\;\;
[d_2]\stackrel{\Gamma_2}{\rightarrow}S_2\times S_2\subset S\times S.$$ An embedding of $\Gamma_2$ as a full subgraph of $\Gamma$ is written $f:\Gamma_2\hookrightarrow \Gamma$. Given an embedding $f$ as above, we define $\Gamma/\Gamma_2\in dgra^{d-d_2}_{S/S_2}$ to be the graph which, as a function, is the composition $$[d-d_2]\cong[d]-{\textrm{im}}(f)\stackrel{\Gamma}{\rightarrow}S\times S \rightarrow (S/S_2)\times (S/S_2).$$ Here the leftmost bijection is the unique order-preserving bijection and the rightmost arrow is given by the canonical projection of $S$ onto $S/S_2=S-S_2+\{S_2\}$ (sending elements of $S_2$ to the element $\{S_2\}$). If $\Gamma_1=\Gamma/\Gamma_2$, $\Gamma_1\in dgra^{d_1}_{S_1+\{v\}}$ (so $S_1=S-S_2$ and we identify the singleton sets $\{v\}$ and $\{S_2\}$), then the embedding and the quotient define a bijection $[d_1]+[d_2]\rightarrow [d]$. This defines an order on $[d_1]+[d_2]$, using the order on $[d]$. This order on $[d_1]+[d_2]$ is related to the lexicographic order given by $[d_1]<[d_2]$ using a unique bijection. Define $\epsilon(\Gamma_2,\Gamma,\Gamma_1)$ to be the sign of that bijection. We may now define a cooperadic cocomposition $$\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(I_1+I_2)\rightarrow \mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(I_1+\{v\})\otimes \mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(I_2)$$ by $$\Gamma\mapsto\sum_{\Gamma_1=\Gamma/\Gamma_2}\epsilon(\Gamma_2,\Gamma,\Gamma_1)\Gamma_1\otimes\Gamma_2.$$ The sum is over all embeddings of some $\Gamma_2$ into $\Gamma$.
The collection $$\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}:=\{\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(\ell)\}$$ carries a cooperad structure. The componentwise linear dual, $\mathfrak{G}_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}:=\{\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(\ell)^*\}$, is an operad.
We define a full subgraph embedding of a graph $\Gamma_2\in dgra^{d_2}_{P_2,Q_2}$ into a graph $\Gamma\in dgra^d_{P,Q}$ exactly as before, except that we now require $P_2\subset P$ and $Q_2\subset Q$ (not just $P_2+Q_2\subset P+Q$). The quotient $\Gamma/\Gamma_2$ is defined as before and regarded as an element of $dgra^{d-d_2}_{P-P_2,Q/Q_2}$. The sign $\epsilon(\Gamma_2,\Gamma,\Gamma_1)$ is also defined as before. With these conventions for subgraphs and quotients we can repeat above definition and get cocomposition maps $$\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(P_1+P_2,Q_1+Q_2)\rightarrow
\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(P_1,Q_1+\{v\})\otimes
\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(P_2,Q_2).$$ The definitions repeat word for word when $\Gamma_2\in dgra^{d_2}_I$, $\Gamma\in dgra^d_{P,Q}$ and $I\subset P$, if we agree on the convention that now $\Gamma/\Gamma_2$ belongs to $dgra^{d-d_2}_{P/I,Q}$, defining cocompositions $$\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(P+I,Q)\rightarrow\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(P+\{v\},Q)\otimes
\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(I).$$
The collection $\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}:=\{\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(\ell),\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(p,q)\}$ carries a cooperad structure. The componentwise linear dual, $\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}:=\{\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(\ell)^*,\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(p,q)^*\}$, is an operad.
With the evident conventions for how to color the new vertex obtained by collapsing an embedded subgraph the same formulas define a cooperad structure on the collection $$\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}:=\{\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(\ell),\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(p,q), \mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}(k,m,n)\}.$$ Its linear dual, denoted $\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}$, is an operad.
de Rham field theory
--------------------
Given a pair of distinct indices $i,j\in[k]+[m]+[n]$ we follow Kontsevich and define a function $$\phi^h_{i,j}:CF_{k,m,n}(\mathbf{H})\rightarrow\mathbf{S}^1,\,
x+\mathbf{R}\rtimes\mathbf{R}_{>0} \mapsto \mathrm{Arg}\biggl(\frac{x_j-x_i}{x_j-\overline{x}_i}\biggr).$$ Here a barred variable denotes the complex conjugate variable. The function is smooth and extends to a smooth function defined on the compactified configuration space. Let $\vartheta$ be the homogeneous normalized volume form on $\mathbf{S}^1$.
Given a graph $\Gamma\in dgra^d_{k,m,n}$, define $$\omega_{\Gamma} := \wedge_{i=1}^d (\phi^h_{s_{\Gamma}(e_i),t_{\Gamma}(e_i)})^*\vartheta.$$ The form $\omega_{\Gamma}$ is a smooth closed differential form of degree $d$ on $\overline{CF}_{k,m,n}(\mathbf{H})$. We extend $\omega$ to a map of dg vector spaces $\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}(k,m,n)\rightarrow\Omega(\overline{CF}_{k,m,n}(\mathbf{H}))$.
Define similarly, for indices $i,j\in[\ell]$, $\phi_{i,j}:C_{\ell}(\mathbf{C})\rightarrow\mathbf{S}^1$ by $$\phi_{i,j}:x+\mathbf{C}\rtimes\mathbf{R}_{>0} \mapsto \mathrm{Arg}(x_j-x_i).$$ The function $\phi$ extends to the compactification. For a graph $\Gamma\in dgra^d_{\ell}$, let $$\omega_{\Gamma} := \wedge_{i=1}^d (\phi_{s_{\Gamma}(e_i),t_{\Gamma}(e_i)})^*\vartheta.$$ This allows us to define maps of dg vector spaces $\omega:\mathfrak{G}^c_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(\ell)\rightarrow\Omega(\overline{C}_{\ell}(\mathbf{C}))$. By identifying $\overline{CF}_{p,q}(\mathbf{C})$ with a subset of $\overline{C}_{p+q}(\mathbf{C})$ and $dgra^d_{p,q}$ with a subset of $dgra^d_{p+q}$ we can use this to define maps of dg vector spaces $\omega:\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(p,q)\rightarrow\Omega(\overline{CF}_{p,q}(\mathbf{C}))$ as well.
In all cases we interpret the form associated to a graph without edges as the function identically equal to $1$.
The de Rham complex functor $\Omega$ is only comonoidal up to quasi-isomorphism with respect to the usual tensor product of dg vector spaces. Hence $\Omega(\overline{CF}(\mathbf{H}))$ is only a cooperad up to quasi-isomorphisms. This inconvenience can be ignored by working with a completed tensor product, regarding it, say, as a cooperad in the category of chain complexes of nuclear Fréchet spaces. Our mapping $\omega:\mathfrak{G}^c_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}\rightarrow\Omega(\overline{CF}(\mathbf{H}))$ is a morphism of cooperads in this category of cooperads.
We shall not prove this statement as it is a consequence of similar statements in [@Mer10].
A representation of the operad of graphs
----------------------------------------
Fix for the remainder of this section a graded vector space $V$, assumed finite-dimensional in each degree.
Define the **formal smooth functions on** $V$, denoted $\mathcal{O}$, to be the completed symmetric algebra on $V^*$. Define the **formal polyvector fields on** $V$, to be denoted $T_{poly}$, as the completed symmetric algebra on $V^*\oplus V[-1]$. Note that $\mathcal{O}$ is a subalgebra of $T_{poly}$.
Let $\tau$ be the image of $id_V$ under $V\times V^*\rightarrow V\otimes V^*[1]\cong (V^*\otimes V[-1])^*$ and regard it as a map $V^*\otimes V[-1]\rightarrow\mathbf{R}$. It extends uniquely to a derivation of $T_{poly}$. This derivation defines an endomorphism (of degree $-1$) of $T_{poly}\otimes T_{poly}$ which we again denote $\tau$. The **Schouten bracket** on $T_{poly}$ is the map $$[\,,\,]_S:=m\circ\tau\circ(id+(21)),$$ where $m$ denotes the product on $T_{poly}$. It is well-known that the Schouten bracket is a (degree $-1$) Lie bracket. Given a finite set $S$ and distinct elements $s,t\in S$, define $\tau_{s,t}$ to be the endomorphism of $T_{poly}^{\otimes S}$ acting as $\tau$ on the “$s$th times $t$th factors” and as the identity on all others.
For a graph $\Gamma\in dgra^d_{k,m,n}$, let $$\Phi_{\Gamma}:= \varepsilon\circ m\circ\bigcirc_{i=1}^d\tau_{s_{\gamma}(e_i),t_{\Gamma}(e_i)}:
T_{poly}^{\otimes k}\otimes T_{poly}^{\otimes m}\otimes \mathcal{O}^{\otimes n}\rightarrow \mathcal{O}.$$ Here $\varepsilon$ is the projection of $T_{poly}$ onto $\mathcal{O}$ defined by the projection $V^*\oplus V[-1]\rightarrow V^*$, we regard $$T_{poly}^{\otimes k}\otimes T_{poly}^{\otimes m}\otimes \mathcal{O}^{\otimes n}\subset T_{poly}^{\otimes k+m+n},$$ and $m:T_{poly}^{k+m+n}\rightarrow T_{poly}$ is the product. For a graph $\Gamma\in dgra^d_{\ell}$ we define $$\Phi_{\Gamma}:= m\circ\bigcirc_{i=1}^d\tau_{s_{\gamma}(e_i),t_{\Gamma}(e_i)}:T_{poly}^{\otimes\ell}\rightarrow T_{poly}.$$ For a graph $\Gamma\in dgra^d_{p,q}$ we use the same formula, $$\Phi_{\Gamma}:= m\circ\bigcirc_{i=1}^d\tau_{s_{\gamma}(e_i),t_{\Gamma}(e_i)}:
T_{poly}^{\otimes p}\otimes T_{poly}^{\otimes q}\rightarrow T_{poly}.$$
One verifies that these definitions define a morphism of dg operads $$\Phi:\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}\rightarrow\mathcal{E}nd(T_{poly},T_{poly},\mathcal{O}).$$
**[$NCG_{\infty}$ formality]{}**
================================
Combining the previous subsections, we have a representation $$\Phi\circ\omega^*:\mathcal{K}(\overline{CF}(\mathbf{H}))\rightarrow \mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{H})}}
\rightarrow\mathcal{E}nd(T_{poly},T_{poly},\mathcal{O}).$$ Since $\mathcal{K}(\overline{CF}(\mathbf{H}))$ is quasi-free the representation consists of a family of maps, one for each generator of $\mathcal{K}(\overline{CF}(\mathbf{H}))$, satisfying some quadratic identities coming from the boundary differential on $\mathcal{K}(\overline{CF}(\mathbf{H}))$. We shall denote the components as follows:
- $\lambda_{\ell}:=\Phi\circ\omega^*([C_{\ell}(\mathbf{C})])\in\mathrm{Map}^{3-2\ell}(T_{poly}^{\otimes\ell},T_{poly})$, for $\ell\geq 2$.
- $\nu_p:=\Phi\circ\omega^*([CF^+_{0,q}(\mathbf{C})])\in \mathrm{Map}^{2-q}(T_{poly}^{\otimes q},T_{poly})$ for $q\geq 2$.
- $\mu_n:=\Phi\circ\omega^*([CF^+_{0,0,n}(\mathbf{H})])\in\mathrm{Map}^{2-n}(\mathcal{O}^{\otimes n},\mathcal{O})$ for $n\geq 2$.
- $\mathcal{V}_{p,q}:=\Phi\circ\omega^*([CF^+_{p,q}(\mathbf{C})])\in\mathrm{Map}^{2-2p-q}(T_{poly}^{\otimes p}\otimes T_{poly}^{\otimes q},T_{poly})$ for $p,q\geq 1$.
- $\mathcal{U}_{k,n}:=\Phi\circ\omega^*([CF^+_{k,0,n}(\mathbf{H})])\in\mathrm{Map}^{2-2k-n}(T_{poly}^{\otimes k}\otimes\mathcal{O}^{\otimes n},\mathcal{O})$ for $k\geq 1$, $n\geq 0$.
- $\mathcal{Z}_{k,m,n}:=\Phi\circ\omega^*([CF^+_{k,m,n}(\mathbf{H})])\in\mathrm{Map}^{1-2k-m-n}(T_{poly}^{\otimes k}\otimes T_{poly}^{\otimes m}\otimes\mathcal{O}^{\otimes n},\mathcal{O})$ for $k\geq 0$, $m\geq 1$, $n\geq 0$.
Recall that the Hochschild cochain complex $C(A,A)$ of an $A_{\infty}$ algebra $A$ is the dg vector space $$\prod_{r\geq 0}\mathrm{Map}(A[1]^{\otimes r},A).$$ (This is the completed Hochschild cochain complex. As our results are of a formal nature we shall always work with completed complexes.) The brace operations on the Hochschild cochains complex are maps $$(\,)\{\dots\}_p : C(A,A)\otimes \bigotimes_{i=1}^p C(A,A)\rightarrow C(A,A),\; p\geq 1,$$ defined for $x\in\mathrm{Map}(A[1]^{\otimes r},A)$, $x_i\in\mathrm{Map}(A[1]^{\otimes r_i},A)$, $1\leq i\leq p\leq r$, $n=r+r_1+\dots+r_p-p$, by $$x\{x_1,\dots,x_p\}_p(a_1,\dots,a_n)
= \sum_{1\leq i_1<\dots<i_p<r} \pm x(a_1,\dots,a_{i_1},x_1(a_{i_1},\dots),\dots,a_{i_p},x_p(a_{i_p},\dots),\dots,a_n).$$ The Gerstenhaber bracket on the Hochschild cochain complex is the operation $$[x,y]_G:=x\{y\}_1\pm y\{x\}_1.$$ It is a graded Lie bracket of degree $-1$ in our grading on the Hochschild cochain complex. Denote by $C^{\geq 1}(A,A)$ the subspace $\prod_{r\geq 1}\mathrm{Map}(A[1]^{\otimes r},A)$. It is a graded Lie subalgebra. Set $(\,)\{\dots\}:=\sum_{p\geq 1}(\,)\{\dots\}_p$ and define $$br:C(A,A)\rightarrow C^{\geq 1}(C(A,A),C(A,A)),x\mapsto ()\{x\}_1+x\{\dots\}.$$ One verifies that this is a map of graded Lie algebras.
An $A_{\infty}$ structure on $A$ is a Maurer-Cartan element $m=d+m_2+\dots$ in $C^{\geq 1}(A,A)$. The differential $[m,\,]_G$ makes the Hochschild cochain complex a dg Lie algebra. It is also an $A_{\infty}$ algebra with $A_{\infty}$ structure the Maurer-Cartan element $\cup^m:=br(m)$ of $C^{\geq 1}(C(A,A),C(A,A))$. When $A$ has a given $A_{\infty}$ structure $m$ we shall usually write $C(m)$ for $C(A,A)$ with differential $[m,\,]_G$.
The interpretation of the components of our representation of $\mathcal{K}(\overline{CF}(\mathbf{H}))$ is that
- $\lambda=\{\lambda_{\ell}\}$ is an $L_{\infty}$ structure on $T_{poly}$.
- $\nu=\{\nu_p\}$ is an $A_{\infty}$ structure on $T_{poly}$.
- $\mu=\{\mu_n\}$ is an $A_{\infty}$ structure on $\mathcal{O}$.
- $\mathcal{V}=\{\mathcal{V}_{p,q}\}$ is an $L_{\infty}$ map $(T_{poly},\lambda)\rightarrow C^{\geq 1}(\nu)$.
- $\mathcal{U}=\{\mathcal{U}_{k,n}\}$ is an $L_{\infty}$ map $(T_{poly},\lambda)\rightarrow C(\mu)$.
- $\mathcal{Z}=\{\mathcal{Z}_{k,m,n}\}$ is a morphism of $A_{\infty}$ algebras $$(T_{poly},\nu,\mathcal{V})\rightarrow(C(\mu),\cup^{\mu},br\circ\mathcal{U})$$ equipped with homotopy actions by $(T_{poly},\lambda)$.
This description is a result of the interpretation of the operad of flag open-closed homotopy algebras. All the component maps have an explicit description as sums over graphs, e.g. $$\mathcal{V}_{p,q} = \sum_{[\Gamma]\in [dgra^{2p+q-2}_{p,q}]} \int_{\overline{CF}^+_{p,q}(\mathbf{C})}\omega_{\Gamma}\Phi_{\Gamma},$$ with $[dgra^{2p+q-2}_{p,q}]$ the set of equivalence classes of graphs under the $\Sigma_{2p+q-2}$-action by permutation of edges. We shall use this description to give a more detailed description of the component maps. The main tool is “Kontsevich’s vanishing lemma”:
[@Kon03] Let $X$ be a complex algebraic variety of dimension $N\geq 1$, and $Z_1, . . . ,Z_{2N}$ be rational functions on $X$, not equal identically to zero. Let $U$ be any Zariski open subset of $X$ such that functions $Z_{\alpha}$ are defined and non-vanishing on $U$, and $U$ consists of smooth points. Then the integral $$\int_{U(\mathbf{C})}\wedge_{\alpha=1}^{2N}d(\mathrm{Arg}(Z_{\alpha}))$$ is absolutely convergent, and equal to zero.
The [$L_{\infty}$]{} structure [$\lambda$]{}
--------------------------------------------
We have $$\lambda_{\ell} = \sum_{[\Gamma]\in[dgra^{2\ell-3}_{\ell}]} \int_{\overline{C}_{\ell}(\mathbf{C})}\omega_{\Gamma}\Phi_{\Gamma}.$$ For $\ell\geq 3$, $C_{\ell}(\mathbf{C})\cong \mathbf{S}^1\times U$, with $U=(\mathbf{C}\setminus\{0,1\})^{\ell-2}\setminus diagonals$. This identification can be obtained by using the translation freedom to fix the point labelled by $1$, say, at the origin of $\mathbf{C}$ and using the dilation freedom to put the point labelled by $2$, say, on the unit circle $\mathbf{S}^1$. Multiplying the remaining points by the inverse of the phase of the point labelled by $2$ gives a point in $U$. Using this description we can reduce every integral $$\int_{\overline{C}_{\ell}(\mathbf{C})}\omega_{\Gamma}$$ to an integral over a circle times an integral of the type appearing in Kontsevich’s vanishing lemma. Hence all weights vanish for $\ell\geq 3$. The configuration space $\overline{C}_2(\mathbf{C})$ is a circle. The set of graphs $dgra^1_2$ contains two elements; the graph with an edge from $1$ to $2$ and the graph with an edge from $2$ to $1$. Both graphs have weight $1$. It follows that $\lambda_2$ is the Schouten bracket. As all higher homotopies $\lambda_{\geq 3}$ vanish, this means $\lambda$ is the usual graded Schouten Lie algebra structure on $T_{poly}$.
The [$A_{\infty}$]{} structure [$\nu$]{}
----------------------------------------
The $A_{\infty}$ structure $\nu$ has components $$\nu_p = \sum_{[\Gamma]\in [dgra^{p-2}_{0,p}]} \int_{\overline{CF}^+_{0,p}(\mathbf{C})}\omega_{\Gamma}\Phi_{\Gamma}.$$ The angle between collinear points is constant, so the differential form associated to a graph containing an edge connecting collinear vertices will be zero; hence no such graphs can contribute. It follows that the only graph which contributes is the graph with two vertices and no edge. The associated differential form is identically equal to one and we evaluate it on the one-point space $\overline{CF}_{0,2}(\mathbf{C})$. It follows that $\nu=\nu_2$ is the usual (wedge) product on $T_{poly}$.
The [$A_{\infty}$]{} structure [$\mu$]{}
----------------------------------------
Arguing as in the preceeding paragraph one deduces that $\mu=\mu_2$ is the usual product on $\mathcal{O}$.
The [$L_{\infty}$]{} map [$\mathcal{V}$]{}
------------------------------------------
Since $$\mathcal{V}_{p,q} = \sum_{[\Gamma]\in [dgra^{2p+q-2}_{p,q}]} \int_{\overline{CF}^+_{p,q}(\mathbf{C})}\omega_{\Gamma}\Phi_{\Gamma}$$ and $\overline{CF}^+_{p,1}(\mathbf{C})\cong\overline{C}_{p+1}(\mathbf{C})$, the argument regarding the $L_{\infty}$ structure $\lambda$ can be repeated to conclude that $\mathcal{V}_{p,1}=0$ for $p\geq 2$, while $$\mathcal{V}_{1,1}:T_{poly}\otimes T_{poly}\rightarrow T_{poly}, X\otimes\xi\mapsto [X,\xi]_S.$$ In other words, $\mathcal{V}_{1,1}$ is the adjoint action $T_{poly}\rightarrow\mathrm{Der}(T_{poly})$ of $T_{poly}$ on itself by derivations of the wedge product.
Using the translation freedom to put the collinear point labelled by $1$ at the origin and the collinear point labelled by $2$ at $1$ identifies $CF^+_{p,2}(\mathbf{C})$ with $(\mathbf{C}\setminus\{0,1\})^p\setminus diagonals$, so that one may again use Kontsevish’s vanishing lemma and conclude that $\mathcal{V}_{p,2}=0$ for all $p\geq 1$.
Reflection of the plane in the line of collinearity induces a diffeomorphism $f$ of $\overline{CF}^+_{p,q}(\mathbf{C})$. (Choosing representative configurations with the collinear points on the real axis identifies $f$ with complex conjugation.) The map $f$ preserves orientation if $(-1)^p$ is even and reverses it if it is odd. For $\Gamma\in dgra^{2p+q-2}_{p,q}$, $f^*\omega_{\Gamma}=(-1)^{2p+q-2}\omega_{\Gamma}=(-1)^q\omega_{\Gamma}$. Thus $$(-1)^p\int_{\overline{CF}^+_{p,q}(\mathbf{C})}\omega_{\Gamma} = (-1)^q\int_{\overline{CF}^+_{p,q}(\mathbf{C})}\omega_{\Gamma},$$ implying the integral is $0$ whenever $p$ and $q$ have different parity, i.e. whenever $p+q$ is odd. This means that the first homotopy to $\mathcal{V}_{1,1}$ is given by $\mathcal{V}_{1,3}$. The angle between collinear points is constant, so the differential form associated to a graph containing an edge connecting collinear vertices will be zero. The set $dgra^3_{1,3}$ contains a unique graph without edges connecting collinear vertices, up to direction and ordering of edges, namely the graph with a free vertex of valence three and three collinear vertices of valence one. Hence there are eight (equivalence classes of) graphs (corresponding to the $2^3$ ways to direct the three edges) contributing to $\mathcal{V}_{1,3}$. Each of these eight equivalence classes has a representative with the edges ordered so that $e_i$ connects the free vertex with the collinear vertex labelled by $i$, $1\leq i\leq 3$. These representatives all have weight $1/24$. It follows that $$\mathcal{V}_{1,3} =\frac{1}{24}
m\circ\bigl(\tau_{1,4}\circ\tau_{1,3}\circ\tau_{1,2} + \tau_{1,4}\circ\tau_{1,3}\circ\tau_{2,1}
+ \tau_{1,4}\circ\tau_{3,1}\circ\tau_{1,2} + \tau_{4,1}\circ\tau_{1,3}\circ\tau_{1,2}$$ $$+ \tau_{4,1}\circ\tau_{3,1}\circ\tau_{1,2} + \tau_{4,1}\circ\tau_{1,3}\circ\tau_{2,1}
+ \tau_{1,4}\circ\tau_{3,1}\circ\tau_{2,1} + \tau_{4,1}\circ\tau_{3,1}\circ\tau_{2,1} \bigr)$$ as a map $T_{poly}^{\otimes 1+3}\rightarrow T_{poly}$. (The first of the four copies of $T_{poly}$ acts on the last three.)
The [$L_{\infty}$]{} map [$\mathcal{U}$]{}
------------------------------------------
The map $\mathcal{U}$ is, by construction, Kontsevich’s Formality Map. Recall that it’s first Taylor component $\mathcal{U}_1=\sum_{n\geq 0}\mathcal{U}_{1,n}$ is the Hochschild-Kostant-Rosenberg quasi-isomorphism.
The map [$\mathcal{Z}$]{} of [$NCG_{\infty}$]{} algebras
--------------------------------------------------------
Since $\overline{CF}^+_{0,1,n}(\mathbf{H})$ is isomorphic to $\overline{CF}^+_{1,0,n}(\mathbf{H})$ and $dgra^n_{0,1,n}$ is isomorphic to $dgra^n_{1,0,n}$, for all $n$, the maps $\mathcal{Z}_{0,1,n}$ coincide with the maps $\mathcal{U}_{1,n}$. Hence the first Taylor component of $\mathcal{Z}$, $$\sum_{n\geq 0}\mathcal{Z}_{0,1,n}:T_{poly}\rightarrow C(\mu),$$ is the Hochschild-Kostant-Rosenberg (HKR) quasi-isomorphism. The higher components of Kontsevich’s Formality Map $\mathcal{U}$ are homotopies measuring the failure of the HKR map to respect the Lie brackets. In the same way, the higher components of $\mathcal{Z}$ are homotopies that keep track of the failure of the HKR map to respect the associative products and the respective actions of $T_{poly}$ by homotopy derivations of said associative products. Since the first component is the HKR morphism, we get the following theorem:
\[maintheorem\] The algebras $(T_{poly},\wedge,\mathcal{V})$ and $(C(\mathcal{O},\mathcal{O}),d_H+\cup,br\circ\mathcal{U})$ are quasi-isomorphic as $A_{\infty}$ algebras with $L_{\infty}$ actions by $(T_{poly},[\,,\,]_S)$. The map $Z=\{Z_{k,m}=\sum_{n\geq 0}\mathcal{Z}_{k,m,n}\}_{k\geq 0,m\geq 1}$ is an explicit such quasi-isomorphism.
This statement implies the following $A_{\infty}$ formality theorem:
The algebras $(T_{poly},\wedge)$ and $(C(\mathcal{O},\mathcal{O}),d_H+\cup)$ are quasi-isomorphic $A_{\infty}$ algebras. The map $A=\{A_m:=\sum_{n\geq 0}\mathcal{Z}_{0,m,n}\}_{m\geq 1}$ is an explicit such quasi-isomorphism.
This result has already been demonstrated, but in a different way, by Shoikhet; see [@Sh].
Induced [$A_{\infty}$]{} maps
-----------------------------
An $NCG_{\infty}$ algebra consists in an $L_{\infty}$ algebra $(L,\lambda)$, an $A_{\infty}$ algebra $(A,\nu)$ and an $L_{\infty}$ morphism $\mathcal{V}:L\rightarrow C^{\geq 1}(\nu)$. Let $\hbar$ be a formal parameter. The map $\mathcal{V}$ induces a map on the sets of Maurer-Cartan elements, $$\mathrm{MC}(L[[\hbar]])\rightarrow\mathrm{MC}(C(\nu)[[\hbar]]), \pi\mapsto \sum_{p\geq 1}\frac{1}{p!}\mathcal{V}_{p,q}((\hbar \pi)^{\otimes p},\,).$$ This gives us, for each Maurer-Cartan element $\pi$ of $L$, an $A_{\infty}$ structure $$\nu^{\mathcal{V}(\pi)}_q:=\nu_q+\sum_{p\geq 1}\frac{1}{p!}\mathcal{V}_{p,q}((\hbar \pi)^{\otimes p},\,), \;q\geq 1,$$ on $A[[\hbar]]$.
If $\mathcal{Z}:(L,A,\lambda,\mathcal{V},\nu)\rightarrow (L,B,\lambda,\mathcal{U},\mu)$ is a morphism of $NCG_{\infty}$ algebras (the same $L_{\infty}$ algebra acting on both and we assume the $NCG_{\infty}$ algebra morphism is the identity on the Lie-color), then, for any Maurer-Cartan element $\pi$ of $L[[\hbar]]$, we get an induced map of $A_{\infty}$ algebras $$\mathcal{Z}^{\pi}: (A[[\hbar]],\nu^{\mathcal{V}(\pi)})\rightarrow(B[[\hbar]],\mu^{\mathcal{U}(\pi)})$$ by $\mathcal{Z}^{\pi}_m:=\mathcal{Z}_{0,m}+\sum_{k\geq 0}\frac{1}{k!}\mathcal{Z}_{k,m}((\hbar\pi)^{\otimes k},\,)$. If $\mathcal{Z}$ is a quasi-isomorphism, then $\mathcal{Z}^{\pi}$ is as well.
Applying this general construction to our representation $\Phi\circ\omega^*$ produces, for any Maurer-Cartan element $\pi\in T_{poly}$ (i.e. a Poisson bivector),
- an $A_{\infty}$ structure $\nu^{\mathcal{V}(\pi)}$ on $T_{poly}[[\hbar]]$ with $\nu^{\mathcal{V}(\pi)}_1+\nu^{\mathcal{V}(\pi)}_2 = \hbar[\pi,\,]_S+\wedge$,
- the $A_{\infty}$ cup product on the Hochschild cochains of $\mathcal{O}[[\hbar]]$ corresponding to the Kontsevich star product $\mu^{\mathcal{U}(\pi)}$ on $\mathcal{O}[[\hbar]]$ defined by $\pi$,
- and an $A_{\infty}$ quasi-isomorphism $\mathcal{Z}^{\pi}:(T_{poly}[[\hbar]],\nu^{\mathcal{V}(\pi)})\rightarrow C(\mu^{\mathcal{U}(\pi)})[[\hbar]]$.
We record this fact as a corollary.
Let $\pi\in T_{poly}$ be a Poisson structure. Then the $A_{\infty}$ algebra $(T_{poly}[[\hbar]],\nu^{\mathcal{V}(\pi)})$ is quasi-isomorphic as an $A_{\infty}$ algebra to the algebra of Hochschild cochains on $\mathcal{O}[[\hbar]]$ equipped with the cup product corresponding to the Kontsevich star product defined by $\pi$. The map $\mathcal{Z}^{\pi}$ is an explicit such quasi-isomorphism.
Homological properties of the exotic [$NCG_{\infty}$]{} algebra structure [$\mathcal{V}$]{}
-------------------------------------------------------------------------------------------
Let $\mathcal{NCG}$ be the two-colored operad of noncommutative Gerstenhaber algebras and let $f:\mathcal{NCG}\rightarrow\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}$ be the map which sends the bracket to the (sum of) graph(s) $e_{12}+e_{21}\in\mathfrak{G}_{{\scriptscriptstyle\overline{CF}(\mathbf{C})}}(2)=\mathfrak{G}_{{\scriptscriptstyle \overline{C}(\mathbf{C})}}(2)$, for $e_{12}$ ($e_{21}$) the graph with vertices $\{1,2\}$ and a single edge from $1$ to $2$ (from $2$ to $1$), sends the product to the graph in $\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(0,2)$ which has two vertices and no edge, and sends the action to the graph in $\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}(1,1)$ which is $e_{12}+e_{21}$ with the vertices in different colors. The composition $\Phi\circ f:\mathcal{NCG}\rightarrow\mathcal{E}nd(T_{poly},T_{poly})$ is the usual structure of NCGA on polyvector fields in terms of the wedge product and the Schouten bracket. The deformation complex of $f$ is the mapping cone $$\mathscr{C}:=\mathrm{Cone}(\mathrm{Def}(\mathcal{L}ie^1_{\infty}\rightarrow \mathfrak{G}_{{\scriptscriptstyle \overline{C}(\mathbf{C})}})[-1]\rightarrow
\mathrm{Def}(\mathcal{A}ss_{\infty}\rightarrow\textstyle{\int}\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}})).$$ See the appendix for notation and further details. The complex $\mathrm{Def}(\mathcal{L}ie^1_{\infty}\rightarrow \mathfrak{G}_{{\scriptscriptstyle \overline{C}(\mathbf{C})}})$ is a directed version of Kontsevich’s graph complex, $GC$, and quasi-isomorphic to it, as shown in [@W]. The operad $\int\mathfrak{G}_{{\scriptscriptstyle \overline{CF}(\mathbf{C})}}$ is a directed version of the operad $\mathcal{G}raphs$ used by Kontsevich in his proof in [@Kon99] of the formality of the little disks operad, and it is quasi-isomorphic to it[@W]. Thomas Willwacher has proved the following:
[@W]
- $H^0(GC)\cong\mathfrak{grt}$ as a graded Lie algebra.
- $H^1(\mathrm{Def}(\mathcal{A}ss_{\infty}\rightarrow\mathcal{G}raphs))\cong\mathfrak{grt}\oplus\mathbf{R}[-1]$ as a vector space, where $\mathbf{R}[-1]$ is spanned by the class of the sum of graphs contributing to $\mathcal{V}_{1,3}$.
- The map $GC[-1]\rightarrow\mathrm{Def}(\mathcal{A}ss_{\infty}\rightarrow\mathcal{G}raphs)$ is injective on cohomology.
This theorem, together with the long exact sequence for our mapping cone, implies that $H^{d+1}(\mathscr{C})\cong H^{d+1}(\mathrm{Def}(\mathcal{A}ss_{\infty}\rightarrow\mathcal{G}raphs))/H^d(GC)$. In particular, $H^1(\mathscr{C})$ is one-dimensional, spanned by the sum of graphs entering $\mathcal{V}_{1,3}$.
Using the representation $\Phi$ we can push this statement to a universal (or, rather, generic) statement about structures on polyvector fields.
The exotic $NCG_{\infty}$ algebra structure $\mathcal{V}$ on polyvector fields is generically not homotopic to the usual such structure. Moreover, it represents the unique infinitesimal deformation of the usual structure.
The $A_{\infty}$ structures $\wedge+\hbar[\pi,\,]$ and $\nu^{\mathcal{V}(\pi)}$ on $T_{poly}[[\hbar]]$ are, generically, not homotopic.
We have to say generically because for some dimensions of the $\mathcal{O}$-module $T_{poly}$ and for some degenerate Maurer-Cartan elements the corollaries might not be true.
**[A Duflo-type theorem]{}**
============================
Kontsevich’s paper [@Kon03] contained a (somewhat sketchy) proof that the tangential morphism of his Formality map, applied to a finite dimensional Lie algebra, defined an isomorphism $H(\mathbf{g},S(\mathbf{g}))\rightarrow H(\mathbf{g}, U(\mathbf{g}))$ of Chevalley-Eilenberg cohomology algebras. This result was later given a detailed proof and generalized to an arbitrary dg Lie algebra of finite type, see [@MT; @PT; @CR]. In this section we discuss a homotopy generalization of this theorem.
Let $\mathbf{g}$ be a graded vector space of finite type (i.e. finite dimensional in each degree). Let $T_{poly}$ be the polyvector fields on $\mathbf{g}[1]$, so $T_{poly}=S(\mathbf{g}^*[-1])\otimes S(\mathbf{g})$. (We shall assume all tensor algebras to be completed.) Identify $T_{poly}$ with $\mathrm{Map}(S(\mathbf{g}[1]),S(\mathbf{g}))$. The graded Lie algebra $$\mathrm{Def}(\mathcal{L}ie,\mathbf{g})[-1]:=\mathrm{Def}(\mathcal{L}ie_{\infty}\stackrel{0}{\rightarrow}\mathcal{E}nd(\mathbf{g}))[-1]
=\mathrm{Map}(S^{\geq 1}(\mathbf{g}[1])),\mathbf{g})$$ embeds into $T_{poly}$ as a Lie subalgebra. We have $\mathcal{O}=S(\mathbf{g}^*[-1])$, and $$C(\mathcal{O},\mathcal{O})=\mathrm{Map}(\mathrm{B}(S^a(\mathbf{g}^*[-1])),S^a(\mathbf{g}^*[-1]))\cong
\mathrm{Map}(S^c(\mathbf{g}[1]),\mathrm{\Omega}(S^c(\mathbf{g}[1]))).$$ Here $\mathrm{B}(\,)$ denotes the (coassociative) bar construction and $\mathrm{\Omega}(\,)$ denotes the (associative) cobar construction.
After the above identifications the following result is a straight-forward corollary to our Main Theorem, \[maintheorem\].
The representation $\Phi\circ\omega^*:\mathcal{K}(\overline{CF}(\mathbf{H}))\rightarrow\mathcal{E}nd(T_{poly},T_{poly},\mathcal{O})$ induces an explicit quasi-isomorphism $\mathrm{Map}(S^c(\mathbf{g}[1]),S^a(\mathbf{g}))\rightarrow\mathrm{Map}(S^c(\mathbf{g}[1]),\mathrm{\Omega}(S^c(\mathbf{g}[1])))$ of $A_{\infty}$ algebras equipped with $L_{\infty}$ actions by the graded Lie algebra $\mathrm{Def}(\mathcal{L}ie,\mathbf{g})$.
As before we can, given a Maurer-Cartan element of $\mathrm{Def}(\mathcal{L}ie,\mathbf{g})$ push this to a quasi-isomorphism of the induced A-infintity structures. The formal parameter $\hbar$ may in the present case be discarded (set to $1$). It’s purpose is only to define filtrations that ensure we never encounter diverging sums, but in the present case one may use weight grading by tensor lengths to define such filtrations. It is a standard argument and we omit the details.
A Maurer-Cartan element $Q$ of $\mathrm{Def}(\mathcal{L}ie,\mathbf{g})$ is precisely an $L_{\infty}$ structure on $\mathbf{g}$. Assume $Q$ given and interpret it as a coderivation of $S^c(\mathbf{g}[1])$, and denote the dg coalgebra $(S^c(\mathbf{g}[1]),Q)$ by $C(\mathbf{g})$. The cobar construction $$\mathrm{\Omega}(C(\mathbf{g}))=:U_{\infty}(\mathbf{g})$$ is the derived universal enveloping algebra of the $L_{\infty}$ algebra $(\mathbf{g},Q)$ introduced by V. Baranovsky in [@B07]. (In the special case of a dg Lie algebra it is quasi-isomorphic to the usual universal enveloping algebra.) Kontsevich’s formality map $\mathcal{U}$ quantizes $Q$ to a differential on $S^a(\mathbf{g}^*[-1])$. Denote $S^a(\mathbf{g}^*[-1])$ equipped with this differential by $C(\mathbf{g},\mathbf{R})$. We have an isomorphism of algebras $$C(C(\mathbf{g},\mathbf{R}),C(\mathbf{g},\mathbf{R}))\cong\mathrm{Map}(C(\mathbf{g}),U_{\infty}(\mathbf{g}))=:C(\mathbf{g},U_{\infty}(\mathbf{g})).$$ However, the induced $A_{\infty}$ structure on $\mathrm{Map}(S^c(\mathbf{g}[1]),S^a(\mathbf{g}))$ is not simply $$C(\mathbf{g},S(\mathbf{g}))=\mathrm{Map}(C(\mathbf{g}),S^a(\mathbf{g})).$$ Instead, we obtain an $A_{\infty}$ algebra $C(\mathbf{g},S(\mathbf{g}))_{exotic}$, which is a (generically) homotopy nontrivial deformation of $C(\mathbf{g},S(\mathbf{g}))$. The induced $A_{\infty}$ quasi-isomorphism is $$\mathcal{Z}^Q:C(\mathbf{g},S(\mathbf{g}))_{exotic}\rightarrow C(\mathbf{g},U_{\infty}(\mathbf{g})).$$
- The cohomologies $H(C(\mathbf{g},S(\mathbf{g}))_{exotic})$ and $H(\mathbf{g},S(\mathbf{g}))$ are isomorphic as associative algebras and the map on cohomology induced by $\mathcal{Z}^Q$ coincides, by construction, with the Duflo-Kontsevich isomorphism. Thus our theorem generalizes the Duflo-Kontsevich statement.
- Since $C(\mathbf{g},S(\mathbf{g}))_{exotic}$ is, generically, not quasi-isomorphic to $C(\mathbf{g},S(\mathbf{g}))$, but–by our theorem–is quasi-isomorphic to $C(\mathbf{g},U_{\infty}(\mathbf{g}))$, it follows that there does not, generically, exist a quasi-isomorphism of $A_{\infty}$ algebras $$C(\mathbf{g},S(\mathbf{g})) \rightarrow C(\mathbf{g},U_{\infty}(\mathbf{g})).$$ In other words, it is impossible to find a universal $A_{\infty}$ lift of the Duflo-Kontsevich isomorphism on Chevalley-Eilenberg cohomologies to the Chevalley-Eilenberg cochain algebras.
There is a canonical isomorphism between $T_{poly}$ on $\mathbf{g}[1]$ and $T_{poly}$ on $\mathbf{g}^*$. Above we used the first graded vector space, for which $\mathcal{O}=S^a(\mathbf{g}^*[-1])$. Application of Kontsevich’s formality to the second case, for which $\mathcal{O}=S^a(\mathbf{g})$, quantizes an $L_{\infty}$ structure $Q\in T_{poly}$ to a (flat) $A_{\infty}$ structure $\star$ on $S(\mathbf{g})[[\hbar]]$. Calaque, Felder, Ferrario and Rossi constructed in [@CFFR] a nontrivial but explicit $A_{\infty}$ $(S(\mathbf{g})[[\hbar]],\star)-C(\mathbf{g},\mathbf{R})[[\hbar]]$-bimodule structure $K_{\hbar}$ on $\mathbf{R}[[\hbar]]$ and they proved that the derived left action $$L:(S(\mathbf{g})[[\hbar]],\star)\rightarrow\mathrm{Map}_{\hbar}(K_{\hbar}[1]\otimes \mathrm{B}(C(\mathbf{g},\mathbf{R}))[[\hbar]],K_{\hbar}[1])$$ is a quasi-isomorphism of $A_{\infty}$ algebras. One may formally set $\hbar=1$ in this quasi-isomorphism, for essentially the same reasons as those which allowed us to do so above then identify the term on the right (above), with the cobar construction $\mathrm{\Omega}(C(\mathbf{g}))$. Thus the result of [@CFFR] implies that the quantization of the symmetric algebra on the $L_{\infty}$ algebra $\mathbf{g}$, $(S(\mathbf{g}),\star)$, is quasi-isomorphic to Baranovsky’s derived universal enveloping algebra of $\mathbf{g}$. A detailed proof of this will be contained in [@AR]. Together with our result this quasi-isomorphism implies that the $A_{\infty}$ algebras $C(\mathbf{g},S(\mathbf{g}))_{exotic}$ and $C(\mathbf{g},(S(\mathbf{g}),\star))$ are quasi-isomorphic, though the quasi-isomorphism is presently not explicit.
[**[$NCG_{\infty}$]{} algebras**]{}
===================================
Let $\mathcal{NCG}$ be the two-colored operad generated by a degree $-1$ Lie bracket $[x_1,x_2]$ in one color, call it $\mathbf{x}$, an associative degree $0$ product $a_1\cdot a_2$ in another color, call it $\mathbf{a}$, and an operation which we denote $x_1\bullet a_1$, of the type $(\mathbf{x},\mathbf{a})\rightarrow\mathbf{a}$, which represents the bracket in derivations of the product. This is the operad of NCGAs.
The operad $\mathcal{NCG}$ is Koszul.
We shall use the rewriting systems method of [@LV]. The rewriting rules are $$\begin{aligned}
(a_1\cdot a_2)\cdot a_3 &\mapsto a_1\cdot (a_2\cdot a_3) \nonumber \\
[[x_1,x_2],x_3] &\mapsto -[[x_2,x_3],x_1]-[[x_3,x_1],x_2] \nonumber \\
x_1\bullet(a_1\cdot a_2) &\mapsto (x_1\bullet a_1)\cdot a_2 + a_1\cdot(x_1\bullet a_2) \nonumber\\
[x_1,x_2] \bullet a_1 &\mapsto x_1\bullet(x_2\bullet a_1) - x_2\bullet(x_1\bullet a_1). \nonumber\end{aligned}$$ The critical monomials are $((a_1\cdot a_2)\cdot a_3)\cdot a_4$, $[[[x_1,x_2],x_3],x_4]$, $x_1\bullet ((a_1\cdot a_2)\cdot a_3)$, $[x_1,x_2] \bullet(a_1\cdot a_2)$ and $[[x_1,x_2],x_3] \bullet a_1$. The first two are known to be confluent as the operads $\mathcal{L}ie^1$ and $\mathcal{A}ss$ are known to be Koszul. The third critical monomial can be rewritten either as $$\begin{aligned}
x_1\bullet ((a_1a_2)a_3) &\mapsto (x_1\bullet(a_1a_2))a_3+(a_1a_2)(x_1\bullet a_3) \nonumber \\
&\mapsto ((x_1\bullet a_1)a_2)a_3+(a_1(x_1\bullet a_2))a_3+a_1(a_2(x_1\bullet a_3)) \nonumber \\
&\mapsto (x_1\bullet a_1)(a_2a_3)+a_1((x_1\bullet a_2)a_3)+a_1(a_2(x_1\bullet a_3)) \nonumber\end{aligned}$$ or as $$\begin{aligned}
x_1\bullet ((a_1a_2)a_3) &\mapsto x_1\bullet(a_1(a_2a_3)) \nonumber \\
&\mapsto (x_1\bullet a)(a_2a_3)+a_1(x_1\bullet(a_2a_3)) \nonumber \\
&\mapsto (x_1\bullet a_1)(a_2a_3)+a_1((x_1\bullet a_2)a_3)+a_1(a_2(x_1\bullet a_3)). \nonumber\end{aligned}$$ Since both ways give the same end result, $x_1\bullet ((a_1a_2)a_3)$ is confluent.
The critical monomial $[x_1,x_2] \bullet(a_1\cdot a_2)$ can be rewritten either as $$\begin{aligned}
[x_1,x_2] \bullet(a_1a_2) &\mapsto ([x_1,x_2]\bullet a_1)a_2+a_1([x_1,x_2]\bullet a_2) \nonumber \\
&\mapsto (x_1\bullet(x_2\bullet a_1))a_2-(x_2\bullet(x_1\bullet a_1))a_2+a_1(x_1\bullet(x_2\bullet a_2))-a_1(x_2\bullet(x_1\bullet a_2)) \nonumber\end{aligned}$$ or $$\begin{aligned}
[x_1,x_2] \bullet(a_1a_2) &\mapsto x_1\bullet(x_2\bullet(a_1a_2))-x_2\bullet(x_1\bullet(a_1a_2)) \nonumber \\
&\mapsto x_1\bullet((x_2\bullet a_1)a_2)+x_1\bullet(a_1(x_2\bullet a_2))-x_2\bullet((x_1\bullet a_1)a_2)-x_2\bullet(a_1(x_1\bullet a_2)) \nonumber \\
&\mapsto (x_1\bullet(x_2\bullet a_1))a_2+(x_2\bullet a_1)(x_1\bullet a_2)+(x_1\bullet a_1)(x_2\bullet a_2) + a_1(x_1\bullet(x_2\bullet a_2)) \nonumber \\
&-(x_2\bullet(x_1\bullet a_1))a_2-(x_1\bullet a_1)(x_2\bullet a_2)-(x_2\bullet a_1)(x_1\bullet a_2)-a_1(x_2\bullet(x_1\bullet a_2)). \nonumber\end{aligned}$$ These two ways to rewrite the monomial agree, so it is also confluent. Confluence of the last critical monomial, $[[x_1,x_2],x_3] \bullet a_1$, is a similar straightforward manipulation and we omit it.
For a Koszul operad $\mathcal{P}$ the operad $\mathcal{M}or(\mathcal{P})$, whose representations are pairs of $\mathcal{P}$-algebras together with a morphism of $\mathcal{P}$-algebras between them, is again Koszul by the results of [@MV09]. An algebra for the resolution $\mathrm{\Omega}(\mathcal{M}or(\mathcal{P})^{\text{!`}})$ consists in two strong homotopy $\mathcal{P}$-algebras and a strong homotopy morphism between them. This general machinery produces a four-colored operad $\mathrm{\Omega}(\mathcal{M}or(\mathcal{NCG})^{\text{!`}})$ of morphisms of $NCG_{\infty}$ algebras. It has two “Lie-colors” and two “Ass-colors”. We can make it into a 3-colored operad by identifying the two Lie colors. (A representation of that new operad will be a morphism of $NCG_{\infty}$ algebras having the same $L_{\infty}$ algebra acting on both $A_{\infty}$ algebras.) This operad includes generators describing an $L_{\infty}$ endomorphism of the Lie-color. Quotient out these generators and get a new three-colored operad $\mathcal{M}or_*(\mathcal{NCG})_{\infty}$. Its representations are morphisms of $NCG_{\infty}$ algebras that have the same $L_{\infty}$ algebra acting on both $A_{\infty}$ algebras, and for which the $L_{\infty}$ endomorphism is the identity. It is easy to see, knowing that $\mathrm{\Omega}(\mathcal{M}or(\mathcal{NCG})^{\text{!`}})\rightarrow\mathcal{M}or(\mathcal{NCG})$ is a quasi-isomorphism, that $\mathcal{M}or_*(\mathcal{NCG})_{\infty}$ is quasi-isomorphic to the operad $\mathcal{M}or_*(\mathcal{NCG})$ which has as representations two NCGAs with the same Lie algebra acting on both associative algebras and a morphism between the NCGAs which is the identity on the Lie algebra. Finally one can note that, in fact, $\mathcal{M}or_*(\mathcal{NCG})_{\infty}=\mathrm{\Omega}(\mathcal{M}or_*(\mathcal{NCG})^{\text{!`}})$.
Consider the operad $\mathcal{NCG}^{(1)}$ of (one-colored) noncommutative Gerstenhaber algebras (chain complexes that are simultaneously a dg Lie algebra, with the bracket of degree $-1$, and an associative algebra, and the Lie bracket acts by derivations of the associative product). Our method to prove Koszulity of $\mathcal{NCG}$ does not repeat mutatis mutandum for $\mathcal{NCG}^{(1)}$. The problem is that one gets a new critical monomial, $[x_1x_2,x_3x_4]$, which is not confluent. This suggests (but does not prove) that $\mathcal{NCG}^{(1)}$ is not Koszul.
The deformation complex of [$\mathcal{NCG}_{\infty}\rightarrow\mathcal{P}$]{}
-----------------------------------------------------------------------------
Recall that we denote the two colors of $\mathcal{NCG}$ by $\mathbf{x}$ and $\mathbf{a}$. Here $\mathbf{x}$ is the “Lie” color and $\mathbf{a}$ is the “Ass” color. Define $\mathcal{NCG}_{\infty}:=\mathrm{\Omega}(\mathcal{NCG}^{\text{!`}})$.
Let $\mathcal{P}$ be a dg operad with colors $\mathbf{x}$ and $\mathbf{a}$ and assume given a morphism of operads $f:\mathcal{NCG}_{\infty}\rightarrow\mathcal{P}$. We shall describe the deformation complex $\mathrm{Def}(f)$.
We shall simplify notation and write $\mathcal{P}(k)$ for $\mathcal{P}(k,0;\mathbf{x})$ and $\mathcal{P}(p,q)$ for $\mathcal{P}(p,q;\mathbf{a})$. As a chain complex, $$\mathrm{Def}(f) =\prod_{k\geq 2} \mathcal{P}(k)_{\Sigma_k}[2-2k] \oplus
\prod_{\substack{p\geq 0,q\geq 1\\p+q\geq 2}} \mathcal{P}(p,q)_{\Sigma_p}\otimes sgn_q[1-2p-q].$$ This chain complex has a degree zero graded Lie bracket defined by taking the commutator of operadic composition. The map $f=(f_k)+(f_{p,q})$ is a Maurer-Cartan-element and the differential on the deformation complex is the internal differential on $\mathcal{P}$ plus the bracket $[f,\,]$. We can give a more suggestive formulation of the deformation complex as follows. The components $(f_k)$ define a morphism $\lambda^f:\mathcal{L}ie^1_{\infty}\rightarrow\mathcal{P}$ and $$\mathrm{Def}(\lambda^f)=\prod_{k\geq 2} \mathcal{P}(k)_{\Sigma_k}[2-2k]$$ with differential (the internal differential on $\mathcal{P}$ plus) $[(f_k),\,]$. Set $${\textstyle{\int}}\mathcal{P}(q):=\prod_p\mathcal{P}(p,q)_{\Sigma_p}[-2p].$$ The collection $\int\mathcal{P}=\{\int\mathcal{P}(q)\}$ has a structure of dg operad. (This can actually be interpreted as a categorical end: a $\Sigma$-bimodule can be regarded as a bifunctor and we take the limit over one argument.) The compositions of the Lie-color in $\mathcal{P}$ define a right action $\bullet$ of $\mathrm{Def}(\lambda^f)$ on $\int\mathcal{P}$ by operadic derivations. Add to the differential the term $[(f_{p,1}),\,]+(\,)\bullet(f_k)$. The remaining mixed components of $f$, i.e. $(f_{p,q})$ with $q\geq 2$, define a morphism $\mu^f:\mathcal{A}ss_{\infty}\rightarrow\int\mathcal{P}$ with $\mu^f_q=(f_{p,q})_{p\geq 0}$. We have $$Def(\mu^f)=\prod_{\substack{p\geq 0,q\geq 1\\p+q\geq 2}} \mathcal{P}(p,q)_{\Sigma_p}\otimes sgn_q[1-2p-q].$$ The components $(f_{p,q})_{p\geq 1}$ define a map of complexes $\rho^f:\mathrm{Def}(\lambda^f)[-1]\rightarrow\mathrm{Def}(\mu^f)$ by $\gamma\mapsto \gamma\circ (f_{p,q})$.
The deformation complex $\mathrm{Def}(f)$ is isomorphic as a chain complex to the mapping cone of $\rho^f$ and as a graded Lie algebra to $\mathrm{Def}(\lambda^f)\ltimes\mathrm{Def}(\mu^f)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The collection ${\mathfrak{C}}_3$ of all triples of commuting contractions forms a family in the sense of Agler, and so has an “optimal” model ${\partial}{\mathfrak{C}}_3$ generated by its extremal elements. A given $T\in{\mathfrak{C}}_3$ is extremal if every $X\in{\mathfrak{C}}_3$ extending $T$ is an extension by direct sum. We show that many of the known examples of triples in ${\mathfrak{C}}_3$ that fail to have coisometric extensions are in fact extremal.'
author:
- |
Edward J. Timko\
Indiana University Department of Mathematics
title: Some examples of extremal triples of commuting contractions
---
Introduction
============
Given $n\in{\mathbb{N}}$, we denote by ${\mathfrak{C}}_n$ the class of all $n$-tuples of commuting contractions. Observe that ${\mathfrak{C}}_n$ is a *family* in the sense of Alger [@Agler1], which is to say that:
1. ${\mathfrak{C}}_n$ is closed with respect to direct sums. That is, given $A^{(j)}\in{\mathfrak{C}}_n$ for every $j\in J$, we have $\left(\bigoplus_{j\in J} A^{(j)}_i\right)_{i=1}^n \in{\mathfrak{C}}_n$;
2. Given $A\in{\mathfrak{C}}_n$ and a unital $*$-representation $\pi$ of the unital C$^*$-algebra generated by $A_1,\dots, A_n$, then $(\pi(A_i))_{i=1}^n \in {\mathfrak{C}}_n$; and
3. ${\mathfrak{C}}_n$ is hereditary. That is, if $A\in{\mathfrak{C}}_n$ and if ${\mathcal{M}}$ is an invariant subspace of ${\mathcal{H}}$ for $A_1,\dots ,A_n$, then $(A_i|{\mathcal{M}})_{i=1}^n\in{\mathfrak{C}}_n$.
An element $T\in{\mathfrak{C}}_n$ is said to be *extremal* if whenever $S\in{\mathfrak{C}}_n$ is an extension of $T$, then $S$ is an extension by direct sum. That is, if ${\mathcal{N}}$ is invariant for $S$ so $T=S|{\mathcal{N}}$, then ${\mathcal{N}}$ is a reducing subspace for $S$. We say that $S$ is a *trivial* extension of $T$ if $S$ is an extension of $T$ by direct sum.
Let ${\mathcal{B}}{\subseteq}{\mathfrak{C}}_n$. We say that ${\mathcal{B}}$ is a *model* for ${\mathfrak{C}}_n$ if
1. ${\mathcal{B}}$ is closed with respect to direct sums and unitary $*$-representations; and
2. given $T\in{\mathfrak{C}}_n$ acting a Hilbert space ${\mathcal{H}}$, there exists $S\in{\mathcal{B}}$ having ${\mathcal{H}}$ as an invariant subspace so that $S|{\mathcal{H}}=T$.
Lastly, the *boundary* of ${\mathfrak{C}}_n$, denoted by ${\partial}{\mathfrak{C}}_n$, is the smallest model for ${\mathfrak{C}}_n$. It follows as a consequence of Proposition 5.9 and 5.10 in [@Agler1] that this family always exists and is generated by the extremal elements of ${\mathfrak{C}}_n$.
In the case that $n=1,2$, the boundary ${\partial}{\mathfrak{C}}_n$ consists of all tuples of commuting coisometries, a consequence of the work of Sz.-Nagy for $n=1$ and Andô for $n=2$ [@SzNagy1]. For $n>2$ this characterization is no longer valid. It may well be the case that no concrete description of the extremal elements of ${\mathfrak{C}}_3$ or ${\partial}{\mathfrak{C}}_3$ is possible. We show that many of the known examples of triples in ${\mathfrak{C}}_3$ that fail to have coisometric extensions are in fact extremal.
Agler’s theory has seen some application. Dritschel and McCullough show in [@DritMcCu1] that if ${\mathcal{F}}$ denotes the family of contractive hyponormal operators, then ${\partial}{\mathcal{F}}={\mathcal{F}}$. In the same article, sufficient conditions are given for an $n$-hyponormal operator to be extremal. In an article by Curto and Lee [@CurtoLee], it is shown that a weakly subnormal operator satisfying the conditions of [@DritMcCu1] must be normal and so extremal for the collection of all weakly subnormal operators. Dritschel, McCullough, and Woerdeman [@DritMcCuWo] give a collection of equivalent conditions for a $\rho$-contraction (for $\rho\leq 2$) to be extremal, ultimately showing for $\rho\in(0,1)\cup(1,2]$ that ${\mathcal{C}}_\rho={\partial}{\mathcal{C}}_\rho$, with ${\mathcal{C}}_\rho$ denoting the class of $\rho$-contractions. In another article by Dritschel and McCullough [@DritMcCu2] it is shown that a family, in “Agler’s sense”, of representations of either an operator algebra or an operator space has boundary representations, as related to the non-commutative Shilov boundary. Finally, in [@RichSund] Richter and Sundberg apply Agler’s theory to the study of row contractions and spherical contractions.
Here is an outline of the material found in this paper. In Section 2 we make some observations that apply to any $n$-tuple of commuting contractions. While these results are only applied in Section 5, they are general enough to merit separate exposition. In Section 3 we study an $n$-tuple of Parrot [@Par1], finding the $n$-tuple to be extremal if and only if a certain subspace is trivial. In Section 4 we prove a triple of Crabb and Davie [@Pisier p. 23] is extremal in ${\mathfrak{C}}_3$. In Section 5 we examine a triple due to Varopoulos [@Pisier p. 86] and show that this triple is extremal only for a relatively narrow range of parameters.
We comment on another triple that has appeared in the literature. In [@LotSte] Lotto and Steger find a triple of commuting, diagonalizable contractions that fail to obey the von Neumann inequality. This triple does not appear to produce extremal elements so its examination has been omitted from this paper.
Before closing this section, I would like to thank Hari Bercovici for his guidance and the helpful criticism he provided in the preparation of this paper, and to thank the referee for his helpful suggestions.
Some General Remarks
====================
Lacking a complete description of the boundary elements, we develop some tools to tell us when we may determine when certain elements are not extremal. For the first lemma, we use the notation $\operatorname{Ran}T:=\bigvee_i \operatorname{ran}T_i$ and $\operatorname{Ker}T:=\bigcap_i \ker T_i$.
\[GR:L1\] Let $T\in{\mathfrak{C}}_n$ operating on a Hilbert space ${\mathcal{H}}$. If $(\operatorname{Ran}T)^\bot\cap\operatorname{Ker}T\neq \{0\}$, then $T$ is not extremal.
Let ${\mathcal{E}}:=(\operatorname{Ran}T)^\bot\cap\operatorname{Ker}T$ and $V:{\mathcal{E}}\to{\mathcal{H}}$ the inclusion map. Define $X$ on ${\mathcal{H}}\oplus{\mathcal{E}}$ by $$X_i:={\begin{pmatrix} T_i & V\\ 0 & 0 \end{pmatrix}} \quad i=0,1,\dots,n.$$ As $T_iV=0$ for each $i$, the $X_i$ commute. Since $VV^*$ is orthogonal to the range of each $T_i$, it follows that $T_iT_i^*+VV^*\leq 1$, and therefore each $X_i$ is a contraction. Since ${\mathcal{E}}\neq 0$, $X_i$ is a non-trivial extension.
\[GR:L2\] If $T\in{\mathfrak{C}}_n$ satisfies $\min_i\|T_i\|<1$, then $T$ is not extremal.
Consider the extension $$X_i={\begin{pmatrix} T_i & \delta_i T_i \\ 0 & \eta_i T_i \end{pmatrix}} \quad i=0,1,\dots,n$$ where $\delta_i,\eta_i\in [0,1]$ are to be determined. We want $X$ to be in ${\mathfrak{C}}_n$. Note that $$X_iX_j={\begin{pmatrix} T_iT_j & (\delta_j+\delta_i\eta_j)T_iT_j \\ 0 & \eta_i\eta_j T_iT_j \end{pmatrix}}.$$ and therefore $X_iX_j=X_jX_i$ when either $\delta_j+\delta_i\eta_j=\delta_i+\delta_j\eta_i$ or $T_iT_j=0$ for each $i,j$. It suffices to set $\eta_i=1-\delta_i$ for each $i$.
Observe now that $$X_i^*X_i={\begin{pmatrix} T_i^*T_i & \delta_i T_i^*T_i\\ \delta_i T_i^*T_i & (\delta_i^2+\eta_i^2)T_i^*T_i \end{pmatrix}}$$ Setting $\beta_i:=1+\delta_i^2+\eta_i^2$, we easily see $$\label{GR:L2:Eq1}
\|X_i\|^2\leq \frac{1}{2}\left[\beta_i+\sqrt{\beta_i^2-4\eta_i^2}\right]\|T_i\|^2.$$ To conclude the proof, we show that the $\delta_i$ can be chosen so that the right-hand side of is at most 1 for each $i$. This is equivalent to insisting $$\beta_i-\|T_i\|^2\eta_i^2\leq \frac{1}{\|T_i\|^2}$$ or equivalently $$\delta_i^2+(1-\|T_i\|^2)(\eta_i^2-1)\leq \|T_i\|^2+\|T_i\|^{-2}-2=\left\{\frac{1-\|T_i\|^2}{\|T_i\|}\right\}^2.$$ Since $(1-\|T_i\|^2)(\eta_i^2-1)\leq 0$, fix $$\delta_i = \min\left\{1,\frac{1-\|T_i\|^2}{\|T_i\|}\right\}.$$ As $\|T_i\|<1$ for some $i$ we have $\delta_i>0$.
Parrot’s Example
================
Parrott provided the first example of a triple of commuting contractions which has no commuting coisometric extension ([@Par1]; see also [@SzNagy1 p. 23]). Let $U_1,\dots,U_n$ be an arbitrary $n$-tuple of unitaries operators on a Hilbert space ${\mathcal{H}}$, and define $$\label{PE:Eq1}
T_i:={\begin{pmatrix} 0 & 0 \\ U_i & 0 \end{pmatrix}} \quad i=1,2,\dots,n$$ acting on ${\mathcal{H}}\oplus{\mathcal{H}}$. It is easily checked that the $T_i$ are commuting partial isometries, and so $T\in{\mathfrak{C}}_n$. When the $U_i$ do not commute “enough”, then $T$ has no extension to an $n$-tuple of commuting coisometries. In particular, if for some $i\neq j$ the commutator $[U_n^{-1}U_i,U_n^{-1}U_j]$ does not vanish, then $T$ has no coisometric extension (here and elsewhere $[X,Y]=XY-YX$). We refer the reader to [@SzNagy1 p. 23] for details in the case $n=3$. A similar criterion determines when $T$ is extremal.
Let $T$ denote the Parrott $n$-tuple defined by unitaries $U_1,\dots,U_n$ acting on a Hilbert space ${\mathcal{H}}$. Then $T$ is extremal if and only if $$\bigcap_{i,j=1}^n \ker[U_n^{-1}U_i,U_n^{-1}U_j]=\{0\}.$$
An extension $X$ of $T$ takes the form $$X_i={\begin{pmatrix} 0 & 0 & A_i \\ U_i & 0 & B_i \\ 0 & 0 & C_i \end{pmatrix}}, \quad i=1,2,\dots,n.$$ As $X_i$ is contractive, ${\begin{pmatrix} U_i^*\\B_i^* \end{pmatrix}}$ is also a contraction, and so $B_i= 0$ for each $i$. Therefore $$X_i^*X_i={\begin{pmatrix} 1 & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & A_i^*A_i+C^*_iC_i \end{pmatrix}}$$ and so $X_i$ is a contraction if and only if $A_i^*A_i+C_i^*C_i\leq 1$. Since $$X_iX_j={\begin{pmatrix} 0 & 0 & A_iC_j \\ U_iU_j & 0 & U_iA_j \\ 0 & 0 & C_iC_j \end{pmatrix}} \quad i,j=1,2,\dots n,$$ commutivity requires $$A_iC_j=A_jC_i, \quad U_iA_j=U_jA_i, \quad [C_i,C_j]=0$$ for all $i,j$. Using the notation $W_j=U_n^{-1}U_j$ for $j=1,2,\dots,n$, the second of these implies $A_j=W_jA_n$ for each $j$ and therefore $$[W_i,W_j]A_n=0, \quad i,j=1,2,\dots,n.$$ Define ${\mathcal{K}}=\bigcap_{i,j}\ker[W_i,W_j]$. Thus $\operatorname{ran}A_n{\subseteq}{\mathcal{K}}$. If ${\mathcal{K}}=0$, then $A_n=0$ hence $A_i=0$ for each $i$ and therefore every extension $X$ is by direct sum; $T$ is extremal.
Conversely, suppose that ${\mathcal{K}}\neq \{0\}$. Let $A_n$ denote the canonical embedding of ${\mathcal{K}}$ into ${\mathcal{H}}$. Then $$X_i={\begin{pmatrix} 0 & 0 & W_iA_n \\ U_i & 0 & 0 \\ 0 & 0 & 0 \end{pmatrix}}, \quad i=1,2,\dots,n$$ defines a non-trivial extension of $T$ in ${\mathfrak{C}}_n$.
The conditions given by seem to unfairly favor $U_n$. The favoritism is in fact superficial. While this can be seen as a corollary of the preceding proposition, one may also directly show that $\bigcap_{i,j=1}^n\ker[U_n^{-1}U_i,U_n^{-1}U_j]=\{0\}$ if and only if $\bigcap_{i,j=1}^n\ker[U_k^{-1}U_i,U_k^{-1}U_j]=\{0\}$ for some $k=1,\dots,n$.
It should be noted that the condition in can indeed be satisfied by some tuple of operators. Consider $g_1,g_2$, generators of the free group on two elements, acting by translation on $\ell^2(\mathbb{F}_2)$. Now consider the triple of unitaries $(g_1,g_2,1)$. The intersection in reduces to $\ker[g_1,g_2]=\{0\}$.
The Crabb-Davie Example
=======================
While Parrot’s example has no coisometric extension in ${\mathfrak{C}}_n$, it nevertheless obeys the von Neumann inequality. That is, if $p$ is an analytic polynomial in three variables, then $$\|p(T_1,T_2,T_3)\|\leq \sup\{|p(z_1,z_2,z_3)|: 0\leq |z_1|,|z_2|,|z_3|<1\}.$$ However, there are triples in ${\mathfrak{C}}_3$ that do not satisfy the von Neumann inequality. A construction of Crabb and Davie [@Pisier p. 23] provides an example which consists of the three $8\times 8$-matrices, $$T_i={\begin{pmatrix} 0 & & & & & & & \\
\delta_{i1} & & & & & & & \\
\delta_{i2} & & & & & & & \\
\delta_{i3} & & & & & & & \\
& -\delta_{i1} & \delta_{i3} & \delta_{i2} & & & & \\
& \delta_{i3} & -\delta_{i2} & \delta_{i1} & & & & \\
& \delta_{i2} & \delta_{i1} & -\delta_{i3} & & & & \\
& & & & \delta_{i1} & \delta_{i2} & \delta_{i3} & 0 \end{pmatrix}} \quad i=1,2,3,$$ where every non-specified entry is 0. These are commuting partial isometries and $$T_iT_i^* = \operatorname{diag}(0,\delta_{i1},\delta_{i2},\delta_{i3},1,1,1,1).$$
The Crabb-Davie triple is extremal.
Let $X\in{\mathfrak{C}}_3$ be an extension of $T$ so $$X_i={\begin{pmatrix} T_i & A_i \\ 0 & B_i \end{pmatrix}}, \quad i=1,2,3,$$ where $A_i\in{\mathcal{L}}({\mathcal{H}},{\mathbb{C}}^8)$ and $B_i\in{\mathcal{L}}({\mathcal{H}})$ for some Hilbert space ${\mathcal{H}}$. In order for $X_i$ to be a contraction, we need in particular that $$T_iT_i^*+A_iA_i^*\leq 1, \quad i=1,2,3.$$ This implies that $\operatorname{ran}A_i{\subseteq}\operatorname{ran}(1-T_iT_i^*)$. Since $$1-T_iT_i^*=\operatorname{diag}(1,1-\delta_{1i},1-\delta_{2i},1-\delta_{3i},0,0,0,0),$$ we can express the $A_i$ as column vectors whose entries linear functionals on ${\mathcal{H}}$; $$A_1={\begin{pmatrix} \eta_1 \\ 0 \\ \phi_1 \\ \psi_1 \\ 0 \\ 0 \\ 0 \\ 0 \end{pmatrix}}, \quad A_2={\begin{pmatrix} \eta_2 \\ \phi_2 \\ 0 \\ \psi_2 \\ 0 \\ 0 \\ 0 \\ 0 \end{pmatrix}}, \quad A_3={\begin{pmatrix} \eta_3 \\ \phi_3 \\ \psi_3 \\ 0 \\ 0 \\ 0 \\ 0 \\ 0 \end{pmatrix}}.$$
Notice that $$X_iX_j={\begin{pmatrix} T_iT_j & T_iA_j+A_iB_j \\ 0 & B_iB_j \end{pmatrix}}, \quad i=1,2,3.$$ Therefore $[X_i,X_j]=0$ for all $i,j$ if and only if $$[B_i,B_j]=0 \quad T_iA_j+A_iB_j=T_jA_i+A_jB_i.$$ The second series of equations can be expressed as equalities of certain column vectors; [$$\label{CD:Eq1}
{\begin{pmatrix} \eta_1{\circ}B_2 \\ \eta_2 \\ \phi_1{\circ}B_2 \\ \psi_1{\circ}B_2 \\ -\phi_2 \\ \psi_2 \\ 0 \\ 0 \end{pmatrix}}= {\begin{pmatrix} \eta_2{\circ}B_1 \\ \phi_2{\circ}B_1 \\ \eta_1 \\ \psi_2{\circ}B_1 \\ \psi_1 \\ -\phi_1 \\ 0 \\ 0 \end{pmatrix}}, \qquad {\begin{pmatrix} \eta_3{\circ}B_1 \\ \phi_3{\circ}B_1 \\ \psi_3{\circ}B_1 \\ \eta_1 \\ \phi_1 \\ 0 \\ -\psi_1 \\ 0 \end{pmatrix}}={\begin{pmatrix} \eta_1{\circ}B_3 \\ \eta_3 \\ \phi_1{\circ}B_3 \\ \psi_1{\circ}B_3 \\ -\phi_3 \\ 0 \\ \psi_3 \\ 0 \end{pmatrix}}, \qquad {\begin{pmatrix} \eta_2{\circ}B_3 \\ \phi_2{\circ}B_3 \\ \eta_3 \\ \psi_2{\circ}B_3 \\ 0 \\ -\psi_3 \\ \phi_3 \\ 0 \end{pmatrix}}={\begin{pmatrix} \eta_3{\circ}B_2 \\ \phi_3{\circ}B_2 \\ \psi_3{\circ}B_2 \\ \eta_2 \\ 0 \\ \phi_2 \\ -\psi_2 \\ 0 \end{pmatrix}}$$ ]{}By stringing together equations from the 5th through 7th rows of , we find $$\psi_1=-\phi_2=\psi_3=-\psi_1,$$ $$\psi_2=-\phi_1=\phi_3=-\psi_2.$$ Thus $\psi_i=\phi_i=0$ for all $i$. From the 2nd through 4th rows of $$\eta_1=\phi_1{\circ}B_2=0, \quad \eta_2=\phi_2{\circ}B_1=0, \quad \eta_3=\phi_3{\circ}B_1=0.$$ Thus $A_i=0$ for each $i$, and so $X$ is a trivial extension of $T$.
We take a moment to show that the Crabb-Davie example does not satisfy the conditions of Lemma \[GR:L1\]. Note that $\operatorname{Ran}T=\{0\}\oplus {\mathbb{C}}^{\oplus 7}$ and $\operatorname{Ker}T=\{0\}^{\oplus 7}\oplus{\mathbb{C}}$. Therefore $(\operatorname{Ran}T)^\bot\cap\operatorname{Ker}T=\{0\}$.
The Varopoulos Example
======================
We need to establish some notation. Let $J$ be a set, and given $\alpha\in J$ and $x\in\ell^2(J)$, let $x(\alpha)$ denote the $\alpha$-component of $x$. Noting that a linear operator from ${\mathbb{C}}$ to $\ell^2(J)$ is uniquely determined by its value at 1, we view the elements of $\ell^2(J)$ as bounded operators ${\mathbb{C}}\to\ell^2(J)$ and the linear functionals on $\ell^2(J)$ as bounded operators $\ell^2(J)\to{\mathbb{C}}$, the operator adjoint $x\mapsto x^*$ mapping between these. Given $x,y\in\ell^2(J)$ we may now write $xy^*$ for the rank one operator $h\mapsto { \langle h , y \rangle}x$, and $y^*x={ \langle x , y \rangle}$. Another operation we define on $\ell^2(J)$ is the conjugation $${\overline{x}}(\alpha)={\overline{x(\alpha)}}, \quad \alpha \in J.$$ Note that ${\overline{x}}^*y={\overline{y}}^*x$.
Another triple that fails to obey the von Neumann inequality is provided by Varopoulos [@Pisier p. 86]. Define the Hilbert space ${\mathcal{H}}={\mathbb{C}}\oplus \ell^2(J)\oplus {\mathbb{C}}$ and let $x_1,x_2,x_3$ be in the unit ball of $\ell^2(J)$. The Varopoulos example consists of the three operators $T_1,T_2,T_3\in{\mathcal{L}}({\mathcal{H}})$ defined by $$\label{TVE:DefOfT}
T_i={\begin{pmatrix} 0 & 0 & 0 \\ x_i & 0 & 0 \\ 0 & {\overline{x}}_i^* & 0 \end{pmatrix}}, \quad i=1,2,3.$$ The $T_i$ commute because ${\overline{x}}_i^*x_j={\overline{x}}^*_jx_i$ for $i,j=1,2,3$. The identity $$\label{TVE:Eq1}
T_iT_i^*=\operatorname{diag}(0,x_ix_i^*,\|x_i\|^2)$$ implies $\|T_i\|=\|x_i\|\leq 1$ for each $i$, and so $T\in{\mathfrak{C}}_3$.
While each $J$ and each triple $x_1,x_2,x_3$ in the unit ball of $\ell^2(J)$ define a $T$ in ${\mathfrak{C}}_3$, only certain choices of $J$ and $(x_1,x_2,x_3)$ produce an extremal triple. Before providing triples that are in the boundary, we show that we may limit our attention to certain special cases. One restriction we immediately make is to limit ourselves to $\|x_i\|=1$ for each $i$. Indeed, Lemma \[GR:L2\] and imply that $T$ cannot be extremal if $\|x_i\|<1$ for some $i$. Under this restriction $\eqref{TVE:Eq1}$ shows each $T_i$ is a partial isometry.
Another immediate restriction we make is on the size of $\ell^2(J)$. Define the subspace ${\mathcal{R}}{\subseteq}\ell^2(J)$ by $$\label{TVE:DefOfR}
{\mathcal{R}}=\operatorname{Span}\{x_1,x_2,x_3,{\overline{x}}_1,{\overline{x}}_2,{\overline{x}}_3\}.$$ If ${\mathcal{R}}$ is a proper subspace of $\ell^2(J)$ then $T$ cannot be extremal. Indeed, $$\ker T_i=\{0\}\oplus \{{\overline{x}}_i\}^\bot\oplus {\mathbb{C}}{\quad\text{and}\quad} \operatorname{ran}T_i=\{0\}\oplus {\mathbb{C}}x_i\oplus {\mathbb{C}}$$ for each $i$, and therefore $$(\operatorname{Ran}T)^\bot\cap\operatorname{Ker}T= \{0\}\oplus {\mathcal{R}}^\bot\oplus \{0\}.$$ Applying Lemma \[GR:L1\] we find that $T$ cannot be extremal if ${\mathcal{R}}^\bot\neq \{0\}$. Therefore we limit our attention to the case ${\mathcal{R}}=\ell^2(J)$. With $r=\dim{\mathcal{R}}$, we note that ${\mathcal{R}}$ is finite dimensional and we fix an orthonormal basis $e_1,\dots,e_r\in{\mathcal{R}}$ with the property that ${\overline{e_i}}=e_i$ for each $i$.
Any extension $X\in{\mathfrak{C}}_3$ of $T$ takes the form $$\label{TVE:ExtDef}
X_i={\begin{pmatrix} 0 & 0 & 0 & \phi_i \\ x_i & 0 & 0 & C_i \\ 0 & {\overline{x}}_i^* & 0 & 0 \\ 0 & 0 & 0 & B_i \end{pmatrix}}$$ acting on ${\mathcal{H}}\oplus{\mathcal{M}}$ for some Hilbert space ${\mathcal{M}}$ where $C_i\in{\mathcal{L}}({\mathcal{M}},{\mathcal{R}})$, $B_i\in{\mathcal{L}}({\mathcal{M}})$, and $\phi_i$ is a linear functional on ${\mathcal{M}}$ for $i=1,2,3$. The third entry of the fourth column is 0 because $\|X_i\|\leq 1$ and $\|x_i\|=1$ for each $i$. A second consequence of the inequality $\|X_i\|\leq 1$ is $$C_iC_i^*\leq 1-x_ix_i^*, \quad i=1,2,3.$$ This implies $x_i^*C_i=0$ for each $i$. The condition that $X_iX_j=X_jX_i$ is equivalent to requiring $$\phi_i{\circ}B_j=\phi_j{\circ}B_i, \quad B_iB_j=B_jB_i$$ $$\label{TVE:CC1}
x_i\phi_j+C_iB_j=x_j\phi_i+C_jB_i,$$ $$\label{TVE:CC2}
{\overline{x}}_i^*C_j={\overline{x}}_j^*C_i$$ for $i,j=1,2,3$, where $x_i\phi_j$ denotes the map $h\mapsto \phi_j(h)x_i$. Observe that $x_i^*C_i=0$ implies $C_i^*x_i=0$, and that ${\overline{x}}_i^*C_j={\overline{x}}_j^*C_i$ is equivalent to $C_j^*{\overline{x}}_i=C_i^*{\overline{x}}_j$ for all $i$ and $j$.
Define $h^{(i)}_j=C_i^*e_j$ and write $x_i=\sum_{\ell=1}^r a_{i\ell}e_\ell$ for $i=1,2,3$ and $j=1,2,\dots,r$. Then $C_i^*x_i=0$ ($i=1,2,3$) and become a homogeneous system of linear equations in the vectors $h^{(i)}_\ell$. $$\label{TVE:Systm1}
\begin{array}{cc}
a_{i1}h^{(i)}_1+ \dots + a_{ir}h^{(i)}_r =0 & (i=1,2,3) \\
{\overline{a_{i1}}}h^{(j)}_1+\dots+{\overline{a_{ir}}}h^{(j)}_r = {\overline{a_{j1}}}h^{(i)}_1+\dots+{\overline{a_{jr}}}h^{(i)}_r & (i,j=1,2,3)
\end{array}$$ Let $\Lambda$ denote the $6\times 3r$ scalar matrix representing this linear system; $$\Lambda=\begin{pmatrix}
a_{11} & \cdots & a_{1r} & & & & & & \\
& & & a_{21} & \dots & a_{2r} & & & \\
& & & & & & a_{31} & \dots & a_{3r} \\
{\overline{a_{21}}} & \dots & {\overline{a_{2r}}} & -{\overline{a_{11}}} & \dots & -{\overline{a_{1r}}} & & & \\
{\overline{a_{31}}} & \dots & {\overline{a_{3r}}} & & & & -{\overline{a_{11}}} & \dots & -{\overline{a_{1r}}} \\
& & & {\overline{a_{31}}} & \dots & {\overline{a_{3r}}} & -{\overline{a_{21}}} & \dots & -{\overline{a_{2r}}}
\end{pmatrix}$$ where every non-specified entry is 0.
\[TVE:Lem\] If $\Lambda$ has a non-trivial kernel, then $T$ is not extremal.
Assume that $\Lambda$ has non-trivial kernel and set ${\mathcal{M}}={\mathbb{C}}$, $\phi_i=0$, $B_i=0$ for $i=1,2,3$. Then the commutivity of $X$ is determined entirely by and contractivity entirely by $C_iC_i^*+x_ix_i^*\leq 1$ for each $i$. As $\Lambda$ has a non-trivial kernel and ${\mathcal{M}}={\mathbb{C}}$, there is a non-zero solution to . Thus $C_i=\sum_{j=1}^r e_jh^{(i)*}_j$ is non-zero. Since the kernel of $\Lambda$ is linear, we may assume $\|C_i\|\leq 1$.
The Varopoulos triple $T$ is extremal if and only if $\dim{\mathcal{R}}=2$ and $\ker\Lambda=\{0\}$.
Suppose first that $T$ is extremal. Then Lemma \[TVE:Lem\] implies $\ker\Lambda= \{0\}$. Since $r>2$, or rather $3r>6$, implies that $\Lambda$ has a non-trivial kernel, it follows that $r\leq 2$. In the case that $r=1$, there are $c_i\in{\mathbb{C}}$ with $|c_i|=1$ so that $x_i=c_ix_1$ for $i=1,2,3$. Setting ${\mathcal{M}}={\mathbb{C}}$ and $B_i=0,C_i=0$, $\phi_i=c_i$ in for each $i$ yields a non-trivial extension.
Conversely, suppose $r=2$ and $\ker\Lambda=\{0\}$, in which case $$\label{TVE:Lambdar2}
\Lambda=\begin{pmatrix}
a_{11} & a_{12} & & & & \\
& & a_{21} & a_{22} & & \\
& & & & a_{31} & a_{32} \\
{\overline{a_{21}}} & {\overline{a_{22}}} & -{\overline{a_{11}}} & -{\overline{a_{12}}} & & \\
{\overline{a_{31}}} & {\overline{a_{32}}} & & & -{\overline{a_{11}}} & -{\overline{a_{12}}} \\
& & {\overline{a_{31}}} & {\overline{a_{32}}} & -{\overline{a_{21}}} & -{\overline{a_{22}}}
\end{pmatrix}$$ where every non-specified entry is 0. We have $\det\Lambda\neq 0$ and this easily implies $\dim\operatorname{Span}\{x_1,x_2,x_3\}>1$. Suppose $X\in{\mathfrak{C}}_3$ is an extension of $T$, written as in . Since $\det\Lambda\neq 0$ it follows that has only the trivial solution, so that $C_i^*e_j=h_j^{(i)}=0$ for all $i,j$ and therefore $C_1=C_2=C_3=0$. Then yields $x_i\phi_j=x_j\phi_i$ for each $i,j$. The set $\{x_1,x_2,x_3\}$ does not generate a space of dimension 1, and therefore $\phi_i=0$ for $i=1,2,3$. We conclude that $X$ is trivial and thus $T$ is extremal.
We demonstrate that the condition $\ker\Lambda=\{0\}$ is not automatically satisfied when $r=2$. Using the matrix representation of $\Lambda$ given in , one finds that $\det\Lambda=1$ for the vectors $(1,0)$, $(1/\sqrt{2},1/\sqrt{2})$, $(0,1)$. On the other hand, when the vectors $x_1,x_2,x_3$ are instead $(1,0)$, $(1,0)$, $(0,1)$, one finds $\det\Lambda=0$.
[99]{} J. Agler, *An abstract approach to model theory.* Surveys of some recent results in operator theory, Vol. II, 1-23, Pitman Res. Notes Math. Ser., 192, Longman Sci. Tech., Harlow 1988.
T. Andô, *On a pair of commutative contractions.* Acta Sci. Math. (Szeged) 24 (1963) 88-90.
R. Curto, W. Lee, *Towards a model theory for 2-hyponormal operators*. Integral Equations Operator Theory 44 (2002), no. 3, 290-315.
M. Dritschel, S. McCullough, *Model theory for hyponormal contractions.* Integral Equations Operator Theory 36 (2000), no. 2, 182-192.
M. Dritschel, S. McCullough, *Boundary representations for families of representations of operator algebras and spaces.* J. Operator Theory 53 (2005), no. 1, 159-167.
M. Dritschel, S. McCullough, H. Woerdeman, *Model theory for $\rho$-contractions, $\rho\leq 2$.* J. Operator Theory 41 (1999), no. 2, 321-350.
B. A. Lotto, T. Steger, *von Neumann’s inequality for commuting, diagonalizable contractions. II.* Proc. Amer. Math. Soc. 120 (1994), no. 3, 897-901.
S. Parrott, *Unitary dilations of commuting contractions.* Pacific J. Math. 34 (1970) 481-490.
G. Pisier, *Similarity problems and completely bounded maps.* Lecture Notes in Mathematics, 1618. Springer-Verlag, Berlin, 1996.
S. Richter, C. Sundberg, *Joint extensions in families of contractive commuting operator tuples.* J. Funct. Anal. 258 (2010), no. 10, 3319-3346.
B. Sz.-Nagy, C. Foias, H. Bercovici, L. Kérchy, *Harmonic analysis of operators on Hilbert space.* 2nd ed. Revised and enlarged ed. Universitext. Spring, New York 2010.
| {
"pile_set_name": "ArXiv"
} |
epsf
=1200 =0 =5.6in =7.5in =0.2in = -0.2in
=cmb10 scaled 1440
barra[[O]{} -3.7pt [l]{} ]{}
barra[1 -1.1pt [l]{}]{}
= cmti10 at 4pt
Analytical Results for the Grand Canonical
Partition Function for Unidimensional Hubbard Model,
Up to Order $\beta^5$.
0.3cm
I.C. Charret[^1][E–mail: [email protected]]{},
*Departamento de Física*
*Instituto de Ciências Exatas*
*Universidade Federal de Minas Gerais*
*Campus da Pampulha*
*Belo Horizonte, M.G., 31270–901 $\quad$ BRAZIL*
0.2cm
E.V. Corrêa Silva[^2][E–mail: [email protected] ]{}
*Centro Brasileiro de Pesquisas Físicas*
*R. Dr. Xavier Sigaud n.$\!\!^{\rm o}$ 150*
*Rio de Janeiro, R.J., 22290-180 $\quad$ BRAZIL*
0.2cm
S.M. de Souza[^3][E–mail: [email protected]]{},
*Departamento de Ciências Exatas*
*Universidade Federal de Lavras*
*C.P.: 37*
*Lavras, M.G., 37200–000 $\quad$ BRAZIL*
0.2cm
M.T. Thomaz[^4][Corresponding author: Dr. Maria Teresa Thomaz; R. Domingos Sávio Nogueira Saad n.$\!\!^{\rm o}$ 120 apto 404, Niterói, R.J., 24210–340, BRAZIL –Phone/Fax: (21) 620–6735; E–mail: [email protected] ]{}
*Instituto de Física*
*Universidade Federal Fluminense*
*Av. Gal. Milton Tavares de Souza s/n.$\!\!^o$*
*Campus da Praia Vermelha*
*Niterói, R.J., 24210–310 $\quad$ BRAZIL*
0.3cm
**Abstract**
0.2cm
We calculate the exact analytical coefficients of the $\beta$–expansion of the grand canonical partition function of the unidimensional Hubbard model up to order $\beta^5$, using an alternative method, based on properties of the Grassmann algebra. The results derived are non–perturbative and no restrictions on the set of parameters that characterize the model are required. By applying this method we obtain analytical results for the themodynamical quantities, in the high–temperature limits, for arbitrary density of electrons in the unidimensional chain.
[**1. Introduction**]{}
Working with fermionic variables seems discouraging, for their non–commuting nature, in a certain way, compells one to a higher degree of care than that required by commutative variables. Nevertheless, Grassmann algebra properties justify their use in many circumstances \[1\]. With that in mind, we have recently described an alternative method to calculate the terms of the $\beta$– expansion of the grand canonical partition function of periodic unidimensional self–interacting fermionic models, in which Grassmann algebra properties play a central role. No auxiliary fields are needed, and we express our results in terms of matrices with commuting elements. The general approach for a periodic unidimensional fermionic model has been applied to the unidimensional Hubbard model up to order $\beta^3$ \[2\]. An important point about the method developed in reference \[2\] is that even though the unidimensional Hubbard model has exact solutions \[3\], the analytical expressions are only known in the half– filled case. This drawback hinders the analytical evaluation of the partition function for the model from the knowledge of its energy spectrum when we are not considering the half–filling case. Takahashi\[4\] derived a closed expression for the grand canonical partition function of the unidimensional Hubbard model, but besides requiring the formulation of some additional hypotesis, a simple closed expression for the grand canonical partition function is only obtained in the strong coupling limit. The literature offers many examples of high temperature expansions of the grand canonical partition function for the Hubbard model in different space dimensions, some of them up to order $\beta^{9}$ \[5\]. However, all these works referer to either some approximation in which one of the characteristic constants of the model has to be much bigger than the other, or some consideration based on important numerical analysis.
Our results are [*analytical*]{} and do not rest upon any additional hypothesis on the constants that characterize the model. We should point out that we do [*not*]{} perform a perturbative expansion valid in the high temperature limit \[5\], but a $\beta$–expansion of the grand canonical partition function \[6\] where we calculate the exact analytical coefficients of the terms up to order $\beta^5$ for any density of electrons in the unidimensional chain.
In the present paper, we will apply the approach developed in reference \[2\] to get the exact terms at orders $\beta^4$ and $\beta^5$ of the grand canonical partition function of the unidimensional Hubbard model. In section 2 we present a review of the results of reference \[2\]. In section 3, we write down the unidimensional Hubbard model and the Grassmann functions necessary in the calculations that follow. In section 4 we obtain the coefficients at orders $\beta^4$ and $\beta^5$ of the $\beta$–expansion of the grand canonical partition function of the unidimensional Hubbard model. We use the grand potencial derived up to order $\beta^4$ to calculate some physical quantities. A certain property of the multivariable Grassmann integrals — namely, its factorization — opens the way to extending the method of reference \[2\] to orders higher than $\beta^3$. This factorization property is presented in Appendix A through one example. In Appendix B we introduce a graphical notation that greatly simplifies the calculations. In Appendix C, we present a table of the necessary multivariable Grassmann integrals. Section 5 is dedicated to our conclusions.
[**2. Review of Previous Results \[2\]**]{}
The grand canonical partition function of any system is given by:
$${\cal Z}(\beta; \mu) = {\rm Tr}( e^{- \beta {\bf K}}), \eqno(2.1)$$
where $\beta = {1 \over kT}$, [*k*]{} is the Boltzmann constant, [*T*]{} is the absolute temperature, and
$${\bf K} = {\bf H} - \mu {\bf N}, \eqno (2.1a)$$
[**H**]{} is the hamiltonian of the system, $\mu$ is the chemical potential and [**N**]{} is the total number of particles operator.
In the high temperature limit, $\beta \ll 1$, ${\cal Z}(\beta; \mu)$ has the expansion
$${\cal Z}(\beta, \mu) = {\rm Tr}[ 1\!\hskip -1pt{\rm I} - \beta {\bf K}] +
\;\; \sum_{ n=2}^{\infty} {(-\beta)^n \over n!}\;\; {\rm Tr}[ {\bf K}^n],
\eqno(2.2)$$
that we call the $\beta$–expansion of the grand canonical partition function.
For any self–interacting fermionic quantum system, in reference \[2\] we used the Grassmann algebra to show that
-0.7cm
$$\eqalignno{ {\rm Tr}[{\bf K}^n] &= \int \prod_{I=1}^{2nN} \, d\eta_I
d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{2nN} \bar{\eta}_I\; A_{I
J} \; \eta_J} \times & \cr
%
%segunda linha
%
& \hskip -1cm \times
{\cal K}^{\normal} (\bar{\eta}, \eta; \nu=0)\; {\cal K}^{\normal}
(\bar{\eta}, \eta; \nu=1) \; \cdots \; {\cal K}^{\normal} (\bar{\eta},
\eta; \nu=n-1), & (2.3) }$$
where the matrix [**A**]{} is given by,
$${\bf A} = \pmatrix { {\bf A}^{\uparrow \uparrow} & \0barra \cr
& & \cr
\0barra & {\bf A}^{\downarrow \downarrow} \cr }, \eqno (2.3a)$$
and
$${\bf A}^{\uparrow \uparrow} = {\bf A}^{\downarrow \downarrow} =
\pmatrix { \1barra_{N\times N} & - \1barra_{N\times N} &
\0barra_{N\times N} & \cdots & \0barra_{N\times N} \cr & & & & \cr
%
%segunda linha da matriz
%
\0barra_{N\times N} & \1barra_{N\times N}&
- \1barra_{N\times N} & \cdots & \0barra_{N\times N}\cr
%
%terceira linha da matriz
%
& & & & \cr \vdots & & & & \vdots \cr
%
% ultima linha da matriz
%
\1barra_{N\times N} & \0barra_{N\times N} &
\0barra_{N\times N} & \cdots & \1barra_{N\times N} \cr}. \eqno(2.3b)$$
Each matrix ${\bf A}^{\sigma \sigma}$ has dimension $ nN \times nN$, $\1barra_{N\times N}$ been the identity matrix in dimension $N\times N$ and $\0barra_{N\times N}$ the null matrix in this dimension. [*N*]{} is the number of space sites and [*n*]{} is the power of the $\beta$ term. The non–null elements of ${\bf A}^{\sigma \sigma}$ , $\sigma= \uparrow$ and $\sigma= \downarrow$, are
$$A_{I J}^{\sigma \sigma} = \cases{ a_{II} = 1, & $I= 1, 2, \cdots,
nN $ \cr & \cr a_{I, I+N} = -1, & $I= 1, 2, \cdots, (n-2)N $ \cr & \cr
a_{(n-1)N +I, I} = 1, & $ I= 1, 2, \cdots, N. $ \cr} \eqno(2.3c)$$
0.5cm
The Grassmann function ${\cal K}^{\normal} (\bar{\eta}, \eta)$ is the kernel of the fermionic operator [**K**]{} in the normal order \[7\]. In writing down eq.(2.3), we have used a particular mapping for the Grassmann generators \[2\], that greatly simplify our calculations:
-0.5cm
$$\eqalignno{ \eta_\uparrow (x_l, \tau_\nu) &\equiv \eta_{\nu N + l} &
(2.4a) \cr
%
\noalign {\hbox{ and}}
%
\eta_\downarrow (x_l, \tau_\nu)
&\equiv \eta_{(n+\nu) N + l}, & (2.4b) \cr }$$
where $ l= 1, 2, \cdots, N$, and $ \nu= 0, 1, \cdots, n-1$. The mappings (2.4a–b) can be summarized as \[8\]:
$$\eta_\sigma (x_l, \tau_\nu) \equiv \eta_{[{(1-\sigma)\over 2} n +
\nu]N + l}\hskip 4pt. \eqno (2.4c)$$
The generators $\bar{\eta}_\sigma (x_l, \tau_\nu)$ have an equivalent mapping.
In reference \[9\] we showed that the Grassmann integrals (2.3) can be written as co–factors of the matrix ${\bf A}^{\sigma \sigma}$. As we have stated before, the calculation of the co–factors of matrix ${\bf A}^{\sigma \sigma}$ gets simpler when we diagonalize it through a similarity transformation
$${\bf P}^{-1} {\bf A}^{\sigma\sigma} {\bf P} = {\bf D}, \eqno(2.5)$$
where the matrix [**D**]{} is,
$${\bf D} = \pmatrix{ \lambda_1 \1barra_{N\times N} &
\0barra_{N\times N} & \cdots & \0barra_{N\times N} \cr
%
%
\0barra_{N\times N} & \lambda_2 \1barra_{N\times N} & \cdots &
\0barra_{N\times N} \cr
%
%
\vdots & & & \vdots\cr
%
%
\0barra_{N\times N} & \0barra_{N\times N} & \cdots & \lambda_n
\1barra_{N\times N} \cr }, \eqno(2.5a)$$
$\lambda_i$, $i= 1, 2, \cdots, n$, are the eigenvalues of matrices ${\bf A}^{\sigma\sigma}$, $\sigma= \uparrow, \downarrow$. At the same time, we transform of the anti–commuting variables,
$$\eta^\prime = {\bf P}^{-1} \, \eta \hskip 1cm {\rm and} \hskip 1cm
\bar{\eta}^\prime = \bar{\eta} \,{\bf P}. \eqno(2.5b)$$
The matrices [**P**]{} and ${\bf P}^{-1}$ have a block–structure, and in reference \[2\] we got the elements of matrix [**P**]{} and its inverse for the particular cases $n=2$ and $n=3$, besides the eigenvalues of matrices ${\bf A}^{\sigma \sigma}$, $\sigma= \uparrow, \downarrow$. However, for any value of [*n*]{}, we have that
$$\eqalignno{ p_{\nu \nu^\prime}^{(n)} &= {1 \over \sqrt{n}}
e^{{i \pi \over n} (2\nu^\prime + 1)(\nu +1)}, & (2.6a) \cr
%
%
\noalign{\hbox{and}}
%
%
q_{\nu^\prime \nu}^{(n)} &= {1 \over \sqrt{n}}
e^{-{i \pi \over n} (2\nu^\prime + 1)(\nu +1)}, & (2.6b) \cr
}$$
with $\nu, \; \nu^\prime = 0, 1, \cdots, n-1$, and
$${\bf P} = \pmatrix{ p_{0 0}^{(n)}\1barra_{N\times N} &
\cdots & p_{0, n-1}^{(n)} \1barra_{N\times N} \cr
%
%
\vdots & & \vdots\cr
%
%
p_{n-1, 0}^{(n)}\1barra_{N\times N} & \cdots & p_{n-1, n-1}^{(n)}
\1barra_{N\times N} \cr } \eqno(2.6c)$$
and
$${\bf P}^{-1} = \pmatrix{ q_{0 0}^{(n)}\1barra_{N\times N} &
\cdots & q_{0, n-1}^{(n)} \1barra_{N\times N} \cr
%
%
\vdots & & \vdots\cr
%
%
q_{n-1, 0}^{(n)}\1barra_{N\times N} & \cdots & q_{n-1, n-1}^{(n)}
\1barra_{N\times N} \cr }. \eqno(2.6d)$$
0.5cm
The diagonal elements of matrix [**D**]{} are:
-0.5cm
$$\lambda_\nu^{(n)} = 1 - e^{{i\pi \over n} (2\nu +1)},
\hskip 1cm \nu = 0, 1, \cdots, n-1. \eqno(2.6e)$$
where the eigenvalues are [*N*]{}–fold degenerated, $N$ being the number of space sites. Due to space translation symmetry, we should note that the elements $p_{\nu \nu^\prime}^{(n)}$ and $q_{\nu \nu^\prime}^{(n)}$ do not carry any space site index.
Once the matrix [**A**]{} has a block–structure as depicted in eq.(2.3), the integrals (2.3) are equal to the product of the integral in the sector $\sigma\sigma = \uparrow\uparrow$ times the integral in the sector $\sigma\sigma = \downarrow\downarrow$. The Grassmann integrals in the sector $\uparrow\uparrow$ have the form:
$$M(L,K) = \int \prod_{i = 1}^{\rm nN} d\eta_i d\bar\eta_i
%
\;
\bar\eta_{l_1} \eta_{k_1} \cdots \bar\eta_{l_m} \eta_{k_m} \hskip 3pt
%
e^{ \sum\limits_{ i, j = 1}^{\rm nN} \bar\eta_i
A_{i j}^{\uparrow\uparrow} \eta_j},
\eqno(2.7)$$
with $ L= \{ l_1, \cdots, l_m\}$ and $K= \{k_1, \cdots,
k_m\}$. The products $\bar{\eta}\eta$ are ordered in such a way that $ l_1< l_2< \cdots< l_m$ and $ k_1< k_2< \cdots< k_m$. From reference \[9\], the result of this type of integrals is equal to:
$$M(L,K) = (-1)^{(l_1 + l_2+ \cdots+ \l_m) + (k_1+ k_2+ \cdots+ k_m)}
A(L,K), \eqno(2.7a)$$
where $A(L,K)$ is the determinant of the matrix obtained from matrix [**A**]{} by deleting the lines $\{ l_1, \cdots, l_m\}$ and the columns: $\{k_1, \cdots, k_m\}$. The Grassmann integrals to be calculated in sector $\downarrow\downarrow$ are of the same type as that of eq.(2.7).
After the similarity transformation (2.5) and the change of variables (2.5b), in a schematic way, the integral (2.7) becomes
$$\eqalignno{ M(L,K) &= \int \prod_{i = 1}^{\rm nN} d\eta_i
d\bar\eta_i
%
\; (\bar\eta {\bf P}^{-1})_{l_1} ({\bf P}\eta)_{k_1}
\cdots (\bar\eta {\bf P}^{-1})_{l_m} ({\bf P} \eta)_{k_m}
\hskip 6pt
%
e^{ \sum\limits_{ i, j = 1}^{\rm nN} \bar\eta_i D_{i
j} \eta_j}, & \cr
%
% proxima linha
%
& & (2.8) \cr }$$
where [**D**]{} is a diagonal matrix whose entries are given by eq.(2.6e).
We should point out that the relations (2.6a)–(2.6e) are valid for any unidimensional self–interacting fermionic model with space translation symmetry.
We have a large number of integrals that contribute to eq.(2.3). For a discussion on some useful symmetries and their use in the reduction of the number of contributing integrals, the reader is referred to \[2\].
[**3. Unidimensional Hubbard Model**]{}
The hamiltonian that describes the Hubbard model in one space dimension is \[10\]:
$${\bf H} = \sum_{ {i=1}\atop{j=1}}^{\rm N} \sum_{\sigma= -1, 1}
t_{ij} {\bf a}_{i\sigma}^{\dagger}{\bf a}_{j\sigma} + U \sum_
{i=1}^{\rm N} {\bf a}_{i\uparrow}^\dagger{\bf a}_{i\uparrow} {\bf
a}_{i\downarrow}^\dagger{\bf a}_{i\downarrow} + \lambda_B \sum_
{i=1}^{\rm N} \sum_{\sigma=-1, 1} \sigma {\bf a}_{i\sigma}^\dagger{\bf
a}_{i\sigma} \eqno (3.1)$$
where ${\bf a}_{i\sigma}^\dagger$ is the creation operator of an electron in site [*i*]{} with spin $\sigma$ and ${\bf a}_{i\sigma}$ is the destruction operator of an electron in site [*i*]{} with spin $\sigma$. The first term on the r.h.s. of eq.(3.1) is the kinetic energy operator. All diagonal elements of $t_{ij}$ are equal, $t_{ii} = {\rm E_0}$, the only non–null off–diagonal terms are $t_{i, i-1}=t_{i, i+1}= t$, where $ i=1, 2, \dots , {\rm N}$, and they contribute to the hopping term. [*U*]{} is the strength of the interaction between the electrons in the same site but with different spins. We have defined $\lambda_B= -{1\over 2} g\mu_B B$, where [*g*]{} is the Landé’s factor, $\mu_B$ is the Bohr’s magneton and $B$ is the constant external magnetic field in the $\hat{z}$ direction.
The periodic boundary condition in space is implemented by imposing that ${\bf a}_{0 \sigma} \equiv {\bf a}_{N \sigma}$ and ${\bf a}_{N+1,
\sigma} \equiv {\bf a}_{1 \sigma}$. Therefore, the hopping terms $t_{1
0} {\bf a}_{1 \sigma}^\dagger {\bf a}_{0 \sigma}$ and $t_{N, N+1} {\bf
a}_{N \sigma}^\dagger {\bf a}_{N+1, \sigma}$ become $t_{1 N} {\bf
a}_{1 \sigma}^\dagger {\bf a}_{N\sigma}$ and $t_{N, 1} {\bf a}_{N
\sigma}^\dagger {\bf a}_{1 \sigma}$ respectively. We point out that the hamiltonian (3.1) is already in normal order.
The kernel of the operator [**K**]{} (eq.(2.1a)) for the unidimensional Hubbard model on a lattice with [*N*]{} space sites, written in terms of the generators $\bar{\eta}_I$ and $\eta_J$, is equal to
$$\eqalignno{ {\cal K}^{\normal} &(\bar{\eta},\eta;\nu)) =
\sum_{l=1}^N \sum_{\sigma =\pm 1} (E_0 + \sigma \lambda_B - \mu) \;
\bar{\eta}_{[{(1-\sigma)\over 2}n +\nu]N + l}\;
\eta_{[{(1-\sigma)\over 2}n +\nu]N + l} + & \cr
%
% segunda linha da expressao
%
& + \sum_{l=1}^N \sum_{\sigma =\pm 1} t [
\bar{\eta}_{[{(1-\sigma)\over 2}n +\nu]N + l}\;
\eta_{[{(1-\sigma)\over 2}n +\nu]N + l+1}+
\bar{\eta}_{[{(1-\sigma)\over 2}n +\nu]N + l}\;
\eta_{[{(1-\sigma)\over 2}n +\nu]N + l-1}] + & \cr
%
%terceira linha da expressao
%
& + \sum_{l=1}^N U \bar{\eta}_{(n+\nu)N + l}\;\;
\eta_{(n+\nu)N + l}\;\; \bar{\eta}_{\nu N + l} \; \; \eta_{\nu N + l},
& (3.2a) \cr }$$
with the periodic spatial boundary condition:
-0.7cm
$$\eqalignno{ t\; \bar{\eta}_{[{(1-\sigma)\over 2} n + \nu]N + N}
\; \eta_{[{(1-\sigma)\over 2} n + \nu]N + N +1}
& \equiv t \; \bar{\eta}_{[{(1-\sigma)\over 2} n +\nu]N + N} \;
\eta_{[{(1-\sigma)\over 2} n + \nu]N + 1} & (3.2b) \cr
%
%
\noalign{\hbox{and}}
%
%
t \; \bar{\eta} _{[{(1-\sigma)\over 2} n + \nu]N + 1} \;
\eta_{[{(1-\sigma)\over 2} n + \nu]N } & \equiv
t \; \bar{\eta}_{[{(1-\sigma)\over 2} n +\nu]N + 1} \;
\eta_{[{(1-\sigma)\over 2} n +\nu]N + N}\;\; , &
(3.2c) \cr }$$
and the anti–periodic boundary condition in $\nu$:
$$\eta_{[{(1-\sigma)\over 2} n + n]N + l} =
- \eta_{[{(1-\sigma)\over 2} n ]N + l}\;\;, \eqno
(3.2d)$$
for $l= 1, 2, \cdots, N$, and $\sigma= \uparrow,
\downarrow$. We have used the mapping (2.4c) to write the previous expressions.
In order to write down the terms that contribute to $Tr[{\bf K}^4]$ and $Tr[{\bf K}^5]$ in a simplified way , we define:
$$\eqalignno{ {\cal E} (\bar{\eta}, \eta; \nu; \sigma) &\equiv
\sum\limits_{l=1}^N \bar{\eta}_{[{(1-\sigma)\over 2}n +\nu]N + l}\;
\eta_{[{(1-\sigma)\over 2}n +\nu]N + l}\;\;; & (3.3a) \cr
%
%segunda linha
%
{\cal T}^{-} (\bar{\eta}, \eta; \nu; \sigma) & \equiv
\sum\limits_{l=1}^N \bar{\eta}_{[{(1-\sigma)\over 2}n +\nu]N + l}\;
\eta_{[{(1-\sigma)\over 2}n +\nu]N + l+1} \;\; ; & (3.3b)\cr
%
\noalign{\hbox{and}}
%
%terceira linha
%
{\cal T}^{+} (\bar{\eta}, \eta; \nu; \sigma) &\equiv
\sum\limits_{l=1}^N \bar{\eta}_{[{(1-\sigma)\over 2}n +\nu]N + l}\;
\eta_{[{(1-\sigma)\over 2}n +\nu]N + l-1}\;\; . & (3.3c)\cr
}$$
We also define
$$\eqalignno{ {\cal E} (\bar{\eta}, \eta; \nu) &\equiv
\sum\limits_{\sigma=\pm 1} E(\sigma) {\cal E}(\bar{\eta}, \eta; \nu;
\sigma), & (3.4a) \cr
%
%segunda linha
%
{\cal T}^{-} (\bar{\eta},
\eta; \nu) &\equiv \sum\limits_{\sigma = \pm 1} t\, {\cal
T}^{-}(\bar{\eta}, \eta; \nu; \sigma), & (3.4b) \cr
%
%terceira linha
%
{\cal T}^{+} (\bar{\eta},
\eta; \nu) &\equiv \sum\limits_{\sigma =\pm 1} t \,{\cal
T}^{+}(\bar{\eta}, \eta; \nu; \sigma), & (3.4c) \cr
%
\noalign{\hbox{and}}
%
%quarta linha
%
{\cal U}(\bar{\eta}, \eta; \nu)
&\equiv \sum\limits_{l=1}^N \bar{\eta}_{(n+\nu)N + l}\;\;
\eta_{(n+\nu)N + l}\;\; \bar{\eta}_{\nu N + l} \; \; \eta_{\nu N + l},
& (3.4d) \cr }$$
where $E(\sigma) \equiv E_0 - \sigma\lambda_B - \mu$. The term ${\cal E} (\bar{\eta}, \eta; \nu)$ represents the diagonal part of the kinetic energy , ${\cal T}^{-} (\bar{\eta}, \eta; \nu)$ and $ {\cal T}^{+}
(\bar{\eta}, \eta; \nu)$ are the hopping terms and ${\cal U}
(\bar{\eta}, \eta; \nu)$ is the fermionic interaction term.
For the unidimensional Hubbard model, the Grassmannian function ${\cal K}^{\normal} (\bar{\eta},\eta;\nu)$ is written as
$${\cal K}^{\normal} (\bar{\eta}, \eta; \nu)= {\cal E} (\bar{\eta},
\eta; \nu)+ {\cal T}^{-} (\bar{\eta}, \eta; \nu)+ {\cal T}^{+}
(\bar{\eta}, \eta; \nu) + {\cal U} (\bar{\eta}, \eta; \nu).
\eqno(3.5)$$
[**4. The Exact Coefficients of the $\beta$–Expansion of the Grand Canonical Partition Function for the Unidimensional Hubbard Model**]{}
In eq.(2.2) we have the $\beta$–expansion of the grand canonical partition function for any quantum system. For the particular case of self–interacting unidimensional fermionic models, the expression of $Tr[{\bf K}^n]$ is given by eq.(2.3).
In reference \[2\], we calculated the exact coefficients of the terms $\beta^2$ and $\beta^3$ of the expression (2.2) for the unidimensional Hubbard model for arbitrary values of the constants $E_0$, $t$, $U$ and $\mu$, that characterize the model, and for any value of the constant external magnetic field $B$.
The evaluation of integrals has been performed by a number of procedures (computer programs) developed by the authors in the symbolic language Maple V.3, that consist in the computational implementation of the method described in reference \[2\]. We have called this package of procedures [GINT]{}.
The method applied in \[2\] greatly simplifies the calculations made in reference \[11\], but memory utilization problems have appeared as we tried to go beyond $n>3$. Luckily, the factorization property of multivariable integrals of type (2.8) allowed us to optimize the performance of the package. In Appendix A we consider one typical Grassmann integral to exemplify the factorization of the sub–graphs.
The procedure [perm ]{} is one of the procedures contained in the package, being a useful tool to calculate the independent non–null terms \[12\] that contribute to $Tr[{\bf K}^4]$ and $Tr[{\bf K}^5]$. In this procedure are implemented the symmetries discussed in reference \[2\]. The procedure [gint]{} has been used to calculate the multivariable Grassmann integrals, taking into account the factorization into sub–graphs. The package can be downloaded through [*ftp*]{} from the site [*http:/www.if.uff.br*]{}.
[**4.1. Calculation of Tr\[${\bf K}^4$\]**]{}
For the case $n=4$, we get from eq.(2.3) that
-0.7cm
$$\eqalignno{ {\rm Tr}[{\bf K}^4] &= \int \prod_{I=1}^{8N} \, d\eta_I
d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{8N} \bar{\eta}_I\; A_{I
J} \; \eta_J} \times & \cr
%
%segunda linha
%
& \hskip -1cm \times
{\cal K}^{\normal} (\bar{\eta}, \eta; \nu=0)\; {\cal K}^{\normal}
(\bar{\eta}, \eta; \nu=1) \;
{\cal K}^{\normal} (\bar{\eta}, \eta; \nu=2)\;
{\cal K}^{\normal} (\bar{\eta}, \eta; \nu=3). & (4.1.1)
}$$
The expressions of ${\bf A}^{\sigma \sigma}$, $p_{\nu \nu^\prime}^{(4)}$, $q_{\nu \nu^\prime}^{(4)}$ and $\lambda_\nu^{(4)}$ are obtained from eqs.(2.3b), (2.6a), (2.6b) and (2.6e).
From eq.(3.5), for the unidimensional Hubbard model, we have that the Grassmann function ${\cal K}^{\normal} (\bar{\eta}, \eta; \nu)$ is equal to
$${\cal K}^{\normal} (\bar{\eta}, \eta; \nu)= {\cal E} (\bar{\eta},
\eta; \nu)+ {\cal T}^{-} (\bar{\eta}, \eta; \nu)+ {\cal T}^{+}
(\bar{\eta}, \eta; \nu) + {\cal U} (\bar{\eta}, \eta; \nu).
\eqno(4.1.1a)$$
We defined a simplified notation,
$$\eqalignno{ < {\cal O}_1 (\nu_1) \cdots {\cal O}_m (\nu_m) > &
\equiv \int \prod_{I=1}^{2nN} \, d\eta_I d\bar{\eta}_I \;\;
e^{\sum\limits_{I,J= 1}^{2nN} \bar{\eta}_I\; A_{I J} \; \eta_J} \;\;
\times & \cr
%
& \hskip 1cm \times \;\; {\cal O}_1
(\bar{\eta},\eta;\nu_1) \cdots {\cal O}_m (\bar{\eta},\eta;\nu_m) &
(4.1.2a) \cr
%
\noalign{\hbox{and}}
%
%segunda linha
%
< {\cal O}_1 (\sigma, \nu_1) \cdots {\cal O}_m (\sigma,\nu_m) >
&\equiv \int
\prod_{I=(1-\sigma)nN+1}^{(3-\sigma)nN}\, d\eta_I d\bar{\eta}_I \;\;
e^{\sum\limits_{I,J= 1}^{2nN} \bar{\eta}_{(1-\sigma)nN+I} \; A_{I J}
\; \eta_{(1-\sigma)nN+J}} \;\; \times & \cr
%
& \hskip 1cm \times
{\cal O}_1 (\bar{\eta},\eta;\nu_1) \cdots {\cal O}_m
(\bar{\eta},\eta;\nu_m), & (4.1.2b) \cr }$$
where ${\cal O}_j (\bar{\eta},\eta;\nu_j)$ are Grassmann functions.
The independent terms that contribute to $Tr[{\bf K}^4]$ are:
-0.7cm
$$\eqalignno{
%primeira linha
%
Tr[{\bf K}^4] &= <{\cal E}_0, {\cal E}_0, {\cal E}_0,{\cal E}_0> +
4 <{\cal U}, {\cal E}_0, {\cal E}_0, {\cal E}_0> +
2 <{\cal U}, {\cal E}_0, {\cal U}, {\cal E}_0> +
8 < {\cal U},{\cal T}^-,{\cal T}^+, {\cal E}_0> + & \cr
%
%segunda linha
%
& + 4 <{\cal U}, {\cal U}, {\cal E}_0, {\cal E}_0> +
4 < {\cal U}, {\cal U}, {\cal U}, {\cal E}_0> +
< {\cal U}, {\cal U}, {\cal U}, {\cal U}> +
4 < {\cal T}^-, {\cal E}_0, {\cal T}^+, {\cal E}_0> + & \cr
%
%terceira linha
%
& + 8 < {\cal T}^-, {\cal U}, {\cal T}^+, {\cal E}_0> +
4 < {\cal T}^-, {\cal U}, {\cal T}^+, {\cal U}> +
8 < {\cal T}^-, {\cal T}^+, {\cal E}_0, {\cal E}_0> + & \cr
%
%quarta linha
%
& + 8 < {\cal T}^-, {\cal T}^+, {\cal U}, {\cal E}_0> +
8 < {\cal T}^-, {\cal T}^+, {\cal U}, {\cal U}> +
2 < {\cal T}^-, {\cal T}^+, {\cal T}^-, {\cal T}^+> + & \cr
%
%quinta linha
%
& + 4 <{\cal T}^-, {\cal T}^-, {\cal T}^+, {\cal T}^+>.
& (4.1.3) \cr
}$$
In order to calculate the terms on the r.h.s. of (4.1.3), we need the result of a set of Grassmann multivariable integrals that are presented on Appendix C. We used the procedure [gint]{} to calculate them. Before using the procedure [gint]{} to calculate the terms in eq.(4.1.3) that include ${\cal E}_0$, $ {\cal T}^-$ and ${\cal T}^+$, we need to explicit the contributions coming from the sectors $\sigma= \uparrow$ and $\sigma=\downarrow$. For example, in the term $< {\cal E}(0), {\cal E}(1), {\cal E}(2), {\cal E}(3) > $ we have 16 terms when we explicitly write down the spin–sectors. However, the number of terms is diminished when we use the symmetries discussed in reference \[2\] and the fact that ${\bf A}^{\uparrow \uparrow}$ = ${\bf A}^{\downarrow \downarrow}$. The application of these symmetries is simplified by using the graphic notation explained in Appendix B.
0.5cm
By taking into account the results of integrals in Appendix C and their contributions to the sum over the space indices in each term of $Tr[{\bf K}^4]$ (eq.(4.1.3)), we finally obtain
-0.3cm
$$\eqalignno{
%linha 1
Tr[{\bf K}^4] &= N^4 \biggl( U \Delta E^3 + {1\over 256} U^4 +
{3\over 8} U^2 \Delta E^2 + \Delta E^4 + {1\over 16}U^3 \Delta E \biggr) +
N^3 \biggl( 3 \Delta E^2 \lambda_B^2 + 3 \Delta E^4 +
& \cr
% linha 2
& \hskip 2cm + 3 t^2 U \Delta E+
{9 \over 2} U \Delta E^3 + {3 \over 16} U^2 \lambda_B^2 + 6t^2 \Delta E^2 +
{9 \over 128} U^4 + {45 \over 16} U^2 \Delta E^2 + & \cr
%linha 3
& + {3 \over 8} t^2 U^2 + {3\over 4} U^3 \Delta E +
{3 \over 2} U \Delta E \lambda_B^2 \biggr) +
N^2 \biggl( {3\over 4} \Delta E^4 + {3\over 16} U^2 \lambda_B^2 +
3 U \Delta E^3 + {51 \over 16} U^2 \Delta E^2 + & \cr
% linha 4
& + 3 t^2 U \Delta E + {3\over 2} \Delta E^2 \lambda_B^2 +
3 t^2 \lambda_B^2 + {3 \over 4} \lambda_B^4+
{9\over 8} t^2 U^2 + 3 t^4 + 3 t^2 \Delta E^2 + {21\over 16} U^3 \Delta E
+ {51 \over 256} U^4 \biggr) + & \cr
% linha 5
& - N \biggl( 3 t^2 \lambda_B^2 + {3\over 2} t^4 + 3 t^2 \Delta E^2 +
{1\over 2} U \Delta E^3 + {5\over 4} t^2 U^2 + {3\over 8} U^2 \lambda_B^2
+ {1\over 4} \Delta E^4 + & \cr
%linha 5
& + 3 t^2 U \Delta E + {3\over 2} \Delta E^2 \lambda_B^2 +
{3\over 8} U^2 \Delta E^2 + {1\over 4} \lambda_B^4 +
{1\over 8} U^3 \Delta E + {3\over 2} U \Delta E \lambda_B^2 +
{3 \over 128} U^4 \biggr). & \cr
%linha 6
& & (4.1.4) \cr
}$$
We use the short notation: $ \Delta E \equiv E_0 - \mu$.
[**4.2. Calculation of Tr\[${\bf K}^5$\]**]{}
For $n=5$, the expression of $Tr[{\bf K}^5]$ obtained from the procedure [perm]{}, where each term is multiplied by the respective constant, is:
-0.7cm
$$\eqalignno{
%linha 1
Tr[{\bf K}^5] &=
5 < {\cal U}, {\cal E}_0, {\cal E}_0, {\cal E}_0, {\cal E}_0> +
5 < {\cal U}, {\cal E}_0, {\cal U}, {\cal E}_0, {\cal E}_0> +
5 < {\cal U}, {\cal U}, {\cal E}_0, {\cal E}_0, {\cal E}_0> + & \cr
%linha 2
& + 5 < {\cal U}, {\cal U}, {\cal E}_0, {\cal U}, {\cal E}_0> +
10 < {\cal U}, {\cal U}, {\cal T}^-, {\cal T}^+, {\cal E}_0> +
5 < {\cal U}, {\cal U} , {\cal U}, {\cal E}_0, {\cal E}_0> + & \cr
%linha 3
& + 5 < {\cal U}, {\cal U}, {\cal U}, {\cal U}, {\cal E}_0> +
< {\cal U}, {\cal U}, {\cal U}, {\cal U}, {\cal U}> +
< {\cal E}_0, {\cal E}_0, {\cal E}_0, {\cal E}_0, {\cal E}_0> + & \cr
%linha 4
& +10 < {\cal T}^-, {\cal T}^+, {\cal T}^-, {\cal T}^+, {\cal E}_0> +
10 < {\cal T}^-, {\cal T}^+, {\cal T}^-, {\cal T}^+, {\cal U}> +
10 < {\cal T}^-, {\cal U}, {\cal U}, {\cal T}^+, {\cal E}_0> + & \cr
%linha 5
& + 10 <{\cal T}^-, {\cal T}^+, {\cal E}_0, {\cal U}, {\cal E}_0> +
10 < {\cal T}^-, {\cal T}^+, {\cal E}_0, {\cal E}_0, {\cal E}_0>+
10 < {\cal T}^-, {\cal T}^+, {\cal U}, {\cal E}_0, {\cal E}_0> + & \cr
%linha 6
& + 10 < {\cal T}^-, {\cal T}^+, {\cal U}, {\cal U}, {\cal E}_0> +
10 < {\cal T}^-, {\cal T}^+, {\cal U}, {\cal U}, {\cal U}> +
10 < {\cal T}^-, {\cal E}_0, {\cal T}^+, {\cal U}, {\cal E}_0> + & \cr
%linha 7
& + 10 < {\cal T}^-, {\cal E}_0, {\cal T}^+, {\cal E}_0, {\cal E}_0> +
10 < {\cal T}^-, {\cal U}, {\cal T}^+, {\cal E}_0, {\cal E}_0> +
10 < {\cal T}^-, {\cal U}, {\cal T}^+, {\cal U}, {\cal E}_0> + & \cr
% linha 8
&+ 10 < {\cal T}^-, {\cal U}, {\cal T}^+, {\cal U}, {\cal U}> +
10 < {\cal T}^+, {\cal T}^-, {\cal T}^-, {\cal T}^+, {\cal E}_0> +
10 < {\cal T}^+, {\cal T}^-, {\cal T}^-, {\cal T}^+, {\cal U}> + & \cr
% linha 9
& + 10 < {\cal U}, {\cal T}^-, {\cal U}, {\cal T}^+, {\cal E}_0> +
10 < {\cal U}, {\cal T}^-, {\cal E}_0, {\cal T}^+, {\cal E}_0> +
10 < {\cal U}, {\cal T}^-, {\cal T}^+, {\cal E}_0, {\cal E}_0> + & \cr
% linha 10
& + 10 < {\cal U}, {\cal T}^-, {\cal T}^+, {\cal U}, {\cal E}_0> +
10 < {\cal T}^-, {\cal T}^-, {\cal T}^+, {\cal T}^+, {\cal E}_0> +
10 < {\cal T}^-, {\cal T}^-,{\cal T}^+, {\cal T}^+, {\cal U}>, & \cr
% linha 11
& & (4.2.1) \cr
}$$
-0.5cm
where we are using the convention (4.1.2a) to write down each term.
In $n=5$ we have seven new type of integrals that do not have an equivalent one for $n<5$. We table those integrals in section C.2 of Appendix C.
After a long but convergent calculation, we get
-0.7cm
$$\eqalignno{
% linha 1
Tr[{\bf K}^5] &= N^5 \biggl( {1 \over 1024} U^5 +
{5\over 256} U^4 \Delta E + {5\over 4} U \Delta E^4 +
{5\over 32} U^3 \Delta E^2 + \Delta E^5 +
{5\over 8} U^2 \Delta E^3 \biggr)
+ & \cr
%linha 2
& + N^4 \biggl( {105\over 16} U^2 \Delta E^3 +
{15\over 16} U^2 \Delta E \lambda_B^2 + {5\over 32} U^3 t^2 +
10 t^2 \Delta E^3 + 5 \Delta E^3 \lambda_B^2 + 5 \Delta E^5 +
{35 \over 4} U \Delta E^4 + & \cr
% linha 3
& + {15 \over 4} U \Delta E^2 \lambda_B^2 +
{15 \over 2} U t^2 \Delta E^2 +
{15 \over 512} U^5 + {5 \over 64} U^3 \lambda_B^2 +
{155\over 64} U^3 \Delta E^2 + {55\over 128} U^4 \Delta E + & \cr
% linha 4
& + {15\over 8} U^2 t^2 \Delta E \biggr) +
N^3 \biggl( 15 t^2 \Delta E^3 +
15 t^2 \Delta E \lambda_B^2 + {195\over 16} U \Delta E^4 +
{45\over 8} U \Delta E ^2 \lambda_B^2 + & \cr
% linha 5
& + {15\over 16} U \lambda_B^4 +
{15 \over 2} \Delta E^3 \lambda_B^2+
{15 \over 4} \Delta E^5 + {15 \over 4} \Delta E \lambda_B^4 +
{15\over 4} U t^4 + {225 \over 16} U^2 \Delta E^3 +
{45\over 16} U^2 \Delta E \lambda_B^2 + & \cr
%linha 6
& + {15\over 4} U t^2 \lambda_B^2 + {75\over 4} U t^2 \Delta E^2 +
15 t^4 \Delta E + {495\over 256} U^4 \Delta E +
{15\over 32} U^3 \lambda_B^2 + {15 \over 2} U^3 \Delta E^2 + & \cr
% linha 7
&+ {75\over 8} U^2 t^2 \Delta E + {195\over 1024} U^5 +
{45\over 32} U^3 t^2 \biggr) +
N^2 \biggl( -{25\over 4} U^2 t^2 \Delta E
+ {5\over 4} U^2 \Delta E^3 - & \cr
%linha 8
& - {15\over 4} U^2 \Delta E \lambda_B^2 + {75\over 512} U^5 -
{15\over 16} U \Delta E^4 - {75\over8} U \Delta E^2 \lambda_B^2 -
{35 \over 16} U \lambda_B^4 - {15\over 2} t^4 \Delta E - & \cr
%linha 9
& - 15 U t^2 \Delta E^2 - {15 \over 2} U t^2 \lambda_B^2 +
{115\over 128} U^4 \Delta E - {15\over 2} \Delta E^3 \lambda_B^2 -
{5\over 4} \Delta E^5 - {5\over 4} \Delta E \lambda_B^4-
15 t^2 \Delta E^3 - & \cr
% linha 10
&- 15 t^2 \Delta E \lambda_B^2 - {45\over 64} U^3 \lambda_B^2 +
{125\over 64} U^3 \Delta E^2 - {15\over 8} U t^4 - {5\over 8} U^3 t^2
\biggr) + & \cr
%linha 11
& + N \biggl( - {5\over 2} U^2 \Delta E^3 -
{15\over 2} U t^2 \Delta E^2 +
{15\over 2} U t^2 \lambda_B^2 - {15 \over 2} U^2 t^2 \Delta E -
{25\over 32} U^4 \Delta E - {65\over 32} U^3 \Delta E^2 + & \cr
% linha 12
& + {5 \over 32} U^3 \lambda_B^2 - {15\over 128} U^5 -
{15\over 8} U^3 t^2
- {5\over 4} U \Delta E^4 + {5\over 4} U \lambda_B^4 \biggr). & \cr
& & (4.2.2) \cr
}$$
We continue to use the notation: $ \Delta E \equiv E_0 - \mu$.
Certainly, the most subtle part in calculating expression (4.2.2) comes from the product of Grassmann integrals for different $\sigma$–sectors when we have Grassmann generators at the same space indice. In this case, we have to suit the conditions satisfied by both integrals and calculate the contribution of the product to the sum over space indices.
[**4.3. The $\beta$–Expansion of the Grand Potential Up to Order $ \beta^4$ and 0.4cm Physical Quantities**]{}
The relation between the grand potential ${\cal W}(\beta; \mu)$ and the grand canonical partition function ${\cal Z}(\beta; \mu)$ is
$${\cal W}(\beta;\mu) =- {1 \over \beta}
\ln {\cal Z}(\beta; \mu). \eqno(4.3.1)$$
The $\beta$–expansion of ${\cal Z}(\beta; \mu)$ up to order $\beta^5$ is (see eq.(2.2)):
-0.5cm
$${\cal Z}(\beta, \mu) \approx
{\rm Tr}[ 1\!\hskip -1pt{\rm I} - \beta {\bf K}] +
{\beta^2 \over 2!} {\rm Tr}[ {\bf K}^2] -
{\beta^3 \over 3!} {\rm Tr}[{\bf K}^3] +
%
%dois ultimos termos da expressao
%
{\beta^4 \over 4!} {\rm Tr}[ {\bf K}^4] -
{\beta^5 \over 5!} {\rm Tr}[{\bf K}^5]. \eqno (4.3.2)$$
The first term on the r.h.s. of (4.3.2) was calculated in reference\[11\], the second and third terms were calculated in reference \[2\] and in its last two terms we substitute the results of eqs.(4.1.4) and (4.2.2).
From eqs.(4.3.1) and (4.3.2), we get the grand potential up to order $\beta^4$, that is,
$$\eqalignno{
%linha 1
{\cal W}(\beta;\mu) & = - N \Bigl\{ {2\over \beta} \ln 2 +
\biggl( - {1\over 16} U t^2 \lambda_B^2 + {1\over 16} U^2 t^2 \Delta E
+ {1\over 1024} U^5 + {1\over 16} U t^2 \Delta E^2 +
{13 \over 768} U^3 \Delta E^2 - & \cr
% linha 2
& - {1 \over 768} U^3 \lambda_B^2 - {1\over 96} U \lambda_B^4 +
{1\over 96} U \Delta E^4 + {5\over 768} U^4 \Delta E +
{1\over 48} U^2 \Delta E^3 + {1\over 64} U^3 t^2 \biggr) \beta^4-
& \cr
% linha 3
& - \biggl( {1\over 8} t^2 U \Delta E + {1\over 16} U \Delta E \lambda_B^2
+ {1\over 96} \Delta E^4 + {1\over 16} t^4 + {1\over 96} \lambda_B^4
+ { 1\over 1024} U^4 + {1\over 48} U \Delta E^3
+ & \cr
% linha 4
& + {1\over 8} t^2 \Delta E^2 +
{1 \over 192} U^3 \Delta E + {5 \over 96} t^2 U^2 +
{1\over 64} U^2 \lambda_B^2 + {1\over 64} U^2 \Delta E^2 +
{1 \over 16} \Delta E^2 \lambda_B^2 + & \cr
% linha 5
& + {1\over 8} t^2 \lambda_B^2 \biggr)\beta^3 +
\biggl( -{U^3\over 64} + {1\over 16} U \lambda_B^2
- {1\over 16} \Delta E^2 U - {1\over 16} \Delta E U^2 \biggr) \beta^2
+ & \cr
%linha 6
& \hskip -0.5cm
+ \biggl( {1\over 4} \Delta E^2 + {1\over 4}
\lambda_B^2 + {1\over 4} \Delta E\; U + {t^2 \over 2} + {3\over 32}
U^2 \biggr) \beta - \big( \Delta E + {U\over 4} \bigr) +
{\cal O}(\beta^5) \Bigr\} . & \cr
& \cr
%linha 7
& & (4.3.3) \cr
}$$
It is important to stress out that the coefficients of the $\beta$–expansion of function ${\cal W}(\beta;\mu)$ are exact for any set of constants: $E_0$, $t$, $U$, $\mu$ and $B$ for the unidimensional Hubbard model. From expression (4.3.3) we can get the strong limit approximation by taking $U \gg t$, as well the atomic limit approximation when $U \ll t$.
From expression (4.3.3), we can derive any physical quantity for the model at thermal equilibrium at high temperature. As examples, we consider the following quantities:
0.5cm
$i)$ specific heat at constant length and constant number of fermions: $C_{L} (\beta)$.
$$C_{L} (\beta) = - k \beta {\partial \over \partial \beta}\biggl[ \beta^2
{\partial {\cal W}(\beta;\mu) \over \partial \beta}\biggr],
\eqno (4.3.4)$$
where $k$ is the Boltzmann constant. From eq.(4.3.3), we get,
-0.5cm
$$\eqalignno{
%linha 1
C_{L} (\beta)& = Nk \biggl\{ \biggl( {5\over 256} U^5 -
{5\over 24} U \lambda_B^4
+ {5 \over 24} U \Delta E^4 + {65\over 192} U^3 \Delta E^2 +
{5\over 12} U^2 \Delta E^3 - {5\over 192} U^3 \lambda_B^2 - & \cr
%linha 2
& - {5\over 4} U t^2 \lambda_B^2 + {5\over 4} U t^2 \Delta E^2 +
{5\over 4} U^2 t^2 \Delta E + {25 \over 192} U^4 \Delta E +
{5 \over 16} U^3 t^2 \biggr) \beta^5 + \biggl( -{3 \over 256} U^4 - & \cr
%linha 3
& - {3\over 2} t^2 \Delta E^2 - {1\over 8} \lambda_B^4 - {3 \over 4} t^4
- {1 \over 8} \Delta E^4 - {5\over 8} t^2 U^2 - {1\over 16} U^3 \Delta E
- {3\over 4} U \Delta E \lambda_B^2 -
{3\over 4} \Delta E^2 \lambda_B^2 - & \cr
%linha 4
& - {3\over 2} t^2 \lambda_B^2 - {3\over 2} t^2 U \Delta E -
{3\over 16} U^2 \lambda_B^2 - {3\over 16} U^2 \Delta E^2 -
{1 \over 4} U \Delta E^3 \biggr) \beta^4 + & \cr
& + \biggl( - {3\over 8} U^2 \Delta E - {3\over 32} U^3 +
{3\over 8} U \lambda_B^2 - {3\over 8} U \Delta E^2 \biggr) \beta^3 + &\cr
% linha 5
& + \biggl( {1\over 2} U \Delta E + t^2 + {3\over 16} U^2 +
{1\over 2} \lambda_B^2 + {1\over 2} \Delta E^2 \biggr) \beta^2
+ {\cal O}(\beta^6) \biggr\}.
& (4.3.4a)
}$$
0.5cm
$ii)$ average energy per site: $<h> (\beta)$.
The simplest way to derive the average energy per site from the grand potential, is to scale the constants: $(E_0, t, U,
\lambda_B) \rightarrow (\alpha E_0, \alpha t, \alpha U,\alpha
\lambda_B)$ and substitute in eq.(3.2a) to obtain ${\cal W}(\beta; \mu; \alpha)$. From the scaled grand potential, we have
$$<h> (\beta) = {1\over N} {\partial {\cal W}(\beta;\mu; \alpha)
\over \partial \alpha} \Big|_{\alpha=1}. \eqno(4.3.5)$$
From eq.(4.3.3), we get that
-0.5cm
$$\eqalignno{
%linha 1
<h> (\beta) & = E_0 + {U\over 4} +
\biggl( - t^2 -{3\over 16} U^2 -{1\over 2} E_0 U + {1\over
4}U \mu - {1\over 2} E_0^2 + {1\over 2} E_0 \mu - {1\over 2}
\lambda_B^2 \biggr) \beta + & \cr
%linha 2
& \hskip -0.5cm +
\biggl( {3\over 16} E_0 U^2 -
{1\over 8} U^2 \mu - {3\over 16} U \lambda_B^2 + {3\over 16} U E_0^2 -
{1\over 4} U E_0 \mu +
{1\over 16} U \mu^2 + {3\over 64} U^3 \biggr) \beta^2 + & \cr
% linha 3
& + \biggl( - {3 \over 16} U E_0^2 \mu + {1\over 256} U^4 + {1\over 4} t^4+
{1\over 24} \lambda_B^4 + {1\over 16} U^2 \lambda_B^2 +
{5 \over 24}t^2 U^2 + {1\over 2} t^2 \lambda_B^2 + & \cr
%linha 4
& + {1 \over 16} U^2 E_0^2 + {1\over 32} U^2 \mu^2 + {1\over 12} U E_0^3
- {1\over 48} U \mu^3 + {1\over 48} U^3 E_0 - {1\over 64} U^3 \mu +
{1 \over 24} E_0^4 - & \cr
% linha 5
& - {3 \over 32} U^2 E_0 \mu + {1\over 8} U E_0 \mu^2 +
{1\over 2} U t^2 E_0 - {3\over 8} U t^2 \mu - {1\over 8} E_0^3 \mu +
{1\over 8} E_0^2 \mu^2 - & \cr
% linha 6
& - {1\over 24} E_0 \mu^3 + {1\over 2} t^2 E_0^2 + {1\over 4} t^2 \mu^2 -
{3\over 4} t^2 E_0 \mu + {1\over 4} U E_0 \lambda_B^2 -
{3\over 8} \lambda_B^2 E_0 \mu - & \cr
% linha 7
& - {3\over 16} U \lambda_B^2 \mu + {1\over 4} \lambda_B^2 E_0^2 +
{1\over 8} \lambda_B^2 \mu^2 \biggr) \beta^3 + & \cr
% linha 8
& + \biggl( - {5\over 1024} U^5 + - {25\over 768} U^4 E_0 -
{5 \over 64} U^3 t^2 + {5\over 96} U \lambda_B^4 +
{5\over 192} U^4 \mu - {5 \over 48} U^2 E_0^3 + & \cr
% linha 9
& + {5 \over 768} U^3 \lambda_B^2
- {5 \over 96} U E_0^4 - {1\over 96} U \mu^4 - {65 \over 768} U^3 E_0^2
+ {1\over 24} U^2 \mu^3 + {1\over 6} U E_0^3 \mu - & \cr
%linha 10
& + {13\over 96} U^3 E_0 \mu
- {3\over 16} U E_0^2 \mu^2 + {1\over 4} U^2 E_0^2 \mu -
{3 \over 16} U^2 E_0 \mu^2 + {1 \over 12} U E_0 \mu^3 -
{5 \over 16} U^2 t^2 E_0 + & \cr
% linha 11
& + {5 \over 16} U t^2 \lambda_B^2 - {5\over 16} U t^2 E_0^2 -
{3\over 16} U t^2 \mu^2 + {1\over 4}t^2 U^2 \mu +{1\over 2} U t^2 E_0 \mu
- {13 \over 256} U^3 \mu^2 \biggr) \beta^4 + & \cr
% linha 12
& \hskip 2cm + {\cal O}(\beta^5). & (4.3.5a)\cr }$$
0.5cm
$iii)$ difference between average numbers of spin up and spin down particles per site: $< n_\uparrow> - <n_\downarrow>$.
From the definition of the grand potential (eq.(4.3.1)), we have that
$$< n_\uparrow> (\beta) - <n_\downarrow> (\beta) = {1 \over N}
{ \partial {\cal W}(\beta;\mu) \over \partial \lambda_B}. \eqno(4.3.6)$$
Up to order $\beta^4$, we get from eq.(4.3.3) that,
-0.5cm
$$\eqalignno{
% linha 1
< n_\uparrow> (\beta) - <n_\downarrow> (\beta) &= -{\lambda_B \over8}
\biggl\{ - \bigl( U \beta + 4 \bigr) \beta + \biggl( \Delta E^2 +
U \Delta E + {1\over 3} \lambda_B^2 + {1\over 4} U^2
+ 2 t^2 \biggr) \beta^3 +
& \cr
%linha 2
& \hskip -1.5cm
+ \biggl( {1\over 3} U \lambda_B^2 + U t^2 + {1\over 48} U^3 \biggr)
\beta^4 + {\cal O}(\beta^5) \biggr\}. & (4.3.6a) \cr
}$$
0.5cm
$iv)$ average of the square of the magnetization per site: $< m_z^2> (\beta) $.
-0.5cm
$$\eqalignno{ < m_z^2> (\beta) &= {\lambda_B \over B^2} <({\bf n}_{i
\uparrow} - {\bf n}_{i \downarrow})^2 > & \cr
%
%segunda linha
%
& =
- \big( {1\over 2} g \mu_B \big)^2 {1\over N} \Big[ {\partial
{\cal W}(\beta;\mu) \over \partial \mu} + 2 {\partial {\cal
W}(\beta;\mu) \over \partial U} \Big], & (4.3.7) \cr }$$
where $B$ is the external magnetic field. From eq.(4.3.3), we obtain that
-0.5cm
$$\eqalignno{
% linha 1
< m_z^2> (\beta) &= {1\over 4} g^2 \mu_B^2 \biggl\{
{1\over 2} + {1\over 8} U \beta +
\bigl( -{1\over 8} U E_0 + {1\over 8} \mu U - {1\over 32} U^2 + {1\over 8}
\lambda_B^2 - {1\over 8} E_0^2 + & \cr
% linha 2
& + {1\over 4} E_0 \mu - {1\over 8} \mu^2 \bigr) \beta^2 -
\biggl( { 1\over 12} U t^2 + {1 \over 384} U^3 \biggr) \beta^3 + & \cr
%linha 3
& + \biggl( {5 \over 1536} U^4 + {15\over 384} U^2 E_0^2 +
{15\over 384} U^2 \mu^2 + {1\over 24} U E_0^3 - {1\over 24} U \mu^3 +
& \cr
%linha 4
& + {7 \over 384} U^3 E_0 - {7 \over 384} U^3 \mu + {1\over 32} t^2 U^2
- {3\over 384} U^2 \lambda_B^2 - {1\over 8} t^2 \lambda_B^2 -
{1 \over 48} \lambda_B^4 - {1\over 8} U E_0^2 \mu + & \cr
% linha 5
& + {1\over 8} U E_0 \mu^2 + {1\over 8} U t^2 E_0 - {1\over 8} U t^2 \mu +
{1\over 48} E^4 + {1\over 48} \mu^4 - {1\over 12}E_0^3\mu +
{1\over 8} E_0^2 \mu^2 - & \cr
%linha 6
&- {1 \over 12} E_0 \mu^3 + {1\over 8} t^2 E_0^2 + {1\over 8}t^2 \mu^2
- {1\over 4} t^2 E_0 \mu - {15 \over 192} U^2 E_0 \mu \biggr) \beta^4
+ {\cal O}(\beta^5) \biggr\} , & (4.3.7a) \cr }$$
where [*g*]{} is the Landé’s factor and $\mu_B$ is the Bohr’s magneton.
0.5cm
$v)$ magnetic susceptibility: $\chi (\beta)$.
$$\chi (\beta) = - \big( {1\over 2} g \mu_B\bigr)^2 \;{1 \over N}
{\partial ^2 {\cal W}(\beta;\mu) \over \partial \lambda_B^2 }.
\eqno (4.3.8)$$
From eq. (4.3.3), we obtain that
$$\eqalignno{
%linha 1
\chi (\beta) &= - \bigl( {1\over 2} g \mu_B\bigr)^2 \;\biggl[
\biggl( {1\over 8} U t^2 + {1\over 8} U \lambda_B^2 + {1\over 384} U^3
\biggr) \beta^4 + \biggl( {1\over 4} t^2 + {1\over 8} \lambda_B^2 +
{1\over 8} U \Delta E + & \cr
%linha 2
& + {1\over 8} \Delta E^2 + {1\over 32} U^2
\biggr) \beta^3 - {1\over 8} U \beta^2 -
{1\over 2} \beta + {\cal O}(\beta^5) \biggr]. & (4.3.8a) \cr
}$$
[**5. Conclusions**]{}
With the implementation of the factorization into sub–graphs of the Grassmann multivariable integrals, we can certainly go beyond the calculation of the term $\beta^4$ of the $\beta$–expansion of the grand potential of the unidimensional Hubbrad model. Even though the physics for $U>0$ and $U<0$ are different, the results of section 4.3 apply equally well for both cases. Recently, dos Santos and Thomaz have applied the results of reference \[2\] to calculate the $\beta$–expansion of the grand canonical partition function of the extended unidimensional Hubbard model up to orde $\beta^3$ \[16\]. But the important point is that the present approach opens the possibility to calculate the first terms of the $\beta$–expansion of the grand canonical partition function of the Hubbard model in two space dimensions, as well as of unidimensional models with impurities. We believe that improvements on the present approach will render a valuable tool for tackling with such problems.
**Appendix A**
**Factorization of Grassmannian Sub–Graphs**
0.5cm For calculating the co–factors of matrices ${\bf A}^{\sigma \sigma}$, $ \sigma= \uparrow, \downarrow$, it helps to have the value of their determinant. For arbitrary $n$, the determinant of these matrices is equal to
$$\det {\bf A}^{\sigma \sigma} = \Bigl[\;\; \prod_{\nu =0}^{n-1}
\lambda_{\nu}^{(n)}\;\; \Bigr] ^N, \eqno (A.1)$$
where $N$ is the number of space sites and $\lambda_{\nu}^{(n)}$ are the [*N*]{}–fold degenerated eigenvalue of ${\bf A}^{\sigma \sigma}$. From eq.(2.6e) we have that
$$\lambda_\nu^{(n)} = 1 - e^{{i\pi \over n} (2\nu +1)}. \eqno(A.2)$$
We should notice that $ \lambda_{(n-1)-\nu}^{(n)} = {\lambda_\nu^{(n)}}^{*}$.
We define:
$$P^{(n)} \equiv \prod_{\nu =0}^{n-1} \lambda_{\nu}^{(n)}. \eqno(A.3)$$
For calculating $P^{(n)}$ we need to consider the cases $n$ even and $n$ odd separately.
For $n$ even, expression (A.3) can be rewritten as,
$$P^{(n)} = \prod_{ \nu =0}^{n-2 \over 2} \; \biggl[
2 - 2 \cos \bigl({\pi\over n} (2\nu +1) \bigr) \;\; \biggr] = 2,
\eqno(A.4)$$
where the last equality is already known \[13\].
For $n$ odd, expression (A.3) can be rewritten as,
$$P^{(n)} = 2 \times 2^{n-1} \;\; \prod_{ \nu =0}^{n-1} \;\;
\sin \biggl( {\pi \over 2n}(2 \nu +1) \biggr). \eqno(A.5)$$
From reference \[13\], we have that \[14\]:
$$2^{n-1} \;\; \prod_{ \nu =0}^{n-1} \;\;
\sin \biggl( {\pi \over 2n}(2 \nu +1) \biggr) = 1, \eqno(A.6)$$
that substituted in eq. (A.5) gives
$$P^{(n)} = 2. \eqno (A.7)$$
From the results (A.4) and (A.7), for any $n$, we get that
$$\eqalignno{
\det {\bf A}^{\sigma \sigma} & = \Bigl[\;\; \prod_{\nu =0}^{n-1}
\lambda_{\nu}^{(n)}\;\; \Bigr] ^N & \cr
%
%
& = 2^N. & (A.8) \cr
}$$
0.5cm
To present the factorization of the Grassmannian integrals, we consider an example and use the graphic notation explained in Appendix B.
Let us consider the integral for fixed space indices $l_1$ and $l_3$, that contributes to $< {\cal E}_0 (\uparrow), {\cal E}_1 (\uparrow),{\cal E}_2 (\uparrow)>$,
-0.7cm
$${\cal I} (l_1, l_1, l_3) =
\int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow\uparrow} \; \eta_J}
\;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_1}\; \eta_{N+l_1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3}$$
= 3cm -2cm $$\eqalignno
{& & (A.9) \cr}$$
where $l_1 \not= l_3$. Under the similarity transformation (2.5), eq.(A.9) becomes
-0.7cm
$$\eqalignno{
%linha 1
{\cal I} (l_1, l_1, l_3) & = \int \prod_{I=1}^{4N}
d\bar{\eta}_I^\prime d{\eta}_I^\prime \;\;
\Bigl[ \sum_{ {\nu_1, \nu_2 =0} \atop {\tau_1,\tau_2=0}}^3 \;\;
q_{\nu_1 0} p_{0 \tau_1} \; q_{\nu_2 1} p_{1 \tau_2}
\bar{\eta}_{\nu_1 N + l_1}^\prime \eta_{\tau_1 N + l_1}^\prime
\bar{\eta}_{\nu_2 N + l_1}^\prime \eta_{\tau_2 N + l_1}^\prime
\;\; \Bigr] \times & \cr
%
%linha 2
%
&\hskip 2cm \times \Bigl[ \sum_{\nu_3 = 0}^3 \; \;
q_{\nu_3 2} p_{2 \nu_3 }
\bar{\eta}_{\nu_3 N + l_3}^\prime \eta_{\nu_3 N + l_3}^\prime \Bigr]
e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; D_{I J} \; \eta_J} & \cr
%
%linha 3
%
& =
\Bigl[ 2^N \; \sum_{ {\nu_1, \nu_2 =0} \atop {\tau_1,\tau_2=0}}^3 \;\;
{ q_{\nu_1 0} p_{0 \tau_1} \; q_{\nu_2 1} p_{1 \tau_2} \over
\lambda_{\nu_1}^{(4)} \lambda_{\nu_2}^{(4)} } \Bigr]
\times
\Bigl[ \sum_{\nu_3 = 0}^3 \; \;
{q_{\nu_3 2} p_{2 \nu_3 } \over \lambda_{\nu_3}^{(4)}} \Bigr],
& (A.10) \cr
}$$
where to write the second equality on the r.h.s. of eq.(A.10), we used the result (A.8).
By explicitly writting down the expressions, we see that,
-0.5cm
$$\eqalignno{
% linha 1
& 2^N \; \sum_{ {\nu_1, \nu_2 =0} \atop {\tau_1,\tau_2=0}}^3 \;\;
{ q_{\nu_1 0} p_{0 \tau_1} \; q_{\nu_2 1} p_{1 \tau_2} \over
\lambda_{\nu_1}^{(4)} \lambda_{\nu_1}^{(4)} } =
\int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow\uparrow} \; \eta_J}
\;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_1}\; \eta_{N+l_1} , & \cr
& & (A.10a)\cr
%
%linha 2
%
\noalign{\hbox{and}}
%
%linha 3
%
& \sum_{\nu_3 = 0}^3 \; \;
{q_{\nu_3 2} p_{2 \nu_3 } \over \lambda_{\nu_3}^{(4)}} = {1 \over 2^N}\;
\int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow\uparrow} \; \eta_J}
\;\;\bar{\eta}_{2N+l_3}\; \eta_{2N+l_3}. & (A.10b)\cr
}$$
Using the graphic representation of Appendix B, we write result (A.10) as,
0.3cm
= 3cm -2.5cm
$$\eqalignno{ & & (A.11)\cr
& & }$$
-0.6cm
The factorization (A.11) comes directly from the fact that the matrix [**D**]{} is diagonal (see eq.(2.5a)) and the result (2.7a). Once the presence of Grassmann generators in the integrand of integrals (2.8) correspond to cutting lines and columns of matrix [**D**]{},then only for cutts at the same space index and any $\nu$–indices we get co–factors of matrix [**D**]{} that are non–zero. In summary, the factorization of the type (A.11) always happens when two or more space indices are different.
In a similar way and by the reasons discussed before, it is simple to show that ${\cal I}(l_1, l_2, l_3)$, where all the space indices $l_i$, $i=1,2,3$, are distinct, is easily written as:
0.5cm
= 3cm -2.5cm
$$\eqno(A.12)$$
**Appendix B**
**Graphic Notation of the Multivariable Grassmann Integrals**
0.5cm
To exemplify our graphic notation for the graphs that contribute to eq. (2.3), for fixed value of $n$, we consider some terms of $Tr[{\bf K}^4]$. This graphic notation is very helpful when we apply the symmetries discussed in reference \[2\] to identify equivalent terms in $Tr[{\bf K}^n]$.
We present the graphic notation through examples and its application to the identification of equivalent integrals.
0.5cm
[**1)**]{}
$$\hskip -4.5cm <{\cal E}(\uparrow,0), {\cal E}(\downarrow,1),
{\cal E}(\uparrow,2), {\cal E}(\uparrow,3)> =
E(\uparrow)^3 E(\downarrow) \times$$
-3cm
$$%\hskip 8.8cm \epsfysize = 3cm \epsfbox{eqb1.eps } \eqno (B.1)$$
1.5cm
[**2)**]{}
$$\hskip -4.5cm <{\cal E}(\downarrow,0) {\cal E}(\uparrow,1)
{\cal E}(\uparrow,2) {\cal E}(\uparrow,3)> =
E(\uparrow)^3 E(\downarrow) \times$$
-3cm
$$%\hskip 8.8cm \epsfysize = 3cm \epsfbox{eqb2.eps } \eqno (B.2)$$
The constants $E(\uparrow)$ and $E(\downarrow)$ are defined just below eq.(3.4d). In order to show that terms (B.1) and (B.2) are equal, we use the invariance of the integrals under a cyclic translation in the temperature parameter $\nu$ in each sector $\sigma
= \uparrow$ and $\sigma = \downarrow$, separately. Therefore,
$$<{\cal E}(\uparrow,0) {\cal E}(\downarrow,1)
{\cal E}(\uparrow,2) {\cal E}(\uparrow,3)> =
<{\cal E}(\downarrow,0), {\cal E}(\uparrow,1),
{\cal E}(\uparrow,2), {\cal E}(\uparrow,3)> \eqno (B.3)$$
[**3)**]{}
$$\hskip -4.5cm
< {\cal U}(0) {\cal T}^- (\uparrow, 1) {\cal T}^+(\uparrow, 2)
{\cal E} (\uparrow, 3) > = Ut^2 E(\uparrow) \times$$
-3cm
$$%\hskip 8.8cm \epsfysize = 3cm \epsfbox{eqb4.eps } \eqno (B.4)$$
0.2cm
Due to the presence of the term ${\cal U}(0)$, the integrals in the two $\sigma$–sectors have one space index $l$ in common.
1.5cm
[**4)**]{}
$$\hskip -4.5cm
< {\cal U}(0) {\cal T}^- (\downarrow, 1) {\cal T}^+(\downarrow, 2)
{\cal E} (\downarrow, 3) > = Ut^2 E(\downarrow) \times$$
-3cm
$$%\hskip 8.8cm \epsfysize = 3cm \epsfbox{eqb5.eps } \eqno (B.5)$$ 0.2cm
The terms (B.4) and (B.5) are equal, up to a multiplicative factor, due to the fact that ${\bf A}^{\uparrow \uparrow} = {\bf A}^{\downarrow \downarrow}$.
0.5cm
The graphic notation was used along all the calculations and permitted us to considerably reduce the number of terms that contribute to the expressions of $Tr[{\bf K}^4]$ and $Tr[{\bf K}^5]$.
**Appendix C**
**Useful Multivariable Grassmann Integrals at $n=4$ and $n=5$**
0.5cm
[**C.1. Useful integrals for n=4**]{}
We need the result of twelve integrals only, to calculate the terms that contribute to (4.1.3). In this Appendix, we present the value of these integrals according to the conditions satisfied by the space indices \[15\].
1)
-0.5cm
$${\cal I}_1^{(4)} (l) \equiv \int \prod_{I=1}^{4N} \, d\eta_I
d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N} \bar{\eta}_I\;
A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_l \;\eta_l \;\; = 2^{N-1}, \eqno(C.1.1)$$
for $ l= 1, 2, \cdots, N$.
2)
-0.5cm
$$\eqalignno{ {\cal I}_2^{(4)} (l_1,l_2) &\equiv \int \prod_{I=1}^{4N} \,
d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N} \bar{\eta}_I\;
A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1}\; \eta_{l_1}
\;\;\bar{\eta}_{N+l_2} \;\eta_{N+l_2} & \cr
%
& & \cr
%
%outras linhas
%
& =\cases{ 2^{N-1}, &
$ l_1=l_2, \hskip 0.5cm l_1= 1, 2, \cdots, N$ \cr 2^{N-2}, & $ l_2\not=
l_1,\hskip 0.5cm l_1, l_2=1, 2, \cdots, N.$\cr } & (C.1.2) \cr }$$
3)
-0.5cm
$$\eqalignno{ {\cal I}_3^{(4)} (l_1,l_2) &\equiv \int \prod_{I=1}^{4N} \,
d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N} \bar{\eta}_I\;
A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1+1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2 -1} & \cr
%
& & \cr
%
% outras linhas
%
& = 2^{N-2}, l_2= l_1 +1, \hskip 0.5cm l_1= 1, 2, \cdots, N.
& (C.1.3) \cr }$$
4)
-0.5cm
$$\eqalignno{ {\cal I}_4^{(4)} (l_1, l_2, l_3) & \equiv \int
\prod_{I=1}^{4N} \, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J=
1}^{4N} \bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1}
\;\eta_{l_1} \;\;\bar{\eta}_{N+l_2} \; \eta_{N+l_2+1}\;\;
\bar{\eta}_{2N+l_3} \; \eta_{2N+l_3 -1} & \cr
%
& & \cr
%
%outras linhas
%
& = \cases{ 2^{N-3}, & $l_3=l_2 +1$\cr
2^{N-2}, & $l_1=l_3= l_2 -1$. \cr}
& (C.1.4) \cr }$$
5)
-0.5cm
$$\eqalignno{ {\cal I}_5^{(4)} (l_1, l_2, l_3) &\equiv \int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{2^{N-3}, & $l_1\not= l_2 \not= l_3$ \cr
2^{N-2}, & $l_1= l_2 $, or, $l_1=l_3$, or, $l_2=l_3 $ \cr
2^{N-1}, & $ l_1 = l_2 = l_3$. \cr }
& (C.1.5) \cr }$$
6)
-0.5cm
$$\eqalignno{ {\cal I}_6^{(4)} (l_1, l_2, l_3) &\equiv \int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1 +1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3 -1} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{2^{N-3}, & $l_3 = l_1 +1$ \cr
2^{N-2}, & $l_2 = l_1 $ and $l_3 = l_1 +1$. \cr }
& (C.1.6) \cr }$$
7)
-0.5cm
$$\eqalignno{ {\cal I}_7^{(4)} (l_1, l_2, l_3) &\equiv \int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2 +1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3 -1} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{2^{N-3}, & $l_2 = l_3 -1$ \cr
2^{N-2}, & $l_1 = l_3 = l_2 +1$. \cr }
& (C.1.7) \cr }$$
8)
-0.5cm
$$\eqalignno{ {\cal I}_8^{(4)} (l_1, l_2, l_3, l_4)
&\equiv \int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3} \;\;\bar{\eta}_{3N+l_4}\; \eta_{3N+l_4} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{ 2^{N-4}, & $l_1\not= l_2 \not= l_3 \not= l_4$ \cr
2^{N-3}, & $l_1 = l_2$, or, $\cdots,{\rm or}, l_3 = l_4$ \cr
2^{N-2}, & $l_1= l_2 $ and $ l_3=l_4$, or, all permutations 2 by 2 \cr
&$l_1=l_2=l_3$, or, all permutations with 3 equal
space indices\cr
2^{N-1}, & $ l_1 = l_2 = l_3 = l_4$. \cr }
& (C.1.8) \cr }$$
9)
-0.5cm
$$\eqalignno{ {\cal I}_9^{(4)} (l_1, l_2, l_3, l_4)
&\equiv \int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1+1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2-1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3+1} \times & \cr
& \hskip 6cm \times {\eta}_{3N+l_4}\; \eta_{3N+l_4 -1} & \cr
%
& & \cr
%
% outras linhas
%
& \hskip -0.5cm
= \cases{ 2^{N-4}, & $l_2 = l_1+1$ and $l_4= l_3+1$, or,
$l_2= l_3 +1$ and $l_4= l_1+1$ \cr
2^{N-2}, & $l_2= l_3+1 $ and $ l_1=l_3$ and $ l_4=l_3+1$. \cr}
& (C.1.9) \cr }$$
10)
-0.5cm
$$\eqalignno{ {\cal I}_{10}^{(4)} (l_1, l_2, l_3, l_4)
&\equiv \int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1-1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2-1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3+1} \times & \cr
& \hskip 6cm \times {\eta}_{3N+l_4}\; \eta_{3N+l_4 +1} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{ 2^{N-4}, & $l_2 = l_3+1$ and $l_1= l_4 + 1$, or,
$l_4= l_2 - 1$ and $l_1= l_3+1$ \cr
2^{N-3}, & $l_2= l_3+1 $ and $ l_1=l_3 +2$ and $ l_4=l_3+1$ \cr
& $l_2 = l_3 +1$ and $l_1 = l_3$ and $ l_4 = l_3 -1$. \cr}
& (C.1.10) \cr }$$
11)
-0.5cm
$$\eqalignno{ {\cal I}_{11}^{(4)} (l_1, l_2, l_3, l_4)
& \equiv \int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3 +1} \times & \cr
& \hskip 5cm \times \;\;\bar{\eta}_{3N+l_4}\; \eta_{3N+l_4 -1} & \cr
%
& & \cr
%
% outras linhas
%
& \hskip -0.5cm
=\cases{ 2^{N-4}, & $ l_4 = l_3 +1$ \cr
2^{N-3}, & $l_2 = l_3+1$ and $l_4 = l_3 +1$, or,
$l_1 =l_2$ and $ l_4= l_3 +1$, \cr
& or, $l_1= l_3 +1$ and $l_4 = l_3 +1$ \cr
2^{N-2}, & $l_1= l_3 + 1 $ and $ l_2=l_3 +1$ and $ l_4= l_3 +1$. \cr }
& (C.1.11) \cr }$$
12)
-0.5cm
$$\eqalignno{ {\cal I}_{12}^{(4)} (l_1, l_2, l_3, l_4)
&\equiv \int \prod_{I=1}^{4N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{4N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2 +1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3} \times & \cr
%
& \hskip 5cm \times \bar{\eta}_{3N+l_4}\; \eta_{3N+l_4 -1} & \cr
%
& & \cr
%
% outras linhas
%
& \hskip -1.5cm
=\cases{ 2^{N-4}, & $ l_4 = l_2 +1$ \cr
2^{N-3}, & $l_1 = l_3$ and $l_4 = l_2 +1$, or,
$l_1 =l_2 +1$ and $ l_4= l_2 +1$, \cr
& or, $l_2= l_3 $ and $l_4 = l_2 +1$ \cr
2^{N-2}, & $l_1= l_2 + 1 $ and $ l_3=l_2$ and $ l_4= l_2 +1$. \cr}
& (C.1.12) \cr }$$
[**C.2. Useful integrals for n=5**]{}
We present here the seven integrals that have no equivalent ones for $n<5$; i.e., integrals that cannot be factorized into any of the integrals for $n<5$ for all the conditions satisfied by the apace indices. In some graphs, we do not have generators in the integrand of integrals of type (2.7) at a given value of $\nu$, as we can see for example in the graphs presented in Appendix B. Those rings in the integrals of type (2.7) that have no associated Grassmann generators in the integrand, we call [*empty rings*]{}. For example, in (B.1) we have one empty ring at $\sigma=\uparrow$ ($\nu =1$), and three empty rings at $\sigma=\downarrow$ ($\nu= 0,1$, and $3$). The integrals for $n=5$ with empty rings give the same results to the equivalent integrals for $ n=4$. We have not demonstrated this property in general form for any $n$, but we have detected it by evaluating these integrals through the procedure [gint]{}.
The seven integrals for $n=5$ and the conditions satisfied by the space indices \[15\] are:
1)
-0.5cm
$$\eqalignno{ {\cal G}_1^{(5)} (l_1, l_2, l_3, l_4, l_5)
& \equiv \int \prod_{I=1}^{5N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{5N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3} \times & \cr
%
& \hskip 3cm
\times \bar{\eta}_{3N+l_4}\; \eta_{3N+l_4}
\;\;\bar{\eta}_{4N+l_5}\; \eta_{4N+l_5} & \cr
%
& & \cr
%
% outras linhas
%
& \hskip -1cm
=\cases{ 2^{N-5}, & $l_1 \not= l_2\not= l_3 \not= l_4 \not= l_5$ \cr
2^{N-4}, & $l_1= l_2$ , or, all permutations with 2 equal space indices\cr
2^{N-3}, & $l_1 = l_2= l_3$, or, all permutations with 3 equal
space indices \cr
& $l_1 = l_2$ and $ l_3 = l_4$, \cr
& or, all permutations 2 by 2 \cr
2^{N-2}, & $l_1= l_2= l_3= l_5 $, or, \cr
& all permutations with 4 equal space indices \cr
& $l_1= l_2= l_3$ and $l_4=l_5$, or, \cr
& all permutations with 2 or 3 equal space indices \cr
2^{N-1}, & $ l_1 = l_2 = l_3 = l_4 = l_5$. \cr }
& (C.2.1) \cr }$$
0.3cm
2)
-0.5cm
$$\eqalignno{ {\cal G}_2^{(5)} (l_1, l_2, l_3, l_4, l_5)
& \equiv \int \prod_{I=1}^{5N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{5N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2+1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3} \times & \cr
%
& \hskip 3cm
\times \bar{\eta}_{3N+l_4}\; \eta_{3N+l_4 -1}
\;\;\bar{\eta}_{4N+l_5}\; \eta_{4N+l_5} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{ 2^{N-5}, & $l_4 = l_2 + 1 $ \cr
2^{N-4}, & $l_4= l_2 + 1$ and two other space indices are equal\cr
2^{N-3}, & $l_4 = l_2+1$ and $l_1 = l_4$ and $l_3 = l_5$, \cr
& or, all permutations 2 by 2\cr
& $l_4 = l_2+1$ and $ l_1 = l_4= l_5$, or, \cr
& all permutations with 3 equal space indices \cr
2^{N-2}, & $l_2 = l_5 -1= l_3$ and $l_1= l_5=l_4$. \cr }
& (C.2.2) \cr }$$
0.3cm
3)
-0.5cm
$$\eqalignno{ {\cal G}_3^{(5)} (l_1, l_2, l_3, l_4, l_5)
& \equiv \int \prod_{I=1}^{5N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{5N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2+1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3 -1} \times & \cr
%
& \hskip 3cm
\times \bar{\eta}_{3N+l_4}\; \eta_{3N+l_4}
\;\;\bar{\eta}_{4N+l_5}\; \eta_{4N+l_5} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{ 2^{N-5}, & $l_3 = l_2 + 1 $ \cr
2^{N-4}, & $l_3= l_2 + 1$ and two other space indices are equal\cr
2^{N-3}, & $l_3 = l_2+1$ and $l_1 = l_4$ and $l_3 = l_5$, or, \cr
& all permutations 2 by 2 \cr
& $l_3 = l_2+1$ and $ l_1 = l_3= l_4$, or, \cr
& all permutations with 3 equal space indices \cr
2^{N-2}, & $l_3 = l_2 + 1$ and $l_1= l_3=l_4 = l_5$. \cr }
& (C.2.3) \cr }$$
0.3cm
4)
-0.5cm
$$\eqalignno{ {\cal G}_4^{(5)} (l_1, l_2, l_3, l_4, l_5)
& \equiv \int \prod_{I=1}^{5N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{5N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1 -1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2+1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3 +1} \times & \cr
%
& \hskip 3cm
\times \bar{\eta}_{3N+l_4}\; \eta_{3N+l_4-1}
\;\;\bar{\eta}_{4N+l_5}\; \eta_{4N+l_5} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{ 2^{N-5}, & $l_1 = l_2 + 1 $ and $ l_4 = l_3 +1$, or, \cr
& $ l_1 = l_3 +1$ and $ l_4= l_2 +1$ \cr
2^{N-4}, & $l_1= l_2 + 1$ and $ l_5= l_3 +1$ and $ l_4 = l_5$, or,\cr
& $l_2 =l_4 =l_1 -1 = l_3 +1$, or, \cr
& $l_2= l_5 = l_1 -1$ and $l_3= l_4-1$, or, \cr
& $l_1 = l_3 +1$ and $l_4= l_5= l_2+1$, or, \cr
& $l_3 = l_5 = l_1 -1$ and $l_2= l_4 -1$, or, \cr
& $l_1 = l_3 = l_4 -1$ and $l_4= l_2 +2$ \cr
2^{N-3}, & $l_1 =l_3 = l_5+1$ and $l_2 = l_5$ and $l_4 = l_2+2 $, or, \cr
& $l_1 = l_5 +1$ and $l_2= l_5$ and $l_3= l_5 -1$ and $l_4=l_5$,
or, \cr
& $l_1= l_3 = l_5 -1$ and $ l_2 = l_5 - 2$ and $l_4=l_5$. \cr}
& (C.2.4) \cr }$$
5)
-0.5cm
$$\eqalignno{ {\cal G}_5^{(5)} (l_1, l_2, l_3, l_4, l_5)
& \equiv \int \prod_{I=1}^{5N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{5N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1+1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3} \times & \cr
%
& \hskip 3cm
\times \bar{\eta}_{3N+l_4}\; \eta_{3N+l_4 -1}
\;\;\bar{\eta}_{4N+l_5}\; \eta_{4N+l_5} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{ 2^{N-5}, & $l_4 = l_1 + 1 $ \cr
2^{N-4}, & $l_4= l_1 + 1$ and two other space indices are equal\cr
2^{N-3}, & $l_4 = l_1+1$ and $l_1 = l_2= l_3 $, or, \cr
& all permutations with 3 equal space indices \cr
& $l_4 = l_1+1$ and $ l_1 = l_2$ and $l_3= l_5$, or, \cr
& all permutations 2 by 2 \cr
2^{N-2}, & $l_4 = l_1 + 1$ and $l_1= l_2=l_3 = l_5-1$. \cr }
& (C.2.5) \cr }$$
0.3cm
6)
-0.5cm
$$\eqalignno{ {\cal G}_6^{(5)} (l_1, l_2, l_3, l_4, l_5)
& \equiv \int \prod_{I=1}^{5N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{5N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1+1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2-1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3+1} \times & \cr
%
& \hskip 3cm
\times \bar{\eta}_{3N+l_4}\; \eta_{3N+l_4 -1}
\;\;\bar{\eta}_{4N+l_5}\; \eta_{4N+l_5} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{ 2^{N-5}, & $l_2= l_1 + 1 $ and $l_4 =l_3 +1$, or, \cr
& $l_4= l_1 +1$ and $ l_2 = l_3+1$ \cr
2^{N-4}, & $l_2= l_1 + 1$ and $l_4= l_5 = l_3 +1$ \cr
& $l_2=l_5 = l_1 +1$ and $l_4= l_3 +1$ \cr
& $ l_4 = l_5= l_1 +1$ and $ l_2= l_3 +1$ \cr
& $l_1 = l_4 -1$ and $ l_3 = l_5 = l_2 -1$ \cr
2^{N-3}, & $l_1=l_3 = l_4 -1$ and $l_2 = l_4 $\cr
2^{N-2}, & $l_1 = l_3 = l_5 - 1$ and $l_2= l_4=l_5$ \cr }
& (C.2.6) \cr }$$
0.3cm
7)
-0.5cm
$$\eqalignno{ {\cal G}_7^{(5)} (l_1, l_2, l_3, l_4, l_5)
& \equiv \int \prod_{I=1}^{5N}
\, d\eta_I d\bar{\eta}_I \;\; e^{\sum\limits_{I,J= 1}^{5N}
\bar{\eta}_I\; A_{I J}^{\uparrow \uparrow}
\; \eta_J} \;\; \bar{\eta}_{l_1} \;\eta_{l_1+1}
\;\;\bar{\eta}_{N+l_2}\; \eta_{N+l_2+1} \;\; \bar{\eta}_{2N+l_3} \;
\eta_{2N+l_3-1} \times & \cr
%
& \hskip 3cm
\times \bar{\eta}_{3N+l_4}\; \eta_{3N+l_4 -1}
\;\;\bar{\eta}_{4N+l_5}\; \eta_{4N+l_5} & \cr
%
& & \cr
%
% outras linhas
%
& =\cases{ 2^{N-5}, & $l_3= l_1 + 1 $ and $l_4 = l_2 +1$, or, \cr
& $l_4= l_1 +1$ and $ l_3 = l_2+1$ \cr
2^{N-4}, & $l_3= l_1 + 1$ and $l_4 = l_5 = l_2 +1$, or, \cr
& $l_3 = l_5 = l_1 +1 $ and $l_4= l_2 +1$, or, \cr
& $ l_3 = l_5 = l_2 +1$ and $l_4 = l_1 +1$, or, \cr
& $ l_4= l_5= l_1 +1$ and $ l_2 = l_3 -1$, or, \cr
& $ l_2 = l_4 = l_3 -1 = l_1 +1$, or
$ l_1 = l_3 = l_4 -1 = l_2 +1 $ \cr
2^{N-3}, & $l_2=l_4 = l_5$ and $l_1 = l_5 -1 $ and $ l_3 = l_5 +1$, or,\cr
& $ l_1 = l_3 = l_5 -1$ and $l_4 = l_5$ and $ l_2 = l_5 -2$,
or, \cr
& $ l_2 = l_4 =l_5 -1$ and $ l_3 = l_5$ and $l_1= l_5 -2$. \cr}
& (C.2.7) \cr }$$
[**Acknowledgements**]{}
The authors thank J. Florencio Jr. for interesting discutions and A.T. Costa Jr. for making the figures. I.C.C thanks FAPMG and E.V.C.S. thanks CNPq for financial support. M.T.T. thanks CNPq and FINEP for partial financial support.
**REFERENCES**
[1.]{} [S. Samuel, J. Math. Phys. [**21**]{} (1980) 2806–2833; C. Itzykson, Nucl. Phys. [**B210**]{} \[FS6\] (1982) 448; V.N. Plechko, Physica A [**152**]{} (1988) 51; V.N. Plechko and I.K. Sobolev, Physica A [**197**]{} (1993) 323; ]{}
[2.]{}[I.C. Charret, S.M. de Souza, E.V. Corrêa Silva and M.T. Thomaz, [*Grand Canonical Partition Function for the Unidimensional Systems: Application to Hubbard Model Up to Order $\beta^3$*]{}, submitted to Jour. of Phys. A,(pre-print cond/mat/ 9607171); ]{}
[3. ]{} [ E.H. Lieb and F.Y. Wu, Phys. Rev. Letters [**20**]{} (1968) 1445; A.A. Ovchinnikov, Sov. Phys. JETP [**30**]{} (1970) 1160; ]{}
[4.]{} [ M. Takahashi, Prog. Theoret. Phys. (Kyoto) [**43**]{} (1970) 1619; ]{}
[5.]{} [ K. Kubo and M. Tada, Prog. Theoret. Phys. [**69**]{} (1983) 1345; C.J. Thompson, Y.S. Young, A.J. Guttmann and M.F. Sykes, J. Phys. A: Math. Gen. [**24**]{} (1991) 1261; J.A. Henderson, J. Oitmaa and M.C.B. Ashley, Phys. Rev. [**B46**]{} (1992) 6328; ]{}
[6.]{} [ H.E. Stanley, [*Introduction to Phase Transitions and Critical Phenomena*]{}, Oxford Univ. Press (1971);]{}
[7.]{} [ C. Itzykson and J.–B. Zuber, [*Quantum Field Theory*]{}, McGraw–Hill (1980); U. Wolf, Nucl. Phys. [**B225**]{} \[FS9\] (1983) 391; ]{}
[8.]{} [We are using the convention: $\sigma= 1 = \uparrow$ and $\sigma= -1 = \downarrow$; ]{}
[9.]{}[ I.C. Charret, S.M. de Souza and M.T. Thomaz, Braz. Jour. of Phys. [**26**]{} (1996) 720; ]{}
[10.]{}[ J. Hubbard, Proc. Roy. Soc. [**A277**]{} (1963) 237; [**A281**]{} (1964) 401; M. Gutzwiller, Phys. Rev. Lett. [**10**]{} (1963) 159; Phys. Rev. [**A137**]{} (1965) 1726;]{}
[11.]{}[ I.C. Charret, E.V. Corrêa Silva, S.M. de Souza and M.T. Thomaz, J. Math. Phys. [**36**]{} (1995) 4100;]{}
[12.]{} [See section 4.1 of reference \[2\] for a discussion on the vanishing Grassmann integrals;]{}
[13.]{} [ I.S. Gradshteyn and I.M. Ryzhik; [*Table of Integrals, Series and Products*]{}, $4^{th}$ edition, Academic Press (1965); expression: 1.396.4;]{}
[14.]{} [ Reference \[11\], expression: 1.392.2;]{}
[15.]{} [ When the space index does not appear among the conditions, we mean that it is distinct from any other space index in the graph;]{}
[16.]{} [ O.R. dos Santos and M.T. Thomaz, private communication.]{}
[^1]: $^1$
[^2]: $^2$
[^3]: $^3$
[^4]: $^4$
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.